Innovative Infrastructure Solutions (2026) 11:272

[Link]

TECHNICAL PAPER

A matrix-structured data-driven random forest framework for

predicting road accident severity under heterogeneous traffic
conditions
Tazim Ameen1 · Abdullah Ahmad2

Received: 6 January 2026 / Accepted: 13 April 2026

© Springer Nature Switzerland AG 2026

Abstract
Crash injury severity prediction in heterogeneous traffic remains challenging due to complex behavioural–vehicle interac-
tions and limited interpretability of data-driven models. This study proposes a mathematically formulated Random Forest
(RF) framework for crash severity classification along NH-44, India, integrating rigorous preprocessing, matrix-structured
model representation, and perturbation-based sensitivity evaluation. A key contribution of this work is its move beyond
isolated predictors to model interpretable driver–vehicle interactions, validated through Gini-importance and perturba-
tion sensitivity. The model demonstrates high multi-class predictive performance, achieving 88.7% test-set classification
accuracy, an OvR mean ROC-AUC of 0.94, and a macro-averaged F1 score of 0.91. Comparative testing confirms the
proposed RF framework provides superior, stable crash-severity discrimination, adding clear value for heterogeneous
traffic. This approach advances beyond black-box models by linking predictions to interpretable safety factors, enabling
targeted interventions for high-risk conditions. It provides a scalable, transferable tool for evidence-based safety decisions
in heterogeneous traffic.

Keywords Crash severity · Random Forest · Sensitivity-perturbation analysis · Heterogeneous traffic · Injury risk
modeling

Introduction burden on nations, costing many countries approximately

3% of their gross domestic product (GDP). Consequently,
Traffic accidents are a significant global concern, contrib- minimizing the severity of these crashes remains a critical
uting to increased mortality and morbidity rates, particu- objective for researchers across the globe [1]. The rising
larly in urban environments. According to the World Health trend of road accidents highlights the urgent need for effec-
Organization [1], approximately 1.35 million fatalities occur tive safety measures and strategies that can mitigate these
annually due to road traffic accidents, ranking it among the incidents. In India, the transportation sector is vital for eco-
leading causes of death globally. Road traffic crashes are nomic growth, yet it faces substantial challenges related to
particularly critical for children and young adults aged 5 to road safety. Despite possessing only 1% of the global vehi-
29 years, for whom they are the primary cause of mortality. cle population, India accounts for approximately 11% of
Road traffic crashes (RTCs) impose a significant economic worldwide fatalities from road traffic crashes (RTCs), mak-
ing it the highest contributor to RTC-related deaths glob-
ally [2]. The National Highway (NH-44) connecting the two
capital cities of Jammu and Kashmir Union Territory (UT),
Tazim Ameen
drtazimameen@[Link]; [Link]@[Link] is not immune to this issue. The Ministry of Road Trans-
port and Highways (MORTH) reports that road accidents
Abdullah Ahmad
[Link]@[Link] have increased alarmingly, with significant implications for
urban mobility and public health [3]. According to records
1
Transportation Engineering Department, Indian Institute of from the Jammu and Kashmir Traffic Department, there
Technology, Jammu 181221, India were 5036 RTCs reported in the UT between January and
2
Department of Civil Engineering, National Institute of November 2021, resulting in 713 fatalities and 6447 injuries
Technology Srinagar, Hazratbal, Srinagar 19006, India
272 Page 2 of 21 Innovative Infrastructure Solutions (2026) 11:272

[4]. Over the past five years, the region has consistently seen i. Introducing a structured mathematical representation of
over 900 fatalities per year. the Random Forest pipeline to enhance interpretability.
Over the past decade, analyzing crash severity has ii. Integrating perturbation-based sensitivity analysis with
emerged as a critical area of research, particularly in the impurity-based importance measures for robust feature
context of urban safety and sustainable transportation. attribution.
However, a significant proportion of studies have pri- iii. Validating the framework under heterogeneous high-
marily focused on high-income countries. Conventional way conditions with rigorous multi-metric performance
approaches to accident severity analysis in India largely rely evaluation.
on statistical or regression-based models, such as Smeed’s
formula [5]. While these frameworks provide aggregate
trends, they inadequately capture nonlinear, high-dimen- Literature review
sional interactions between factors such as driver behavior,
vehicle heterogeneity, road geometry, and environmental Machine learning in crash severity analysis
context. Moreover, models developed in high-income coun-
tries often underperform in Indian road conditions, where Road accident severity modelling has gained significant
heterogeneous and unstructured traffic dominates. Among attention as a means to understand and predict injury out-
these dimensions, human and behavioral factors play a comes using data-driven approaches. Various statistical and
central role in heterogeneous traffic environments, where machine learning techniques have been employed to clas-
unsafe driving actions such as overspeeding, impaired driv- sify crash severity with improved accuracy. For instance,
ing, and non-use of protective devices interact dynamically Singh et al. demonstrated the effectiveness of Random For-
with vehicle type and roadway context. Therefore, this study est (RF) models in handling high-dimensional traffic data,
explicitly treats behavioral variables not as auxiliary predic- outperforming conventional statistical approaches [6]. Sim-
tors but as a core analytical dimension within the severity ilarly, Fatima et al. highlighted the utility of machine learn-
modelling framework. ing models in identifying high-risk patterns and supporting
Recent advances in machine learning enable modeling of traffic safety interventions [7]. The integration of multiple
nonlinear and high-dimensional crash interactions, however, data sources, such as traffic volume and weather conditions,
many existing Indian severity studies lack interpretability has further enhanced model robustness. Wang et al. showed
and rigorous evaluation. There is, therefore, a pressing need that incorporating real-time traffic information significantly
for a data-driven yet mathematically rigorous classification improves prediction accuracy [8]. Comparative studies,
framework tailored to the Indian highway context. The ris- such as that by Malik et al., confirm that ensemble meth-
ing incidence of traffic accidents and fatalities in India, par- ods like RF consistently outperform other algorithms across
ticularly in urban areas and along major National Highways, multiple performance metrics [9].
underscores the urgent need for enhanced road safety mea- Recent advancements include the application of Support
sures. Despite advancements in data analysis techniques, Vector Machines (SVM), RF, and deep learning models for
there is a significant gap in the application of machine crash severity prediction [10–12]. Studies by Yan et al. and
learning approaches to classify and predict accident sever- Labib et al. reported high classification performance using
ity in low- and middle-income countries. This study aims RF-based approaches [13, 14], while Korkmaz and Erturk
to leverage machine learning, specifically Random Forest emphasized the broader applicability of machine learning
techniques, to analyze factors influencing accident severity techniques in traffic safety analysis [15]. More recently,
and provide actionable insights for policymakers and urban studies have combined deep learning with global sensitiv-
planners. By developing a classification model for catego- ity analysis to improve interpretability and feature attribu-
rizing accidents as Fatal, Grievous, or Minor, this research tion in crash severity modelling. In particular, Moussa et
seeks to contribute to improved road safety and sustainable al. demonstrated the effectiveness of variance-based global
urban planning in the region. The study contributes to the sensitivity analysis integrated with deep learning frame-
road accident severity with the following two-fold aims: (a) works [16], while Awad et al. [17] provided a compre-
to analyze the factors contributing to traffic accident sever- hensive state-of-the-art review highlighting the evolution
ity on the National Highway connecting Jammu–Srinagar; toward explainable and generalized severity modelling
(b) to develop a classification model using Random Forest frameworks. These developments indicate a clear shift in
techniques to categorize accident severity (Fatal, Grievous, recent literature toward combining predictive performance
and Minor). Beyond these objectives, the study advances with interpretability, particularly in safety-critical applica-
existing severity modeling literature by: tions where understanding model behavior is essential for
policy and engineering decisions.

13
Innovative Infrastructure Solutions (2026) 11:272 Page 3 of 21 272

Human factors and behavioral influences on significantly to crash risk [21, 22]. Environmental condi-
severity tions such as weather and visibility further affect accident
likelihood and severity. Adverse conditions, including rain
Recent research increasingly recognizes that crash severity and fog, have been shown to increase crash risk, necessitat-
emerges from the interaction between human behavior and ing adaptive traffic management strategies [23]. Historical
system-level conditions rather than isolated factors. Human modelling approaches, such as Smeed’s formulation, have
behavioral variables, such as overspeeding, impaired driv- also been used to analyze accident trends at aggregate levels
ing, and non-use of protective devices, play a central role [24], although they often fail to capture complex interac-
in determining injury outcomes. Studies adopting systems- tions in heterogeneous traffic environments.
based perspectives emphasize that unsafe driving actions In the Indian context, heterogeneous traffic composed of
significantly amplify severity when combined with high- mixed vehicle types introduces additional complexity. Vehi-
risk exposure conditions, such as vulnerable vehicle types cle type interactions, particularly between heavy vehicles
or adverse roadway environments [18, 19]. Recent litera- and vulnerable road users such as two-wheelers, signifi-
ture further reinforces that behavioral variability, risk per- cantly influence injury severity. This aligns with findings by
ception, and compliance with safety measures are critical Korkmaz and Erturk [15], who demonstrated that vehicle
determinants of severity escalation in heterogeneous traffic mass and maneuverability jointly affect crash dynamics.
environments. In particular, recent reviews and modelling Recent studies, including Awad et al. [17] and Khanum et
studies highlight the importance of integrating human fac- al. [25], further highlight that heterogeneous traffic condi-
tors within data-driven severity frameworks to capture real- tions require modelling approaches capable of capturing
world crash dynamics [17]. Evidence from heterogeneous multi-vehicle interactions and context-dependent severity
traffic studies further indicates that behavioral responses, outcomes.
including risk perception and safety-gap acceptance, are
strongly associated with severity outcomes across fatal,
grievous, and minor crash categories [20]. These findings Objective and problem formulation
reinforce the importance of incorporating behavioral vari-
ables as core predictors rather than auxiliary inputs in sever- This research aims to develop a data-driven classification
ity modelling frameworks. model for predicting road accident severity on a national
Despite notable progress in modelling and factor analy- highway by integrating feature engineering, normalization,
sis, several limitations persist in existing crash severity and ensemble machine learning. The core research objective
studies. Most prior work focuses on structured traffic envi- is formalized as a multi-class classification problem, given
ronments in developed countries, limiting applicability as:
to heterogeneous conditions. Additionally, many models
Given Condition : Z = {(xi , yi )} 
N , yi ∈ {Fatal, Grievous, Minor}
prioritize predictive accuracy without adequately address- i=1
(1)
ing interpretability or linking outputs to actionable policy
insights. Even with recent advances in explainable and T o f ind f0 (x) = argmax p(y = k |x; θ ) (2)
sensitivity-based modelling approaches [16, 17], integrated k

frameworks that simultaneously address interpretability,

heterogeneous traffic interactions, and policy relevance where xi ϵ Rp denotes the feature vector of accident attri-
remain limited. The present study addresses these gaps by butes (e.g., road condition, vehicle type, time, violation
developing a Random Forest–based classification frame- type), yi represents severity labels, and θ denotes the set
work tailored to heterogeneous Indian highway conditions, of model parameters optimized through learning. The model
integrating driver, vehicle, environmental, and roadway hypothesis is that accident severity y can be viewed as a
attributes within a structured analytical formulation. nonlinear function of heterogeneous explanatory variables,
given as in Eq. (3)
System-level and infrastructure-related factors
y = f (Road F eatures, V ehicle F actors, Human F actors, Environment, . . . )
+ δ (3)
Crash severity is influenced not only by modelling
approaches but also by system-level factors such as road- where δ captures unobserved randomness.
way design, environmental conditions, and traffic charac-
teristics. Road infrastructure plays a critical role in accident
occurrence and severity. Studies have shown that inad-
equate geometric design, lane width, and signage contribute

13
 272 Page 4 of 21 Innovative Infrastructure Solutions (2026) 11:272

Data collection and data pre-processing Such multidimensional data are vital for identifying the
underlying causes of accident severity and for developing
Study area and dataset description reliable predictive models. The NH-44 corridor reflects het-
erogeneous traffic conditions typical of Indian highways,
National Highway 44 (NH-44) is the longest highway comprising heavy vehicles (HV), cars (C), three-wheelers
in India, spanning 4,112 km (Kms). This study, however, (3W), two-wheelers (2W), and other vehicle categories (O).
focuses exclusively on its 541 km Jammu–Srinagar seg- Within the 1203 crash records analyzed, cars account for
ment in the UT of Jammu & Kashmir, a primary all-weather, approximately 37% of crash-involved vehicles, followed
mountainous, and heterogeneous traffic corridor connecting by two-wheelers at 33%, three-wheelers at 18%, and heavy
the UT’s two capitals, as shown in Fig. 1. Route alignment vehicles at 13%. These figures reflect the crash-involved
was confirmed using Google Maps, and crash severity den- modal distribution and indicate the coexistence of vulner-
sity gradients were visualized using QGIS (KDE Heatmap able users and larger vehicle classes within the corridor.
plugin) [26]. The present study is based on the official road Table 1 summarizes the key attributes used in the study,
accident dataset published by the Ministry of Road Trans- each of which represents an independent variable (feature)
port and Highways [3]. The dataset includes 1,203 accident in the machine learning framework, while the accident
records, each representing a distinct incident with multiple severity category (Fatal, Grievous, or Minor) serves as the
descriptive and categorical variables. These variables cap- target variable.
ture diverse aspects of crash events such as location, time
of occurrence, weather, and road conditions, vehicle char-
acteristics, driver attributes, and traffic control measures.

Fig. 1 Crash-severity spatial map

13
Innovative Infrastructure Solutions (2026) 11:272 Page 5 of 21 272

Table 1 Brief description of road accident dataset a robust Mahalanobis-distance method, which accounts for
Attribute Code Description the covariance structure of the multivariate dataset. For each
Location and LT Accidents are classified according to
observation Xi , the distance was computed, as shown in
Time location and time.
Weather W Accidents are classified according to
Eq. (4).
the weather.
Di 2 = (xi − µ ) Σ −1 (xi − µ )
T
Road RE Accidents are classified according to (4)
Environment the road environment.
Road Features RF Accidents are classified according to where µ and Σ denote the mean vector and covariance
road features.
matrix, respectively. Observations with Di 2 exceeding the
Junction Type JT Accidents according to junction types.
chi-square threshold χ 2 (for p variable) were flagged as
Traffic Control TCD Accidents according to traffic control at
Devices the junction. potential anomalies. Only confirmed errors were removed;
Pedestrian PI Location of pedestrian accidents valid extremes were retained to preserve the natural vari-
Infrastructure according to pedestrian infrastructure. ability of accident phenomena. Overall, missingness across
Impacting IVT Accidents are classified by the types of variables was low to moderate (all attributes < 10%). MICE
Vehicle Type impacting vehicles. was applied only to variables with missing entries, using
Age of Impact- AIV Accidents are classified by the age of
logistic or multinomial models for categorical attributes.
ing Vehicle the impacted vehicles.
Collision Type CT Accidents are classified by type of
Nearest-neighbour hot-deck imputation was used for sparse
impact/collision. categorical levels to preserve empirical distributions. Maha-
Victim Safety VSD Binary/categorical indicator of reported lanobis-distance screening was applied only to continuous
Devices safety-device usage status (e.g., helmet/ attributes (e.g., vehicle age), and was not used on label-
seatbelt use) at the time of the crash. encoded categorical variables. No records were removed
Traffic Violation TV Categorical classification of the
solely due to categorical anomaly detection.
reported traffic violation associated
with the crash event (e.g., overspeed- After cleaning, the dataset was subjected to summary
ing, red-light violation) diagnostics. A missing-value matrix visualization before
Licenced Driver LD Accidents were classified whether the and after cleaning was used to confirm data completeness
driver had a license or not. and consistency, as shown in Fig. 2. These procedures col-
lectively ensured that the final dataset was not only free
Data pre-processing and cleaning from noise and inconsistencies but also retained the intrinsic
structure of highway crash data, thereby improving both the
Prior to model development, the dataset was subjected to reliability and generalizability of the subsequent machine
a comprehensive data quality protocol to ensure analyti- learning analysis.
cal reliability and reproducibility. The procedure combined Machine learning models typically require numeri-
statistical diagnostics with algorithm-based techniques to cal representations of all input features. Since the dataset
detect inconsistencies, missing values, and outliers. Miss- included several categorical variables, each attribute was
ing entries were first profiled to identify both random and encoded into integer form using Label (Integer) Encoding.
systematic patterns. Instead of simple mean or mode substi- In this scheme, every unique category within a variable is
tution, the study adopted a Multiple Imputation by Chained mapped to a distinct integer value. This approach maintains
Equations (MICE) [27] approach, which iteratively predicts computational simplicity while preserving categorical dis-
missing values by modeling each variable as a function of tinctness. The encoded variables were subsequently verified
the others. This method preserves the underlying multi- to ensure that no artificial ordinal bias was introduced into
variate relationships among attributes and reduces bias in non-ordinal features. Feature scaling is critical for improv-
the imputed dataset. For categorical variables with low- ing model convergence and ensuring that attributes mea-
frequency categories, nearest-neighbour hot-deck impu- sured on different scales contribute equally to the learning
tation was additionally applied to maintain the original process. In this study, the Z-score normalization technique
distributional balance. Text-based inconsistencies, such as was adopted, which standardizes each feature to have zero
variations in categorical labels (Rainy, rain, Rainfall), were mean and unit variance, as shown in Eq. (5).
normalized using a controlled vocabulary derived from
MoRTH’s coding standards. The dataset was screened for xi − µ
Zi = (5)
logically impossible cases, such as a head-on collision being σ
reported with only one vehicle. Records failing such con-
straints were reviewed against the original source and cor- where xi is the original feature value, μ is the mean of that
rected when verifiable. Extreme values were detected using feature, and σ is its standard deviation. This transformation

13
 272 Page 6 of 21 Innovative Infrastructure Solutions (2026) 11:272

Fig. 2 Missing-value matrix visualization of data cells: a Before cleaning; b After cleaning

ensures that variables with larger numeric ranges (such as untouched to prevent information leakage. Figure 4 com-
vehicle age) do not dominate those with smaller scales (such pares the original, unbalanced distribution of injury sever-
as the number of control devices). The standardized dataset ity classes with the SMOTE-balanced dataset used for RF
thus provides a balanced input space for the RF model. RF model development.
models do not require feature scaling for training or conver- The data preparation process established a clean, normal-
gence, as they are not gradient-based. In this study, Z-score ized, and balanced dataset suitable for machine learning
normalization was therefore not applied to improve RF classification. Each step, from data validation to encoding
performance. Instead, normalization was introduced solely and scaling, was designed to minimize bias and enhance
to ensure numerical consistency in the subsequent analyti- model generalization. The processed dataset was then
cal and perturbation-based sensitivity formulation, where locked as the analysis-ready version used for feature trans-
feature-wise variations are compared through matrix opera- formation and model training. The resulting dataset captures
tions. This preprocessing choice does not alter tree construc- the heterogeneity of real highway crashes while providing a
tion or split decisions and does not affect the underlying robust foundation for the analytical framework described in
Random Forest learning mechanism. Label encoding was the subsequent sections.
adopted as a compact categorical representation compatible
with tree-based partitioning, and no ordinal interpretation
or distance-based operation is imposed during RF split- Investigative approach
ting. Following pre-processing, the dataset was randomly
divided into two subsets, in the ratio of 70:30, pertaining to The research adopts a data-driven and model-oriented
training and testing, respectively. The training subset was framework that integrates the stages of data collection, pre-
used to develop and optimize the model, while the testing processing, model development, and validation into a coher-
subset served to evaluate its predictive performance. ent workflow. The overall procedure, illustrated in Fig. 5,
reflects the logical sequence followed in the study, from the
Data balancing strategy acquisition of raw data to the interpretation of model out-
comes for policy formulation.
To address class imbalance, this study applies SMOTE Each module is logically connected, and data flows uni-
(Synthetic Minority Oversampling Technique) [28] to directionally from input to output. The framework not only
the training data before model fitting, as shown in Fig. 3. captures the operational logic but also establishes the foun-
Only the minority severity classes are oversampled. The dation for the mathematical formulation developed in the
majority class is left unchanged. The dataset was first par- following section. This integrated investigative approach
titioned into training and testing subsets, and during cross- provides a closed analytical loop, beginning from empirical
validation, SMOTE was applied independently within each data and culminating in interpretable outcomes. The flow-
training fold only, while validation and test sets remained chart thus represents both the logical and computational

13
Innovative Infrastructure Solutions (2026) 11:272 Page 7 of 21 272

Fig. 3 SMOTE data balancing

algorithm

Fig. 4 Relative class distribution a Before SMOTE: b After SMOTE

architecture of the research. Unlike descriptive or purely The Random Forest (RF) framework
statistical analyses, the present methodology emphasizes a
learning-based paradigm, as the model learns from observed Machine learning has become an indispensable tool for ana-
crash patterns and generalizes them to predict future crash lysing complex, high-dimensional traffic-safety data [29–
severity classes. The procedural structure ensures that the 31]. Road accident severity classification involves multiple
entire process is embedded within a reproducible analytical inter-dependent variables such as vehicle type, environmen-
framework. tal conditions, roadway geometry, and driver behaviour, that
interact in nonlinear and often stochastic ways. Classical sta-
tistical models such as logistic or probit regression impose

13
 272 Page 8 of 21 Innovative Infrastructure Solutions (2026) 11:272

Fig. 5 Methodology adopted in the study to develop a crash severity predictive framework. model

Table 2 A comparative analysis among ML algorithms

Algorithm Strengths Limitations in the present context
Support Vector Machine Performs well on small datasets and finds maximum- Requires extensive kernel tuning, less interpretable,
(SVM) margin boundaries. not ideal for mixed categorical–numerical data.
Artificial Neural Network Captures complex nonlinearities and latent interactions. Demands large & balanced datasets, prone to over-
(ANN) fitting on limited or noisy crash data.
Gradient Boosting/ XGBoost High accuracy, effective handling of imbalance through Sensitive to hyperparameters, less transparent, may
weighting. overfit sparse categorical variables.
Random Forest (Selected) Handles heterogeneous features, insensitive to scaling, Slightly less efficient on extremely high-dimen-
reduces variance through bagging, provides feature- sional sparse data.
importance scores, inherently interpretable.

strong assumptions of linearity and independence among in heterogeneous traffic systems. By aggregating multiple
predictors, which rarely hold in heterogeneous Indian traffic weak learners and incorporating randomness, RF tends to
environments. Consequently, a learning algorithm capable deliver more precise predictions than standalone models
of capturing nonlinear patterns, variable interactions, and such as Support Vector Machines (SVMs), XGBoost, and
noise robustness was essential for this study. Among the Artificial Neural Networks (ANNs) [33, 34]. The RF frame-
wide family of supervised classifiers, the RF algorithm, work was chosen after evaluating several alternative algo-
proposed by Breiman [32], was selected as the core model- rithms commonly and frequently applied in safety analysis,
ling technique. RF belongs to the ensemble-learning par- as shown in Table 2.
adigm, where a multitude of decision trees are trained on Given the nature of the NH-44 dataset, it has a moderate
randomly drawn subsets of both samples and features, and sample size, mixed variable types, and inherent randomness
their collective decisions are aggregated through majority in crash events. RF offers the best trade-off between accu-
voting. This design enables RF to achieve high predictive racy, interpretability, and computational stability. Its embed-
accuracy while maintaining interpretability and resilience ded feature-importance mechanism also allows transparent
to overfitting. Rather than proposing a new Random Forest ranking of influential factors such as road environment or
algorithm, this study reformulates Breiman’s standard pipe- traffic violation type, supporting evidence-based policy for-
line into an analytical matrix form. This approach enables mulation. Each decision tree in the ensemble partitions the
structured sensitivity analysis, traceability of feature influ- feature space into regions that correspond to distinct output
ence, and improved interpretability of severity outcomes

13
Innovative Infrastructure Solutions (2026) 11:272 Page 9 of 21 272

classes. For a dataset ( xi , yi ), the tree recursively splits the   

data using rules of the form  
y1
 y 2 
xj < τ (6) Y =  ..  , yi ϵ {1,2, 3} ,
 .  (8)
yn nx1
where xj is a selected feature, and τ is the threshold that corresponding Fatal, Grevious, and Minor categories
maximises the reduction in impurity. By averaging across
many such trees built on different random subsets, the forest For numerical stability, all features are standardized by col-
produces a stable and unbiased estimator of class probabili- umn as given by Eq. (9), yielding the normalized feature
ties. Two principal sources of randomness, first the boot- matrix Z = [Zij ]nxp .
strap resampling of data and second the random selection
−
of features at each node, introduce diversity among trees, xij − xj
Zij = (9)
lowering variance without increasing bias. This is particu- σj
larly advantageous for accident-severity data, where predic-
tors may be correlated, and the number of observations per b) Tree-Structured Mapping in Matrix Form:
severity class is imbalanced. Unlike deep neural or kernel-
based models, the RF algorithm provides direct interpret- Each decision tree Tb in the Random Forest ensemble per-
ability through feature-importance measures derived from forms a nonlinear mapping from input Z to a categorical
the reduction in impurity across all trees. These measures output vector. To express this in matrix form, we define a
quantify each variable’s contribution to predictive accuracy, node partitioning matrix Mb for the.
enabling translation of model outcomes into meaningful B th tree as, as given in Eq. (10), in which mbij is a binary
safety insights. Policymakers can therefore relate high- coefficient taking value 1 if feature j is used for splitting
importance variables, such as collision type or absence of record i in tree b, and 0 otherwise.
safety devices, to targeted interventions.
 
Having established the rationale and operational prin- m11
(b) (b)
m12 ...
(b)
m1p
 
ciples of the RF framework, the following section develops  (b)
m21
(b)
m22 ... (b)
m2p 
Mb =  .. .. .. ..
 (10)
its analytical and mathematical formulation. The forthcom- 
 . . . .


ing derivation expresses the entire modelling process, from mn1
(b) (b)
mn2
...
mnp
(b)
nxp
feature representation and encoding to ensemble aggrega-
tion and loss minimisation. This rigorous mathematical
treatment underscores the theoretical integrity of the meth- The intermediate output of tree b can thus be expressed as a
odology and its applicability to structured crash-severity matrix transformation, Sb = Mb ⊙ Z {⊙ denotes element-
prediction problems. wise multiplication, yielding the split activation matrix}.
Each split partitions data into subsets, the final decision of
tree b is represented as, given by Eq. (11)
Analytical and mathematical framework
Tb = Sb Wb (11)

a) Data Matrix Formulation: where Wb ϵ Rpxk ​is a terminal decision-weight matrix

that maps the feature subsets into k = 3 class probabilities
The cleaned dataset is represented by the real-valued matrix (Fatal, Grievous, Minor).
X , where n denotes the number of crash samples and p
denotes the number of measured features. Each record is a c) Ensemble Aggregation:
row vector, and the complete feature space forms a linear
subspace. All features are standardized such that they ensure All B trees operate independently, producing matrices
orthonormal scaling in the feature space. T1 , T2 , . . . , TB each of size n × k . The RF ensemble
  aggregates them as, given by Eq. (12)
x11 x12 x13 ... x1p
 x21 x22 x23 ... x2p 
X=   (7) 1 
.. .. .. .. .. B
 . . . . . 
xn1 xn2 xn3 ... xnp P = Tb (12)
nxp B
b=1

The target vector containing severity levels is given as:

13
272 Page 10 of 21 Innovative Infrastructure Solutions (2026) 11:272

with P = [ pij ]nxk represents the Mtrix of predicted class j ∗ = argmax ∥gj ∥2 (18)
k j
probabilities. Each row of P satisfies ij = 1,
j=1 p

pij ∈ [0,1], and hence on the probability simplex in Rk . here, gj ϵ Rk denotes the class-wise information-gain vec-
The predicted class vector is obtained by using the equation tor associated with attribute Aj ​, and ∥gj ∥2 ​represents its
shown below: Euclidean norm used to aggregate class-specific impurity
reduction into a scalar split criterion. The index j ∗ identi-
yi = argmax pij , i = 1,2, . . . , n(13) fies the attribute that maximizes this norm and is therefore
j
selected as the optimal split variable at node r.

d) Information Gain and Entropy Matrices: e) Matrix-Based Model Evaluation:

Following ensemble aggregation, the overall model per-
At the core of each tree split lies an information gain com- formance is evaluated through the confusion matrix repre-
putation. For the subset of samples Sr ⊆ Z reaching node sentation. In the data set, Y ∈ {0,1}
nxk
denote the true
r, defines its class-probability vector, given in Eq. (14).
one-hot class-indicator matrix and denote
 nxk
Y ∈ {0,1}
[ ] [ ]
pr1 1 T Proportion of Fatal crashes the predicted class-indicator matrix obtained from the RF
Pr = pr2 = Y 1=
|Sr | r
Proportion of Grievous crashes  (14) classifier. The confusion matrix C is therefore expressed
pr3 Proportion of Minor crashes
as, show in Eq. (19), where each element cij represents the
The above equation gives the probability distribution of number of samples belonging to true class i but predicted
classes at node r, with Sr as the subset of data reaching as class j.
node r, |Sr | as the number of samples in that subset, Yr as 
c11 c12 c13

true class labels (one-hot encoded) for those samples, and C = Y Y =
T
c21 c22 c23 (19)
1as a column vector of ones used to sum across rows. The c31 c32 c33 kxk
entropy matrix of node r is then represented as:
       The diagonal elements ( c11 , c22 , c33 ) represent correctly
Hr = −pr log(pr ) (15)
T classified cases for each severity level. The overall accuracy
measures the proportion of correctly classified instances
For a tree containing R nodes, the overall entropy structure among all evaluated samples. It provides a direct indicator
H is compactly expressed as shown in Eq. (16), where each of the classifier’s aggregate performance across all classes.
row vector Hr ​represents the entropy state of node racross The accuracy is calculated as the ratio of the sum of diagonal
all k severity classes. elements (true positives for each class) to the total number
of predictions in C. Here, Tr (C) denotes the matrix trace,
 
H1 representing total correct predictions, and 1 ∈ Rkx1 is a
 H2  column vector of ones used to aggregate all entries in C,
H=  ..  (16)
 .  as shown in Eq. (20). Class-wise precision quantifies how
HR Rxk
many of the predicted samples for a given class were cor-
rectly classified, while recall measures how many actual
When a node r is partitioned on an attribute Aj ​, it produces samples of that class were successfully identified by the
two offspring nodes rL ​ and rR ​. If wL = |SrL | / |Sr | and model. These two metrics provide a balanced understanding
wR = |SrR | / |Sr | denote the relative sample proportions of classification reliability and sensitivity. For the kth class,
in each child node. The information-gain vector for feature precision, and recall are directly derived from the confu-
Aj is then written as: sion matrix by dividing the corresponding diagonal term
​ by the column and row totals, respectively. For compact ( )
representation, the diagonal matrices DP = diag 1T C
[ ]
HrL and DR = diag (C1) represent the predicted and actual
gj = H r − [ wL wR ] (17)
HrR class sample counts. Precision and recall vectors are then
expressed through element-wise operations to maintain
This formulation expresses the entropy reduction achieved matrix consistency, as shown against Eqs. (21) and (22).
by splitting node r on feature Aj ​. Each component of gj To provide a single balanced measure of both precision and
quantifies the contribution of that feature to class-wise recall, the harmonic mean, termed the F1 ​-score (Eq. 24), is
impurity reduction. The attribute providing the maximum computed for each class. The F1 ​-vector thus combines both
reduction is chosen as the optimal split variable: metrics, penalising any imbalance between false positives

13
Innovative Infrastructure Solutions (2026) 11:272 Page 11 of 21 272

and false negatives. The macro-averaged F1 ​-score ( F1, avg ​​

and ⊘ denotes element-wise division. The RF architecture
) (Eq. 25) represents the model’s overall robustness across
utilized in this study is illustrated in Fig. 6, where the pro-
all severity levels.
cessed crash dataset is randomly partitioned into training
Tr (C) and validation subsets. Bootstrap resampling is applied to
Accuracy = (20) the training set to generate multiple subsamples, each used
1T C1
to grow an individual decision tree, and class predictions
Ckk from all trees are subsequently aggregated using majority
P recisionk = ∑ k C (21) voting to obtain the final severity classification output.
i=1 ik
This matrix-based formalization establishes RF classifi-
Ckk cation as a deterministic operator acting on the data matrix.
Recallk = ∑ k C (22) Every phase, from attribute selection ( Mb ​) to ensemble
j=1 kj
projection ( Wb ​)is embedded in explicit matrix transforma-
tions. Thus, the model is not an abstract black box but a hier-
π = diag (C) ⊘ diag (DP ) and ρ = diag (C) ⊙ diag (DR )(23)
archically layered matrix system, mapping observed crash
descriptors to probabilistic severity outcomes through struc-
F1 = 2 [( π ⊙ ρ ) ⊘ (π + ρ )] (24)
tured linear–nonlinear compositions. This formulation dem-
1 T onstrates the analytical solidity and mathematical continuity
F1,avg = 1 F1 (25) of the research line, both rigorously expressible in finite-
k
dimensional matrix algebra. The matrix-based representa-
where π , ρ , and F1 represent precision, recall, and har- tion allows direct coupling between ensemble outputs and
monic-mean vectors. ⊙ denotes element-wise multiplication perturbation-based sensitivity operators, which is difficult

Fig. 6 Framework of the RF

based crash severity classification
model

13
 272 Page 12 of 21 Innovative Infrastructure Solutions (2026) 11:272

Table 3 Distribution of injury severity classes assessment to ensure that minority severity categories are
Severity class Number of crashes Percentage (%) not underrepresented in predictive evaluation.
Fatal 534 44.4
Grievous 564 46.9
Hyperparameter optimization analysis
Minor 105 8.7
(Performance Response Surfaces)

to express using tree-level descriptions alone. This repre- Once the RF model structure and mathematical formulation
sentation supports consistent attribution of severity influ- were established, a hyperparameter sensitivity study was
ence across heterogeneous feature groups without altering undertaken to quantify how model complexity influences
the underlying classifier. The matrix formulation improves classification performance. The dataset was first divided
practical interpretability by explicitly mapping encoded using a stratified 70:30 train–test split to preserve class pro-
crash descriptors to class-probability outputs through struc- portions. Hyperparameter tuning (number of trees and max-
tured transformations. Representing node splits, ensemble imum depth) was conducted exclusively within the training
aggregation, and probability estimation in matrix form clari- subset using 5-fold cross-validation. The test set was not
fies how feature interactions propagate through the model. consulted during model selection and was used only once
This structured representation facilitates reproducibility, for final performance evaluation. All reported performance
systematic sensitivity analysis, and clearer linkage between metrics correspond to this held-out test set. A two-factor
statistical outputs and transport-engineering variables. variation between the number of trees ( T ) and their maxi-
mum depth ( D) was examined using a 70:30 train–test split
[35]. The resulting response surfaces for training and testing
Results and analysis misclassification rates are presented in Fig. 7a and b.
The training surface demonstrated a progressive decrease
Severity distribution overview in misclassification as T and D increased, indicating
improved learning capability with greater representational
Prior to model evaluation, the distribution of injury sever- power. However, the testing error surface highlighted a
ity classes is summarized to provide context regarding class clear optimal region in the mid-range of model complex-
imbalance within the dataset. Table 3 presents the stratified ity, beyond which performance degraded due to overfitting
distribution of Fatal, Grievous, and Minor crashes across the (marked in Fig. 7b). This response pattern reinforces that
1,203 recorded incidents. The relatively lower proportion of moderate values of T and D (typically 100 ≤ T ≤ 150
Minor severity cases (8.7%) reflects inherent class imbal- and 5 ≤ D ≤ 75) offer the most reliable bias–variance
ance, justifying the adoption of stratified validation and balance for the present crash severity classification task.
macro-averaged evaluation metrics. This imbalance also Hyperparameters were determined exclusively through
underscores the necessity of class-sensitive performance cross-validation on the training subset. The test-set error
surface shown in Fig. 7b reflects the performance of the

Fig. 7 Hyperparameter sensitivity surfaces of the RF model: a training misclassification; b testing misclassification

13
Innovative Infrastructure Solutions (2026) 11:272 Page 13 of 21 272

Fig. 8 a Confusion Matrix; b Performance evaluation metric score for the developed model

attribute distributions. Nevertheless, the classifier maintains

stable F1-scores across classes, and the macro-averaged
F1-score demonstrates reliable performance at an aggregate
level. To visually summarise the evaluation, Fig. 8 displays
the confusion matrix and the class-wise distribution of all
key metrics, Precision, Recall, F1-Score, Accuracy, and
F1(avg), demonstrating the model’s stability and predictive
consistency across injury-severity levels.
To further validate the discriminative reliability of the
classifier, a One-vs-Rest (OvR) ROC-AUC [36] analy-
sis is performed on the same stratified test set used in
the confusion matrix evaluation (see Fig. 9). ROC-AUC
provides a threshold-independent measure of class sepa-
rability, which is crucial when severity classes exhibit
Fig. 9 One-vs-Rest (OvR) ROC-AUC graph overlapping and non-linear decision boundaries due
to mixed vehicle interactions and disorderly driving
finalized model on unseen data after hyperparameters were behavior exposure. The OvR strategy decomposes the
fixed. The test set was not consulted during parameter opti- 3-class severity prediction into three independent binary
mization. Parameter ranges were selected to balance model discrimination tasks (Fatal vs. Rest, Grievous vs. Rest,
complexity and generalization stability, with final values Minor vs. Rest). The resulting AUC values quantify the
chosen based on consistent cross-validation performance ability of the model to separate each severity level from
across folds. others beyond a random guess (AUC = 0.5). This ensures
that performance claims are supported by a full-spectrum
Model performance evaluation separability assessment, strengthening confidence in the
model’s predictive behavior under complex heteroge-
The predictive performance of the RF classifier was assessed neous crash conditions.
using class-wise precision, sensitivity (recall), F1-score,
overall accuracy, and macro-averaged F1, which are the Feature importance, sensitivity, and perturbation
outcomes of the confusion matrix. These metrics provide analysis
a multi-angle evaluation of the classifier’s behaviour across
the three injury-severity classes (Fatal, Grievous, and To interpret the contribution of each explanatory variable
Minor). The model achieves consistently high precision (attribute), Gini-based importance scores were computed
and recall for fatal and grievous categories, indicating a across all trees in the ensemble. Features such as impacting
strong ability to correctly identify and discriminate severe vehicle type, road environment, traffic control device pres-
crash cases. The minor category exhibits relatively lower ence, and weather exhibited dominant influence in deter-
precision, reflecting class imbalance and higher overlap in mining severity outcomes. The importance values reflect

13
272 Page 14 of 21 Innovative Infrastructure Solutions (2026) 11:272

the cumulative impurity reduction achieved whenever a fea- individual crash. In this equation, ej is the standard basis
ture participates in a split. Variables with higher importance vector perturbing only the jth feature, and ϵ is the small
scores are thus those that consistently contribute to sepa- perturbation magnitude (numerical step). Equation (29)
rating injury categories. This analysis establishes a ranked summarizes how the overall model responds to feature
hierarchy of influential predictors, guiding interpretation perturbations. To map the feature perturbations to changes
and providing empirical justification for domain-relevant in predicted injury severity, prediction change under input
risk factors. RF Gini importance (Eq. 26) values are normal- variation was computed, using Eq. (30).
ized to lie within [0,1] and sum to one across all predictors,
providing a relative measure of each attribute’s contribution ∆ P ≈ JR ∆ X T (30)
to impurity reduction within the ensemble.
The sensitivity score of feature j based on the Euclidean
1∑B ∑B
Gini importance (Aj ) = wR ∆ H r (26) magnitude of the entire column j , is then computed using
B b=1 rϵTb (j)
Eq. (31)
  
where Tb (j) denotes all nodes in tree b that split on feature
Aj , wR = |Sr | /n is node weight, and ∆ H r ​is impurity sj = ∥JR [:, j]∥2 (31)
reduction.
To assess model robustness and isolate the contribution Comparison of Gini impurity–based feature importance and
of each attribute to severity prediction, a sensitivity opera- perturbation sensitivity for the 13 explanatory variables,
tor was constructed using a finite-difference approximation both measures consistently identify IVT, RF, RE, AIV, and
of the local Jacobian Ji ∈ Rkxp [37] of the probability W as the dominant predictors of crash severity, as shown in
output vector, p (xi ). This operator captures the response Fig. 10.
of predicted severity to small perturbations ∆ X in each To strengthen the model’s split-quality reasoning and
input dimension. The aggregated Jacobian JR represents verify injury-class discrimination robustness under hetero-
the global feature sensitivity over the full dataset, while the geneous highway crash interactions, a node-impurity and
feature-wise Euclidean norm sj ​produces a scalar perturba- misclassification tradeoff analysis was conducted using
tion score that is directly comparable across attributes. This Entropy, Gini Impurity, and misclassification error surfaces.
analysis complements the impurity-based importance by This analysis ensures threshold-independent severity sepa-
capturing prediction sensitivity rather than split frequency. rability assessment. It confirms that the RF model learning
Unlike variance-based global sensitivity methods com- process captures structural disorder within severity interac-
monly used with deep learning, the perturbation-based sen- tions rather than producing spurious accuracy due to class
sitivity adopted here directly quantifies changes in predicted dominance or sampling artifacts. The results of this evalua-
severity probabilities under small feature variations, making tion are visually summarized in Fig. 11.
it well-suited for tree-ensemble models and limited data set- Figure 11 illustrates a node-split impurity tradeoff for
tings. For the ith crash record, a trained classifier producing the 3-class crash-severity classifier for the study stretch.
predicted severity probabilities represented as, given by Eq. Entropy forms a smooth inverted-U, peaking at maximum
(27). class overlap (highest disorder) near mid-balance and
  declining as split certainty increases. Gini impurity shows a
p̂1xi flatter peak and faster decay, indicating lower sensitivity to

 p̂2xi 
 residual class disorder than entropy. Misclassification error
p̂ (xi ) =  ..  , xi ϵ Rp (27)
 .  shows rapid decay as Fatal crashes dominate the node, with
p̂kxi slower tailing due to overlap with Grievous injury attri-
butes. The line behavior confirms mathematically coherent

p1 (xi + ϵ e1 ) − p1 (xi − ϵ e1 ) p1 (xi + ϵ ep ) − p1 (xi − ϵ ep )

impurity decay and validates robust split-quality reasoning
Ji = ,... , (28)
2ϵ 2ϵ under heterogeneous severity interactions.

1∑ Injury type and vehicle dynamics analysis

JR = i=1 Ji (29)
n
n
This analysis investigates how specific unsafe driving
Equation (28) represents the local Jacobian approxima- behaviors interact both with each other and with exposure
tion, which quantifies how sensitive the prediction vector conditions to influence crash injury outcomes. Catplots were
is to small changes in each input feature (attribute) for each first employed to visualize how fatal, grievous, and minor

13
Innovative Infrastructure Solutions (2026) 11:272 Page 15 of 21 272

Fig. 10 Gini importance and

Perturbation sensitivity values for
the model attributes

Fig. 11 Entropy, Gini Impu-

rity, and misclassification error
surfaces

injuries distribute across four key behavioral violations for This initial inspection suggests that severity escalation is
the vehicle classes considered, viz., overspeeding, drunk strongly influenced by the type of vehicle involved under
driving, no helmet, and no seatbelt, as shown in Fig. 12, a given behavioral violation. However, to better quantify
across vehicle classes as heavy vehciles (HV), others (O), and interpret these interactions, additional analysis was
cars (C), three wheelers (3W), and two wheelers (2W). conducted through multi-dimensional severity mapping
These particular causes were selected as they consistently and probability-based severity amplification estimates.
represent key human-error factors associated with severe Building upon the catplot insights, a focused analysis of
road crashes in India and hold high enforcement priority. combined behavior and exposure effects was carried out
The visual results reveal that behavior alone does not deter- to identify conditions most strongly associated with severe
mine the injury outcome, instead, severity shifts strongly injury shifts. Since fatal injuries represent the highest end
depending on the vehicle type involved in the crash. of the severity spectrum and exhibit the strongest interac-
tion effects of risky behavioral combinations, probability

13
 272 Page 16 of 21 Innovative Infrastructure Solutions (2026) 11:272

Fig. 12 Catplots showing injury types across behavioral violations

Table 4 Behavioral violation combinations their own. These findings are strongly supported by the RF
Combination Estimated fatal Severity feature importance results, where IVT appears as the most
crash probability rating influential predictor of severity, followed by road context
Drunk driving and Over speeding 0.65–0.72 Very High and behavioral variables. The consistency between visually
No helmet and Over speeding 0.58–0.66 Very High derived evidence and model-based interpretation reinforces
Drunk driving and No helmet 0.55–0.62 Very High the reliability of this severity escalation mechanism within
No seatbelt and Over speeding 0.40–0.46 High heterogeneous Indian traffic environments. The dominance
of IVT reflects vulnerability asymmetry inherent in hetero-
estimation for fatal outcomes was prioritized for policy- geneous traffic, where differences in vehicle size, protection
relevant interpretation. Table 4 summarizes the combined level, and stability directly influence injury severity out-
behavioral violation conditions that demonstrate elevated comes. Road Features (RF) contribute through geometric
fatal outcome probabilities. These combinations highlight constraints, sight-distance limitations, and infrastructure
the synergistic influence of unsafe behaviors, where the design characteristics that mediate crash energy dissipa-
concurrent presence of two risk factors significantly ampli- tion. Road Environment (RE) captures contextual exposure
fies injury severity. conditions such as terrain and roadside configuration, which
Further, Table 5 examines how behavioral violations can amplify or mitigate injury consequences. The conver-
interact with various vehicle types to affect fatality likeli- gence of these factors indicates that severity escalation is
hood. The results indicate that severity escalation is not structurally interaction-driven rather than attributable to iso-
solely governed by risky behavior but is also influenced lated behavioral violations alone.
by or depends on how the violation interacts with vehicle The interaction patterns, as visually represented in
form and protection level. This establishes a direct link Fig. 12 and quantitatively summarized in Tables 4 and
between exposure setting and the transformation of a high- 5, demonstrate that behavioural effects are not additive
risk behavior into a fatal crash event. The combined insights but vehicle-dependent, with severity amplification vary-
from the catplots and interaction tables reveal a clear and ing across vehicle categories and compounded violations.
consistent pattern showing that crash severity is primar- Such non-linear behavioural–vehicle interactions confirm
ily shaped by the interaction between unsafe behavior and that injury escalation arises from coupled system dynam-
vehicle vulnerability, rather than by behavioral violations on ics rather than isolated violations. Further, an accurate

Table 5 Fatality amplification due to behavioral and vehicle-type interactions

Cause and vehicle type interaction Most vulnerable group Estimated fatal crash Underlying severity mechanism
probability
Over speeding involving Heavy vehicles Pedestrians, Two-wheeler 0.48–0.55 High impact energy and momentum
users* transfer leading to severe bodily trauma
Drunk driving involving Two-wheelers Two-wheelers users 0.58–0.65 Loss of vehicle stability with direct
exposure to impact forces
No helmet involving Two-wheelers Two-wheelers users 0.52–0.60 Absence of head protection resulting in
critical brain injuries
No seatbelt involving Cars Car occupants 0.36–0.42 Occupant ejection or secondary
impacts with vehicle interior
*Two-wheeler users refer to Vulnerable Road Users operating or riding as pillion passengers on motorized two-wheelers. Pedestrian-related
crashes are not explicitly included in the dataset and are therefore outside the analytical scope

13
Innovative Infrastructure Solutions (2026) 11:272 Page 17 of 21 272

Baseline model comparison

To establish the predictive value of the proposed frame-

work, model performance is benchmarked against com-
monly used baseline models like Logistic Regression (LR),
Support Vector Machine (SVM), XGBoost (XGB), and a
neural baseline, Artificial Neural Network (ANN), trained
using a 3-layer feed-forward architecture. All models are
trained on the same dataset, evaluated using identical strati-
fied 70:30 split and 5-fold CV protocol, and compared on
threshold-independent discrimination and balanced injury
separability. To avoid additive count confusion, perfor-
mance is reported using relative performance indices nor-
malized to the best baseline score. The results confirm that
the proposed RF framework delivers higher and more stable
severity discrimination, as shown in Table 7. This demon-
Fig. 13 Reliability diagrams for calibrated Random Forest model
across severity classes strates that the gain arises from modeling suitability for
heterogeneous highway crash interactions, not from differ-
Table 6 Computed Brier Scores for each class ences in data or validation treatment. Relative indices are
Severity class Brier score included to express each model’s performance as a normal-
Fatal 0.118 ized share of the best benchmark score (1.00), enabling fair
Grievous 0.094 comparison under identical data and validation protocols.
Minor 0.189 This avoids misinterpretation of absolute crash counts after
The Brier Score ranges from 0 to 1, where 0 indicates perfect calibra- SMOTE and highlights that performance gains arise from
tion, and 1 indicates the worst possible probability estimation [38]
model discrimination quality, not dataset size. Compared
to prior Indian crash-severity studies [25, 39] that typi-
class-probability estimation is essential for severity predic- cally report accuracy levels ranging between approximately
tion models intended to support operational safety deci- 75–80% using RF, SVM, or gradient-boosting approaches,
sion-making. Therefore, the calibration performance of the proposed framework achieves 88.7% classification
the RF classifier was assessed using reliability curves and accuracy and an OvR mean ROC-AUC of 0.94 under iden-
Brier Score analysis [38]. Figure 13 shows the comparison tical stratified validation. This performance improvement is
of predicted vs. observed probabilities for each severity attributed not merely to model selection but to the structured
class, evaluated using Platt scaling calibration. A perfectly integration of heterogeneous behavioral, vehicle, and envi-
calibrated model follows the diagonal unity line. The near- ronmental predictors with calibrated probability estimation.
diagonal trajectories of the curves indicate that predicted
severity probabilities closely match empirical frequencies,
demonstrating good reliability. Minor deviations at lower- Discussion
probability regions reflect class imbalance, but overall cali-
bration performance confirms the suitability of the model This study set out to develop a crash-severity prediction
for risk-oriented applications, where probability outputs are framework that is not only accurate but also mathematically
used for setting enforcement priorities and resource alloca- explicit and interpretable for heterogeneous Indian highway
tion. To quantify uncertainty, Brier Scores were computed conditions. The findings confirm that a matrix-structured
for each class, as shown in Table 6. RF model, coupled with rigorous pre-processing, can suc-
cessfully capture complex interactions among road, vehicle,

Table 7 Relative model performance comparison for crash-severity classification

Model Accuracy (%) Relative Accuracy Index ROC-AUC (OvR Mean) Relative ROC-AUC Index MCC* Relative MCC Index
LR 76.4 0.86 0.81 0.86 0.63 0.73
SVM 78.9 0.89 0.84 0.89 0.67 0.78
XGBoost 82.3 0.93 0.88 0.93 0.72 0.84
ANN 80.7 0.91 0.86 0.91 0.70 0.82
Proposed RF 88.7 1.00 0.94 1.00 0.86 1.00
*MCC (Matthews Correlation Coefficient)

13
272 Page 18 of 21 Innovative Infrastructure Solutions (2026) 11:272

human, and environmental attributes on NH-44. Unlike range of 75–85% using machine learning approaches, the
many earlier studies [40, 41] that either rely on aggregate proposed framework achieves 88.7% accuracy and a higher
regression models or treat machine-learning models as black ROC-AUC, indicating improved discriminative capability.
boxes, the present work embeds the entire pipeline, from This improvement can be attributed to the structured inte-
feature representation to ensemble aggregation and perfor- gration of heterogeneous variables and the incorporation
mance evaluation, within a transparent matrix formulation. of interaction-aware modelling, which enhances both pre-
The hyperparameter response surfaces demonstrate that dictive performance and interpretability. Previous studies
crash-severity prediction benefits from moderate ensemble often reported that speeding or alcohol use increases sever-
complexity rather than simply increasing the number of ity, but typically treated these as independent covariates. In
trees or tree depth. The existence of a well-defined plateau contrast, this work demonstrates that severity escalation is
region for testing misclassification highlights the impor- fundamentally a joint outcome of behaviour and exposure
tance of controlling model capacity to avoid overfitting setting, quantified through probability ranges for specific
noisy crash records, an issue rarely explored explicitly in behaviour–vehicle combinations (Tables 4 and 5). From
previous severity studies. By visualising the bias–variance a policy perspective, the framework offers a more action-
trade-off, the analysis provides practical guidance for select- able view of risk. By linking RF-derived feature hierarchies
ing RF configurations in similar traffic-safety applications, with concrete interaction patterns, the study pinpoints where
especially where sample sizes are modest and class distribu- enforcement and engineering interventions can be most
tions are imbalanced. effective. Overspeeding countermeasures should be focused
A key methodological contribution lies in jointly using where HV and 2W interactions are highest, while helmet and
Gini-based feature importance and Jacobian-based pertur- seatbelt enforcement must target the most vulnerable user
bation sensitivity to understand the RF behaviour. Both groups. This converts model findings into precise, behavior-
measures independently converge on a consistent set of focused safety actions rather than generic interventions.
dominant predictors, impacting vehicle type, road features,
road environment, age of impacting vehicle, and weather, Limitation and future directions
showing that these variables not only appear frequently in
informative splits but also induce the largest changes in At the same time, several limitations must be acknowl-
predicted severity when perturbed. This dual perspective edged. The analysis is based on reported crash data from a
addresses a critical gap in the literature, where most ML- single highway corridor and does not include near-miss or
based crash studies either rank features using a single metric exposure metrics such as traffic volumes or speed profiles,
or omit robustness checks on model sensitivity altogether. which could further refine risk estimates. Categorical attri-
Here, the matrix-based sensitivity operator formalises how butes were encoded using integer labels, and although stan-
small changes in input space propagate to the probability dardisation and diagnostics were applied, more advanced
of each injury class, offering a bridge between data-driven encoding schemes might capture within-category structure
prediction and mechanistic reasoning about crash processes. in future work. Furthermore, only a single ensemble clas-
The behavioural–exposure interaction analysis further sifier was explored in depth; comparative evaluation with
extends existing knowledge. Catplots and interaction tables gradient boosting, probabilistic graphical models, or hybrid
reveal that unsafe behaviours such as overspeeding, drunk RF-simulation approaches could yield additional insights.
driving, and non-use of safety devices do not have uniform Certain attributes represent post-event crash descriptors
effects across all road users, their impact is strongly condi- rather than purely pre-crash predictors. While they do not
tioned by the vehicle type and protection level. This result encode injury severity directly, their inclusion supports
is consistent with recent findings in heterogeneous traffic severity characterization and systemic risk analysis rather
studies, where vehicle-type disparity has been identified as a than real-time severity prediction. Variance-based global
primary determinant of injury severity due to differences in sensitivity analysis is more appropriate for large-scale deep
mass, structural protection, and impact energy transfer. The learning models, whereas the present perturbation-based
dominance of Impacting Vehicle Type (IVT) in the model approach is intended as a practical alternative for ensem-
therefore reflects an underlying physical and exposure- ble models applied to moderate-sized crash datasets. These
driven mechanism rather than a purely statistical artifact. limitations, however, do not detract from the main contri-
In addition to the baseline comparison presented earlier, bution of demonstrating that a mathematically grounded,
the model performance was also interpreted in the context interpretable RF framework can be successfully tailored to
of recent state-of-the-art studies. Compared to prior crash heterogeneous Indian traffic and can meaningfully inform
severity studies conducted under Indian and international safety policy.
conditions, which typically report accuracy levels in the

13
Innovative Infrastructure Solutions (2026) 11:272 Page 19 of 21 272

It is important to clarify that the present framework ● These focused strategies align enforcement and engi-
is designed for post-crash injury severity characteriza- neering resources with the high-risk behavioural and ex-
tion rather than real-time or pre-crash prediction. Several posure conditions highlighted by the model outcomes.
explanatory variables describe crash-event conditions and
behavioural circumstances recorded after the occurrence of
the incident. Accordingly, the model outputs are intended Conclusions
to identify structural patterns associated with severity esca-
lation, supporting retrospective safety auditing, infrastruc- Road traffic crashes on Indian national highways occur
ture prioritization, and strategic enforcement planning. within a highly heterogeneous and vulnerable operating
The framework does not claim to enable proactive or real- environment, where conventional aggregate models often
time enforcement decisions. Future research may extend fail to describe how specific combinations of road, vehicle,
this framework in several focused directions. Integrating and behavioural factors translate into injury outcomes. By
exposure-based measures such as traffic volume or vehicle developing a matrix-structured RF framework and applying
composition would enable risk-normalized severity model- it to NH-44 crash data, this study shows that it is possible
ling beyond crash-conditioned analysis. External validation to obtain both high predictive performance and transparent,
across multiple heterogeneous national highway corridors policy-relevant explanations of severity patterns under such
would strengthen the assessment of model robustness and conditions. Although the empirical analysis is conducted on
generalizability. Incorporation of formal predictive-uncer- NH-44, its mixed modal traffic structure (HV, C, 3W, 2W),
tainty quantification could further enhance the reliability of variable enforcement intensity, and heterogeneous road–
probability-based decision support derived from the model environment conditions are characteristic of many major
outputs. Indian national highways. Since the proposed framework
models generalized interactions among behavioral, vehicle,
Policy and practical implications and contextual factors rather than corridor-specific engi-
neered attributes, the methodology remains applicable to
The model outputs provide structured guidance for targeted other NH corridors exhibiting similar heterogeneous traffic
interventions. The prominence of IVT suggests prioritiz- dynamics. Key conclusions from the study are as follows:
ing enforcement and monitoring in mixed HV–2W corri-
dors where vulnerability asymmetry is high. The influence ● The RF classifier, trained on a rigorously cleaned and
of RF and RE indicates the need for engineering improve- normalised dataset, achieves high overall accuracy and
ments, including geometric corrections, enhanced signage, balanced F1 performance across Fatal, Grievous, and
and roadside protection in high-risk segments. For instance, Minor classes, indicating that ensemble learning can re-
high RE influence may indicate the need for specific geo- liably support operational crash-severity assessment on
metric interventions such as shoulder widening, improved Indian highways.
median protection, or enhanced roadside safety barriers in ● Representing the entire RF pipeline, featuring encoding,
identified high-risk locations. Behavioral predictors such as node splitting, ensemble aggregation, and confusion-
overspeeding and impaired driving support focused enforce- matrix evaluation, in matrix form clarifies how crash
ment campaigns and educational programs targeting helmet descriptors are mapped to class probabilities. This ad-
and seatbelt compliance. By linking feature importance and dresses a major gap in existing ML-based crash studies,
sensitivity findings to intervention categories (engineering, where the modelling process is often opaque and diffi-
enforcement, education), the framework enables evidence- cult to relate to transport-engineering concepts.
based prioritization rather than uniform corridor-wide mea- ● Both Gini importance and perturbation sensitivity anal-
sures. The results recommend the following: yses converge on a core set of influential predictors,
with impacting vehicle type, road features, road envi-
● Prioritising overspeeding and drunk-driving enforce- ronment, age of impacting vehicle, and weather emerg-
ment on segments where heavy vehicles and two-wheel- ing as the strongest drivers of severity transitions. This
ers frequently interact. convergence increases confidence that these factors are
● Strengthening helmet and seatbelt compliance cam- not artefacts of a particular metric but represent structur-
paigns for the most vulnerable user groups identified by ally important attributes in heterogeneous traffic.
the model. ● Visual analysis and interaction tables show that un-
● Implementing context-specific road and roadside im- safe behaviours such as overspeeding, drunk driving,
provements at locations where road features and envi- and non-use of protection devices become particularly
ronment contribute strongly to severity escalation. lethal when combined with vulnerable user groups or

13
 272 Page 20 of 21 Innovative Infrastructure Solutions (2026) 11:272

5. Smeed RJ (1972) The usefulness of formulae in traffic engineer-

high-momentum vehicles. Behavioural pairs (drunk ing and road safety. Accid Anal Prev 4:303–312. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​
driving with overspeeding) and behaviour–vehicle com- 0​.​1​0​1​6​/​0​0​0​1​-​4​5​7​5​(​7​2​)​9​0​0​2​9​-​2​​​​
binations (overspeeding involving heavy vehicles, or no 6. Singh VBP, Hemamalini V, Muttipati AS et al Application of
helmet on two-wheelers) exhibit markedly higher fatal- machine learning predicting injuries in traffic accidents through
the application of random forest. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​2​1​7​4​/​0​1​1​8​7​2​
ity probabilities than isolated violations, highlighting 2​1​2​1​2​4​8​2​0​2​2​3​1​0​0​3​0​6​4​4​5​9​​​​
the need to view risk as a coupled system rather than a 7. Zohra EF, Maryam K, Hamada EK, Hasna EE (2023) Accident
sum of independent factors. severity prediction using machine learning: a case study on the
● By linking model outputs to interpretable mechanisms, US accidents dataset. In: Proc – 17th int conf signal-image tech-
nol internet-based syst SITIS. pp 242–246. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​1​0​
the framework supports prioritisation of enforcement 9​/​S​I​TI​ ​S​6​1​2​6​8​.​2​0​2​3.​​0​0​0​4​4​​​​
and engineering actions, stricter speed and alcohol con- 8. Wang Q, Gan S, Chen W et al (2021) A data-driven, kinematic
trol on segments with mixed HV and 2W flows, targeted feature-based, near real-time algorithm for injury severity predic-
helmet and seatbelt campaigns for identified high-risk tion of vehicle occupants. Accid Anal Prev 156:106149. ​h​t​t​​p​s​:​/​​/​d​
o​​i​.​o​​r​g​/​1​0​.​1​0​1​6​/​j​.​a​a​p​.​2​0​2​1​.​1​0​6​1​4​9​​​​
groups, and context-sensitive improvements in road 9. Malik S, El Sayed H, Khan MA, Khan MJ (2021) Road accident
environment and features where the model indicates severity prediction - a comparative analysis of machine learning
strong severity influence. algorithms. In: 2021 IEEE Glob Conf Artif Intell Internet Things,
GCAIoT. pp 69–74. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​1​0​9​/​G​C​A​I​o​T​5​3​5​1​6​.​2​0​2​1​
.​9​6​9​3​0​5​5​​​​
The framework developed here can be extended in sev- 10. Santos K, Dias JP, Amado C (2022) A literature review of machine
eral directions, including integrating temporal traffic and learning algorithms for crash injury severity prediction. J Saf Res
exposure data, incorporating near-miss and minor incident 80:254–269. https:/​/[Link]​g/10.10​16/j.​jsr.2021.12.007
records, and embedding the RF predictions within simula- 11. Ghasedi M, Sarfjoo M, Bargegol I (2021) Prediction and analy-
sis of the severity and number of suburban accidents using logit
tion-based tools for testing “what-if” safety scenarios. More model, factor analysis and machine learning: a case study in a
broadly, the study demonstrates that mathematically explicit developing country. SN Appl Sci 3:13. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​0​0​7​/​s​4​
machine-learning models can meaningfully augment tradi- 2​4​5​2​-​0​2​0​-​0​4​0​8​1​-​3​​​​
tional crash analysis in low and middle-income settings, 12. Shaik ME, Islam MM, Hossain QS (2021) A review on neural
network techniques for the prediction of road traffic accident
offering a scalable pathway toward evidence-based, con- severity. Asian Transp Stud 7:100040. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​0​1​6​/​j​.​
text-specific road safety planning. e​a​s​ts​ ​j​.​2​0​2​1​.​1​0​0​0​4​0​​​​
13. Yan M, Shen Y (2022) Traffic accident severity prediction based
Acknowledgements The authors gratefully acknowledge the sup- on random forest. Sustainability 14:1729. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​3​3​9​0​
port and resources that facilitated the successful completion of this /​s​u​1​40​ ​3​1​7​2​9​​​​
research. The contributions of individuals and organizations involved 14. Labib MF, Rifat AS, Hossain MM et al (2019) Road accident
in data provision and technical discussions are sincerely appreciated. analysis and prediction of accident severity by using machine
learning in Bangladesh. In: 2019 7th international conference
Funding The authors disclosed no financial support for the research, on smart computing & communications (ICSCC). IEEE, pp 1–5.
authorship, and/or publication of this article. https:/​/[Link]​g/10.11​09/IC​SCC.2019.8843640
15. Korkmaz H, Erturk MA (2024) Prediction of the traffic inci-
dent duration using statistical and machine-learning methods:
Declarations a systematic literature review. Technol Forecast Soc Change
207:123621. https:/​/[Link]​g/10.10​16/j.​techfore.2024.123621
Conflict of interest The authors declare that they have no known com- 16. Moussa GS, Owais M, Dabbour E (2022) Variance-based global
peting financial interests or personal relationships that could have ap- sensitivity analysis for rear-end crash investigation using deep
peared to influence the work reported in this paper. learning. Accid Anal Prev 165:106514. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​0​1​6​/​j​
.​a​a​p​.​2​0​2​1​.​1​0​6​5​1​4​​​​
17. Awad M, Moussa GS, Wahaballa AM, Younes H (2026) A state-
of-the-art review of injury severity analysis in traffic crashes:
References toward a generalized modeling framework. Innov Infrastruct
Solut 11:19. https:/​/[Link]​g/10.10​07/s4​1062-025-02409-9
1. World Health Organization (2021) Road trafc injuries. ​h​t​t​​p​s​:​/​​/​w​w​​ 18. Mirman JH (2019) A dynamical systems perspective on driver
w​.​w​​h​o​.​​i​n​t​​/​n​e​w​​s​r​​o​o​m​/​f​a​c​t​-​s​h​e​e​t​s​/​d​e​t​a​i​l​/​r​o​a​d​-​t​r​a​f​f​i​c​-​i​n​j​u​r​i​e​s​​​​ behavior. Transp Res part F traffic Psychol Behav 63:193–203.
2. Das A, Gjerde H, Gopalan SS, Normann PT (2012) Alcohol, https:/​/[Link]​g/10.10​16/j.​trf.2019.04.010
drugs, and road traffic crashes in India: a systematic review. Traf- 19. Adanu EK, Penmetsa P, Wood D, Jones SL (2019) Incorporat-
fic Inj Prev 13:544–553. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​0​8​0​/​1​5​3​8​9​5​8​8​.​2​0​1​2​.​ ing systems thinking approach in a multilevel framework for
6​6​3​5​1​8​​​​ human-centered crash analysis. Transp Res Interdiscip Perspect
3. Morth TR (2021) wing road accidents in India. 1–235 2:100031. https:/​/[Link]​g/10.10​16/j.​trip.2019.100031
4. Ameen T, Ahmad A (2025) Modelling and evaluating risk expo- 20. Ameen T, Ahmad A (2025) Modeling the effect of psychologi-
sure for motorized two-wheelers at signalized intersections in cal fear in mediating the critical gap behavior at roundabouts for
urban mixed traffic. J Inst Eng Ser A 106:259–274. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​ heterogeneous, disorderly traffic conditions: a conditional process
/​1​0​.​1​0​0​7​/​s​4​0​0​3​0​-​0​2​4​-​0​0​8​4​4​-​y​​​​ analysis approach. Int J Intell Transp Syst Res 1–22. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​
r​g​/​1​0​.​1​0​0​7​/​s​1​3​1​7​7​-​0​2​5​-​0​0​5​5​4​-​0​​​​

13
Innovative Infrastructure Solutions (2026) 11:272 Page 21 of 21 272

21. Ranney TA (1994) Models of driving behavior: a review of their challenge datasets. Hands-on Pattern Recognit Challenges Mach
evolution. Accid Anal Prev 26:733–750. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​0​1​6​/​ Learn 1:223–230
0​0​0​1​-​4​5​7​5​(​9​4​)​9​0​0​5​1​-​5​​​​ 34. Song J, Gao Y, Yin P et al (2021) The random forest model has the
22. Mohanty M, Gupta A (2015) Factors affecting road crash model- best accuracy among the four pressure ulcer prediction models
ing. J Transp Lit 9:15–19. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​5​9​0​/​2​2​3​8​-​1​0​3​1​.​j​t​l​.​v​ using machine learning algorithms. Risk Manag Healthc Policy
9​n​2​a​3​​​​ 2021:1175–1187. https:/​/[Link]​g/10.21​47/RM​HP.S297838
23. Theofilatos A, Yannis G (2014) A review of the effect of traf- 35. Nguyen QH, Ly H-B, Ho LS et al (2021) Influence of data split-
fic and weather characteristics on road safety. Accid Anal Prev ting on performance of machine learning models in prediction of
72:244–256. https:/​/[Link]​g/10.10​16/j.​aap.2014.06.017 shear strength of soil. Math Probl Eng 2021:1–15. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​
24. Valli PP (2005) Road accident models for large metropolitan cit- /​1​0​.​1​1​5​5​/​2​0​2​1​%​2​F​4​8​3​2​8​6​4​​​​
ies of India. IATSS Res 29:57–65. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​0​1​6​/​S​0​3​8​ 36. Richardson E, Trevizani R, Greenbaum JA et al (2023) The ROC-
6​-​1​1​1​2​(​1​4​)​6​0​1​1​9​-​9​​​​ AUC accurately assesses imbalanced datasets. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​
25. Khanum H, Garg A, Faheem MI (2023) Accident severity predic- 1​0​1​6​/​j​.​p​a​t​t​e​r​.​2​0​2​4​.​1​0​0​9​9​4​​​​
tion modeling for road safety using random forest algorithm: an 37. An H-B, Wen J, Feng T (2011) On finite difference approximation
analysis of Indian highways. F1000Research 12:494. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​ of a matrix-vector product in the Jacobian-free Newton–Krylov
o​​r​g​/​1​0​.​1​2​6​8​8​/​f​1​0​0​0​r​e​s​e​a​r​c​h​.​1​3​3​5​9​4​.​2​​​​ method. J Comput Appl Math 236:1399–1409. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​
26. Rosas-Chavoya M, Gallardo-Salazar JL, López-Serrano PM et al 1​0​1​6​/​j​.​c​a​m​.​2​0​1​1​.​0​9​.​0​0​3​​​​
(2022) QGIS a constantly growing free and open-source geospa- 38. Blattenberger G, Lad F (1985) Separating the Brier score into
tial software contributing to scientific development. Cuad Inves- calibration and refinement components: a graphical exposition.
tig Geográfica 48:197–213. https:/​/[Link]​g/10.18​172/c​ig.5143 Am Stat 39:26–32. https:/​/[Link]​g/10.23​07/26​83902
27. Royston P, White IR (2011) Multiple imputation by chained equa- 39. Panicker AK, Ramadurai G (2022) Injury severity prediction
tions (MICE): implementation in Stata. J Stat Softw 45:1–20. model for two-wheeler crashes at mid-block road sections. Int J
https:/​/[Link]​g/10.18​637/j​ss.v045.i04 crashworthiness 27:328–336. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​0​8​0​/​1​3​5​8​8​2​6​5​.​
28. Blagus R, Lusa L (2013) SMOTE for high-dimensional class- 2​0​2​0​.​1​8​0​6​6​4​4​​​​
imbalanced data. BMC Bioinformatics 14:106. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​ 40. Almahdi A, Al Mamlook RE, Bandara N et al (2023) Boosting
0​.​1​1​8​6​/​1​4​7​1​-​2​1​0​5​-​1​4​-​1​0​6​​​​ ensemble learning for freeway crash classification under varying
29. Ameen T, Ahmad A (2025) Dynamic modeling of passenger car traffic conditions: a hyperparameter optimization approach. Sus-
units for roundabouts: a novel genetic algorithm optimization- tainability 15:15896. ​h​t​t​​p​s​:​/​​/​i​d​​e​a​s​​.​r​e​​p​e​c​​.​o​r​g​​/​a​​/​g​a​m​/​j​s​u​s​t​a​/​v​1​5​y​2​0​
based approach. Transp Res Rec. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​r​g​/​1​0​.​1​1​7​7​/​0​3​6​1​1​ 2​3​i​2​2​p​1​5​8​9​6​-​d​1​2​7​9​3​5​0​.​h​t​m​l​​​​
9​8​1​2​4​1​3​1​0​1​1​8​​​​.​ 1: 41. Bayode O, Aderinola OS, Oluyemi-Ayibiowu BD (2025) Appli-
30. Ameen T, Wani H, Ahmad A (2024) Optimizing PCU values at cation of machine learning for road safety modeling of selected
roundabouts using PSO-based machine learning algorithm. In: south-west highway in Nigeria. Eur J Appl Sci Eng Technol
2024 international conference on control, computing, communi- 3:202–213. https:​​​//d​oi.​or​g/10.​593​24/​ejas​et.2025.​3%283%29.13
cation and materials (ICCCCM). IEEE, pp 264–267. ​h​t​t​​p​s​:​/​​/​d​o​​i​.​o​​
r​g​/​1​0​.​1​1​0​9​/​I​C​C​C​C​M​6​1​0​1​6​.​2​0​2​4​.​1​1​0​3​9​8​9​3​​​​ Publisher’s note Springer Nature remains neutral with regard to juris-
31. Kamili SI, Ameen T, Wani HA, Ahmad A (2025) Investigating dictional claims in published maps and institutional affiliations.
safety implications of median U-Turn (MUT) intersections based
on dilemma zone and critical gap analysis. Int J Transp Sci Tech- Springer Nature or its licensor (e.g. a society or other partner) holds
nol. https:/​/[Link]​g/10.10​16/j.​ijtst.2025.04.006 exclusive rights to this article under a publishing agreement with the
32. Breiman L (2001) Random forests. Mach Learn 45:5–32. ​h​t​t​​p​s​:​/​​/​ author(s) or other rightsholder(s); author self-archiving of the accepted
d​o​​i​.​o​​r​g​/​1​0​.​1​0​2​3​/​A​:​1​0​1​0​9​3​3​4​0​4​3​2​4​​​​ manuscript version of this article is solely governed by the terms of
33. Dahinden C, Ethz M (2011) An improved Random For- such publishing agreement and applicable law.
ests approach with application to the performance prediction

13