Case Studies on Transport Policy 20 (2025) 101440

Contents lists available at ScienceDirect

Case Studies on Transport Policy

journal homepage: [Link]/locate/cstp

Human cost, machine insight: A data-driven analysis of Australian

road crashes
Ali Soltani a,b,c,* , Saeid Afshari d , Mohammad Amin Amiri e
a
FHMRI, Flinders University, Bedfork Park, South Australia 5042, Australia
b ́
LE STUDIUM Loire Valley Institute for Advanced Studies, Orleans, France
c ́
CEDETE Research Center, University of Orleans, ́
Orleans, France
d
Department of Computer Engineering, University of Isfahan, Iran
e
Faculty of Mathematics and Statistics, University of Isfahan, Iran

A R T I C L E I N F O A B S T R A C T

Keywords: In Australia, road crash injuries continue to be a serious public health issue. Machine learning is used in this
Road safety study to analyse injury data from road crashes between 2011 and 2021 that was taken from the national hos­
Hospitalized injuries pitalized injury database. We investigate how the number of injuries and duration of stay for road users are
Crash severity
affected by variables such as gender, age, seasonal variation, collision type, and location (urban vs. regional).
Machine Learning
Road safety measures are informed by patterns and relationships found in the data by machine learning models.
Autoregressive time series model
Australia Hospitalizations have been trending upward between 2011 and 2019, with a pause in 2020 due to COVID-19
lockdowns. In all categories, men sustain more injuries than women, though the number varies according to
age and geography. The type of road user also affects collision patterns. The time-series projections demonstrate
that the goal of zero fatalities in 2050 will not be achieved under the business-as-usual scenario. The findings
highlight the necessity of focused interventions predicated on collision trends and demographics. This includes
better infrastructure design, increased surveillance, and customized safety measures.

1. Introduction spent on road trauma each year, with hospital stays costing $239,000
(Connections, 2017).
Despite continuous efforts to lower fatalities and injuries, road traffic The introduction of the National Road Safety Strategy in 1992
crashes continue to be a major problem in Australia (Hassouna & Prin­ marked a significant milestone in coordinated efforts to improve road
gle, 2019; Henley & Harrison, 2016; Kavalamthara et al., 2019). safety across the country. This strategy has been periodically updated to
Australia has four fatalities and ninety serious injuries per day from incorporate new safety targets and measures, reflecting the evolving
crashes, placing it among the top ten OECD countries for crashes per landscape of road safety challenges and technological advancements
100,000 people (OECD, 2021). (Mason et al., 2022) have identified (Green et al., 2022). Australia’s Zero-Fatality Target aims to eliminate
several high-risk groups, such as females, children aged 0–14, and all road crash deaths by 2050, driven by the National Road Safety
Aboriginal and Torres Strait Islander peoples residing in remote areas. Strategy. This ambitious goal focuses on a comprehensive approach that
Hospitalization rates have increased over the last ten years despite in­ includes enhancing road infrastructure, advancing vehicle safety tech­
terventions, and there has been no decline in cases that pose a threat to nology, enforcing stricter traffic laws, and promoting safer road-user
life (BITRE, 2023; Connections, 2017). An estimated $29.7 billion is behaviors. Crash causes include a variety of behaviors, such as

Abbreviations: AI, Artificial Intelligence; AIC, Akaike Information Criterion; ANRAM, Australian National Risk Assessment Model; ARIMA, Autoregressive Inte­
grated Moving Average; BIC, Bayesian Information Criterion; BITRE, Bureau of Infrastructure and Transport Research Economics; COVID, Coronavirus Disease 2019;
DUI, Driving Under the Influence; GWR, Geographically Weighted Regression; IOA, Index of Agreement; ITS, Intelligent Transport Systems; IV, Instrumental Var­
iable; KNN, K-Nearest Neighbors; MAE, Mean Absolute Error; MLP, Multi-Layer Perceptron; MSE, Mean Squared Error; NDVI, Normalized Difference Vegetation
Index; NMSE, Normalized Mean Squared Error; NRSAP, National Road Safety Action Plan; NRSS, National Road Safety Strategy; OECD, Organization for Economic
Co-operation and Development; RMSE, Root Mean Squared Error; SARIMA, Seasonal Autoregressive Integrated Moving Average; SHAP, SHapley Additive exPla­
nations; SVM, Support Vector Machine; XGBoost, Extreme Gradient Boosting.
* Corresponding author at: FHMRI, Flinders University, Bedfork Park, South Australia 5042, Australia.
E-mail addresses: [Link]@[Link] (A. Soltani), [Link]@[Link] (S. Afshari), aminss07@[Link] (M.A. Amiri).

[Link]
Received 13 October 2024; Received in revised form 19 March 2025; Accepted 29 March 2025
Available online 30 March 2025
2213-624X/© 2025 The Author(s). Published by Elsevier Ltd on behalf of World Conference on Transport Research Society. This is an open access article under the
CC BY license ([Link]
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

speeding, exhaustion, drug and alcohol abuse, and disregard for the use This study enhances road safety research through methodological,
of seat belts and helmets (Hojati et al., 2011; Pammer et al., 2021). conceptual, and empirical differences, addressing research gaps. Unlike
Delineation issues, poor surfaces, and roadside hazards are examples of prior Australian work (Arteaga et al., 2020; Duong et al., 2024; Komol
poor road conditions that also have an impact (Sadeghi & Goli, 2024). et al., 2021), this research integrates four innovations for a more holistic
Particularly when it comes to heavy vehicles and motorcycles, speeding, understanding of Australian traffic crashes: (1) Analyzing data from all
exhaustion, and intoxication have a substantial impact on crash severity Australian jurisdictions and road user types (including vulnerable road
(Abrari Vajari et al., 2022; Devlin & Fitzharris, 2013; Hassouna & users such as pedestrians, cyclists, and motorcyclists) captures diverse
Pringle, 2019; Seyfi et al., 2023). Crash risk is also influenced by crash risks across urban, rural, and remote environments, addressing
external variables such as road infrastructure, lighting, and weather spatial inequities and challenging urban-centric biases. (2) Utilizing the
(Paul et al., 2022; Soltani and Qadikolaei, 2024). Research indicates a BITRE injury dataset (2011–2021), a hospital admissions-based source,
connection between Australia’s crash rates and fuel prices (Ahangari, mitigates underreporting of vulnerable road user crashes (Harrison
2015; Burke and Teame, 2018). Australia is committed to safer roads, as et al., 2023), providing a more accurate assessment of severity
evidenced by its National Road Safety Strategy 2021–30, which sets a (Mandacaru et al., 2017) compared to police reports (e.g.,
goal of a 50 % reduction in fatalities and a 30 % decrease in serious (Siamidoudaran & İşçioğlu, 2019; Vadhwani & Thakor, 2023). Com­
injuries by 2030. bined with the BITRE fatality dataset (1989–2024), this offers unprec­
Enhancing road safety requires being able to predict the frequency edented granularity. (3) Employing a continuous crash severity scale,
and severity of crashes. According to (Astarita et al., 2023), predictive this research pioneers advanced machine learning algorithms like
models aid in the identification of crash causes, collision types, and XGBoost and an autoregressive time series model incorporating non-
driver behaviors, allowing for proactive crash prevention measures. temporal features (e.g., demographics, infrastructure). This approach
With algorithms like XGBoost, Random Forest, SVM, KNN, and Deep surpasses traditional methods (e.g., logistic regression) by achieving
Learning offering varying degrees of accuracy, machine learning has greater predictive accuracy and revealing complex variable interactions,
become a crucial tool in crash analysis (Çelik & Sevli, 2022; Ozcan et al., like the combined influence of rurality, weather, and behavior on injury
2022). In studies (Boyagoda & Nawarathna, 2022; Prasad et al., 2022), outcomes. (4) The 35-year fatality dataset (1989–2024) enables longi­
supervised learning techniques like Random Forest, SVM, Decision Tree, tudinal analysis of policy impacts (e.g., seatbelts, speed cameras) and
and KNN analyze crash data to classify injury severity, achieving accu­ emerging risks (e.g., distracted driving). Our autoregressive model,
racies of about 84 %. As shown by (Toğan et al., 2022) and (Habibzadeh incorporating socioeconomic and infrastructural covariates, provides
et al., 2022) in construction and rural road crash predictions, machine actionable insights for targeted interventions.
learning is not limited to traffic crashes. (Shan et al., 2023) used stacking
models to achieve over 84 % precision. Random Forest, Logistic 2. Literature review
Regression, and Gradient Boosting were compared by (Santos et al.,
2024) to predict the severity of injuries in motorbike crashes. With sophisticated algorithms like AdaBoost highlighting age-
(Kumeda et al., 2019) found that the J48 classification was the most specific risks, age is a significant factor in traffic crashes. Despite hav­
accurate for predicting crashes, while (Kim et al., 2021) found that SVM ing lower crash rates, drivers 75 years of age and older have higher fa­
performed better for senior drivers than other methods. (Sameen & tality rates, so special interventions are needed (J. Lee et al., 2023).
Pradhan, 2017; Siamidoudaran & İşçioğlu, 2019; Vadhwani & Thakor, Children and young adults (20–29) are especially prone to serious in­
2023) introduced neural networks and factor analysis; to predict injury juries, especially when they are pedestrians (Alicioglu et al., 2022; Park
severity, XGBoost was optimized, and (Bi & Sadia, 2023) highlighted et al., 2013; Zhang et al., 2021). Adolescent crash victims’ substance
Random Forest, Naive Bayes, and AdaBoost as popular options. While abuse highlights the importance of early interventions even more (Mitra
(Vanitha, 2023) used Decision Tree, Random Forest, and Logistic et al., 2012).
Regression for crash prediction, (Cicek et al., 2023) discovered that MLP Crash rates are also influenced by cultural background and gender.
was the most accurate method. Research indicates that drivers 75 years of age and older do not have
The objective of this research is to forecast the frequency and disproportionately high crash risks (Tuncalı Yaman et al., 2022).
severity of traffic crashes by looking at a variety of influencing factors, Nonetheless, the number of road traffic crashes among younger adults
including locational traits, seasonal conditions, and demographic vari­ (ages 22–29) is higher (Theofilatos & Yannis, 2019), necessitating
ables. Additionally, the study aims to assess how well various machine educational interventions. To meet the safety requirements of various
learning models predict these things. Below is a presentation of the age groups, customized strategies are imperative (Alicioglu et al., 2022).
research questions (RQs): Although older drivers—especially women—tend to be more cautious,
their injuries are more severe (J. Lee et al., 2023). Seat belt use reduces
• RQ1: How do Ensemble models compare to other regression models injury severity in all age groups, but teen drivers face particular risk
in terms of predictive accuracy, model complexity, and performance factors, such as vehicle type, road conditions, and gender (Duddu et al.,
consistency when forecasting Australian crash frequency and 2019).
severity across different road user categories? Differences in crash severity between genders exist. Female drivers
• RQ2: How do demographic factors, collision types, and environ­ are more at risk from inexperience and car problems, while male drivers
mental conditions differentially influence crash frequency versus are more likely to speed up and break the law (Razi-Ardakani et al.,
crash severity, and how do these influences vary across different road 2018). In collisions involving cars with comparable safety ratings,
user groups in Australia? women are more likely to suffer injuries (Fu & Lee, 2022). While females
• RQ3: How do time-series forecasting models compare in predicting are more vulnerable at intersections, males are more likely to be
long-term trends in road crash outcomes in Australia? involved in serious crashes at night and in rainy weather (Rusli et al.,
2020). According to (Lopez-Valdes et al., 2022) improving outcomes
The Bureau of Infrastructure and Transport Research Economics, requires gender-specific safety systems. Crash frequency is also influ­
which offers extensive road safety crash data series covering monthly, enced by seasonal variations. According to (Khorshidi et al., 2016),
quarterly, and annual intervals from 2011 to 2021, served as the main summer is the busiest time of year for traffic-related injuries, while
source of data for this study. The serious non-fatal hospitalized injuries winter usually sees fewer crashes. Higher rates are also a result of hol­
from traffic crashes form the basis of this data series. Furthermore, as a iday seasons and higher travel volumes (Shahid et al., 2015). Crash
supplemental database, the historic road traffic crash deaths reported by severity is significantly influenced by location (Soltani et al., 2024;
police from 1989 to 2024 were obtained from BITRE (BITRE, 2024). Soltani and Askari, 2014, 2017). Road conditions and crash types affect

2
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

the severity of injuries, with rural travelers suffering more severe in­ Table 1
juries than urban travelers (Adeyemi et al., 2023; Hao & Kamga, 2017; Target variable and independent variables description.
Park & Bae, 2020). Because recovery services are frequently insufficient Variable name Description Mean Std.
in rural areas, traffic calming and driver education are essential
Age07 The person in age of 0 to 7 years old 0.06 0.24
(Morling, 2015). Research indicates that trauma outcomes differ in rural Age1725 The person aged 17 to 25 years old 0.18 0.38
areas due to a higher frequency of off-road crashes and spinal cord in­ Age2639 The person ages 26 to 39 years old 0.19 0.39
juries (Heathcote et al., 2022). The time of day is a significant factor in Age4064 The person of 40 to 64 years old 0.22 0.41
the severity of crashes in the rural–urban periphery, according to (Wu Age6574 The person aged 65 to 74 years old 0.13 0.33
Age75plus People over 75 years old 0.11 0.31
et al., 2021). Making educated policy decisions requires an under­ Age816 The person of 8 to 16 years old 0.12 0.32
standing of crash behavior in both urban and rural areas (Cabrera-Arnau April The crash happened in April 0.08 0.28
et al., 2020). August The crash happened in August 0.08 0.27
(Ahmed et al., 2023) investigate the application of explainable ma­ BeddystoCC Bed Days to Crash Case, the ratio of the 5.56 10.84
(Crash duration of stay in the hospital to the
chine learning techniques to predict road crashes and identify key
Severity) number of people involved in the crash.
contributing factors using New Zealand Crash data from 2010 to 2020. Cardriver The crash involved a car driver 0.23 0.42
The authors utilize various machine learning models to analyze data Carpngr The crash involved the person who was a car 0.22 0.41
from a specific region, focusing on understanding the relationships be­ passenger
tween different variables and their impact on road safety. The study Cntrprtcrtrkvn Collision type: the crash counterpart was 0.26 0.44
car, truck or van
emphasizes the importance of explainability in machine learning, Cntrprtfxdstnry Collision type: the crash counterpart was 0.17 0.37
allowing stakeholders to interpret the results and make informed de­ fixed stationary
cisions for improving road safety measures. Cntrprthvyvhcl Collision type: the crash counterpart was a 0.10 0.30
heavy vehicle
Cntrprtnclsn Collision type: the crash counterpart was not 0.21 0.41
3. Materials and methods
a collision
Cntrprtother Collision type: the crash counterpart was 0.14 0.34
3.1. Data collection and preprocessing other
Cntrprtpdlc Collision type: the crash counterpart was 0.03 0.18
We have used two different data sets, BITRE injury dataset and pedal cyclist
Cntrprttrn Collision type: the crash counterpart was 0.01 0.07
BITRE fatality dataset. BITRE injury dataset which provides summary of tram or train
hospitalized injuries (injury resulting in confirmed admission to hospi­ Covidprd The crash happened during Covid 2019 0.11 0.31
tal) is available from 2011 to 2021 according to (BITRE, 2023). This lockdown
dataset includes information such as month and year of admission, December The crash happened in December 0.08 0.28
February The crash happened in February 0.08 0.27
gender, age group, number of admitted cases, days of non-fatal hospi­
Female The gender of involved person was female 0.44 0.50
talization and type of crash, region and reason for admission, and type of Hospitalization The number of people who are admitted to 6.75 10.74
discharge (death or treatment) in an aggregated scale. Table 1 shows the (Crash hospitals due to road crashes.
list and description of target variables used in this study. data fields Frequency)
details and the description of independent variables and their descrip­ January The crash happened in January 0.08 0.28
July The crash happened in July 0.08 0.27
tive statistics are shown in Table 1 as well. BITRE fatality dataset which
June The crash happened in June 0.08 0.27
provides details of road traffic crash fatalities in Australia is available Majorcities Crash happened in major cities 0.56 0.50
from 1989 to 2024 (BITRE, 2024). Road injury data collected by the Male The gender of involved person was male 0.56 0.50
Bureau of Infrastructure, Transport and Regional Economics is used to March The crash happened in March 0.09 0.28
May The crash happened in May 0.08 0.28
forecast the number of fatalities in Australia. The target variable which
Mtrcyclist The crash involved a person who was a 0.21 0.41
is used “Number of Fatalities” which shows the number of people who motorcycle driver
died due to crashes. A road death or fatality is a person who dies within November The crash happened in November 0.08 0.28
30 days of a crash because of injuries received in that crash. October The crash happened in October 0.09 0.28
To address feature dependency analysis, we have utilized a correla­ Pdlcyclist The person involved was a pedal cyclist 0.23 0.42
Pedestrian The crash involved a pedestrian 0.11 0.32
tion matrix to illustrate the relationships between features in our data­
Regional A crash happened in regional area 0.44 0.50
set. This can help us understand which features are dependent on each September The crash happened in September 0.08 0.27
other. We provided a visual representation of this matrix in Fig. 1. This
approach helps to highlight any potential dependencies among the
features considered in our analysis. where the response variable represents counts of events or occurrences,
such as the number of customer arrivals, website visits, or product sales.
3.2. Methods The model assumes that the response variable follows Poisson distri­
bution, which is appropriate for non-negative integer values. Poisson
For regression analysis, the following algorithms have been applied: regression is useful for analyzing count data and estimating the effects of
XGBoost, Poisson Regression, and Ensemble Regression models. predictor variables on the rate or frequency of events (Consul & Famoye,
XGBoost (Extreme Gradient Boosting): XGBoost is an optimized and 1992).
efficient implementation of the gradient boosting algorithm. It is widely Ensemble Regression model: Ensemble regression models are
used for regression tasks and has gained popularity for its speed and powerful tools for improving predictive accuracy and robustness by
performance (Jin et al., 2024). XGBoost works by building an ensemble leveraging the strengths of multiple base models. They are particularly
of weak learners (typically decision trees) in a sequential manner, where useful when individual models have complementary strengths and
each new model corrects the errors made by the previous ones. This weaknesses, leading to better overall performance when combined. This
iterative process leads to a strong predictive model that can handle model is used to combine the above-mentioned regression models to
complex relationships in the data and reduce overfitting. XGBoost also improve predictive performance. Each base regression model is trained
includes regularization techniques to prevent overfitting and improve on the training data using different subsets of features or data samples.
generalization (Chen et al., 2015). This diversity allows for capturing different aspects of the underlying
Poisson Regression: this is a type of generalized linear model used for data patterns. Once the individual base models are trained, their pre­
data that follows Poisson distribution. It is commonly used in situations dictions are combined by averaging to make the final ensemble

3
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Fig. 1. Correlation matrix to illustrate the relationships between features.

prediction. The hyperparameters of the individual base models and the hospitalized injuries did not significantly correlate with the length of
ensemble method are fine-tuned to optimize performance. stay indicates that other factors—such as changes in crash types, ad­
Furthermore, we have projected Australia’s fatality rate by 2050 vancements in medical treatments, or improvements in car safety fea­
using the Autoregressive Integrated Moving Average (ARIMA) and tures—have an impact on collision severity and healthcare efficiency. A
Seasonal Autoregressive Integrated Moving Average (SARIMA) models. logarithmic equation of y = 2813ln(x) + 32877, R2 = 0.81 – where y
We can forecast future values by using previous data patterns and the represents the number of injuries, and x represents the year − explains
time series analysis approach. Because it can handle both seasonal the temporal trend of hospitalization frequency. This model demon­
fluctuations and non-stationary trends—data that fluctuate over time, strates a non-linear pattern in the incidence of hospitalizations, sug­
the SARIMA model is especially well-suited for this purpose. gesting complex interactions between many factors that impact severe
Table 2 provides a comparison and a better clarification our rationale road crash injuries and consequent hospitalizations. We will investigate
for model selection in the context of the BITRE injury dataset. this non-linearity later in detail.
Fig. 3 uses heatmaps to investigate seasonal trends in data on traffic
4. Comparative description crashes. The research considers crash frequency (a) and crash severity
(b). The safest month for hospital admissions was April 2020, which was
4.1. Temporal and seasonal trends in line with the early COVID-19 limits. Other low-risk months were
February 2013, February 2011, July 2012, and September 2012. On the
The analysis in this part is only descriptive and comparative using other hand, March was always a high-risk month, especially in
two major indices as previously introduced in Table 1. An analysis of the December 2020, 2017, 2019, and 2021. Pre-pandemic March to May
number of non-fatal serious injuries sustained in road incidents in (2016–2019) was the riskiest period, with injuries declining during
Australia between 2011 and 2019 shows a steady increase in crash in­ COVID-19 in 2020 but increasing again in 2021.
juries, indicating an increase in severe crashes. Examining traffic crash The pattern of injury severity changed, becoming more severe after
injuries in Australia between 2013 and 2021 uncovers remarkable pat­ July 2015. The most serious collisions occurred in April 2020, February
terns. As depicted in Fig. 2 the number of hospitalizations due to crash 2016, and July 2015, whilst the safest ones occurred in March 2014 and
injuries saw a consistent upward trend from 2013 to 2019, indicating a December 2013. There was also no discernible trend in the total number
possible increase in the occurrence of severe crashes. Nonetheless, a of deaths, although Saturdays had the highest death toll (247 in 2012),
significant decrease was observed in 2020, which was brought on by followed by Sundays. Fewer deaths occurred during the week, peaking
COVID-19 lockdowns that altered driving patterns, traffic volume, and from Monday to Thursday and then slightly increasing on Fridays.
road user behavior. When drivers resumed their pre-lockdown habits by Variations in the number of deaths each year might be caused by
2021, crash frequency and severity reached or surpassed pre-pandemic changes in vehicle safety, law enforcement, road safety campaigns, and
levels. Remarkably, over time, the length of hospital stays (BeddystoCC), population density (Intini et al., 2018; Steiner et al., 2014). The year
a measure of the severity of injuries, did not change. The lengthier 2020, with the implementation of COVID-19 laws, is a clear departure.
hospital stays in 2020 suggested that patients had more serious injuries All days saw a significant decrease in deaths over the previous year.
even if there were fewer injuries overall. The fact that the number of Reduced travel because of lockdown procedures and remote work, as

4
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Table 2 Table 2 (continued )

Comparison of applied algorithms. ALGORITHM Strengths Weaknesses
ALGORITHM Strengths Weaknesses
− Suitable for data with both − Requires careful tuning of
XGBoost − Handles complex, non-linear − Requires careful trend and seasonality (e.g., seasonal parameters (P, D, Q,
relationships well. hyperparameter tuning. monthly or yearly fatality m).
− Robust to outliers and − Less interpretable compared rates).
missing data. to simpler models (e.g., linear − Robust for medium to long- − Computationally intensive
models). term forecasting. for large datasets or high-
− Includes regularization to − May overfit if not properly frequency data.
prevent overfitting. regularized or if data is noisy. − Can model complex time- − Limited to univariate time-
− High predictive accuracy. − Computationally expensive series patterns effectively. series data (cannot
for very large datasets. incorporate external
− Scalable and efficient. − Struggles with extrapolation predictors).
beyond the training data ​ − Sensitive to outliers and
range. missing data.
− Supports parallel processing − May not perform well with
for faster training. very sparse data.
− Can handle both categorical ​ well as decreased traffic levels because of social gathering prohibitions,
and numerical features are the main causes of this decline (Adanu et al., 2021; J. Lee et al.,
effectively.
2023; Li & Zhao, 2022). The pattern of weekends continuing to be the
− Provides feature importance ​
scores for interpretability. most dangerous persisted despite the general drop. Even though the
Poisson − Specifically designed for − Assumes mean equals number was fewer than in previous years, Saturdays continued to have
Regression count data (e.g., injury variance (may not hold for the most deaths, while Sundays showed a steady trend. This highlights
frequency). overdispersed data). the inherent risks associated with traveling at the weekends, even when
− Interpretable coefficients. − Limited ability to capture
complex, non-linear
there is less traffic.
relationships.
− Simple and computationally − Not suitable for continuous 4.2. Severity level
efficient. outcomes (e.g., injury
severity).
− Works well for data following − Requires data
Fig. 4 shows the distribution of significant non-fatal road injuries
Poisson distribution. transformation for non- across different road user categories in Australia, categorized by hos­
Poisson distributed count data. pitalization length.
− Easy to implement and − Sensitive to outliers in Car drivers and passengers exhibit similar trends, with the majority
widely supported in statistical predictors.
needing one-day or fewer hospital stays (60 % and 62 %, respectively),
software.
− Suitable for small datasets. − May struggle with zero- while smaller numbers require two to five days (24 % and 23 %) or six or
inflated data (excessive zeros). more days (16 % and 15 %). Motorcyclists and pedal cyclists had a
− Provides confidence intervals ​ slightly greater percentage of moderate and lengthy hospital admissions,
for coefficients. with 31 % hospitalized for 2–5 days and 24 % for six or more days,
Ensemble − Combines strengths of − Computationally expensive
compared to 30 % and 11 % for the same durations, respectively. Pe­
Regression multiple models (e.g., XGBoost due to combining multiple
and Poisson Regression). models. destrians had the greatest proportion of lengthy hospitalization (33 %
− Reduces overfitting and − More complex to implement for six or more days), indicating more serious injuries in this group than
improves generalization. and interpret. others. Overall, the data shows that injury severity and recovery time
− Can handle both discrete and − Requires careful selection
vary by road user category, with pedestrians and motorcyclists suffering
continuous outcomes and weighting of base models.
effectively.
the most serious injuries.
− Robust to noise and outliers − May suffer from “black box”
in data. nature, reducing 4.3. Gender, age, and location impacts
interpretability.
− Flexible in integrating diverse − Risk of overfitting if base
models (e.g., linear, tree-based, models are too complex or not
There is a gender difference in non-fatal hospitalizations connected
etc.). properly regularized. to traffic crashes, with around 64.8 % of men and 35.2 % of females
− Often achieves state-of-the- − Requires significant (Fig. 5). 5.56 days is the average amount of time that a crash case re­
art predictive performance. expertise to tune and optimize. quires hospitalization. In comparison to females (5.12), males had a
ARIMA − Effective for time-series − Limited to univariate time-
significantly greater mean number of bed days per case (5.90), which
forecasting. series data (cannot
incorporate external may indicate more serious injuries or longer recovery times. According
predictors). to age analysis, hospitalizations are most common among those aged
− Handles non-stationary data − Requires stationary data 40–64 (32.7 %), followed by people aged 26–39 (24.0 %). Hospital stays
through differencing. (may need differencing or are higher for older age groups: the greatest mean bed days per case
transformation).
− Simple and interpretable for − Struggles with long-term
(9.33) are seen in patients 75 years of age and beyond, followed by 65 to
univariate time-series data. forecasting due to error 74 years of age (7.72). This implies that elderly people suffer from more
accumulation. serious injuries or need more time to recover. The youngest group
− Works well for short-term − Not suitable for data with (8–16) has the lowest mean bed days per case (3.46), which suggests
forecasting. strong seasonal patterns
that their injuries are less serious and that their recovery periods are
(requires SARIMA).
− Widely used and supported in − Sensitive to outliers and shorter. Age-related increases in crash severity are seen, with the
statistical software. missing data. greatest BeddystoCC ratio (5.99) in the 40–64 age range. Because there
SARIMA − Extends ARIMA to handle − More complex than ARIMA are fewer deaths in the 65 + age group than in the 40–64 age group,
seasonal patterns in time-series due to additional seasonal older people have longer hospital stays but lower severity indices.
data. parameters.
Regional regions account for 29.6 % of hospitalizations, whereas
metropolitan areas account for 70 % of them. Compared to victims of
urban areas (5.33), individuals from regional crashes require somewhat
more bed days per case (5.85), which may indicate more serious injuries

5
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Fig. 2. Traffic crash injuries in Australia between 2013 and 2021.

or longer recovery times. The modest BeddystoCC ratio (5.33) of colli­ numerous fractured bones, and brain trauma.
sions in large cities suggests less serious wrecks and shorter hospital Drivers of cars engaged in crashes with stationery or permanent
stays. As opposed to big cities, rural locations had a higher BeddystoCC objects suffer injuries that are rather severe, requiring an average of 7.45
ratio (5.85), which may indicate lengthier stays because of restricted hospital bed days per crash event. This is probably because crashes with
access to healthcare and a greater total crash severity. stationary objects like trees, poles, or concrete barriers entail high
speeds and forces that frequently cause casualties including head
4.4. Collision types and their difference in frequency and severity trauma. It is crucial to understand that these results might change based
on several variables, including the exact kind of crash (side vs. frontal
The crosstabulation of non-fatal hospitalization and severity for contact), the speed of impact, and the usage of safety gear (seatbelts,
various car make and collision types are displayed in Fig. 6. It shows helmets). Geographical considerations may also have an impact on the
intriguing parallels and divergences in crash trends across different types and severity of collisions, such as the primary road types (rural vs.
types of road users. The blue circles represent the frequency of injuries, urban).
while the red circles indicate the severity of injuries. The darker shades
of blue or red correspond to higher values, meaning a deeper hue sig­ 5. Machine learning result
nifies greater frequency or severity. Additionally, the size of the circles
reflects the magnitude of either injury frequency or severity. Thus, While the comparative analyses in the previous section offer insights,
larger and darker blue circles indicate a higher frequency of injuries, a deeper analysis is needed, and we used various algorithms to explore
whereas larger and darker red circles represent greater injury severity. crash factors and compare algorithmic effectiveness, to predict crashes,
Car collisions involving other cars, pick-up trucks, or vans are common and thus inform targeted interventions. Fig. 7 shows a flowchart and the
for both drivers and passengers (about 55 % for drivers and 54 % for steps that have been taken for machine learning analysis.
passengers). Furthermore, both groups experience a considerable num­
ber of crashes with immovable objects, roughly 24 % for drivers and 21 5.1. Hyperparameter tuning
% for passengers. Passengers in cars are, nevertheless, marginally more
often than drivers to be involved in non-collision transport crashes (18 % Hyperparameter tuning is the process of optimizing different pa­
vs. 13 %). Both motorcyclists and pedal bikers are quite concerned about rameters that govern the algorithm’s behavior (Alibrahim & Ludwig,
non-collision transport events. 2021). While random search selects hyperparameters at random from a
Collisions with heavier and larger cars also disproportionately injure given distribution, grid search tries every conceivable combination of
pedestrians and pedal bikers, frequently leading to severe injuries and hyperparameters within a certain range (Bergstra & Bengio, 2012). It is
extended hospital admissions. Remarkably, compared to other vehicles, crucial to define an appropriate initial search space for hyperparameter
pedestrians are more likely to sustain serious injuries in crashes with optimization, as this significantly influences the model’s performance.
trams. This emphasizes the need for specific safety measures, like better Table 2 lists the selection and description of the hyperparameter set that
pedestrian infrastructure near tram lines and more public awareness produced the best results on the validation set. For further reading on
campaigns emphasizing safe pedestrian behavior around trams, to pro­ hyperparameter tuning and its impact on model performance, we
tect this vulnerable road user group, particularly in Australian cities recommend consulting (Alibrahim & Ludwig, 2021; Bergstra & Bengio,
with tram networks’ urban cores. These results highlight how crucial it 2012). The following explanation outlines the hyperparameter adjust­
is to consider the unique requirements and vulnerabilities of various ment for Poisson regression and XGBoost techniques. The key hyper­
road users when creating strategies and treatments for road safety. The parameters include (Table 3):
top three crash types with the highest degree of crash injury severity, as
determined by the statistics, are: Pedestrians who get hit by buses or 1) Number of Estimators: This parameter specifies the total number of
large public vehicles sustain the most serious injuries, averaging 13.08 trees in the ensemble model. A higher number of estimators can lead
hospital bed days per crash event. This is probably because of the large to better performance but may also increase the risk of overfitting. It
weight and size disparity between big trucks and pedestrians. Motor­ Set to 500, While a higher number can improve performance, it is
cyclists who are involved in crashes with buses or large vehicles suffer important to monitor for potential overfitting.
very serious injuries, averaging 12.69 hospital bed days per mishap case. 2) Max Depth: This parameter determines the maximum depth of each
Because there are no barriers to protect them and motorcycle riders are tree in the forest. Deeper trees can capture more complex patterns in
exposed, they are more vulnerable to injuries including road rash, the data but may also lead to overfitting. It Configured at 10, this

6
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Fig. 3. Crash injury trends by month and year, frequency (a); severity (b).

parameter determines the maximum depth of each tree. A depth of 4) Regularization Penalty (L1 or L2): Regularization techniques help
10 allows the model to capture complex patterns without excessively prevent overfitting by penalizing large coefficients. L2 regularization
increasing the risk of overfitting. discourages large weights while allowing all features to contribute,
3) Learning Rate (Eta): The learning rate controls the contribution of whereas L1 regularization can result in sparse models by setting
each tree to the final model. A typical range for the learning rate is some coefficients to zero. We have chosen L1 regularization to pre­
between 0 and 1, with lower values often leading to better general­ vent overfitting by promoting sparsity in the model. This approach
ization by requiring more trees to converge to the optimal solution. helps in setting some coefficients to zero, effectively selecting rele­
The learning rate is set to 0.05. This value controls the contribution vant features.
of each tree to the final model, balancing convergence speed and 5) Solver: The algorithm used to optimize the loss function during
generalization. A lower learning rate typically enhances model per­ training. Common solvers for Poisson regression include Newton
formance by requiring more trees to reach optimal results. Conjugate Gradient (NCG) and Limited-memory Broyden-Fletcher-
Goldfarb-Shanno (LBFGS). The optimization algorithm selected is

7
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Fig. 4. Length of stay as the severity level across different road user categories.

Fig. 5. The distribution of injury frequency and severity against gender, age, and location.

LBFGS, which is effective for handling large datasets and optimizing achieves the best performance (NMSE = 0.24, IOA = 0.96, R2 = 0.87)
the loss function during training. compared to XGBoost (NMSE = 0.28, R2 = 0.85) and Poisson (NMSE =
0.31, R2 = 0.83), with the lowest AIC (7587) and BIC (7837). For pedal
cyclists, the Ensemble model has superior performance (lower NMSE of
5.2. Predictive performance of regression algorithms
0.44 and higher R2 of 0.69) compared to XGBoost (NMSE = 0.59, R2 =
0.59) and Poisson (NMSE = 0.42, R2 = 0.70), although Poisson has a
The results of the ensemble model, Poisson, and XGBoost regressions
slightly better R2 but higher AIC/BIC. For pedestrians, the Ensemble
for non-fatal hospitalization frequency and the severity target variable model has better NMSE (0.32) than XGBoost (0.44) and is comparable to
(BeddystoCC) are discussed here. Using five key criteria, an evaluation
Poisson (0.30), with higher R2 (0.74) than XGBoost (0.65) but slightly
of the three models (XGBoost, Poisson, and Ensemble) shows that the lower than Poisson (0.75), and the lowest AIC/BIC (3131/3357).
Ensemble model is the most accurate and consistent at predicting hos­
Turning to Severity analysis, the Ensemble model again demon­
pitalization rates for different types of road users (Table 4). strates superior performance across road users. For car drivers, the
For Hospitalization analysis, we compare XGBoost, Poisson, and the
Ensemble model leads with NMSE = 0.20, IOA = 0.96, and R2 = 0.85,
Ensemble model using NMSE, IOA, R2, AIC, and BIC. For car drivers, the while Poisson performs worse (NMSE = 0.32, R2 = 0.75), and the
Ensemble model excels with the lowest NMSE (0.20), highest IOA
Ensemble model has the lowest AIC (17646) and BIC (17900). For car
(0.96), and highest R2 (0.85), alongside the lowest AIC/BIC (9489/
passengers, the Ensemble model outperforms XGBoost and Poisson with
9744), while Poisson struggles with higher NMSE (0.35) and lower R2
lower NMSE (0.35) and higher R2 (0.60), and the lowest AIC/BIC
(0.75). For car passengers, the Ensemble model again leads with lower
(13034/13286). Among motorcyclists, the Ensemble model achieves the
NMSE (0.34) and higher R2 (0.61) compared to XGBoost (NMSE = 0.43,
best results (NMSE = 0.25, IOA = 0.96, R2 = 0.87), compared to
R2 = 0.53) and Poisson (NMSE = 0.42, R2 = 0.52), and significantly
XGBoost (NMSE = 0.29, R2 = 0.85), with the lowest AIC/BIC (14851/
lower AIC/BIC (5341/5593) compared to XGBoost (6013/6265) and
15100). For pedal cyclists, the Ensemble model performs as well as
Poisson (6040/6292). Among motorcyclists, the Ensemble model

8
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Fig. 6. Collision types for different road users: (a) frequency, (b) severity.

Poisson (NMSE = 0.44, R2 = 0.69), but has lower AIC (16308) and BIC different road users (e.g., car drivers, passengers, motorcyclists, cyclists,
(16562). For pedestrians, the Ensemble model performs equally well as and pedestrians). According to (Lundberg & Lee, 2017), SHAP values are
Poisson (NMSE = 0.32, R2 = 0.74), but has lower AIC (7317) and BIC a unique metric that explains the output of machine learning models and
(7542). In summary, for Hospitalization, the Ensemble model consis­ sheds light on the relative contributions of each feature. The Regression
tently outperforms XGBoost and Poisson, especially for car passengers model’s mean SHAP values for the frequency and severity across five
and motorcyclists. For Severity, the Ensemble model remains the best road user groups are shown in Fig. 8. Based on these figures, we can
option, significantly outperforming others for motorcyclists and car draw some comparative insights by contrasting the factors influencing
drivers. While Poisson performs comparably for pedestrians and pedal crash frequency and crash severity across different road users in
cyclists, the Ensemble model is more efficient. Overall, the Ensemble Australia. The comparisons reveal important nuances in how different
model demonstrates the best performance across all metrics. factors affect the likelihood and severity of crashes depending on the
type of road user involved.
5.3. SHAP values For motorcyclists, car drivers, and pedal cyclists, being male is the
most significant factor influencing crash frequency. This aligns with data
For answering the second RQ of how do demographic factors, colli­ showing that male drivers and riders tend to exhibit riskier driving be­
sion types, and environmental conditions differentially influence crash haviors. In contrast, car passengers and pedestrians show a more
frequency versus crash severity, and how do these influences vary across balanced gender influence, with males having less weight in predicting
different road user groups in Australia, we have used the SHAP (SHapley crash frequency. The influence of being male remains significant for
Additive exPlanations) values. This question reflects the two layers of motorcyclists, pedal cyclists, and car drivers, indicating that males are
analysis in your study: (1) the contrast between factors affecting crash not only more involved in crashes but are also at a higher risk for severe
frequency versus severity, and (2) the variations in these factors among outcomes. However, for pedestrians, while males still have higher crash

9
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Fig. 7. The flowchart and the steps taken for machine learning analysis.

reflecting their higher risk-taking behaviors and vulnerability in high-

Table 3
speed collisions.
Regression models’ parameters.
For car drivers, car passengers, and pedal cyclists, collisions
Model Selected values for Possible values for parameters involving other vehicles (CntrprtCrTrkVn) and stationary objects
parameters
(CntrprtFxdStnry) are major contributors. Motorcyclists frequently
XGBoost max_depth = 10 Maximum depth of tree= experience non-collision incidents (CntrprtNClsn), likely due to loss of
n_estimators = 500eta (4,8,10,16,32)Number of estimators= control or single-vehicle crashes, which are prevalent on rural roads.
(learning rate) = 0.05 (1000,…,200,100)Learning rate=
(0–1)
Heavy vehicles (CntrprtHvyVhcl) have a substantial impact on severity
Poisson Solver = LBFGS Solver= (NCG, LBFGS)Penalty for pedestrians, car drivers, and motorcyclists, as these collisions are
Regression Regularization (L1 or L2) often fatal or lead to severe injuries. For pedestrians, severity is also
Penalty = L1 influenced by collisions with pedal cyclists (CntrprtPdlC), highlighting
the vulnerability of pedestrians in diverse crash scenarios.
severity, older age groups (Age75plus) and heavy vehicles have a larger Crashes in major cities are more common for pedal cyclists, car
impact on injury severity. passengers, and pedestrians due to higher traffic density and exposure in
Younger individuals (Age1725) are more frequently involved in urban environments. For motorcyclists, crashes are more frequent in
crashes across all categories, especially car drivers and motorcyclists. regional areas, where higher speeds and road conditions may increase
This reflects the inexperience and higher risk-taking behavior common the risk. Severity increases in regional areas for motorcyclists and car
in younger drivers. For car passengers, both younger (Age1725) and drivers due to higher speed limits and less immediate access to emer­
middle-aged individuals (Age4064) show higher crash frequency, gency services. For pedestrians, urban settings still influence severity,
possibly due to higher exposure rates. Older age groups (Age75plus) but heavy vehicle involvement remains a critical factor regardless of
significantly impact severity for pedestrians and pedal cyclists, as these location.
groups are more vulnerable to serious injuries. Young drivers (Age1725) Based on the above result, males are at a higher risk of both frequent
are associated with severe crashes for car drivers and motorcyclists, and severe crashes, particularly when operating vehicles (car drivers,
motorcyclists, cyclists). For passengers and pedestrians, gender plays a

10
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Table 4
Performance metrics for XGBoost, Poisson, and ensemble model for crash frequency and severity.

less dominant role in predicting severity, with age and vehicle type help achieve the goal of zero fatalities.
becoming more critical. Age is a critical factor, but its influence varies: A comparison of the accuracy of several models in predicting the
younger individuals drive up crash frequency, especially when in control number of deaths from traffic crashes is shown in Table 6, which shows
of vehicles, whereas older individuals are more affected by crash that the SARIMA model obtains the lowest mean absolute error (MAE) of
severity, particularly when they are passengers or pedestrians. This type all the models tested (12.276), suggesting a reduced average discrep­
of collision significantly affects both frequency and severity. Non- ancy between the model’s predicted and actual fatality counts. None­
collision incidents are more common for motorcyclists, while heavy theless, the ARIMA model appears to be more adept at capturing greater
vehicle collisions drastically increase severity for all road users due to variances between expected and actual values, as seen by its lower RMSE
the increased force of impact. The frequency of crashes is more influ­ (73.0722) and MSE (5339.549). Because of its perhaps simpler struc­
enced by urban exposure, whereas severity is heavily impacted by the ture, the ARIMA model (366.614) is preferred by the AIC values. The
risks associated with rural driving, especially for vulnerable road users optimal model may ultimately rely on whether we consider model
like motorcyclists and car drivers. complexity (AIC) and give priority to minimizing average errors (MAE)
or capturing greater variances (MSE, RMSE). The Autoregressive model
5.4. Predictive performance of forecasting algorithms is a daily model from the COVID-19 period in 2019. Training data was
used from the beginning of 2020 to the end of 2022, while the first two
In this section, we discuss forecasting algorithm results for road crash months of 2023 were used for testing. The model then made predictions
fatalities in Australia over time. Fig. 9 shows ARIMA (Fig. 9a) and for the following 10 months. The resulting MAE of 2.16 and RMSE of
SARIMA (Fig. 9b) results forecasting the number of fatalities. Also, a 2.72 indicate a reasonable fit to the data, suggesting the model’s po­
forecast for the COVID 2019 period by an Autoregressive time-series tential for forecasting future values, although further refinement and
model was developed to predict values based on daily data (Fig. 9c). validation with additional datasets may be beneficial to improve pre­
The data used for this part was BITRE fatality data from 1989 to 2024 dictive performance. A monthly comparison of actual and predicted
(BITRE, 2024). The graph shows a general downward trend in Austral­ values during the test period showed 105 actual versus 119 predicted for
ia’s traffic fatalities. the first month, and 88 actual versus 85 predicted for the second month.
Table 5 provides a few forecast samples. This table compares the
ARIMA and SARIMA models’ predictions for road deaths in Australia 6. Discussion
between 2020 and 2050. Although neither model foresees reaching the
zero deaths target of the Australian Road Safety Vision 2050, both This study investigates road injuries using machine learning to
expect a continuing drop in fatalities. By 2050, the number of deaths is analyze BITRE data, revealing critical patterns that call for targeted
expected to drop from 1034 in 2020 to 525 in the ARIMA model and interventions in road safety.
from 1043 in 2020 to 514 in the SARIMA model.
These estimates demonstrate that the goal of zero deaths will not be 6.1. Hospitalization trends and vulnerable road users
achieved in the case of business as usual. More investigation and effort
are required to meet the Vision 2050 target. This includes investigating Non-fatal hospitalization rates increased between 2011 and 2019,
shifts in safety technology, such as autonomous vehicles and improved however, they sharply decreased in 2020 because of COVID-19 lock­
road infrastructure, and promoting behavioral changes among road downs (Soltani et al., 2023a; Soltani et al., 2023b; Soltani et al., 2023b).
users through education and stricter law enforcement. Integrating these Rebounding in 2021, however, points to the necessity of caution as
factors into future models could provide more accurate forecasts and traffic returns to normal. 2020 will see longer hospital admissions due to

11
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Fig. 8. Mean SHAP values for (a) hospitalization and (b) severity target variables across five road user’s categories.

12
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Fig. 8. (continued).

more serious collisions, which highlight aspects such as improved car outcomes due to their physical frailty, making age a key determinant of
safety, medical advancements, and changed driving habits during crash severity in these groups. The analysis highlights that collision
lockdowns. The higher hospitalization rates among male drivers un­ types have distinct effects on crash frequency and severity for different
derscore the necessity of gender-specific initiatives. The age group of road users. Collisions involving cars, trucks, or vans (CntrprtCrTrkVn)
40–64 had the greatest injury rate, followed by those of 26–39, therefore are a leading cause of both frequent and severe crashes for car drivers,
more study is needed to understand their driving habits. Due of the car passengers, pedal cyclists, and pedestrians. For motorcyclists, non-
severity of their injuries, older folks (75+) stayed longer. March 2021, collision incidents (CntrprtNClsn), such as single-vehicle crashes,
2017, and 2019 are examples of risky months that can help direct significantly contribute to crash frequency, especially in regional areas
enforcement and awareness initiatives. While pedestrian incidents de­ where road conditions and higher speeds present additional risks.
mand improved infrastructure and public awareness efforts, crashes Conversely, collisions with heavy vehicles (CntrprtHvyVhcl) are a crit­
involving large vehicles or stationary objects highlight the need for safe ical factor in increasing severity across all road users, particularly pe­
driving around these types of vehicles. destrians, who are at a greater risk of fatal outcomes when involved in
such crashes.
The influence of regional versus urban environments further differ­
6.2. Model performance and feature selection
entiates the factors affecting crash frequency and severity. Urban set­
tings (MajorCities) are associated with a higher frequency of crashes for
When it came to forecasting hospitalisation rates across various road
pedal cyclists, car passengers, and pedestrians, likely due to the higher
user categories, the Ensemble model proved to be the most dependable
density of road users and complex traffic environments. However, mo­
and precise. Based on measures such as NMSE, IOA, and R-squared
torcyclists and car drivers in regional areas face increased crash severity,
values, its performance was superior to those of other models. The
driven by factors such as higher speed limits, road conditions, and
Ensemble model consistently performed well across a range of assess­
delayed access to emergency services. The impact of the COVID period
ment measures by skillfully utilizing the advantages of several models.
(CovidPrd) has also been significant across all road users, influencing
The capacity of XGBoost, Poisson Regression, and the Ensemble model
both crash frequency and severity. For car drivers and passengers,
to predict crash severity was also evaluated. With the lowest NMSE,
changes in traffic volumes and behaviors during the pandemic led to
greatest IOA and R-squared values, and lowest AIC and BIC values across
fluctuations in crash patterns. The severity of crashes during this period
all crash severity categories, the Ensemble model was superior once
was notably higher for pedestrians and cyclists, possibly due to
more. The Ensemble model consistently outperforms the XGBoost and
increased speeds on less crowded roads and a surge in inexperienced
Poisson models, demonstrating its superiority in crash severity
road users engaging in cycling for recreation during lockdowns.
prediction.

6.3. Differences between crash frequency and severity, for different road 6.4. Policy implications
users
The findings of this study provide critical insights into the factors
SHAP values provided important insights into the variables affecting influencing crash frequency and severity across different road user
hospitalization rates and crash severity. The results show that gender categories in Australia. These insights underscore the need for targeted,
and age play a substantial role across all road users, but with varying evidence-based road safety policies that address the unique risks faced
effects. For car drivers, motorcyclists, and pedal cyclists, being male is a by specific demographic groups, road user types, and geographic re­
dominant factor in increasing both crash frequency and severity. This gions. This is especially pertinent in the context of Australia’s commit­
aligns with Australian road safety data indicating that male drivers and ment to achieving zero road fatalities by 2050, an ambitious target that
riders are more likely to engage in risky driving behaviors. In contrast, remains unmet according to recent projections. Below, we outline spe­
for car passengers and pedestrians, gender is less influential on fre­ cific policy recommendations informed by the study’s findings, aligned
quency and severity, with other factors like age and collision types with Australia’s National Road Safety Strategy (NRSS) and National
having a more significant impact. Road Safety Action Plan (NRSAP).
Age is another critical factor, though its influence varies depending Targeted Interventions for High-Risk Groups: Male drivers and riders
on the road user category. Younger individuals (Age1725) are consis­ are disproportionately involved in severe crashes, particularly in
tently associated with higher crash frequency, particularly for car regional areas (Rabelo-da-Ponte et al., 2021). To address this, Australia
drivers and motorcyclists, reflecting their higher risk-taking behavior should intensify enforcement of speeding and drink-driving laws, espe­
and inexperience. On the other hand, older individuals (Age75plus), cially in high-risk regions. Public awareness campaigns, such as the
especially pedestrians and pedal cyclists, are more vulnerable to severe Towards Zero initiative, should be tailored to resonate with male

13
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Fig. 9. Forecasting the number of fatalities: (a) arima model (b) sarima model (c) covid lockdown period.

drivers, emphasizing the consequences of risky behaviors like speeding pedestrian barriers, and longer traffic light intervals in areas with high
and impaired driving. Younger drivers (aged 17–25) are over­ elderly populations. Vehicle safety standards should also be updated to
represented in crash frequency, particularly in urban areas. Graduated include features like pedestrian detection systems to protect older road
Licensing Systems (GLS) should be strengthened by incorporating users (Jurewicz & Excel, 2016).
mandatory training on risk perception and defensive driving. Addi­ Regional vs. Urban Road Safety Strategies: The study highlights that
tionally, incentives for adopting advanced driver assistance systems crashes in regional areas tend to be more severe due to higher speeds,
(ADAS) in vehicles driven by young people could mitigate crash risks poorer road conditions, and delayed access to emergency services
(Keall et al., 2022). Older pedestrians and cyclists (aged 75 + ) are more (Jurewicz & Excel, 2016). Investments in road infrastructure, such as
vulnerable to severe injuries due to physical frailty. Policies should focus widening shoulders, installing guardrails, and improving lighting on
on improving pedestrian infrastructure, such as dedicated crossings, rural roads, are critical. Additionally, expanding telemedicine and

14
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Table 5 higher severity indices in regional areas underscore the need for
Forecasting the number of fatalities by ARIMA and SARIMA models. improved post-crash care. Investments in regional healthcare infra­
Forecasting technique Interval Number of fatalities structure, including trauma centers and air ambulance services, are
essential to reduce fatalities and improve recovery outcomes (Jurewicz
5 years intervals by arima 2020 1034
2025 1016 & Excel, 2016). Older road users and motorcyclists, who experience
2030 930 longer recovery times, should have access to specialized rehabilitation
2035 911 programs. Policies should also support research into injury prevention
2040 750 technologies, such as improved helmet designs for motorcyclists
2045 672
2050 525
(Wolkow et al., 2019).
monthly interval for 2050 January 43 Intergovernmental Collaboration and Funding: Achieving uniform road
by sarima February 41 safety measures across Australia requires stronger collaboration be­
March 44 tween federal, state, and local governments (Fleisher et al., 2016; Green
April 46
et al., 2022). A coordinated approach to funding, resource allocation,
May 42
Juan 39 and policy implementation is essential to address regional disparities in
July 39 road safety outcomes. Continued investment in road safety research,
August 45 particularly in emerging areas like autonomous vehicles and ITS, is
September 41 critical to achieving the goals of the NRSS and Vision 2050 (Fildes,
October 43
November 40
2015).
December 47 Advancing Australia’s Zero Fatality Vision: Despite significant im­
autoregresive covid period 10 months of 2023 974 provements in road safety over the past decades, recent projections
suggest that Australia’s zero fatality target by 2050 is unlikely to be
achieved. More than 1,300 people die annually on Australian roads,
emergency response capabilities in regional areas could reduce post-
with fatalities expected to persist at significant levels over the next 25
crash fatalities. In contrast, urban settings experience higher crash fre­
years (Soltani et al., 2025). Clear trends indicate that weekends,
quencies, particularly for cyclists and pedestrians. Policies should pri­
particularly Saturdays, see higher fatality rates due to increased social
oritize the development of Intelligent Transport Systems (ITS), such as
travel and alcohol consumption. Additionally, December remains the
real-time traffic monitoring and adaptive traffic signals, to manage
deadliest month, driven by holiday-related travel surges. Geographic
complex urban traffic environments (Davey, 2015). Dedicated cycling
disparities persist, with urban areas such as Sydney and Melbourne ac­
lanes and pedestrian zones should be expanded in major cities like
counting for 35–40 % of fatalities due to high traffic volumes and
Sydney, Melbourne, and Brisbane.
complex intersections, while rural regions experience more severe
Vehicle and Infrastructure Safety Improvements: Collisions involving
crashes due to higher speeds and poor infrastructure. To move closer to
heavy vehicles are a leading cause of severe injuries, particularly for
the Zero Fatality Vision, Australia must enforce dynamic speed limits in
pedestrians and motorcyclists. Mandating the adoption of Autonomous
high-risk zones such as school areas and holiday corridors, expand 80
Emergency Braking (AEB) and blind-spot detection systems in heavy
km/h zones on rural highways, and launch targeted gender-specific and
vehicles could significantly reduce these incidents (Mackenziea & Van
age-specific safety initiatives, including mandatory refresher courses for
den Berga, 2015; Newstead et al., 2020). Additionally, urban planning
drivers over 65. Infrastructure upgrades such as median barriers,
should prioritize separating heavy vehicle traffic from vulnerable road
improved signage, and dedicated cyclist pathways must be prioritized.
users, especially in areas with high pedestrian activity. Pedestrians are at
The adoption of autonomous vehicles and AI-driven traffic systems
high risk of severe injuries in collisions with trams, particularly in cities
should be accelerated to reduce crashes caused by human error.
like Melbourne. Infrastructure improvements, such as raised platforms
Furthermore, expanding public transport networks and subsidizing off-
and pedestrian barriers at tram stops, should be implemented. Public
peak travel options can reduce dependence on car travel, particularly
awareness campaigns should also educate pedestrians on safe behaviors
among high-risk groups.
around tram lines (Jurewicz & Excel, 2016).
Temporal and Seasonal Risk Management: The study identifies March
7. Conclusions
to May as the riskiest period for crashes, with weekends (particularly
Saturdays) being the most dangerous. Policymakers should allocate
7.1. Summary
additional enforcement resources, such as increased police patrols and
speed cameras, during these high-risk periods (Iranitalab & Khattak,
This research delivers some insights into serious road traffic injuries
2017). Public awareness campaigns should also emphasize the dangers
through the application of advanced machine learning techniques. By
of weekend driving, particularly in regional areas. The reduction in
analyzing hospitalized injuries and severe outcomes across diverse road
crashes during the COVID-19 lockdowns highlights the impact of
user groups, we have identified key demographic factors, temporal
reduced traffic volumes on road safety. Policies promoting flexible work
trends, and collision patterns, as well as assessed the accuracy of fore­
arrangements, such as remote work, could have long-term benefits for
casting algorithms.
road safety by reducing peak-hour traffic congestion (Kavalamthara
Nationwide, hospitalizations for serious crash injuries steadily rose
et al., 2019; Pammer et al., 2021).
from 2011 to 2019 before sharply declining in 2020 due to COVID-19
Post-Crash Care and Rehabilitation: The longer hospital stays and
lockdowns. Despite fewer crashes in 2020, injury severity increased,

Table 6
Performance of models in forecasting the number of fatalities.
test results Performance metrics

Model Actual Forecast MAE MSE RMSE AIC

ARIMA 1097 1034 69.701 5339.549 73.0722 366.614

SARIMA 1097 1043 12.276 260.623 16.143 3152.503
autoregresive 193 204 2.16 7.40 2.72 −
covid 2019 period

15
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

with longer hospital stays. Pedestrians suffered the most severe injuries, and emerging road and vehicle safety technologies.
with 33 % requiring ≥ 6 days of hospitalization, compared to 16 % for
car drivers, and collisions with heavy vehicles (e.g., buses) caused the 9. Declarations
highest severity (13.08 mean bed days for pedestrians). Males accounted
for 64.8 % of non-fatal hospitalizations and had longer recovery times Data and Code availability: The data for this research is sourced
(5.90 vs. 5.12 bed days for females), while elderly individuals (75 + ) from the open-source BITRE database, provided by the Commonwealth
had the longest stays (9.33 days) but lower severity indices due to higher of Australia. The codes and scripts developed for this research are
mortality rates. While 70 % of crashes occurred in urban areas, regional available from the corresponding author.
crashes had higher severity (BeddystoCC = 5.85 vs. 5.33) due to delayed AI Statement: Some AI-based English editing tools such as QuillBot,
emergency response and poorer infrastructure, with motorcyclists facing Gemini, and Grammarly are used for paraphrasing and to improve
elevated risks in regional zones. The Ensemble model outperformed English.
XGBoost and Poisson regression, achieving the lowest NMSE (e.g., 0.20
for car drivers) and highest R2 (0.85) in predicting crash frequency and
CRediT authorship contribution statement
severity, with SHAP analysis validating its robustness. SHAP analysis
identified male gender, young drivers (17–25), and collisions with cars/
Ali Soltani: Writing – review & editing, Writing – original draft,
trucks as top predictors of crash frequency, while heavy vehicles and
Validation, Methodology, Formal analysis, Data curation, Conceptuali­
rural locations disproportionately increased severity. March consistently
zation. Saeid Afshari: Writing – original draft, Methodology, Formal
had the highest crash frequency (pre- and post-pandemic), and week­
analysis. Mohammad Amin Amiri: Visualization, Software, Method­
ends (Saturdays) recorded the most fatalities pre-COVID (247 deaths in
ology, Formal analysis.
2012), while severity spiked post-July 2015. ARIMA/SARIMA models
projected fatalities declining from 1,034 (2020) to 525 (2050) but
failing to meet Australia’s Zero-Fatality Target by 2050, underscoring
Declaration of competing interest
the need for enhanced policies. Aboriginal and Torres Strait Islander
peoples in remote areas, elderly pedestrians, and young male drivers
The authors declare that they have no known competing financial
(17–25) were disproportionately affected, with remote Indigenous
interests or personal relationships that could have appeared to influence
communities facing 2–3 × higher severity rates due to healthcare and
the work reported in this paper.
infrastructure inequities.
These findings emphasize the need for targeted interventions: urban
Acknowledgments
(pedestrian infrastructure upgrades and heavy vehicle safety measures);
regional (improved emergency response and road maintenance); de­
We acknowledge the Kaurna people, and respect the Traditional
mographic (gender-specific campaigns for young male drivers and
Custodians and all Aboriginal and Torres Strait Islander peoples.
enhanced vehicle safety for the elderly); and policy (integrating machine
learning, e.g., Ensemble models, into the National Road Safety Strategy
(NRSS) to prioritize high-risk groups and collision types). The method­ References
ological shift to hospital data and advanced algorithms provides a
Abrari Vajari, M., Aghabayk, K., Sadeghian, M., Moridpour, S., 2022. Modelling the
blueprint for global road safety research. injury severity of heavy vehicle crashes in Australia. Iran. J. Sci. Technol., Trans. Civil
Eng. 46 (2), 1635–1644. [Link]
8. Research limitations and directions for future studies Adanu, E.K., Brown, D., Jones, S., Parrish, A., 2021. How did the COVID-19 pandemic
affect road crashes and crash outcomes in Alabama? Accid. Anal. Prev. 163, 106428.
[Link]
This study sheds light on the intricate nature of serious injuries Adeyemi, O., Bukur, M., Berry, C., DiMaggio, C., Grudzen, C.R., Konda, S.,
sustained in traffic crashes and advances our knowledge of crash dy­ Adenikinju, A., Cuthel, A., Bouillon-Minois, J.-B., Akinsola, O., 2023. Substance use
and pre-hospital crash injury severity among US older adults: A five-year national
namics, demographic factors, and forecasting methods. Additionally, it cross-sectional study. PLoS One 18 (10), e0293138. [Link]
highlights the necessity of evidence-based treatments to enhance patient [Link].0293138.
outcomes and road safety in Australia. Because the study uses BITRE Ahangari, H., 2015. A Comprehensive Comparative Assessment of Road Safety in Developed
Countries (Publication Number 917.). University of Connecticut].
data, questions concerning the accuracy, completeness, and represen­
Ahmed, S., Hossain, M.A., Ray, S.K., Bhuiyan, M.M.I., Sabuj, S.R., 2023. A study on road
tativeness of the data are raised. A detailed study requires more exten­ accident prediction and contributing factors using explainable machine learning
sive linked data from hospital registries and police crash records (Soltani models: analysis and performance. Transp. Res. Interdiscip. Perspect. 19, 100814.
et al., 2024; Harrison et al., 2023). Long-term trends may be missed by Alibrahim, H., Ludwig, S.A., 2021. Hyperparameter optimization: Comparing genetic
algorithm against grid search and bayesian optimization. 2021 IEEE Congress on
the 2011–2021 period, and the analysis’s dependability may be Evolutionary Computation (CEC).
impacted by data imbalances across road user groups and collision Alicioglu, G., Sun, B., Ho, S.S., 2022. An injury-severity-prediction-driven accident
severity as well as unobserved variables and confounding factors (Das prevention system. Sustainability 14 (11), 6569. [Link]
su14116569.
et al., 2023). Due to these constraints, the dataset has been consolidated Arteaga, C., Paz, A., Park, J., 2020. Injury severity on traffic crashes: A text mining with
into spatial units. an interpretable machine-learning approach. Saf. Sci. 132, 104988.
Future research should explore key areas such as regional compari­ Astarita, V., Haghshenas, S.S., Guido, G., Vitale, A., 2023. Developing new hybrid grey
wolf optimization-based artificial neural network for predicting road crash severity.
sons to assess how road infrastructure, laws, and socioeconomic factors Transp. Eng. 12, 100164. [Link]
influence crash patterns (Kamargianni et al., 2016). It is critical to focus Bergstra, J., Bengio, Y., 2012. Random search for hyper-parameter optimization. J. Mach.
on vulnerable groups when creating targeted initiatives to lower injury Learn. Res. 13 (2).
Bi, N., Sadia, H., 2023. Accident severity detection using machine learning. Int. J. Eng.
rates and improve safety, such as children, the elderly, low-income Manage. Res. 13 (3), 203–208. [Link]
communities, and those living in rural areas. By adding variables BITRE. (2023). Road Trauma Australia 2020 Statistical Summary [Link]
including weather, road conditions, and driver behavior, machine au/publications/ongoing/hospitalised-injury.
BITRE. (2024). Australian Road Deaths Database [Link]
learning can enhance collision prediction models (Grigorev et al.,
safety/fatal_road_crash_database.
2024a). Furthermore, road safety might be revolutionized by intelligent Boyagoda, L.S., Nawarathna, L.S., 2022. Analysis and Prediction of Severity of United States
transportation systems and vehicle automation (Matin and Dia, 2022; Countrywide Car Accidents Based on. In: Machine Learning Techniques 2022 7th
Soltani et al., 2021; Zali et al., 2022). To promote evidence-based policy International Conference on Information Technology Research (ICITR). [Link]
org/10.1109/ICITR57877.2022.9993371.
and increase road safety for everybody, future research must make Burke, P.J., Teame, A., 2018. Fuel prices and road deaths in Australia. Econ. Papers: J.
extensive use of advanced methodology, a wide range of data sources, Appl. Econ. Policy 37 (2), 146–161. [Link]

16
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Cabrera-Arnau, C., Prieto Curiel, R., Bishop, S.R., 2020. Uncovering the behaviour of Jurewicz, C., Excel, R., 2016. Application of a crash-predictive risk assessment model to
road accidents in urban areas. R. Soc. Open Sci. 7 (4), 191739. [Link] prioritise road safety investment in Australia. Transp. Res. Procedia 14, 2101–2110.
10.1098/rsos.191739. [Link]
Çelik, A., Sevli, O., 2022. Predicting traffic accident severity using machine learning Kamargianni, M., Li, W., Matyas, M., Schäfer, A., 2016. A critical review of new mobility
techniques. Türk Doğa Ve Fen Dergisi 11 (3), 79–83. [Link] services for urban transport. Transp. Res. Procedia 14, 3294–3303. [Link]
tdfd.1136432. 10.1016/[Link].2016.05.277.
Chen, T., He, T., Benesty, M., Khotilovich, V., Tang, Y., Cho, H., Chen, K., Mitchell, R., Kavalamthara, P.J., Cyril, S., Liu, Y., Saverimuttu, V., 2019. Road traffic accidents:
Cano, I., & Zhou, T. (2015). Xgboost: extreme gradient boosting. R package version reviewing the efficacy of road safety measures in New South Wales, Australia. WIT
0.4-2, 1(4), 1-4. Trans. Built Environ. 1, 29–39. [Link]
Cicek, E., Akin, M., Uysal, F., Topcu Aytas, R., 2023. Comparison of traffic accident Keall, M.D., Watson, L.M., Rampollard, C., Newstead, S.V., 2022. Association between
injury severity prediction models with explainable machine learning. Transp. Lett. 15 Australasian New Car Assessment Program pedestrian ratings and injury severity in
(9), 1043–1054. [Link] real-life crashes in different speed limit areas. J. Road Safety 33 (4), 32–54. https://
Connections, E. (2017). Cost of road trauma in Australia 2015 (Report for the Australian [Link]/10.3316/informit.797485501821616.
Automobile Association, Issue. E. Connections. [Link] Khorshidi, A., Ainy, E., Nazari, S.S.H., Soori, H., 2016. Temporal patterns of road traffic
content/uploads/2018/03/AAA-ECON_Cost-of-road-trauma-summary-report_Sep- injuries in Iran. Arch. Trauma Res. 5 (2). [Link]
[Link]. Kim, S., Lym, Y., Kim, K.-J., 2021. Classifying severity of senior driver accidents in
Consul, P., Famoye, F., 1992. Generalized Poisson regression model. Commun. Stat.- capital regions based on machine learning algorithms. J. Digital Convergence 19 (4),
Theory and Methods 21 (1), 89–109. 25–31. [Link]
Das, S., Tamakloe, R., Zubaidi, H., Obaid, I., Rahman, M.A., 2023. Bicyclist injury Komol, M.M.R., Hasan, M.M., Elhenawy, M., Yasmin, S., Masoud, M., Rakotonirainy, A.,
severity classification using a random parameter logit model. Int. J. Transp. Sci. 2021. Crash severity analysis of vulnerable road users using machine learning. PLoS
Technol. 12 (4), 1093–1108. [Link] One 16 (8), e0255828.
Davey, W. (2015). Evaluating the effectiveness of Intelligent Transport Systems at reducing Kumeda, B., Zhang, F., Zhou, F., Hussain, S., Almasri, A., Assefa, M., 2019. Classification
speeds through roadwork sites in Australia School of Civil Engineering and Surveying]. of road traffic accident data using machine learning. In: Algorithms 2019 IEEE 11th
[Link] International Conference on Communication Software and Networks (ICCSN).
Devlin, A., & Fitzharris, M. (2013, August 28–30, 2013). An analysis of single-vehicle [Link]
fatality crashes in Australia at various blood alcohol concentrations Proceedings of the Lee, Guldmann, J.-M., von Rabenau, B., 2023. Impact of driver’s age and gender, built
2013 Australasian Road Safety Research, Policing and Education Conference, environment, and road conditions on crash severity: a logit modeling approach. Int.
Brisbane, Australia. [Link] J. Environ. Res. Public Health 20 (3), 2338. [Link]
%20-%20Alcohol%20and%[Link]. Lee, J., Liu, H., Abdel-Aty, M., 2023. Changes in traffic crash patterns: Before and after
Duddu, V.R., Kukkapalli, V.M., Pulugurtha, S.S., 2019. Crash risk factors associated with the outbreak of COVID-19 in Florida. Accid. Anal. Prev. 190, 107187. [Link]
injury severity of teen drivers. IATSS Res. 43 (1), 37–43. [Link] org/10.1016/[Link].2023.107187.
iatssr.2018.08.003. Li, J., Zhao, Z., 2022. Impact of COVID-19 travel-restriction policies on road traffic
Duong, Q., Gilbert, H., Nguyen, H., 2024. A novel framework for crash frequency accident patterns with emphasis on cyclists: A case study of New York City. Accid.
prediction: Geographic support vector regression based on agent-based activity Anal. Prev. 167, 106586. [Link]
models in Greater Melbourne. Accid. Anal. Prev. 207, 107747. Lopez-Valdes, F.J., Duprey, S., Forman, J., Svensson, M.Y., 2022. Understanding age and
Fildes, B., 2015. Traffic medicine and road safety: the Australian perspective. J. Local sex-related differences in the biomechanics of road traffic associated injuries through
Global Heath Sci. 2015 (2), 80. [Link] population diversity analyses. Front. Bioeng. Biotechnol. 10, 869356. [Link]
Fleisher, A., Wier, M.L., Hunter, M., 2016. A vision for transportation safety: framework 10.3389/fbioe.2022.869356.
for identifying best practice strategies to advance vision zero. Transp. Res. Rec. 2582 Lundberg, S.M., Lee, S.-I., 2017. A unified approach to interpreting model predictions.
(1), 72–86. [Link] Adv. Neural Inf. Proces. Syst. 30.
Fu, W., Lee, J., 2022. Relationship between vehicle safety ratings and drivers’ injury Mackenziea, J., Van den Berga, A., 2015. Establishing a testing capability for the
severity in the context of gender disparity. Int. J. Environ. Res. Public Health 19 (10), assessment of Autonomous Emergency Braking (AEB) and Forward Collision
5885. [Link] Warning (FCW) in Australia Australasian Road Safety Conference, 1st, 2015, Gold
Green, M., Muir, C., Oxley, J., Sobhani, A., 2022. Safe System in road safety public Coast, Queensland, Australia, [Link]
policy: a case study from Victoria, Australia. IATSS Res. 46 (2), 171–180. [Link] Mackenzie/publication/283007357_Establishing_a_testing_capability_for_the_
org/10.1016/[Link].2021.11.006. assessment_of_Autonomous_Emergency_Braking_AEB_and_Forward_Collision_
Grigorev, A., Saleh, K., Ou, Y., & Mihaita, A.-S. (2024). Integrating Large Language Warning_FCW_in_Australia/links/5625dd6508aeabddac91d533/Establishing-a-
Models for Severity Classification in Traffic Incident Management: A Machine testing-capability-for-the-assessment-of-Autonomous-Emergency-Braking-AEB-and-
Learning Approach. arXiv preprint arXiv:2403.13547. doi: 10.48550/ [Link].
arXiv.2403.13547. Mandacaru, P.M.P., Andrade, A.L., Rocha, M.S., Aguiar, F.P., Nogueira, M.S.M.,
Habibzadeh, M., Ameri, M., Ziari, H., Kamboozia, N., Sadat Haghighi, S.M., 2022. Girodo, A.M., Pedrosa, A.A.G., de Oliveira, V.L.A., Alves, M.M.M., Paixão, L.M.M.M.,
Presentation of machine learning approaches for predicting the severity of accidents 2017. Qualifying information on deaths and serious injuries caused by road traffic in
to propose the safety solutions on rural roads. J. Adv. Transp. 2022. [Link] five Brazilian capitals using record linkage. Accid. Anal. Prev. 106, 392–398. https://
10.1155/2022/4857013. [Link]/10.1016/[Link].2017.06.018.
Hao, W., Kamga, C., 2017. Difference in rural and urban driver-injury severities in Mason, H.M., Leggat, P.A., Voaklander, D., Franklin, R.C., 2022. Road traffic fatalities in
highway–rail grade crossing accidents. Int. J. Inj. Contr. Saf. Promot. 24 (2), 174–182. rural and remote Australia from 2006 to 2017: The need for targeted action. Aust. J.
[Link] Rural Health 30 (2), 252–263. [Link]
Harrison, J. E., Watson, A., Flavel, J., 2023. A National Approach to Measuring Non-fatal Matin, A., Dia, H., 2022. Impacts of connected and automated vehicles on road safety
Crash Outcomes Stage 2: An Australian Approach. Star. [Link] and efficiency: A systematic literature review. IEEE Trans. Intell. Transp. Syst. 24 (3),
publications/road-safety/ap-r694-23. 2705–2736. [Link]
Hassouna, F., Pringle, I., 2019. Analysis and prediction of crash fatalities in Australia. The Mitra, S., Mahato, S. K., & Bandyopadhyaya, R. (2012). Development and application of
Open Transp. J. 13 (1). [Link] crash injury severity model in hot-spot detection Proceedings of Transportation
Heathcote, K., Devlin, A., McKie, E., Cameron, P., Earnest, A., Morgan, G., Gardiner, B., Research Board Annual Conference, Washington DC, USA. [Link]
Campbell, D., Wullschleger, M., Warren, J., 2022. Rural and urban patterns of severe View/1131042.
injuries and hospital mortality in Australia: An analysis of the Australia New Zealand Morling, D., 2015. Driving down injuries and deaths on country roads. 23(1), 58-59. doi:
Trauma Registry: 2015–2019. Injury 53 (6), 1893–1903. [Link] 10.1111/ajr.12181.
injury.2022.03.044. Newstead, S., Budd, L., & Stephens, A. (2020). The Potential Benefits of Autonomous
Henley, G., Harrison, J. E., 2016. Trends in serious injury due to road vehicle traffic crashes, Emergency Braking Systems in Australia. [Link]
Australia: 2001 to 2010. Australian Institute of Health and Welfare. pdf_file/0003/2093511/The-Potential-Benefits-of-Autonomous-Emergency-Braking-
Hojati, A. T., Charles, P., Ferreira, L., & bin Kabit, M. R. (2011). An analysis of traffic [Link].
incidents on an Australian urban road network (Vol. 21). Universiti Malaysia Sarawak. OECD. (2021). OECD Secretary-General's Report to Ministers 2021. doi: doi:
[Link] 10.1787/8cd95b77-en.
2011_Hojati_Charles_Ferreira_binKabit.pdf. Ozcan, I., Aydin, H., Cetinkaya, A., 2022. Comparison of classification success rates of
Intini, P., Berloco, N., Colonna, P., Ranieri, V., Ryeng, E., 2018. Exploring the different machine learning algorithms in the diagnosis of breast cancer. Asian Pac. J.
relationships between drivers’ familiarity and two-lane rural road accidents. A multi- Cancer Prev. 23 (10), 3287. [Link]
level study. Accid. Anal. Prev. 111, 280–296. [Link] Pammer, K., Freire, M., Gauld, C., Towney, N., 2021. Keeping safe on Australian roads:
aap.2017.11.013. overview of key determinants of risky driving, passenger injury, and fatalities for
Iranitalab, A., Khattak, A., 2017. Comparison of four statistical and machine learning indigenous populations. Int. J. Environ. Res. Public Health 18 (5), 2446. [Link]
methods for crash severity prediction. Accid. Anal. Prev. 108, 27–36. [Link] org/10.3390/ijerph18052446.
10.1016/[Link].2017.08.008. Park, J., Kwak, H., Kho, S., Kim, D., 2013. Comparison of injury severity by age group a
Jin, C.J., Li, J., Wu, C., Li, D., Jiang, R., 2024. The travel behaviors before and after case of Korean expressways. Asian Transport Stud. 2 (4), 325–334. [Link]
lockdown: Case study on Shanghai, 2022. Case Studies on Transport Policy 17, 10.11175/eastsats.2.325.
101220. [Link] Park, S.-H., Bae, M.-K., 2020. Exploring the determinants of the severity of pedestrian
Jin, C.J., Luo, Y., Wu, C., Song, Y., Li, D., 2024. Exploring the Pedestrian Route Choice injuries by pedestrian age: a case study of Daegu Metropolitan City, South Korea. Int.
Behaviors by Machine Learning Models. ISPRS International Journal of Geo- J. Environ. Res. Public Health 17 (7), 2358. [Link]
Information 13 (5), 146. [Link]

17
A. Soltani et al. Case Studies on Transport Policy 20 (2025) 101440

Paul, A.K., Boni, P.K., Islam, M.Z., 2022. A data-driven study to investigate the causes of Soltani, A., Azmoodeh, M., Qadikolaei, M.R., 2023a. Road crashes in Adelaide
severity of road. In: Accidents 2022 13th International Conference on Computing metropolitan region, the consequences of COVID-19. Journal of Transport & Health
Communication and Networking Technologies (ICCCNT). [Link] 30, 101581. [Link]
ICCCNT54827.2022.9984499. Soltani, A., Harrison, J.E., Ryder, C., Flavel, J., Watson, A., 2024. Police and hospital
Prasad, V., Vellivel, P., Chandru, R., 2022. Machine Learning Algorithms to Model Crash data linkage for traffic injury surveillance: A systematic review. Accident Analysis &
Severity and Collision (s). In: 2022 IEEE North Karnataka Subsection Flagship Prevention 197, 107426. [Link]
International Conference (NKCon). [Link] Soltani, A., Mansourihanis, O., RoohaniQadikolaei, M., Zaroujtaghi, A., 2024. Two
NKCon56289.2022.10126528. decades of geospatial evolution: tracing the analytical journey towards data-driven
Rabelo-da-Ponte, F.D., Scherer, J.N., Roglio, V., Borges, E.N., Galland, F., Sousa, T., road crash prevention. Applied Spatial Analysis and Policy 17 (3), 1301–1334.
Passos, I.C., Von Diemen, L., Kessler, F., Pechansky, F., 2021. From phone use to [Link]
speeding and driving under influence: identifying clusters of driving risk behaviors Soltani, A., Azmoodeh, M., & Qadikolaei, M. R., 2023b. Post COVID-19 transformation in
as an opportunity for targeted interventions. J. Psychiatr. Res. 143, 556–562. https:// the frequency and location of traffic crashes involving older adults. Transportation
[Link]/10.1016/[Link].2020.11.025. Research Record. Volume 2678, Issue 12. doi:10.1177/03611981231163866.
Razi-Ardakani, H., Ariannezhad, A., Kermanshah, M., 2018. A study of sex differences on Soltani, A., Qadikolaei, M.R., 2024. Space-time analysis of accident frequency and the
road crash severity. In: Proceedings of the 3rd International Conference on Civil, role of built environment in mitigation. Transport policy 150, 189–205. [Link]
Structural and Transportation Engineering (ICCSTE’18). [Link] org/10.1016/[Link].2024.02.006.
iccste18.113. Steiner, W., Leisch, F., Hackländer, K., 2014. A review on the temporal pattern of
Rusli, R., Masirin, M.I.M., Hamid, N.A., 2020. Differences level of injury severity for male deer–vehicle accidents: Impact of seasonal, diurnal and lunar effects in cervids.
and female drivers in traffic crashes in Sabah. Int. J. Road Safety 1 (1), 35–41. Accid. Anal. Prev. 66, 168–181. [Link]
[Link] Theofilatos, A., Yannis, G., 2019. Exploring crash injury severity on urban motorways by
Sadeghi, P., Goli, A., 2024. Investigating the impact of pavement condition and weather applying finite mixture models. Transp. Res. Procedia 41, 480–487. [Link]
characteristics on road accidents. Int. J. Crashworthiness 1–17. [Link] 10.1016/[Link].2019.09.079.
10.1080/13588265.2024.2348269. Toğan, V., Mostofi, F., Ayözen, Y.E., Behzat Tokdemir, O., 2022. Customized AutoML: an
Sameen, M.I., Pradhan, B., 2017. Severity prediction of traffic accidents with recurrent automated machine learning system for predicting severity of construction
neural networks. Appl. Sci. 7 (6), 476. [Link] accidents. Buildings 12 (11), 1933. [Link]
Santos, K., Firme, B., Dias, J.P., Amado, C., 2024. Analysis of motorcycle accident injury Tuncalı Yaman, T., Bilgic, E., Fevzi Esen, M., 2022. Analysis of traffic accidents with
severity and performance comparison of machine learning algorithms. Transp. Res. fuzzy and crisp data mining techniques to identify factors affecting injury severity.
Rec. 2678 (1), 736–748. [Link] J. Intell. Fuzzy Syst. 42 (1), 575–592. [Link]
Seyfi, M.A., Aghabayk, K., Karimi Mamaghan, A.M., Shiwakoti, N., 2023. Modeling the Soltani, A., Azmoodeh, M., & Qadikolaei, M. R., 2023b. Post COVID-19 transformation in
motorcycle crash severity on nonintersection urban roadways in the Australian state the frequency and location of traffic crashes involving older adults. Transportation
of victoria using a random parameters logit model. J. Adv. Transp. 2023. [Link] Research Record. Volume 2678, Issue 12. doi:10.1177/03611981231163866.
org/10.1155/2023/2250590. Vadhwani, D., Thakor, D., 2023. Predictive analysis of injury severity of person across
Shahid, S., Minhans, A., Puan, O.C., Hasan, S.A., Ismail, T., 2015. Spatial and temporal angle crashes using machine learning models. Int. J. Crashworthiness 28 (4),
pattern of road accidents and casualties in Peninsular Malaysia. Jurnal Teknologi 76 523–536. [Link]
(14), 57–65. [Link] Vanitha, R., 2023. Prediction of road accidents using machine learning algorithms.
Shan, Y., Zhang, X., Lin, Y., & Ding, T. (2023). Accident Injury Severity Prediction Research Middle East J. Appl. Sci. Technol. (MEJAST) 6 (2), 64–75. [Link]
Based on Stacking Model 2023 4th International Conference on Computer Engineering jsr.2018.12.003.
and Application (ICCEA), doi: 10.1109/ICCEA58433.2023.10135421. Wolkow, A.P., Rajaratnam, S.M., Anderson, C., Howard, M.E., Mansfield, D., 2019.
Siamidoudaran, M., İşçioğlu, E., 2019. Injury severity prediction of traffic collision by Recommendations for current and future countermeasures against sleep disorders
applying a series of neural networks: the City of London Case Study. Promet-Traffic & and sleep loss to improve road safety in Australia. Intern. Med. J. 49 (9), 1181–1184.
Transport. 31 (6), 643–654. [Link] [Link]
Soltani, A., Afshari, S., Amiri, M.A., 2025. Time-series projecting road traffic fatalities in Wu, B., Wang, X., Liu, T., Dong, N., Li, Y., 2021. Exploring factors contributing to crash
Australia: Insights for targeted safety interventions. Injury 56 (3), 112166. https:// injury severity in the rural-urban fringe of the central city. J. Adv. Transp. 2021,
[Link]/10.1016/[Link].2025.112166. 1–10. [Link]
Soltani, A., Ananda, D., Rith, M., 2021. University students’ perspectives on autonomous Zali, N., Amiri, S., Yigitcanlar, T., Soltani, A., 2022. Autonomous vehicle adoption in
vehicle adoption: Adelaide case study. Case Studies on Transport Policy 9 (4), developing countries: Futurist insights. Energies 15 (22), 8464. [Link]
1956–1964. [Link] 10.3390/en15228464.
Soltani, A., Askari, S., 2014. Analysis of intra-urban traffic accidents using Zhang, Z., Jiang, Y., Chen, Z., Xiong, Y., 2021. Analyzing influencing factors of crash
spatiotemporal visualization techniques. Transport and Telecommunication 15 (3), injury severity incorporating FARS data. J. Intell. Fuzzy Syst. 41 (4), 5053–5063.
227–232. [Link] [Link]
Soltani, A., Askari, S., 2017. Exploring spatial autocorrelation of traffic crashes based on
severity. Injury 48 (3), 637–647. [Link]

18