International Journal of Electrical and Computer Engineering (IJECE)

Vol. 14, No. 6, December 2024, pp. 6129~6139

ISSN: 2088-8708, DOI: 10.11591/ijece.v14i6.pp6129-6139  6129

A novel approach for imbalanced instance handling toward

better preterm birth classification

Himani Deshpande1, Leena Ragha2

1
Department of Artificial Intelligence and Data Science, Thadomal Shahani Engineering College, Mumbai, India
2
Department of Computer Science, BLDEA's V.P. Dr. P.G. Halakatti College of Engineering and Technology, Vijayapur, India

Article Info ABSTRACT

Article history: Preterm birth (PTB) is a major cause of child and mother mortality, a PTB
classification model can assist in assessing the health condition ahead of
Received Apr 17, 2024 time and help avoid complications during childbirth. Mother’s significant
Revised Jul 14, 2024 feature (MSF) dataset created for this study has features derived from
Accepted Aug 6, 2024 mother’s physical, lifestyle, social and stress attributes. MSF dataset consists
of 119 features of 1,000 mothers with 172 preterm and 828 full-term
deliveries, resulting in issues of dataset imbalance namely class
Keywords: inseparability and classification bias. To overcome the imbalance issue, a
novel algorithm named majority penalizing minority upsampling (MPMU) is
Class imbalance proposed. MPMU forms clusters looking into the degree of dataset
Clustering imbalance, it analyses the composition of each cluster individually and
Majority class computes the varied penalty for majority class instances. It further balances
Minority class dataset composition by oversampling minority class instances. MPMU
Oversampling processed dataset is further used to train the proposed 6L-ANN network
Penalty which finds the probability of occurrence of PTB. The proposed model has
shown efficient results on MSF sub-datasets with precision values ranging
from 0.90 to 0.97, area under the curve (AUC) between 0.86 to 0.99, and
prediction accuracy ranging from 93.04% to 99.47%. Experiment results
show that a mother’s lifestyle and stress features have a strong influence on
the childbirth outcome.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Himani Deshpande
Department of Artificial Intelligence and Data Science, Thadomal Shahani Engineering College
Bandra (w), Mumbai, Maharashtra, India
Email: [Link]@[Link]

1. INTRODUCTION
Maternal health and childbirth experience have always been a priority concern across the globe.
There are many complications that mother and child can go through during childbirth, one such complication
is when the childbirth takes place before 37 weeks of pregnancy which is termed preterm birth (PTB). PTB is
a leading cause of underdeveloped baby organs, maternal mortality and child mortality [1]. In today’s era
with rapid societal and lifestyle changes, the PTB cases are increasing rapidly. A PTB classification model
can help to assess the health condition of a woman ahead of time to prevent any future health complications.
Pregnancy outcomes are closely associated with the health of the mother [2]. A large number of health issues
that women face is associated with their lifestyle, relationship with family, social surroundings, stress level,
and physical health. All these factors influence their reproductive health as well [3] and hence information
associated with the mother’s mental and physical health can be used for the prognosis of PTB. Exploring
machine learning methods and tools on the mother’s health dataset can be of great help towards finding
useful patterns and facts for a better understanding of PTB. Considering the size of the population, medical

Journal homepage: [Link]

6130  ISSN: 2088-8708

domain datasets are found to be highly imbalanced with fewer instances of diseased patients [4]. Lack of
density in imbalanced datasets frequently results in learning models failing to discover unusual patterns and
having classification bias towards the majority class, thus resulting in minority class misclassification
[5], [6]. This makes imbalanced data handling a high-priority issue that needs to be addressed for efficient
classification of health conditions.
Observing the recent work in the maternal, childbirth domain and specifically in the PTB domain
[7], firstly it was observed that majorly the researchers and experts are from the medical domain and they
have focused on PTB classification models by using existing statistical methodologies and have not focused
on dataset enhancement techniques and modern heuristic classification methods. Secondly, there is no
standard dataset and researchers are relying on small local datasets with majority data consisting of obstetric
features and test results and have not worked on the influencing factors that affect these test results. To fill
this gap a dataset titled mother’s significant feature (MSF) is created for this study, through rigorous
interaction with gynecologists with the aim to derive the influencing features depicting mother’s physical
health, lifestyle, social surroundings and stress.
Looking into the imbalanced dataset handling methods across different domains, it is observed that
researchers have come up with different techniques which are broadly divided into three categories namely
pre-processing, cost-sensitive learning and ensemble learning methods [8], [9]. Pre-processing methods work
on strengthening the minority and suppressing the majority samples before classification. The dataset can be
pre-processed based on instances or features. Instance-based methods either over-sample minority instances
or under-sample majority instances or both [10]–[12]. Many researchers have used cluster-based fuzzy
weighing [13] and approaches where resampling of clusters is proposed [14], [15], these researchers have
focused on changing the size of clusters by over-sampling and under-sampling and have not worked on
understanding the impact of instances falling under different clusters towards class separability. Feature
selection methods pre-process datasets by observing features, these methods are broadly divided into three
categories, namely filter methods, wrapper methods and embedded methods [16], [17]. These methods look
into the strength of a feature individually or in combination with other features towards efficient
classification. The issue with pre-processing methods is that it is difficult to decide on the change in the
distribution of the database to suit the given problem [8].
Cost-sensitive methods are used to solve categorization problems when the costs of incorrect
classification differ depending on the instances and features. These techniques panelize to reduce losses faced
due to misclassification [18]. As the cost is assessed considering the imbalance in the dataset, these methods
are found to be efficient in handling imbalanced datasets. Cost can be incorporated into the classifier’s
training process as a cost matrix or threshold. To create a cost-sensitive classifier, researchers have followed
methodologies like modifying the goal function using a weighing strategy [19]–[21], changing the decision
thresholds, the learning process [22], the encoding network [23], and one-class approach [24], [25]. Although
these techniques are simple to use, overfitting problems could arise [8]. The intricacy of cost calculation may
be one of the main reasons why cost-sensitive approaches are less popular than resampling methods, even
though these methods are effective [26]. When a single classifier fails to work efficiently on an imbalanced
dataset, multiple classifiers are considered while deciding the final outcome [27]. These types of methods are
called ensemble methods. Bagging and Boosting are two different kinds of ensemble approaches [28]. The
majority of ensemble models have taken cost-sensitive and resampling techniques into account [26].
Literature suggests that dataset imbalance can be handled by processing datasets or customizing prediction
algorithms, the former being a more common choice among researchers. This research aims to overcome
issues in the maternal domain by creating a dataset depicting mother’s physical, lifestyle, social and stress-
related features. Further, a novel imbalance handling method is proposed to overcome the bias and
inseparability issues found in the created dataset.

2. METHOD
Researches in the PTB domain are based on limited observations from medical records and have not
focused on the lifestyle, stress and social aspects of a woman’s life that may also contribute to her maternal
health. For the same reason, this research delves into the examination of both the mental and physical health
of mothers, leading to the creation of a corresponding dataset. Observing the current population, it is realized
that the dataset majorly consists of full-term birth (FTB) cases and few PTB cases, thus resulting in
imbalanced dataset formation. Towards this, a novel clustering-based algorithm named majority penalizing
minority upscaling (MPMU) is proposed to handle the data imbalance by focusing on minimizing the impact
of majority dominance that may influence the classification results. Further, a neural network named
6L-ANN is proposed that classifies the instance outcome being preterm or full-term. Figure 1 shows the
architecture of the proposed model.

Int J Elec & Comp Eng, Vol. 14, No. 6, December 2024: 6129-6139
Int J Elec & Comp Eng ISSN: 2088-8708  6131

Figure 1. Architecture of proposed model

2.1. Mother’s significant feature dataset

With the aim to work towards the betterment of mother and child, a dataset titled MSF dataset is
created. MSF dataset gives an insight into the mother’s mental and physical health conditions across three
phases (teenage, after marriage, during pregnancy) of her reproductive age. The majority of the women who
delivered babies between February 2018 and September 2019 at D. Y. Patil Hospital, located in the Mumbai
Metropolitan Region of India, were interviewed. MSF dataset consists of 119 features of 1,000 women,
which are further divided into four categories namely physical (22 features), social (19 features), lifestyle
(68 features), and stress (10 features). Out of all the records 172 women had Preterm delivery while 828 had
full-term delivery. Details about the planning, challenges and procedures followed to create the MSF dataset
are mentioned in our previous work [29]. MSF dataset is partially available online [30]. Our previously
proposed methods on the MSF dataset include models like Variation and Information based random forest
(VIBRF) [31], [32] which focuses on relevant feature selection and random forest based fuzzy feature
weighing for class imbalance (RFFWCI) [33] which weighs features for imbalance handling. This study,
proposes MPMU method which realizes the contribution of instances of the dataset towards imbalance issue.
To validate the performance of the proposed MPMU method, seven standard imbalanced datasets from the
Keel repository [34] are used for experimentation.

2.2. Proposed imbalance handling method

As discussed in the literature under section 2, existing cluster-based approaches have relied on
resampling and have looked upon the clusters individually and not in relation to the dataset as a whole. The
proposed imbalance handling method MPMU provides a novel complete solution where clusters are analyzed
individually as well as in relation to the overall dataset and instances are weighted by investigating the
cluster, they fall in. MPMU consists of two major components namely instance weighing (IW) and random
over sampling (ROS), which are explained further in this section.

2.2.1. Majority penalizing minority upscaling

The proposed MPMU algorithm aims at understanding the mixing up of majority class instances
with minority class instances. MPMU algorithm works on majority instances with an IW method and on
minority instances by randomly upscaling them. MPMU algorithm uses k-means clustering to group similar
instances together, considering intra-cluster variance as expressed in (1) for 𝑘 clusters and 𝑛 instances with 𝑐𝑗
as the centroid of the cluster.

(𝑗)
𝑗 = ∑𝑘𝑗=1 ∑𝑛𝑖=1 ∥ 𝑥𝑖 − 𝑐𝑗 ∥2 (1)

Weights of majority instances are decided based on the composition of the cluster they fall in, thus
the number of clusters 𝑀 to be formed reflects the proportion of majority class instances in the whole dataset
i.e. probability of occurrence of majority class as shown in (3) of Table 1, where 𝑃(𝑀𝑖𝑛) is the probability of
occurrence of minority instance and 𝑃(𝑀𝑎𝑗) is the probability of occurrence of majority instance in the
dataset. 𝑀 is directly proportional to the dataset imbalance, with more imbalance in the dataset greater
number of clusters will be formed, same can be seen in Table 2. A perfectly balanced dataset will have two
clusters with very minimal mix. Figure 2 shows the flowchart for the proposed architecture using MSF
dataset, MPMU and 6L-ANN. Input to the model is the MSF dataset and output is a fuzzy value depicting the
probability of occurrence of PTB and full-term birth. In Figure 2, the blue color boxes show the steps
involved for majority instance handling i.e. IW, the green color box depicts the work done for the minority
class and the grey color boxes denote the processing of the dataset as a whole. MPMU algorithm is
mentioned in Table 1.
A novel approach for imbalanced instance handling toward better preterm birth … (Himani Deshpande)
6132  ISSN: 2088-8708

Table 1. MPMU algorithm

Input: Dataset Y with ‘X’ instances and ‘F’ features
Output: Processed majority and minority instances
1. Calculate dataset imbalance ‘DI’
𝐴
𝐷𝐼 = (2)
𝐵

2. Calculate the number of clusters ‘M’, to be formed.

3. Split dataset Y into ‘M’ clusters.

𝑀 = 𝑃(𝑀𝑖𝑛 𝑈 𝑀𝑎𝑗) − 𝑃(𝑀𝑖𝑛) (3)

4. Create vector V_CMC[] of size X.

5. for N=1 to M (each of the M clusters)
calculate within Cluster Imbalance ‘CIN’
𝐶𝑁𝐴
𝐶𝐼𝑁 = (4)
𝐶𝑁𝐵

6. for i=1 to X (each of the X instances)

• identify cluster ‘N’ within which instance ‘i’ falls.
• identify CIN for cluster ‘N’
if (CIN <  or CIN < DI)
CMCN=0
else
CMCN=|CIN–DI| (5)
• V_CMC[i]=CMCN
7. for i=1 to X (across X instances)
for j=1 to F (across F features)
if (Minority instance)
Y[i][j]=Y[i][j]
else
Y[i][j]=Y[i][j]–(*V_CMC[i]) (6)
8. Randomly oversample minority instances such that A=B

Table 2. Datasets analysis

Dataset No. of features No. of instances Dataset imbalance No. of clusters
1 Physical 22 1,000 4.82 7
2 Lifestyle 68 1,000 4.82 7
3 Stress 10 1,000 4.82 7
4 Social 19 633 5.95 7
5 Physical+Lifestyle+Stress 100 1,000 4.82 7
6 Pima India 7 768 1.86 4
7 vehicle2 13 988 2.88 6
8 segment0 10 5,472 6.02 7
9 ecoli3 7 336 8.6 8
10 Page-blocks0 8 514 8.79 8
11 Yeast2_vs_4 8 768 9.08 8
12 vowel0 19 2,308 9.98 8

2.2.2. Majority instance weighing

A major contribution of the MPMU algorithm is to penalize the majority class instances through the
IW method. Formed clusters are analyzed to identify the majority class instances which are close to minority
class instances, thus difficult to identify and causing misclassification. The IW method subdues the identified
majority class instances by penalizing them. The penalty for each majority class instance is decided by
looking into the composition of the cluster it falls in. The penalty is not the same for all the majority class
instances as the aim is to subside only those majority class instances which are inseparable from minority
class instances. Custer imbalance is calculated for each of the ‘𝑀’ clusters. Cluster imbalance 𝐶𝐼𝑁 of the 𝑁 𝑡ℎ
cluster is calculated as shown in (4) in Table 1, where 𝐶𝑁𝐴 is the number of minority class instances in the
𝑁 𝑡ℎ cluster and 𝐶𝑁𝐵 is the number of majority class instances in the 𝑁 𝑡ℎ cluster respectively. dataset
imbalance (DI) is calculated using (2) in Table 1, where 𝐴 is the number of minority class instances and 𝐵 is
the number of majority class instances in the dataset. The imbalance of the dataset as a whole plays a
significant role in deciding the penalty for the majority class instances. If the imbalance within the cluster is
greater than the imbalance of the dataset, it is considered to be defying the dataset composition and thus, a
penalty will be applied to the majority instances of such clusters. The proposed algorithm penalizes the
majority class instances only if the ‘𝐶𝐼’ value of the cluster into which the instance is falling, is greater than

Int J Elec & Comp Eng, Vol. 14, No. 6, December 2024: 6129-6139
Int J Elec & Comp Eng ISSN: 2088-8708  6133

the value of ‘𝐷𝐼’. Cluster misclassification cost (CMC) for each of the ‘𝑀’ clusters are different and is
decided based on the variation between the cluster imbalance and dataset imbalance. 𝐶𝑀𝐶𝑁 value for the 𝑁 𝑡ℎ
cluster is calculated using (5) in Table 1. If the cluster imbalance is very low i.e. below a certain threshold i.e.
‘’, it shows that the cluster constitutes mainly of the majority class and the presence of the minority class
can be due to noise or outliers. In this case, majority class instances will not be penalized. For a dataset with
‘𝑋’ instances, vector 𝑉_𝐶𝑀𝐶[] of size 𝑋, is used to store the CMC values corresponding to each of the
dataset instances looking into the cluster it falls into. The CMC value of 𝑁 𝑡ℎ cluster is calculated using (6) of
Table 1, considering the imbalance of the cluster and dataset. As shown in (6), while penalising the 𝑗𝑡ℎ
feature of the 𝑖 𝑡ℎ instance,  value is used to normalize the difference between the value of features and
CMC value. For the MSF dataset, optimal value of  and  is found to be 0.7 and 0.3 after empirical
analysis.

Figure 2. Flowchart of proposed imbalanced instance handling method

2.2.3. Cluster formation

Figure 3, shows an example for the formation of MPMU clusters and how the positive class
instances are penalized. In Figure 3, red circles denote minority class instances and the black +(𝑝𝑙𝑢𝑠) sign
denotes majority instances. In Figure 3, cluster 1 and cluster 4 have instances only from one class, thus there
is no penalty on these instances. Cluster 2 has a very low imbalance i.e. below threshold , thus no penalty on
majority class instances. Clusters 3 and 5 have more imbalance as compared to the overall dataset, the
majority class instances of these clusters are penalized by looking at the composition of the cluster and
calculating the CMC value as expressed in (6). Different shades of grey for majority class instances (+) in
Figure 3, depicts the different intensity of the penalty, lighter shade means a higher penalty. MPMU
algorithm does not penalize minority instances.

A novel approach for imbalanced instance handling toward better preterm birth … (Himani Deshpande)
6134  ISSN: 2088-8708

Figure 3. Cluster formation

2.2.4. Minority instance handling

Even after penalizing the majority class instances, the composition of the dataset remains inclined
towards the majority class. To balance the composition of the dataset, the MPMU algorithm oversamples the
minority class instances in the dataset, such that the number of minority class instances gets equal to the
number of majority class instances as shown in step 8 of Table 1. For the MSF dataset, samples are collected
by interviewing women from similar geographic and demographic locations. Keeping the same in mind, for
over-sampling, no synthetic instances are generated instead it is done by random duplication of minority
samples, thus even after over-sampling the instance, the feature value remains true to the original population
of the MSF dataset.

2.3. 6L_ANN
Variations across different hyperparameters of artificial neural network (ANN) models using
multi-layer perceptron (MLP) with back propagation are used in the experiments. Based on the empirical
analysis, 6L_ANN, a six-layered ANN architecture is selected for classification which consists of two
drop-out layers and two activation layers along with an output and an input layer. Rectified linear activation
unit (ReLU) activation function is used for hidden layers and the output layer uses the sigmoid activation
function. MSF dataset processed using the proposed MPMU method is further used to train the 6L_ANN
network to classify PTB and FTB outcomes.

3. RESULTS AND DISCUSSION

This section of the paper presents the classification results using the proposed MPMU method and
6L-ANN network. Multiple experiments are performed to understand the nature of the proposed methods
with different classifiers and datasets. To check the imbalance handling efficiency of the MPMU algorithm,
along with the proposed 6L-ANN, five popular classification algorithms, namely, random forest (RF),
decision tree (DT), gaussian NB (GNB), k-nearest neighbors (KNN) (𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑟 = 3), and MLP are used for
experimentation. Researchers suggest that while working on an imbalanced dataset area under the curve
(AUC) curve, precision value and prediction accuracy together need to be considered to evaluate the
performance of the classification task [35], [36]. Prediction accuracy evaluates the overall performance of a
classifier, AUC value asses the capability to distinguish between the classes, and precision measures the PTB
class. Thus, this study considers precision, AUC and prediction accuracy for evaluating the performance of
the proposed method.

3.1. Datasets analysis

For experimentation, along with MSF dataset, 7 standard datasets are used. Table 2, shows the
dataset details in terms of dataset size and imbalance. The first five rows of Table 2, shows MSF datasets
information and the remaining 7 rows have information about the other standard datasets used. There are
missing records under social features of the MSF dataset, for the same reason only 633 instances of it are
considered during experimentation. As shown in Table 2, standard datasets are selected with variations in
terms of imbalance ratio, number of features and number of instances so that the experimental results validate
the versatility of the proposed MPMU method. The last column of Table 2 shows the count of clusters
formed as expressed as ‘𝑀’ in (3), of the MPMU algorithm. For experimentation, all the datasets are
preprocessed for handling missing values and have been normalized, such that for each record, the sum of the

Int J Elec & Comp Eng, Vol. 14, No. 6, December 2024: 6129-6139
Int J Elec & Comp Eng ISSN: 2088-8708  6135

square of normalized values is equal to one. A total of 70% of the whole dataset is taken for training while
30% is considered for testing.

3.2. Experimental results on MSF dataset

Multiple experiments are performed on the MSF dataset to understand the effectiveness of
penalizing (IW) and oversampling steps of the MPMU algorithm. Figure 4 shows the classification results on
five different MSF sub-datasets as represented in the first five rows of Table 2. Experiments are conducted
with seven combinations of data processing methods and classifiers. Figure 4 shows the results of instance
weight (IW) and MPMU method in combination with MLP, GNB, KNN, RF, DT and the proposed 6L-ANN
classifier. In Figure 4, the MPMU model is evaluated with 6L-ANN classifier. There were many missing
values in the social sub-dataset for the same reason it was not included while conducting combined MSF
dataset experimentation. Figures 4(a) to 4(c) show the precision value, AUC and prediction accuracy using
the proposed IW step and MPMU method on MSF sub-datasets.

(a)

(b)

(c)

Figure 4. Classification results on MSF Sub-datasets (physical, lifestyle, stress, social and combined MSF
features (physical, lifestyle and stress features)) using the proposed IW step and proposed MPMU method
with classifiers namely MLP, GNB, KNN, RF, DT and proposed 6L-ANN (a) demonstrate classification
results in terms of precision, (b) prediction results in terms of AUC value, and (c) prediction accuracy of
classification

MSF dataset is an imbalanced dataset and results reflect classification bias, same can be justified by
the results in Figure 4. While pre-processing the datasets using IW, despite higher prediction accuracy and a
decent AUC value, the precision value is very low for classifiers like GNB, KNN and DT for all MSF
datasets, this shows that the results are inclined toward the minority class. Considering the prediction of PTB

A novel approach for imbalanced instance handling toward better preterm birth … (Himani Deshpande)
6136  ISSN: 2088-8708

as a priority concern over the prediction of FTB, thus counting precision as the most crucial evaluation
parameter, it could be concluded that RF is giving the best results for all the IW pre-processed MSF datasets,
and the second-best results are achieved by 6L-ANN network for all the MSF sub-datasets. It is further
observed that the IW step alone is not as effective as the MPMU method which combines majority instance
weighing and minority oversampling, as there is a significant improvement in results for all MSF sub-
datasets while pre-processing using the proposed MPMU method and further using 6L-ANN network as the
classifier.
Experimental results in Figure 4, validate the strength of the proposed model i.e. MPMU with
6L-ANN classifier, precision between 0.90 to 0.97 is achieved on MST sub-datasets, an AUC value ranging
between 0.86 to 0.99 is observed, and prediction accuracy of 93.00% to 99.47% is achieved. Looking into the
results in Figure 4, it is observed that mother’s lifestyle and stress features are achieving better classification
results for PTB prediction. Third-best classification results are achieved using social features and physical
features classification results are found to be the least accurate ones. One of the probable reasons for inferior
results for social features would be a smaller dataset and higher-class imbalance. This signifies that the
mother’s mental health features are closely related to the pregnancy outcome in terms of the classification of
PTB and FTB. Looking into the results of the proposed MPMU method in Table 3, with standard classifiers
and with the results in Figure 4 with 6L-ANN classifier shows that among all the classifiers MPMU is giving
the best results using 6L-ANN classifier. The experimental results on different sub-datasets of MPMU as
shown in Figure 4, support the hypothesis that a mother’s mental health plays a crucial role in predicting the
PTB birth outcome. As seen in Figure 4, most of the classifiers are giving better predictions using the
lifestyle and stress feature sub-dataset. Using lifestyle features precision is achieved in the range of 0.23 to
0.97, AUC value is achieved between 0.72 to 0.99 and prediction accuracy is in the range of 55%99%. Using
stress features precision is achieved in the range 0.25 to 0.98, AUC value is achieved between 0.62 to 0.99
and prediction accuracy in the range 59% to 100%.
To understand the performance of MPMU method over existing data imbalance solutions,
experiments are conducted using algorithms proposed by researchers in the past. Table 3 demonstrates the
results for the same. The data processing methods used for implementation are scaled values (SV), clustered
SMOTE (CS) [37], random over sampling (ROS) [38] and the proposed MPMU algorithm.
Being imbalanced MSF dataset is found to be performing extremely poorly in terms of precision and
AUC value, even after scaling the values using SV method, classifiers fail to perform the same can be seen in
the results of Table 3. The cluster-based instance weighing and oversampling approach of MPMU is capable
of handling majority bias and inseparability issues of MSF datasets the same can be seen in the results of
Table 3. There is a significant improvement in the results of MPMU method with all the used classifiers
when compared to SV, CS, and ROS imbalance handling methods. On average precision value is improved
by 48%, AUC by 26% and prediction accuracy by 16%, when using MPMU against SV. Precision value is
improved by 52%, AUC is improved by 32% and prediction accuracy is improved by 26%, comparing the
results of MPMU against CS. Against ROS method MPMU results are improved by 54%, 31% and 16% for
precision, AUC and prediction accuracy. Greater improvement in precision values using MPMU validates the
success of MPMU in identifying PTB instances correctly. Among the used standard classifiers RF is showing
the best results for all the selected imbalance handling methods on MSF sub-datasets.

Table 3. Comparative results of different imbalance handling methods on MST sub-datasets

MPMU

MPMU

MPMU

MPMU

MPMU
ROS

ROS

ROS

ROS

ROS
SV

SV

SV

SV

SV
CS

CS

CS

CS

CS

Physical Precision 0.18 0.19 0.19 0.42 0.22 0.21 0.2 0.8 0.2 0.2 0.2 0.27 0.29 0.29 0.21 0.44 0.23 0.23 0.23 0.53
AUC 0.53 0.47 0.55 0.76 0.6 0.45 0.59 0.93 0.52 0.51 0.52 0.66 0.64 0.52 0.49 0.82 0.52 0.5 0.52 0.78
Accuracy 71.7 74.5 72.8 86.2 83.9 80.9 82.8 91.6 77.6 65.1 77 83.2 18.4 18.3 31.3 58.7 66.5 66.5 66.5 79.2
Lifestyle Precision 0.17 0.19 0.17 0.62 0.21 0.2 0.21 0.87 0.17 0.17 0.17 0.23 0.16 0.16 0.15 0.42 0.21 0.21 0.21 0.69
AUC 0.49 0.51 0.5 0.84 0.54 0.51 0.54 0.95 0.49 0.51 0.49 0.55 0.5 0.51 0.49 0.83 0.52 0.48 0.52 0.84
Accuracy 69.3 69.3 71.2 91 82.2 75.3 82.4 93.2 79.2 63.7 79 79.8 16.8 16.8 26.5 72.8 70 70 70 86
Stress Precision 0.2 0.2 0.19 0.69 0.21 0.21 0.21 0.76 0.2 0.19 0.2 0.41 0.23 0.23 0.2 0.25 0.24 0.24 0.24 0.39
AUC 0.53 0.52 0.53 0.87 0.51 0.49 0.51 0.88 0.54 0.5 0.53 0.71 0.58 0.52 0.51 0.59 0.56 0.52 0.56 0.69
Accuracy 81.1 58.6 80.3 91.1 79.5 58.8 79.1 92.1 77.3 62.1 74.7 83.8 80.4 59.9 68.7 78.7 59.5 59.5 59.5 62.5
Social Precision 0.17 0.17 0.16 0.52 0.18 0.18 0.17 0.81 0.15 0.14 0.15 0.18 0.17 0.18 0.15 0.26 0.17 0.17 0.17 0.4
AUC 0.54 0.48 0.52 0.82 0.51 0.48 0.51 0.94 0.53 0.53 0.53 0.54 0.51 0.48 0.43 0.66 0.51 0.54 0.51 0.75
Accuracy 75.6 73.5 74.7 89.7 81.9 76.7 80.9 93.3 82.7 69 80.8 80.1 59.7 34.2 62.3 75.6 66.8 66.8 66.8 76.4
Lifestyle+ Precision 0.2 0.2 0.19 0.24 0.25 0.22 0.26 0.53 0.17 0.16 0.17 0.21 0.24 0.24 0.18 0.26 0.21 0.21 0.21 0.76
Physical+ AUC 0.53 0.5 0.53 0.62 0.58 0.52 0.59 0.77 0.5 0.48 0.5 0.53 0.58 0.47 0.46 0.66 0.53 0.49 0.53 0.88
Stress Accuracy 72.2 72.2 70.3 77.3 81.8 80.1 81.7 84.6 79 62.9 78.4 78.6 24.7 25 52.6 37.6 72.3 72.3 72.3 89.5

Int J Elec & Comp Eng, Vol. 14, No. 6, December 2024: 6129-6139
Int J Elec & Comp Eng ISSN: 2088-8708  6137

3.3. Experimental results on standard datasets

For validation of the efficiency of the proposed MPMU algorithm, experiments are conducted on
seven standard datasets using five different classifiers. These datasets are commonly used by researchers in
the imbalanced handling literature. The number of clusters formed using MPMU, for each of the datasets are
as shown in Table 2. It is observed that the MPMU algorithm is overshining other algorithms for most of the
datasets and classifier combinations. Results in Table 4, depict that the proposed MPMU method works well
with large size datasets like vowel0 and segment0. For smaller datasets like Pima India and ecoli3 the final
results are not very good but there is a remarkable improvement in results when compared with SV, CS and
ROS imbalance handling methods. While comparing the results of SV over the results of the MPMU
algorithm, improvement in results has been observed between 0% to 21% for precision, between 0% to 10%
for AUC and between 0% to 17% for prediction accuracy. The results in the table convey that along with the
MSF dataset, the proposed methods work equally well for other imbalanced datasets with varied sizes and
imbalances.

Table 4. Comparative results of different imbalance handling methods on standard imbalanced datasets
Datasets Decision Tree Random Forest KNN Gaussian NB MLP
SV CS ROS MPMU SV CS ROS MPMU SV CS ROS MPMU SV CS ROS MPMU SV CS ROS MPMU
ecoli3 0.47 0.32 0.39 0.68 0.72 0.68 0.65 0.86 0.46 0.4 0.37 0.66 0.63 0.6 0.65 0.8 0.66 0.66 0.66 0.81
Pima India 0.47 0.49 0.46 0.47 0.71 0.7 0.7 0.73 0.56 0.57 0.56 0.58 0.67 0.66 0.67 0.72 0.7 0.7 0.7 0.78
Precision

yeast-2_vs_4 0.65 0.57 0.56 0.67 0.92 0.93 0.92 0.94 0.63 0.64 0.59 0.88 0.72 0.76 0.72 0.83 0.84 0.84 0.84 0.88
page-blocks0 0.72 0.71 0.7 0.92 0.94 0.94 0.93 0.99 0.78 0.78 0.72 0.8 0.58 0.54 0.58 0.55 0.89 0.89 0.89 0.92
segment0 0.95 0.94 0.95 0.95 1 1 1 1 0.99 0.98 0.97 1 0.92 0.8 0.92 0.99 1 1 1 1
vowel0 0.85 0.79 0.82 0.9 0.99 1 0.99 1 1 1 0.99 1 0.88 0.58 0.87 0.92 1 1 1 1
vehicle2 0.85 0.86 0.85 0.89 1 0.99 1 1 0.93 0.94 0.88 0.96 0.62 0.54 0.61 0.76 0.99 0.99 0.99 1
ecoli3 0.79 0.71 0.49 0.89 0.95 0.94 0.53 0.95 0.87 0.86 0.5 0.87 0.9 0.9 0.5 0.91 0.95 0.95 0.51 0.98
Pima India 0.67 0.68 0.5 0.68 0.82 0.83 0.5 0.85 0.74 0.74 0.49 0.77 0.81 0.8 0.51 0.84 0.83 0.83 0.49 0.87
yeast-2_vs_4 0.88 0.86 0.52 0.92 1 1 0.51 0.99 0.86 0.86 0.51 0.95 0.9 0.88 0.5 0.9 0.94 0.94 0.43 0.96
AUC

page-blocks0 0.91 0.91 0.49 0.98 0.99 0.99 0.49 1 0.93 0.93 0.51 0.93 0.93 0.89 0.5 0.91 0.98 0.98 0.49 0.99
segment0 0.98 0.98 0.5 0.99 1 1 0.5 1 0.99 0.99 0.5 1 0.98 0.96 0.48 1 1 1 0.51 1
vowel0 0.97 0.95 0.51 0.97 1 1 0.5 1 1 1 0.51 1 0.98 0.9 0.47 0.99 1 1 0.49 1
vehicle2 0.94 0.95 0.49 0.96 1 1 0.53 1 0.98 0.98 0.53 0.99 0.82 0.71 0.47 0.9 1 1 0.5 1
ecoli3 91.1 87.3 89.94 95.88 94.1 92.9 91.8 95.88 91.8 89.8 88.8 93.24 75.1 75.7 73.4 81.69 89.8 89.8 89.8 91.59
Pima India 69.5 70.6 68.55 69.84 75.3 76.1 75.5 78.21 71.1 71 69.2 74.46 74.4 73.3 72.8 77.42 74.8 74.8 74.8 79.44
Accuracy

yeast-2_vs_4 95.5 94.1 94.09 95.91 95.9 96.1 95.9 97.53 95.3 95.2 94 97.74 31.3 49 27.7 33.98 93.1 93.1 93.1 91.61
page-blocks0 96.6 96.6 96.48 99.11 97.6 97.6 97.6 99.41 96.1 96.1 95.3 96.02 89.7 90 89.1 89.31 94.8 94.8 94.8 94.81
segment0 99.2 99.1 99.21 99.26 99.7 99.8 99.7 99.76 99.4 99.3 99.3 99.88 82.2 88.8 79.9 98.92 99.7 99.8 99.7 98.57
vowel0 98.5 97.8 98.15 98.99 99.4 99.6 99.6 99.61 99.9 99.9 99.8 99.72 92.7 89.2 86.8 95.74 99.8 99.8 99.8 99.77

4. CONCLUSION
This work proposes a novel imbalance handling model towards the prediction of pre-term birth
using the created MSF dataset, which consists of lifestyle, social, physical and stress features of 1,000
mothers. The proposed MPMU imbalance handling method works on class inseparability and classification
bias issues by subduing the majority class instances following a cluster-based penalty approach and
strengthening the minority instances by random oversampling. It has been observed that the proposed model
consisting of MPMU and 6L-ANN classifier, is highly efficient in classifying PTB and FTB outcomes on
MSF datasets with precision ranging from 0.90 to 0.97, AUC between 0.86 to 0.99, and prediction accuracy
falling under 93% to 99.47%. MPMU imbalance handling method has been validated on other imbalanced
datasets too. MPMU shows a superior performance when compared to other Imbalance handling methods on
MSF as well as other datasets. Thus, this research work succeeds in proposing an efficient cluster-based
imbalanced instance handling method. Experimental results on the mental and physical health-related
features of the mother, show that the lifestyle of the mother and Stress-related features are major contributors
towards analyzing her maternal health which defines her mental health. Thus, this study, concludes that a
mother’s mental health is a major contributor towards understanding the pre-term birth possibilities.

REFERENCES
[1] T. Strunk, A. Currie, P. Richmond, K. Simmer, and D. Burgner, “Innate immunity in human newborn infants: prematurity means
more than immaturity,” The Journal of Maternal-Fetal and Neonatal Medicine, vol. 24, no. 1, pp. 25–31, Jan. 2011, doi:
10.3109/14767058.2010.482605.
[2] M. A. Arfandi, A. Ansariadi, R. Amiruddin, W. Wahiduddin, A. U. Salmah, and A. Salam, “Preterm birth risk in mother with
hypertensive disorders of pregnancy,” International Journal of Public Health Science, vol. 12, no. 2, pp. 590–597, 2023, doi:
10.11591/ijphs.v12i2.22599.
[3] G. Homan, J. Litt, and R. J. Norman, “The FAST study: Fertility assessment and advice targeting lifestyle choices and
behaviours: A pilot study,” Human Reproduction, vol. 27, no. 8, pp. 2396–2404, 2012, doi: 10.1093/humrep/des176.
[4] S. S. Kaddi and M. M. Patil, “Ensemble learning based health care claim fraud detection in an imbalance data environment,”

A novel approach for imbalanced instance handling toward better preterm birth … (Himani Deshpande)
6138  ISSN: 2088-8708

Indonesian Journal of Electrical Engineering and Computer Science, vol. 32, no. 3, pp. 1686–1694, 2023, doi:
10.11591/ijeecs.v32.i3.pp1686-1694.
[5] G. Naik, D. Siroya, M. Nisar, B. Shah, and H. Deshpande, “Analysis on the efficacy of ANN on small imbalanced datasets,”
Springer Proceedings in Mathematics and Statistics, vol. 403, pp. 129–137, 2023, doi: 10.1007/978-3-031-16178-0_11.
[6] M. H. B. M. Razali, R. Bin Saian, Y. B. Wah, and K. R. Ku-Mahamud, “A class skew-insensitive ACO-based decision tree
algorithm for imbalanced data sets,” Indonesian Journal of Electrical Engineering and Computer Science, vol. 21, no. 1,
pp. 412–419, 2021, doi: 10.11591/ijeecs.v21.i1.pp412-419.
[7] M. Son and E. S. Miller, “Predicting preterm birth: Cervical length and fetal fibronectin,” Seminars in Perinatology, vol. 41,
no. 8, pp. 445–451, 2017, doi: 10.1053/[Link].2017.08.002.
[8] P. Branco, L. Torgo, and R. P. Ribeiro, “A survey of predictive modeling on imbalanced domains,” ACM Computing Surveys,
vol. 49, no. 2, 2016, doi: 10.1145/2907070.
[9] F. Nieto-Del-amor et al., “Combination of feature selection and resampling methods to predict preterm birth based on
electrohysterographic signals from imbalance data,” Sensors, vol. 22, no. 14, 2022, doi: 10.3390/s22145098.
[10] M. H. Saeed and J. I. Hama, “Cardiac disease prediction using AI algorithms with SelectKBest,” Medical and Biological
Engineering and Computing, vol. 61, no. 12, pp. 3397–3408, 2023, doi: 10.1007/s11517-023-02918-8.
[11] W. A. Kusuma, N. Noviana, L. S. Hasibuan, and M. Nurilmala, “Improving DNA barcode-based fish identification system on
imbalanced data using SMOTE,” Telkomnika (Telecommunication Computing Electronics and Control), vol. 15, no. 3,
pp. 1230–1238, 2017, doi: 10.12928/TELKOMNIKA.v15i3.5011.
[12] B. Zhu, X. Pan, S. vanden Broucke, and J. Xiao, “A GAN-based hybrid sampling method for imbalanced customer classification,”
Information Sciences, vol. 609, pp. 1397–1411, 2022, doi: 10.1016/[Link].2022.07.145.
[13] S. Ali, P. Chourasia, Z. Tayebi, B. Bello, and M. Patterson, “ViralVectors: compact and scalable alignment-free virome feature
generation,” arXiv preprint arXiv: 2304.02891, Apr. 2023.
[14] G. Andresini, A. Appice, and D. Malerba, “Dealing with class imbalance in android malware detection by cascading clustering
and classification,” Studies in Computational Intelligence, vol. 880, pp. 173–187, 2020, doi: 10.1007/978-3-030-36617-9_11.
[15] R. Choudhary and S. Shukla, “A clustering based ensemble of weighted kernelized extreme learning machine for class imbalance
learning,” Expert Systems with Applications, vol. 164, 2021, doi: 10.1016/[Link].2020.114041.
[16] H. H. Htun, M. Biehl, and N. Petkov, “Survey of feature selection and extraction techniques for stock market prediction,”
Financial Innovation, vol. 9, no. 1, 2023, doi: 10.1186/s40854-022-00441-7.
[17] P. Dhal and C. Azad, “A comprehensive survey on feature selection in the various fields of machine learning,” Applied
Intelligence, vol. 52, no. 4, pp. 4543–4581, 2022, doi: 10.1007/s10489-021-02550-9.
[18] J. W. Wan and M. Yang, “Survey on cost-sensitive learning method,” Ruan Jian Xue Bao/Journal of Software, vol. 31, no. 1,
pp. 113–136, 2020, doi: 10.13328/[Link].005871.
[19] A. Telikani, A. H. Gandomi, K. K. R. Choo, and J. Shen, “A cost-sensitive deep learning-based approach for network traffic
classification,” IEEE Transactions on Network and Service Management, vol. 19, no. 1, pp. 661–670, 2022, doi:
10.1109/TNSM.2021.3112283.
[20] F. Cheng, J. Zhang, and C. Wen, “Cost-sensitive large margin distribution machine for classification of imbalanced data,” Pattern
Recognition Letters, vol. 80, pp. 107–112, 2016, doi: 10.1016/[Link].2016.06.009.
[21] R. Gunawan, H. A. Ghani, N. Khamis, J. Al Amien, and E. Ismanto, “Deep learning approach to DDoS attack with imbalanced
data at the application layer,” Telkomnika (Telecommunication Computing Electronics and Control), vol. 21, no. 5,
pp. 1060–1067, 2023, doi: 10.12928/TELKOMNIKA.v21i5.24857.
[22] G. Casañola-Martin et al., “Exploring different strategies for imbalanced ADME data problem: case study on Caco-2 permeability
modeling,” Molecular diversity, vol. 20, no. 1, pp. 93–109, 2016.
[23] J. Chen, Z. Wu, and J. Zhang, “Driving safety risk prediction using cost-sensitive with nonnegativity-constrained autoencoders
based on imbalanced naturalistic driving data,” IEEE Transactions on Intelligent Transportation Systems, vol. 20, no. 12,
pp. 4450–4465, Dec. 2019, doi: 10.1109/TITS.2018.2886280.
[24] M. G. Madden and S. S. Khan, “One-class classification: taxonomy of study and review of techniques,” Knowledge Engineering
Review, vol. 20, no. 2, pp. 117–125, 2004.
[25] H. J. Lee and S. Cho, “The novelty detection approach for different degrees of class imbalance,” Lecture Notes in Computer
Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), pp. 21–30, 2006, doi:
10.1007/11893257_3.
[26] G. Haixiang, L. Yijing, J. Shang, G. Mingyun, H. Yuanyue, and G. Bing, “Learning from class-imbalanced data: Review of
methods and applications,” Expert Systems with Applications, vol. 73, pp. 220–239, May 2017, doi: 10.1016/[Link].2016.12.035.
[27] L. Rokach, “Ensemble-based classifiers,” Artificial Intelligence Review, vol. 33, no. 1–2, pp. 1–39, 2010, doi: 10.1007/s10462-
009-9124-7.
[28] N. H. A. Malek, W. F. W. Yaacob, Y. B. Wah, S. A. Md Nasir, N. Shaadan, and S. W. Indratno, “Comparison of ensemble hybrid
sampling with bagging and boosting machine learning approach for imbalanced data,” Indonesian Journal of Electrical
Engineering and Computer Science, vol. 29, no. 1, pp. 598–608, 2023, doi: 10.11591/ijeecs.v29.i1.pp598-608.
[29] H. Deshpande and L. Ragha, “Realizing mother’s features influential on childbirth experience, towards creation of a dataset,”
Data Science, pp. 143–167, 2022, doi: 10.1201/9781003283249-10.
[30] H. Deshpande and L. Ragha, “Mother’s significant feature (MSF) dataset,” IEEE DataPort, 2021.
[31] H. Deshpande and L. Ragha, “Mother’s lifestyle feature relevance for NICU and preterm birth prediction,” ITM Web of
Conferences, vol. 40, 2021, doi: 10.1051/itmconf/20214003039.
[32] L. Ragha and H. S. Deshpande, “A hybrid random forest-based feature selection model using mutual information and F-score for
preterm birth classification,” International Journal of Medical Engineering and Informatics, vol. 15, no. 1, 2023, doi:
10.1504/ijmei.2023.10051207.
[33] H. Deshpande and L. Ragha, “Random forest based fuzzy feature weighing model for imbalance class distribution towards
preterm-birth classification,” SSRN Electronic Journal, 2021, doi: 10.2139/ssrn.3867354.
[34] L. S. Ramos et al., “KEEL data-mining software tool: Data set repository, integration of algorithms and experimental analysis
framework,” Journal of Multiple-Valued Logic and Soft Computing, 2011.
[35] H. He and E. Garcia, “Learning from imbalanced data,” IEEE Transactions on Knowledge and Data Engineering, 2009.
[36] B. Mohamed, D. H. Kheliouane, and T. A. Alitouche, “Evaluation measures for models assessment over imbalanced data sets,”
Journal of Information Engineering and Applications, vol. 3, no. 10, pp. 27–38, 2013.
[37] F. Last, G. Douzas, and F. Bacao, “Oversampling for imbalanced learning based on k-means and SMOTE,” arXiv preprint arXiv:
1711.00837, Nov. 2017, doi: 10.1016/[Link].2018.06.056.

Int J Elec & Comp Eng, Vol. 14, No. 6, December 2024: 6129-6139
Int J Elec & Comp Eng ISSN: 2088-8708  6139

[38] M. Hayaty, S. Muthmainah, and S. M. Ghufran, “Random and synthetic over-sampling approach to resolve data imbalance in
classification,” International Journal of Artificial Intelligence Research, vol. 4, no. 2, 2021, doi: 10.29099/ijair.v4i2.152.

BIOGRAPHIES OF AUTHORS

Himani Deshpande a dedicated scholar, holds a Ph.D. degree in computer

science from the University of Mumbai in the year 2022. With more than a decade of
experience in academics, she serves as an assistant professor in the Department of Artificial
Intelligence and Data Science at Thadomal Shahani Engineering College, Mumbai, India. She
has an industry experience of 3 years as software engineer. Her expertise spans diverse
domains, including machine learning, deep learning, and generative AI. Beyond her academic
pursuits, she channels her expertise towards addressing crucial societal issues, notably in
women’s health. She can be contacted at email: [Link]@[Link].

Leena Ragha a distinguished Professor at BLDEA’s V. P. Dr. P. G. Halakatti

College of Engineering and Technology, Vijayapur, India, brings 32 years of enriching
teaching experience. Renowned for her expertise in image processing, data science, and deep
learning, she actively contributes to pioneering research endeavors. She is author of 50 plus
research articles in renowned journals and conferences. She is equally adept at administrative
duties, ensuring a holistic contribution to academia. She can be contacted at email:
[Link]@[Link].

A novel approach for imbalanced instance handling toward better preterm birth … (Himani Deshpande)