Received 22 February 2024, accepted 16 March 2024, date of publication 22 March 2024, date of current version 6 May 2024.

Digital Object Identifier 10.1109/ACCESS.2024.3380895

An Automated Approach for Predicting Road

Traffic Accident Severity Using Transformer
Learning and Explainable AI Technique
OMAR IBRAHIM ABOULOLA 1 , EBTISAM ABDULLAH ALABDULQADER 2 ,
AISHA AHMED ALARFAJ3 , SHTWAI ALSUBAI 4 , AND TAI-HOON KIM 5 , (Member, IEEE)
1 College of Computer Science and Engineering, University of Jeddah, Jeddah 21589, Saudi Arabia
2 Department of Information Technology, College of Computer and Information Sciences, King Saud University, Riyadh 11421, Saudi Arabia
3 Department of Information Systems, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, P.O. Box 84428, Riyadh
11671, Saudi Arabia
4 Department of Computer Science, College of Computer Engineering and Sciences, Prince Sattam bin Abdulaziz University, Al-Kharj 11942, Saudi Arabia
5 School of Electrical and Computer Engineering, Chonnam National University, Yeosu-si, Jeollanam-do 59626, Republic of Korea

Corresponding authors: Aisha Ahmed Alarfaj ([email protected]) and Tai-Hoon Kim ([email protected])
This work was supported by Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia, through Princess Nourah bint
Abdulrahman University Researchers Supporting Project under Grant PNURSP2024R348.

ABSTRACT Traffic accidents continue to be a significant cause of fatalities, injuries, and considerable
disruptions on our highways. Understanding the underlying factors behind these incidents is crucial for
improving safety on road networks. While recent studies have highlighted the usefulness of predictive
modeling in uncovering factors leading to accidents, there remains a gap in explaining the inner workings
of complex machine learning and deep learning models and how various features influence accident
prediction. This lack of transparency may lead to these models being perceived as black boxes, potentially
undermining trust in their findings among stakeholders. The primary aim of this research is to develop
predictive models using diverse transfer learning techniques and shed light on the most influential factors
using Shapley values. In predicting injury severity in accidents, we employ Multilayer Perceptron (MLP),
Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Residual Networks (ResNET),
EfficientNetB4, InceptionV3, Extreme Inception (Xception), Visual Geometry Group (VGG19), AlexNet,
and MobileNet. Among these models, MobileNet emerges with the highest accuracy at 0.9817. Furthermore,
by comprehending how different features impact accident prediction models, researchers can deepen their
understanding of the factors contributing to accidents and devise more effective interventions for their
prevention.

INDEX TERMS Intelligent transportation system, road accidents severity, MobileNet, explainable AI (XAI).

I. INTRODUCTION economic setbacks. To efficiently alleviate the harm and

The escalating number of automobiles on roads has become losses stemming from road traffic accidents (RTAs), under-
a pressing issue in contemporary society, coinciding with a standing the underlying factors of these events and the extent
parallel increase in traffic accidents, resulting in substantial of resulting injuries is crucial. Considering the increasing
human and economic tolls [1]. Annually, a significant intricacy of road systems and the growing number of vehicles,
number of individuals suffer injuries or fatalities in vehicular it is vital to embrace a data-centric strategy for examining
accidents worldwide, resulting in substantial human and patterns in accidents and pinpointing risk elements. Continual
investigation and comprehension of RTA causes, coupled
The associate editor coordinating the review of this manuscript and with the implementation of effective preventive measures, are
approving it for publication was Kah Phooi (Jasmine) Seng . paramount [2].

2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.
61062 For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/ VOLUME 12, 2024
O. I. Aboulola et al.: Automated Approach for Predicting Road Traffic Accident Severity

A recent study conducted by the World Health Orga- losses worldwide, there is a critical demand for effective
nization (WHO) emphasizes the worldwide importance of predictive models to aid in accident severity prediction and
RTAs, contributing to more than 1.3 million fatalities each prevention efforts. Traditional approaches often lack the
year. Globally, vehicular accidents stand as the leading predictive power and interpretability required for accurate
cause of mortality among adolescents and young adults [3]. and actionable insights. Hence, the motivation for this
The intensity of road incidents acts as a pivotal gauge for research lies in harnessing the capabilities of advanced
injuries associated with accidents, with diverse elements machine learning techniques, such as transfer learning
contributing to incidents of varying seriousness [4]. Despite with MobileNet architecture, coupled with Explainable AI
advancements, there has been no substantial reduction in (XAI) techniques like SHAP. By leveraging state-of-the-art
traffic accident fatalities and injuries over the past two methodologies, the research aims to develop more accurate,
decades. Predictive models offer a proactive approach to interpretable, and actionable models for traffic accident
addressing accident factors, potentially curbing fatalities, severity prediction. Ultimately, the overarching goal is to
reducing costs, and enhancing comprehension. Weather empower stakeholders, including transportation planners,
conditions have been discussed in relation to various road law enforcement agencies, and policymakers, with the tools
types [5], alongside other significant factors such as lighting and insights needed to implement targeted interventions and
conditions, road class, vehicle count, and traffic volume [6]. enhance road safety outcomes.
The primary objective of accident data analysis is to
identify key factors influencing the occurrence of RTAs, B. NOVELTY OF THE PROPOSED WORK
thereby addressing critical road safety concerns. The efficacy The study introduces novel elements to traffic accident
of accident prevention strategies hinges on the authenticity of severity prediction, contributing to road safety. Innovations
collected and estimated data, as well as the appropriateness of include adopting MobileNet for transfer learning, providing
chosen analytical methodologies [7]. Selecting suitable data accuracy and efficiency. A comprehensive model compari-
analysis methods is vital for discerning accident causes in son explores diverse models, enhancing understanding and
specific zones or study locations and accurately predicting highlighting transfer learning adaptability. The inclusion of
daily accident probabilities or assessing safety levels for the SHapley Additive exPlanations (SHAP) model addresses
different road user groups in those areas [8]. Consequently, model interpretability, offering transparency in predicting
research quality is contingent upon methodological selection. severity. Rigorous generalizability testing confirms the
Machine learning approaches have been utilized to predict model’s robustness beyond the initial dataset, emphasizing
traffic accidents [9], with Zhang et al. [10] utilizing gener- practical applicability. The research advances predictive
alized random forests to estimate heterogeneous treatment modeling in road safety, offering an accurate, interpretable,
effects in road safety studies, providing comprehensive and widely applicable solution for mitigating traffic acci-
information to local authorities and policymakers to enhance dent impacts. This study makes the following noteworthy
speed camera program effectiveness. Some researchers have contributions:
employed statistical methods [11], reinforcement learning
• Utilization of a MobileNet model based on transfer
approaches [12], hybrid models [13], and deep learning
learning, exhibiting exceptional accuracy in road traffic
models [14]. Zhao et al. [15] employed a deep convolutional
accident severity prediction.
neural network and random forest for accident risk prediction.
• Conducting experiments on three deep learning models
(Multilayer Perceptron (MLP), Convolutional Neu-
A. RESEARCH GAP AND MOTIVATION ral Network (CNN), and Long Short-Term Mem-
Despite numerous attempts to investigate accident- ory (LSTM)) and five transfer learning models
contributing factors, limited attention has been given to (ResNET, EfficientNetB4, InceptionV3, Xception, and
explaining black box models [16]. The authors explored MobileNet).
five machine learning models alongside explainable machine • Demonstrating the significance of various features
learning [17]. This study aims to develop an accident injury through the utilization of the SHapley Additive exPla-
severity prediction model based on transfer learning and nations (SHAP) model.
identify major contributing factors using an explainable • Testing the proposed model on another dataset to
approach. The US accident dataset (2016-2021) is employed validate its generalizability.
for predicting traffic accident severity, with the objective The organization of this research is outlined as follows:
of creating an automated system for categorizing accident Section II offers a glimpse into previous studies in this
severity. The motivation behind the research on Traffic domain. Section III unveils the suggested methodology and
Accident Severity Prediction Using MobileNet Transfer delineates the deep learning and transfer learning models.
Learning Model and SHAP XAI Technique stems from the Section IV lays out the evaluation of the proposed approach,
pressing need to enhance road safety and mitigate the adverse encompassing experimental outcomes and pertinent discus-
impacts of traffic accidents. With traffic accidents remaining sions. Ultimately, Section V functions as the culmination of
a significant cause of fatalities, injuries, and economic this study.

VOLUME 12, 2024 61063

 O. I. Aboulola et al.: Automated Approach for Predicting Road Traffic Accident Severity

II. RELATED WORK and predicted severity levels of traffic accidents with
In the past few years, the utilization of machine learning explanations. They conducted experiments on a Chinese
has surged in predicting the severity of accidents due to its traffic accident dataset. Although many researchers have
capacity to reveal concealed associations and offer more pre- utilized machine learning and deep learning for predicting
cise insights compared to conventional statistical approaches. traffic accident severity, few have conducted comparative
These statistical approaches for forecasting accident severity analyses of different deep learning methods. Additionally,
face limitations like diminished accuracy and impractical limited research has explored contributing factors using
estimation. Researchers have turned to machine learning and explanations. The incorporation of explanation techniques
deep learning methodologies to enhance prediction efficacy. enhances transparency, interpretability, explanatory capacity,
This section presents an overview of previous methodologies domain knowledge integration, and scientific coherence of
employed for forecasting traffic accident severity. models [27], addressing the common perception of prediction
Regarding traffic accident characteristics, Gan et al. [18] methods as black boxes. This research work makes use of
utilized a random forest approach to identify eight attributes the SHAP XAI technique to explain to what extent which
in traffic accident data for predicting accident severity. feature contributes to making specific target class predictions.
Factors such as engine capacity, time of day, vehicle age, Therefore, this study selects five distinct transfer learning
month, day of week, driver age group, vehicle movement, and methods and compares their predictive capabilities. As a
speed restrictions were considered. In another research [19], novel contribution, an explainable technique is applied to
several machine learning models including Naive Bayes forecast the most influential factors contributing to accidents
(NB), Random Forest (RF), Adaptive Boosting (ADA), and in the proposed models. A summary of prior studies is
Logistic Regression (LR) were evaluated for forecasting the presented in Table 1.
severity of injuries in RTAs. The highest results have been
achieved by RF with 75% accuracy. III. DATASET AND METHODOLOGY
In Saudi Arabia, Aldhari et al. [20] suggested a machine This section provides a comprehensive overview of the
learning approach to predict road accident severity, employ- dataset utilized, the deep learning models, transfer learning
ing three models: RF, LR, and XGBoost, with SHAP used to models employed, and the parameters utilized for evaluating
address bias concerns. XGBoost achieved the 71% accuracy the performance of these models in predicting the severity of
score for binary classification and 94% for multi-class traffic accidents. Additionally, the experimental framework
classification. The authors [21] devised a method to fore- adopted is depicted in Figure 1.
cast the seriousness of accidents employing deep-learning
methodologies such as Multilayer Perceptron (MLP), and A. DATASET
Recurrent Neural Network (RNN). The highest results have This study utilizes accident data spanning a five-year period
been achieved by the RNN model with 71.7% of accuracy. (2016-2020) from New Zealand, sourced from the Crash
Wahab and Jiang [22] proposed a basic CART model for Analysis System (CAS) maintained by the Te Manatū Waka
predicting the seriousness of motorcycle accidents, along Ministry of Transport. The dataset is publicly available
with the use of PART and MLP models. The CART attained a through the open data portal. Two distinct sets of data were
73.8% accuracy, while the PART model scored 73.45%. The obtained from the CAS system, containing details related
authors [23] suggested a deep neural network for predicting to individuals involved, vehicles, and accident particulars.
RTAs for the Internet of Vehicles, employing various models. These datasets, referred to as the ‘person’ dataset and the
The DNN surpassed alternative models and demonstrated ‘accident’ dataset, were amalgamated to form a comprehen-
proficiency in both the initial and subsequent phases of sive dataset focusing on factors contributing to accidents.
clustering. Initially, the combined dataset comprised 378,820 entries
Jamal et al. [24] introduced a network utilizing var- across 101 columns. However, several columns, deemed
ious machine learning models to improve the result in irrelevant to accident causation, were excluded from the
forecasting the severity of road accidents. The XGBoost study. For instance, a column containing information regard-
model outperformed others in classwise accuracy and overall ing neighboring police stations was deemed superfluous.
prediction performance, achieving an outstanding 93% As a result, 36 features relevant to various aspects of
accuracy. Manzoor et al. [25] proposed RFCNN, an ensemble accidents were selected for analysis, including crash type,
learning model combining machine and deep learning crash location characteristics, environment, vehicle types,
models to identify relevant reasons for RTA’s severity. The and personal factors influencing accident severity. Accident
proposed RFCNN model demonstrated good accuracy on types are categorized based on severity into four levels,
the 20 most relevant characteristics. Bahiru et al. [26] as outlined in Table 2.
evaluated multiple machine learning methods including
ID3, CART, J48, and NB with 96% accuracy by J48 B. MULTILAYER PERCEPTRON (MLP)
model. Cicek et al. [17] employed various machine learning The multilayer perceptron model [28] represents a significant
models with explanations to predict accident severity, while advancement over the original perceptron model devised by
Yang et al. [16] applied deep learning for multitasking Rosenblatt. Unlike the perceptron, which was constrained

61064 VOLUME 12, 2024

O. I. Aboulola et al.: Automated Approach for Predicting Road Traffic Accident Severity

TABLE 1. Overview of prior studies on predicting accident severity.

TABLE 2. Description of the dataset.

to handling linearly separable problems in basic logic, employs secure interaction protocols for privacy-preserving
the multilayer perceptron introduces multiple layers of feature extraction.
functional neurons, enabling it to address nonlinear separable
problems. Its architecture comprises fully interconnected D. LONG SHORT-TERM MEMORY (LSTM)
layers, facilitating the organized flow of information. The
Long Short-Term Memory (LSTM) [30] represents a special-
training utilizes the error back-propagation algorithm to
ized recurrent neural network (RNN) architecture developed
minimize the cumulative error on the training set, typically
to overcome traditional RNNs’ limitations in handling
measured using mean-square error (MSE) for each sample.
long-term dependencies within sequential data. LSTMs excel
in tasks involving sequences, including natural language
processing, time series analysis, and speech recognition.
C. CONVOLUTIONAL NEURAL NETWORK (CNN) Key features include mitigating the vanishing gradient
CNN [29] is a deep neural network specifically engineered for problem in standard RNNs, memory cells for information
tasks such as image recognition, classification, and segmen- storage and erasure, and gating mechanisms (input, forget,
tation. It incorporates convolution, non-linear activation, and and output gates) to regulate data flow. LSTMs employ
pooling layers to extract features. Stacked CNNs are utilized activation functions to analyze incoming data and train
for specialized tasks, such as detecting parasites in infected using ‘‘Backpropagation through Time’’ (BPTT). Variants
cell images. The architecture is multi-layered, with each layer like Bidirectional LSTMs and simpler Gated Recurrent Unit
applying filters or kernels to input data to generate feature (GRU) networks have also been introduced. LSTMs find
maps. Convolutional layers’ outputs are concatenated and applications in various domains, from language modeling
passed into fully connected layers for further analysis. CNN and stock price prediction to speech recognition and image
has become a standard in medical domain classification and captioning.

VOLUME 12, 2024 61065

 O. I. Aboulola et al.: Automated Approach for Predicting Road Traffic Accident Severity

Over time, adaptations like ResNetV2 and Wide ResNets

have significantly enhanced both performance and efficiency.

F. EFFICIENTNETB4
EfficientNetB4 [32] belongs to the EfficientNet family of
neural networks, known for their exceptional performance in
image classification tasks while remaining computationally
efficient. Balancing model size, computational requirements,
and accuracy, EfficientNetB4 employs a systematic approach
to scale neural network architectures, achieving an optimal
balance of depth, width, and resolution through compound
scaling. It utilizes depth-wise separable convolutions and
squeeze-and-excite blocks to improve efficiency and feature
capture. Demonstrating top-tier accuracy on benchmarks
like ImageNet while maintaining computational efficiency,
EfficientNetB4 is widely used for transfer learning tasks.

G. INCEPTIONV3
InceptionV3 [33] has been widely utilized for image
recognition tasks, achieving high accuracy with numerous
convolutional, pooling, and activation layers. Its architecture
incorporates inception modules, facilitating the network
to acquire unique feature representations across various
scales. Techniques, like factorized 1 × 1 convolutions and
normalizing batches, are employed to reduce parameters and
improve training process. While versatile across various tasks
and datasets, InceptionV3 can be computationally intensive
and memory-consuming.

H. XCEPTION
Xception [34], derived from ‘‘extreme inception,’’ is a deep
CNN architecture proposed by François Chollet in 2017.
It extends Inception’s concepts by employing depthwise sep-
FIGURE 1. Architecture of the proposed framework. arable convolutions, which are more efficient. Consisting of
depthwise and pointwise convolutions, depthwise separable
convolutions reduce computational complexity. Known for
its deep architecture capable of learning complex features,
E. RESNET Xception excels in image classification accuracy.
Residual Networks or ResNets [31], represent a robust and
innovative type of deep neural network design that has had I. ALEXNET
a significant impact on computer vision and deep learning AlexNet, a pioneering convolutional neural network (CNN)
since their inception in 2015 by Kaiming He et al. Developed design, initially gained acclaim for its remarkable achieve-
to address the vanishing gradient problem hindering deep ments in image classification endeavors, notably excelling
network training, ResNets’ essential innovation lies in the on the ImageNet dataset [35]. Nevertheless, its utility
‘‘residual block.’’ This block incorporates two key routes: the transcends conventional image classification realms, show-
identity path, reflecting the original input and transferring casing effectiveness in addressing challenges associated
it directly to the output, and the residual path, applying with categorical data classification. Harnessing its adeptness
a series of convolutional layers and non-linear activations in hierarchical feature extraction, AlexNet demonstrates
to the input. Skip connections, also known as shortcut proficiency in processing categorical data inputs, capturing
connections, facilitate gradient flow during training, enabling nuanced relationships embedded within the data. Through
the training of extremely deep networks with hundreds tailored adaptations of AlexNet’s architecture and training
or thousands of layers without performance degradation. methodologies, researchers have effectively applied it to a
ResNets have excelled in various image classification tasks, wide array of categorical data classification tasks, span-
particularly in the ImageNet Large Scale Visual Recognition ning text classification, sentiment analysis, and medical
Challenge, and are widely employed in transfer learning. diagnosis.

61066 VOLUME 12, 2024

O. I. Aboulola et al.: Automated Approach for Predicting Road Traffic Accident Severity

J. VGG19 TABLE 3. Results of Deep Learning Models for Traffic Accident Severity
Detection.
VGG19, an iteration of the VGG (Visual Geometry Group)
framework, stands out for its intricate convolutional neural
network structure comprising 19 layers [36]. Initially devised
for image classification endeavors, VGG19 has transcended
traditional computer vision realms, showcasing its prowess
in navigating challenges associated with categorical data
classification. Its deep-seated architecture facilitates the and recall, provides an equitable evaluation of the model’s
extraction of intricate hierarchical features from intricate comprehensive effectiveness.
categorical datasets, enabling precise differentiation among
diverse classes. Leveraging transfer learning and fine-tuning IV. RESULTS AND DISCUSSION
methodologies, VGG19 can be tailored to suit various In this investigation, open-source libraries like Keras and
categorical data classification tasks, including those in natural TensorFlow were employed for the creation of pre-trained
language processing, sentiment analysis, and medical diagno- models. The Python language and the Anaconda platform
sis, often delivering cutting-edge performance benchmarks. were utilized for the examination of traffic accident severity
through transfer learning algorithms. A Dell Poweredge T430
server, furnished with a GPU, managed the computational
K. MOBILENET requirements of the dataset. This server featured eight cores,
MobileNet [37] is designed for embedded devices with lim- 32GB of RAM, and sixteen logical processors. The paper
ited processing capabilities, efficiently balancing accuracy suggests employing transfer learning techniques to address
and model size. Its key innovation lies in the utilization the challenge of predicting traffic accidents, and various
of depthwise separable convolutions, dividing convolutions scientific approaches will be employed to assess the efficacy
into depthwise and pointwise stages to significantly reduce and significance of the proposed methodology.
computational costs and model size. MobileNet comes in
multiple versions, including MobileNetV1, V2, and V3, A. RESULTS OF DEEP LEARNING MODELS FOR TRAFFIC
each improving performance and efficiency. Widely used in SEVERITY PREDICTION
mobile and embedded applications like object recognition Table 3 presents a comparative evaluation of the performance
and image classification, MobileNet showcases an innovative of three deep learning models—MLP, CNN, and LSTM—in
approach to creating efficient yet high-performing convolu- predicting traffic accident severity. Results indicate that CNN
tional neural networks. outperformed other deep learning models across evaluation
measures, achieving 0.8937 accuracy, 0.8663 precision,
0.8867 recall, and 0.8719 F1 score. Following CNN, MLP
L. EVALUATION PARAMETERS achieved 0.8727 accuracy, 0.8229 precision, 0.8361 recall,
This study employs various evaluation criteria, including and 0.8258 F1 score. LSTM performed least effectively
accuracy, F1 score, recall, and precision, to assess the among the models, with an accuracy score of 0.8127. The
effectiveness of transfer learning models. Additionally, CNN model demonstrated superior performance in predicting
confusion matrices are utilized to evaluate these algorithms’ traffic accident severity compared to other deep learning
performance. A confusion matrix, also known as an error models.
matrix, provides a tabular representation illustrating the
classifier’s performance on testing data, providing a graphical B. RESULTS OF TRANSFER LEARNING MODELS FOR
representation of the algorithm’s efficacy. TRAFFIC SEVERITY PREDICTION
Positive instances correctly predicted are denoted as True Table 4 provides a detailed analysis of various transfer
Positives (TP), whereas accurate predictions for the negative learning models’ performance in predicting traffic accident
class are represented by True Negatives (TN). False Positives severity. It showcases the performance of different trans-
(FP) indicate erroneous predictions for the positive class fer learning models, including ResNET, EfficientNetB4,
when the actual class is negative, and False Negatives (FN) InceptionV3, Xception, and MobileNet. Results highlight
signify inaccurate predictions for the negative class when the MobileNet as the top performer, achieving the highest
true class is positive. accuracy at 0.9817, alongside 0.9834 precision, 0.9891 recall,
The model’s comprehensive predictive precision is ascer- and 0.9848 F1 score. In contrast, Xception ranks lower in
tained by appraising the ratio of accurate predictions to the precision (0.8263) and F1 score (0.8519), suggesting areas
entire instances in the dataset. Precision gauges the ratio of for improvement. InceptionV3 and EfficientNetB4 achieved
correctly predicted positive instances among all instances accuracy scores of 0.9248 and 0.9367, respectively. ResNET
labeled as positive by the model. Recall alternatively exhibited the second-highest results with 0.9527 accuracy,
termed the genuine positive rate or sensitivity, measures the 0.9625 precision, 0.9819 recall, and 0.9767 F1 score.
model’s efficacy in precisely capturing positive cases. The These findings offer valuable insights for researchers
F1 measure, symbolizing the harmonic mean of precision and practitioners exploring transfer learning for traffic

VOLUME 12, 2024 61067

 O. I. Aboulola et al.: Automated Approach for Predicting Road Traffic Accident Severity

TABLE 4. Outcomes from Transfer Learning Models in Detecting Traffic TABLE 5. Results of MobileNet for traffic accident severity detection on
Accident Severity. US accidents.

axis. The X-axis illustrates the influence of these features on

the model’s output. Each point on the SHAP plot corresponds
to a data point from the training dataset. When an X-axis
value is positioned to the left of 0, it indicates an observation
that shifts the target value in a negative direction, whereas a
value to the right of 0 shifts it positively.
As illustrated in Figure 2, the type of road emerges
as the most influential factor in the model’s performance.
Specifically, higher road category values such as ‘Vehicle
track’ and ‘Motorway’ on the left side exert a negative impact
on the model, consistent with the observation that a majority
of accidents occur in rural and urban areas. Moreover,
Figure 2 demonstrates that drug consumption is associated
with more severe accidents. The Shapley value linked to
the drug-related feature increases proportionally with drug
consumption. In essence, as levels of drug consumption
escalate, so do the probabilities of accidents and the severity
of resulting injuries.

D. VALIDATION AND GENERALIZATION OF THE

PROPOSED APPROACH
Besides validating using an alternative dataset, this study
FIGURE 2. Impact of features on the performance of proposed model. aimed to demonstrate the effectiveness of the proposed
approach through experiments on a distinct dataset named
the ‘‘US accidents (2016-2021)’’ dataset [41], which
accident severity prediction, with MobileNet emerging as encompasses nearly 2.8 million records documenting traffic
a particularly promising candidate for further real-world incidents occurring across 46 states in the United States
applications. between February 2016 and December 2021. This dataset
provides extensive geographic coverage, spanning various
C. SHAP EXPLANATION locations across the country, and spans a five-year timeframe,
SHAP (SHapley Additive exPlanations) [38] is a widely making it valuable for analyzing regional and seasonal
adopted technique for elucidating the predictions made by differences in accident patterns. It consists of 47 variables,
machine learning models, encompassing even black-box detailing the causes contributing to automobile accidents,
models such as deep neural networks employed in transfer including information on accident locations, timings, weather
learning. The predominant variant of SHAP utilized for conditions, road conditions, and accident severity.
explicating black-box models is known as ‘‘Kernel SHAP.’’ The temporal intricacy in learning models primarily
With its robust processing capabilities and visualization revolves around the instructional stage, gauging the duration
tools, researchers have increasingly leveraged SHAP for necessary for adjusting the model’s characteristics relying on
investigating road safety [39], [40]. the provided input information. This intricacy is impacted
In this investigation, Shapley values for each feature by various factors, encompassing the architectural intricacy
are computed using the Python Shap library to unveil the of the model, the size of the instructional dataset, and
importance of features in the black-box transfer learning the optimization method employed. Table 6 juxtaposes the
model. SHAP underscores the significance of features in durations for training and testing across the utilized models.
predicting water quality. While SHAP’s feature relevance Remarkably, the computational duration for the MobileNet
surpasses that of traditional methods, its standalone usage model is exceptionally effective, enduring only 200 seconds,
only yields limited additional insights. markedly less than the training periods of other transfer learn-
The SHAP plot organizes features in a descending order, ing models employed in this investigation. Impressively, this
showcasing their significance. Features with high importance increased efficiency in training time does not compromise the
are depicted in red at the top of the plot, while those with low model’s accuracy, as it consistently outperforms individual
importance are represented in blue at the bottom along the Y- models in terms of predictive accuracy.

61068 VOLUME 12, 2024

O. I. Aboulola et al.: Automated Approach for Predicting Road Traffic Accident Severity

TABLE 6. Time complexity of transfer learning models (in seconds). 3) Improved Traffic Management: Traffic accident sever-
ity prediction can aid in optimizing traffic manage-
ment strategies. Real-time information about potential
accidents and their severity can help traffic manage-
ment authorities reroute traffic, adjust signal timings,
or implement temporary traffic controls to minimize
congestion and maintain smooth traffic flow.
4) Insurance Risk Assessment: Insurance companies can
leverage predictive models of accident severity to
TABLE 7. Findings of k-fold cross-validation. assess the risk associated with insuring vehicles and
drivers. By accurately estimating the likelihood and
severity of accidents, insurers can tailor insurance
premiums more accurately, leading to fairer pricing for
policyholders.
5) Fleet Safety Management: For businesses with vehicle
fleets, predictive models of accident severity can
inform fleet safety management strategies. By identify-
ing routes or driving conditions associated with higher
accident severity, fleet managers can implement train-
To further evaluate the effectiveness of the proposed ing programs, safety protocols, or route optimizations
method, K-fold cross-validation is incorporated as an addi- to reduce the risk of accidents and protect both drivers
tional step for performance evaluation. The results from and assets.
the 5-fold cross-validation are presented in Table 7. These 6) Public Awareness and Education: Transparent and
findings showcase the exceptional efficacy of the suggested explainable predictive models of traffic accident
methodology in terms of precision, F1 score, accuracy, and severity can also serve as educational tools for the
recall when contrasted with alternative models. Notably, the general public. By providing insights into the factors
low standard deviation values indicate consistent and stable influencing accident severity, such as road conditions,
performance across different folds, reinforcing confidence in weather, or time of day, these models can raise
the trustworthiness and reliability of MobileNet. awareness about safe driving practices and encourage
responsible behavior among road users.
E. PRACTICAL IMPLICATIONS OF PROPOSED
FRAMEWORK
F. THEORETICAL CONTRIBUTIONS OF THE PROPOSED
The practical implications of utilizing the MobileNet transfer
MODEL
learning model and SHAP XAI technique for traffic accident
severity prediction extend across various domains, including The theoretical contributions of the research lie in the
road safety management, emergency response, traffic control, advancement of predictive modeling techniques, the inte-
insurance, fleet management, and public education. By lever- gration of explainable AI methods, the validation of model
aging these technologies, stakeholders can work towards interpretability, the understanding of feature importance,
reducing the frequency and severity of traffic accidents, the application of transfer learning in road safety, and the
ultimately saving lives and minimizing the societal and demonstration of model generalization capabilities. These
economic impact of road accidents. contributions collectively enhance the body of knowledge
in the field of traffic accident severity prediction and pave
1) Enhanced Road Safety Measures: By accurately
the way for further advancements in AI-driven road safety
predicting traffic accident severity, authorities can
solutions.
implement targeted road safety measures in high-risk
areas. For instance, if the model predicts a high severity 1) Advancement in Predictive Modeling Techniques: The
of accidents on a particular road segment, authorities research contributes to the advancement of predictive
can focus on improving road infrastructure, installing modeling techniques by integrating transfer learning
additional safety signs, or enforcing speed limits to with the MobileNet architecture. Transfer learning
mitigate the risk of accidents. leverages knowledge gained from pre-trained models,
2) Optimized Emergency Response: Accurate predictions enhancing the efficiency and effectiveness of predictive
of accident severity enable emergency response teams models for traffic accident severity prediction.
to allocate resources more effectively. By anticipating 2) Integration of Explainable AI (XAI) Techniques: By
the severity of accidents, emergency services can incorporating SHAP (SHapley Additive exPlanations)
dispatch appropriate personnel and resources to the XAI technique, the research enhances the interpretabil-
scene promptly, potentially reducing response times ity and transparency of predictive models. This inte-
and improving outcomes for accident victims. gration provides insights into the underlying factors

VOLUME 12, 2024 61069

 O. I. Aboulola et al.: Automated Approach for Predicting Road Traffic Accident Severity

FIGURE 3. Transfer learning models performance comparison.

influencing accident severity predictions, contributing robustness in the context of traffic accident severity
to the development of more transparent and trustworthy prediction.
AI systems.
3) Validation of Model Interpretability: The research G. DISCUSSION
validates the effectiveness of SHAP XAI technique in The discussion section undertakes a comprehensive evalua-
interpreting complex machine learning models such as tion of the study’s findings and the implications of enhanced
MobileNet. This contributes to the growing body of forecast accuracy for the severity of the accidents. The
literature on model interpretability, demonstrating the findings underscore the efficacy of deep learning approaches,
applicability of SHAP in explaining predictions from particularly transfer learning models, in improving the exact-
deep learning models. ness of predicting the severity of road traffic accidents. These
4) Understanding Feature Importance: Through the SHAP findings carry significant implications for various stake-
XAI technique, the research facilitates a deeper under- holders, including transportation planners, policymakers, and
standing of the importance of different features in law enforcement agencies. The proposed MobileNet model
predicting traffic accident severity. This understanding can facilitate the implementation of targeted and efficient
can inform future research on feature engineering, measures to reduce both the frequency and severity of traffic
model optimization, and the development of more accidents by enabling more precise prediction of accident
robust predictive models. severity. Figure 3 illustrates the comparison of transfer
5) Application of Transfer Learning in Road Safety: The learning models, clearly demonstrating the superiority of the
application of transfer learning in the context of road proposed MobileNet.
safety and accident severity prediction represents a Employing the MobileNet transfer learning model for
novel contribution to the field. By leveraging pre- predicting road accident severity offers numerous advantages.
trained models, researchers can adapt state-of-the-art MobileNet arrives pre-trained with a wealth of knowledge
techniques from computer vision to address real-world from extensive image datasets, thereby reducing training
problems in road safety management. time and mitigating the risk of overfitting. Its capability to
6) Demonstration of Model Generalization: The research generalize effectively to new and diverse accident images
demonstrates the generalization capabilities of the enhances predictive accuracy, while its efficient perfor-
MobileNet transfer learning model across different mance even on resource-constrained devices is noteworthy.
datasets and real-world scenarios. This contributes MobileNet’s consistent performance, resilience to image
to the understanding of model performance and variations, and potential for fine-tuning render it a dependable

61070 VOLUME 12, 2024

O. I. Aboulola et al.: Automated Approach for Predicting Road Traffic Accident Severity

choice. Furthermore, its suitability for mobile deployment more effective measures to prevent accidents. By doing so,
and potential to achieve state-of-the-art results make it a this research not only enhances the accuracy of severity
practical and high-accuracy solution for accident severity prediction but also fosters transparency, interpretability, and
prediction. trustworthiness of learning models. This is indispensable
Predicting accident severity holds significant real-world for stakeholders and decision-makers striving to implement
applications across various domains. It contributes to enhanc- evidence-based actions to bolster road safety, ultimately
ing transportation safety, optimizing emergency response, leading to a reduction in the impact of traffic accidents on
assessing insurance risks, managing traffic, and improv- highways. The future work direction of this research work
ing fleet safety. Additionally, it plays a vital role in is to make use of correlational and convergence feature
autonomous vehicle development, supports public health engineering techniques to select the most appropriate factors
research, informs traffic engineering and urban planning that really affect the prediction of traffic accident severity in
initiatives, and aligns with smart city endeavors. Beyond a real-time environment.
road transportation, its applicability extends to aviation,
maritime, and industrial safety, aiding in mitigating the REFERENCES
impact of accidents, improving decision-making processes, [1] M. K. Gebru, ‘‘Road traffic accident: Human security perspective,’’ Int. J.
and ultimately saving lives and resources. Peace Develop. Stud., vol. 8, no. 2, pp. 15–24, 2017.
[2] N. Klinjun, M. Kelly, C. Praditsathaporn, and R. Petsirasan, ‘‘Identification
of factors affecting road traffic injuries incidence and severity in Southern
1) COMPARISON WITH STATE-OF-THE-ART Thailand based on accident investigation reports,’’ Sustainability, vol. 13,
For comparison purposes, two studies are selected. no. 22, p. 12467, 2021.
[3] Road Safety, World Health Organization, Geneva, Switzerland, 2020.
Yang et al. [16] applied the DNN model to detect injury [4] A. Chand, S. Jayesh, and A. B. Bhasi, ‘‘Road traffic accidents: An overview
severity and utilized the layer-wise relevance propagation of data sources, analysis techniques and contributing factors,’’ Mater.
(LRP) method to elucidate prediction outcomes. They Today, Proc., vol. 47, pp. 5135–5141, Jan. 2021.
[5] F. Malin, I. Norros, and S. Innamaa, ‘‘Accident risk of road and weather
conducted experiments using the Chinese traffic accident conditions on different road types,’’ Accident Anal. Prevention, vol. 122,
dataset. On the other hand, Cicek et al. [17] employed pp. 181–188, Jan. 2019.
various models (DT, NB, MLP, SVM, NN, and ANN- [6] V. Nuri Sumantri, A. I. Rifai, and F. Ferial, ‘‘Impact of inter-urban street
lighting on users perception of road safety behavior: A case of jalan
MLP) and achieved 76.90% accuracy on the NHTSA- majalengka-rajagaluh,’’ Citizen, J. Ilmiah Multidisiplin Indonesia, vol. 2,
USA dataset. They utilized the Shapley technique to no. 5, pp. 703–711, Dec. 2022.
extract significant features. However, the NHTSA-USA [7] C. Gutierrez-Osorio and C. Pedraza, ‘‘Modern data sources and techniques
for analysis and forecast of road accidents: A review,’’ J. Traffic
dataset, while valuable for studying traffic accidents, Transp. Eng., vol. 7, no. 4, pp. 432–446, Aug. 2020.
presents limitations such as underreporting, inconsistent [8] D. Lord and F. Mannering, ‘‘The statistical analysis of crash-frequency
reporting, missing data, geographical and temporal biases, data: A review and assessment of methodological alternatives,’’ Transp.
limited context, privacy concerns, data incompleteness, data Res. A, Policy Pract., vol. 44, no. 5, pp. 291–305, Jun. 2010.
[9] K. Santos, J. P. Dias, and C. Amado, ‘‘A literature review of machine
imbalances, data collection bias, and evolving data standards. learning algorithms for crash injury severity prediction,’’ J. Saf. Res.,
In contrast, the current study employs transfer learning vol. 80, pp. 254–269, Feb. 2022.
models on the US traffic dataset and utilizes Shapley values [10] Y. Zhang, H. Li, and G. Ren, ‘‘Estimating heterogeneous treatment effects
in road safety analysis using generalized random forests,’’ Accident Anal.
to elucidate predictive outcomes. Shapley values, grounded Prevention, vol. 165, Feb. 2022, Art. no. 106507.
in cooperative game theory, offer a more interpretable and [11] Y. Yang, K. He, Y.-P. Wang, Z.-Z. Yuan, Y.-H. Yin, and M.-Z. Guo,
versatile framework for explaining feature contributions in ‘‘Identification of dynamic traffic crash risk for cross-area freeways based
on statistical and machine learning methods,’’ Phys. A, Stat. Mech. Appl.,
learning models. The current study demonstrates superior vol. 595, Jun. 2022, Art. no. 127083.
performance, achieving robust results with 98.17% accuracy. [12] B. B. Elallid, N. Benamar, A. S. Hafid, T. Rachidi, and N. Mrani,
‘‘A comprehensive survey on the application of deep and reinforcement
learning approaches in autonomous driving,’’ J. King Saud Univ.-Comput.
V. CONCLUSION Inf. Sci., vol. 34, no. 9, pp. 7366–7390, Oct. 2022.
Traffic accidents persist as a significant threat, resulting [13] L. Li, Y. Lin, B. Du, F. Yang, and B. Ran, ‘‘Real-time traffic incident
in loss of lives, injuries, and substantial disruptions on detection based on a hybrid deep learning model,’’ Transp. A, Transp. Sci.,
vol. 18, no. 1, pp. 78–98, 2022.
roadways. Understanding the underlying factors leading [14] K. Pawar and V. Attar, ‘‘Deep learning based detection and localization of
to these accidents is crucial for enhancing safety across road accidents from traffic surveillance videos,’’ ICT Exp., vol. 8, no. 3,
transportation networks. This study harnesses various transfer pp. 379–387, Sep. 2022.
[15] H. Zhao, X. Li, H. Cheng, J. Zhang, Q. Wang, and H. Zhu, ‘‘Deep learning-
learning techniques and elucidates the most influential based prediction of traffic accidents risk for Internet of Vehicles,’’ China
factors through the application of Shapley values. The Commun., vol. 19, no. 2, pp. 214–224, Feb. 2022.
research delves into predicting accident severity using mod- [16] Z. Yang, W. Zhang, and J. Feng, ‘‘Predicting multiple types of
els including Multilayer Perceptron (MLP), Convolutional traffic accident severity with explanations: A multi-task deep learning
framework,’’ Saf. Sci., vol. 146, Feb. 2022, Art. no. 105522.
Neural Network (CNN), Long Short-Term Memory (LSTM), [17] E. Cicek, M. Akin, F. Uysal, and R. Topcu Aytas, ‘‘Comparison of
Residual Networks (ResNET), EfficientNetB4, InceptionV3, traffic accident injury severity prediction models with explainable machine
Extreme Inception (Xception), and MobileNet. Among these learning,’’ Transp. Lett., vol. 15, no. 9, pp. 1043–1054, Oct. 2023.
[18] J. Gan, L. Li, D. Zhang, Z. Yi, and Q. Xiang, ‘‘An alternative method for
models, MobileNet emerges with the highest accuracy of traffic accident severity prediction: Using deep forests algorithm,’’ J. Adv.
98.17%. This insight lays the groundwork for devising Transp., vol. 2020, pp. 1–13, Dec. 2020.

VOLUME 12, 2024 61071

 O. I. Aboulola et al.: Automated Approach for Predicting Road Traffic Accident Severity

[19] R. E. AlMamlook, K. M. Kwayu, M. R. Alkasisbeh, and A. A. OMAR IBRAHIM ABOULOLA received the
Frefer, ‘‘Comparison of machine learning algorithms for predicting traffic Bachelor of Science degree in computer sci-
accident severity,’’ in Proc. IEEE Jordan Int. Joint Conf. Electr. Eng. Inf. ence from KAU, in 2001, the master’s degree
Technol. (JEEIT), Apr. 2019, pp. 272–276. in information science from the University of
[20] I. Aldhari, M. Almoshaogeh, A. Jamal, F. Alharbi, M. Alinizzi, and Indiana, Bloomington, USA, in 2009, and the
H. Haider, ‘‘Severity prediction of highway crashes in Saudi Arabia using master’s degree in information technology and
machine learning techniques,’’ Appl. Sci., vol. 13, no. 1, p. 233, Dec. 2022. the Ph.D. degree in information systems and
[21] M. Sameen and B. Pradhan, ‘‘Severity prediction of traffic accidents with
technology from Claremont Graduate University
recurrent neural networks,’’ Appl. Sci., vol. 7, no. 6, p. 476, Jun. 2017.
[22] L. Wahab and H. Jiang, ‘‘Severity prediction of motorcycle crashes with (CGU), USA, in 2013 and 2018, respectively. His
machine learning methods,’’ Int. J. Crashworthiness, vol. 25, no. 5, master’s thesis was related to the technology of
pp. 485–492, Sep. 2020. banking. His Ph.D. dissertation aimed to design an innovative assistive
[23] D.-J. Lin, M.-Y. Chen, H.-S. Chiang, and P. K. Sharma, ‘‘Intelligent traffic technology to help retail companies to predict optimum locations for their
accident prediction model for Internet of Vehicles with deep learning businesses. He is currently an Associate Professor with the Information
approach,’’ IEEE Trans. Intell. Transp. Syst., vol. 23, no. 3, pp. 2340–2349, Systems and Technology Department, College of Computer Science and
Mar. 2022. Engineering (CCSE), University of Jeddah.
[24] A. Jamal, M. Zahid, M. T. Rahman, H. M. Al-Ahmadi, M. Almoshaogeh,
D. Farooq, and M. Ahmad, ‘‘Injury severity prediction of traffic crashes
with ensemble machine learning techniques: A comparative study,’’ Int. J.
Injury Control Saf. Promotion, vol. 28, no. 4, pp. 408–427, Oct. 2021.
[25] M. Manzoor, M. Umer, S. Sadiq, A. Ishaq, S. Ullah, H. A. Madni,
and C. Bisogni, ‘‘RFCNN: Traffic accident severity prediction based on EBTISAM ABDULLAH ALABDULQADER
decision level fusion of machine and deep learning model,’’ IEEE Access,
received the Ph.D. degree in HCI from Newcastle
vol. 9, pp. 128359–128371, 2021.
University, U.K. She is currently an experi-
[26] T. K. Bahiru, V. S. Manjula, T. B. Akele, E. A. Tesfaw, and T. D. Belay,
‘‘Mining road traffic accident data for prediction of accident severity,’’ in enced Assistant Professor with the Information
Proc. Int. Conf. Intell. Data Commun. Technol. Internet Things (IDCIoT), Technology Department, King Saud University
Jan. 2023, pp. 606–612. (KSU), Saudi Arabia. She is the Founder of the
[27] R. Roscher, B. Bohn, M. F. Duarte, and J. Garcke, ‘‘Explainable machine ArabHCI.org Community. Her research interests
learning for scientific insights and discoveries,’’ IEEE Access, vol. 8, include social computing, participatory design,
pp. 42200–42216, 2020. CSCW, and community-driven research.
[28] M. T. C. Olmedo, M. Paegelow, J.-F. Mas, and F. Escobar, ‘‘Multi-
layer perceptron (MLP),’’ in Geomatic Approaches for Modeling Land
Change Scenarios. Cham, Switzerland: Springer, 2018, pp. 451–455, doi:
10.1007/978-3-319-60801-3_27.
[29] J. Wu, ‘‘Introduction to convolutional neural networks,’’ Nat. Key Lab
Novel Softw. Technol., vol. 5, no. 23, p. 495, 2017.
[30] R. C. Staudemeyer and E. Rothstein Morris, ‘‘Understanding LSTM—A
AISHA AHMED ALARFAJ received the Ph.D. degree in computer science
tutorial into long short-term memory recurrent neural networks,’’ 2019,
from Newcastle University, U.K., in 2021. Since 2012, she has been a
arXiv:1909.09586.
[31] F. He, T. Liu, and D. Tao, ‘‘Why ResNet works? Residuals generalize,’’ Faculty Member with the Information Systems Department, Princess Nourah
IEEE Trans. Neural Netw. Learn. Syst., vol. 31, no. 12, pp. 5349–5362, bint Abdulrahman University, where she is currently an Assistant Professor.
Dec. 2020. Her research interests include human–computer interaction, UX/UI, social
[32] R. Pillai, N. Sharma, and R. Gupta, ‘‘Fine-tuned EfficientNetB4 transfer commerce, sharing economy, and trust.
learning model for weather classification,’’ in Proc. 3rd Asian Conf. Innov.
Technol. (ASIANCON), Aug. 2023, pp. 1–6.
[33] N. Dong, L. Zhao, C. H. Wu, and J. F. Chang, ‘‘Inception v3 based cervical
cell classification combined with artificially extracted features,’’ Appl. Soft
Comput., vol. 93, Aug. 2020, Art. no. 106311.
[34] F. Chollet, ‘‘Xception: Deep learning with depthwise separable convo- SHTWAI ALSUBAI received the Ph.D. degree
lutions,’’ in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR), in computer science from The University of
Jul. 2017, pp. 1800–1807. Manchester, London. He is currently an Assis-
[35] A. Krizhevsky, I. Sutskever, and G. E. Hinton, ‘‘ImageNet classification
tant Professor with the Computer Engineering
with deep convolutional neural networks,’’ Commun. ACM, vol. 60, no. 6,
Department, Prince Sattam bin Abdulaziz Univer-
pp. 84–90, May 2017.
[36] K. Simonyan and A. Zisserman, ‘‘Very deep convolutional networks for
sity, Saudi Arabia. His research interests include
large-scale image recognition,’’ 2014, arXiv:1409.1556. computer vision, optimization techniques, and
[37] D. Sinha and M. El-Sharkawy, ‘‘Thin MobileNet: An enhanced MobileNet performance enhancement.
architecture,’’ in Proc. IEEE 10th Annu. Ubiquitous Comput., Electron.
Mobile Commun. Conf. (UEMCON), Oct. 2019, pp. 0280–0285.
[38] S. M. Lundberg and S.-I. Lee, ‘‘A unified approach to interpreting model
predictions,’’ in Proc. Adv. Neural Inf. Process. Syst., vol. 30, 2017,
pp. 1–10.
[39] A. B. Parsa, A. Movahedi, H. Taghipour, S. Derrible, and
A. K. Mohammadian, ‘‘Toward safer highways, application of XGBoost
and SHAP for real-time accident detection and feature analysis,’’ Accident
TAI-HOON KIM (Member, IEEE) received the M.S. and Ph.D. degrees
Anal. Prevention, vol. 136, Mar. 2020, Art. no. 105405.
[40] C. Yang, M. Chen, and Q. Yuan, ‘‘The application of XGBoost
in electrics, electronics, and computer engineering from Sungkyunkwan
and SHAP to examining the factors in freight truck-related crashes: University, Seoul, South Korea, and the Ph.D. degree in information science
An exploratory analysis,’’ Accident Anal. Prevention, vol. 158, Aug. 2021, from the University of Tasmania, Hobart, Australia, in December 2011.
Art. no. 106153. He is currently a Professor with Chonnam National University, Gwangju,
[41] Soobhan Moosavi. (May 2021). US Accidents (2016–2021). South Korea. His research interests include statistical analysis, image
Accessed: Feb. 10, 2023. [Online]. Available: https://www.kaggle. processing, and system design.
com/sobhanmoosavi/us-accidents

61072 VOLUME 12, 2024