ArticlePDF Available

A Cross-Cultural Crash Pattern Analysis in the United States and Jordan Using BERT and SHAP

Electronics

January 2025
14(2):272

DOI:10.3390/electronics14020272

License
CC BY 4.0

Authors:

Shadi Jaradat

Queensland University of Technology

Mohammed Elhenawy

Queensland University of Technology

Alexander Paz

University of Nevada, Las Vegas

Taqwa Alhadidi

Al-Ahliyya Amman University

Show all 6 authorsHide

Understanding the cultural and environmental influences on roadway crash patterns is essential for designing effective prevention strategies. This study applies advanced AI techniques, including Bidirectional Encoder Representations from Transformers (BERT) and Shapley Additive Explanations (SHAP), to examine traffic crash patterns in the United States and Jordan. By analyzing tabular data and crash narratives, the research reveals significant regional differences: in the USA, vehicle overturns and roadway conditions, such as guardrails, are major factors in fatal crashes, whereas in Jordan, technical defects and driver behavior play a more critical role. SHAP analysis identifies “driver” and “damage” as pivotal terms across both regions, while country-specific terms such as “overturn” in the USA and “technical” in Jordan highlight regional disparities. Using BERT/Bi-LSTM models, the study achieves up to 99.5% accuracy in crash severity prediction, demonstrating the robustness of AI in traffic safety analysis. These findings underscore the value of contextualized AI-driven insights in developing targeted, region-specific road safety policies and interventions. By bridging the gap between developed and developing country contexts, the study contributes to the global effort to reduce road traffic injuries and fatalities.

Available via license: CC BY 4.0

Content may be subject to copyright.

Academic Editor: Ping-Feng Pai

Received: 16 December 2024

Revised: 28 December 2024

Accepted: 8 January 2025

Published: 10 January 2025

Citation: Jaradat, S.; Elhenawy, M.;

Paz, A.; Alhadidi, T.I.; Ashqar, H.I.;

Nayak, R. A Cross-Cultural Crash

Pattern Analysis in the United States

and Jordan Using BERT and SHAP.

Electronics 2025,14, 272. https://

doi.org/10.3390/electronics14020272

Licensee MDPI, Basel, Switzerland.

This article is an open access article

distributed under the terms and

conditions of the Creative Commons

Attribution (CC BY) license

(https://creativecommons.org/

licenses/by/4.0/).

Article

A Cross-Cultural Crash Pattern Analysis in the United States and

Jordan Using BERT and SHAP

Shadi Jaradat 1, 2, * , Mohammed Elhenawy 1,2 , Alexander Paz 3, Taqwa I. Alhadidi 4, Huthaifa I. Ashqar 5

and Richi Nayak 2,6

1Centre for Accident Research and Road Safety-Queensland, Queensland University of Technology, 130

Victoria Park Rd, Kelvin Grove, QLD 4059, Australia; mohammed.elhenawy@qut.edu.au

Centre for Data Science, Queensland University of Technology, Gardens Point, Brisbane, QLD 4000, Australia;

r.nayak@qut.edu.au

3School of Civil Engineering, Queensland University of Technology, Gardens Point, Brisbane, QLD 4000,

Australia; alexander.paz@qut.edu.au

4Civil Engineering Department, Al-Ahliyya Amman University, Amman 19328, Jordan;

t.alhadidi@ammanu.edu.jo

5Civil Engineering Department, Arab American University, 13 Zababdeh, Jenin P.O Box 240, Palestine;

huthaifa.ashqar@aaup.edu

6School of Computer Science, Queensland University of Technology, Gardens Point,

Brisbane, QLD 4000, Australia

*Correspondence: shadi.jaradat@hdr.qut.edu.au

Abstract: Understanding the cultural and environmental inﬂuences on roadway crash pat-

terns is essential for designing effective prevention strategies. This study applies advanced

AI techniques, including Bidirectional Encoder Representations from Transformers (BERT)

and Shapley Additive Explanations (SHAP), to examine trafﬁc crash patterns in the United

States and Jordan. By analyzing tabular data and crash narratives, the research reveals sig-

niﬁcant regional differences: in the USA, vehicle overturns and roadway conditions, such

as guardrails, are major factors in fatal crashes, whereas in Jordan, technical defects and

driver behavior play a more critical role. SHAP analysis identiﬁes “driver” and “damage”

as pivotal terms across both regions, while country-speciﬁc terms such as “overturn” in the

USA and “technical” in Jordan highlight regional disparities. Using BERT/Bi-LSTM mod-

els, the study achieves up to 99.5% accuracy in crash severity prediction, demonstrating the

robustness of AI in trafﬁc safety analysis. These ﬁndings underscore the value of contextu-

alized AI-driven insights in developing targeted, region-speciﬁc road safety policies and

interventions. By bridging the gap between developed and developing country contexts,

the study contributes to the global effort to reduce road trafﬁc injuries and fatalities.

Keywords: BERT; SHAP; cross-cultural; LSTM; trafﬁc safety; machine learning

1. Introduction

Road crash injuries are projected to become the ﬁfth leading cause of mortality globally

by 2030, highlighting the pressing need for effective prevention strategies in transportation

safety management [

]. Accurate forecasting of trafﬁc crashes is a critical component of

these strategies, as it enables the identiﬁcation of potential risks and the development of

targeted interventions [

]. Crash pattern analysis, which involves examining the factors

contributing to road crashes, plays a vital role in identifying these risks. However, it is

essential to recognize that crash patterns and contributing factors can vary signiﬁcantly

across cultural and geographical contexts. Traditional crash analysis methods often rely

on statistical models that struggle to capture the nuanced and non-linear relationships

Electronics 2025,14, 272 https://doi.org/10.3390/electronics14020272

Electronics 2025,14, 272 2 of 35

present in crash narratives. Additionally, there is a lack of cross-cultural comparisons that

leverage advanced AI techniques to address these gaps. Studying crash patterns in diverse

regions provides valuable insights into how cultural, social, and environmental factors

inﬂuence road safety, enabling the formulation of more tailored and effective prevention

measures [4].

This study focuses on crash patterns in two culturally and infrastructurally distinct

regions, the United States and Jordan, to explore how varying socio-economic, legal, and

environmental factors shape trafﬁc safety outcomes. The United States, as a developed

nation with extensive road networks and advanced trafﬁc safety measures, presents a

markedly different context compared to Jordan, a developing country facing challenges

such as older vehicle ﬂeets, variable enforcement of trafﬁc laws, and unique cultural driving

behaviors. By analyzing these contrasting settings, the study aims to uncover shared and

region-speciﬁc factors inﬂuencing road safety, enhancing understanding of diverse trafﬁc

safety challenges.

By leveraging state-of-the-art natural language processing (NLP) techniques, this

research performs a detailed analysis of crash narratives from the two regions, focusing

on fatal and non-fatal crashes. Bidirectional Encoder Representations from Transformers

(BERT) are employed for topic modeling and classiﬁcation tasks, while Shapley Additive

Explanations (SHAP) enhance model interpretability. This approach ﬁlls a critical research

gap by providing a detailed, AI-driven comparative analysis of crash patterns across

developed and developing countries, enabling a more comprehensive understanding of

global trafﬁc safety challenges. By addressing gaps in existing research, this study uncovers

novel insights into cultural, infrastructural, and legal factors inﬂuencing crash outcomes

and develops actionable recommendations for improving road safety in both developed

and developing countries. The objectives of this study are threefold:

•

To analyze crash patterns in the USA and Jordan, identifying cultural, infrastructural,

and legal factors that contribute to crash outcomes.

•

To apply BERT-based models and SHAP to enhance predictive accuracy and model

transparency, uncovering critical risk factors.

•

To provide region-speciﬁc policy recommendations for road safety, including infras-

tructure improvements, vehicle safety standards, and emergency response strategies.

By integrating AI-driven methods and focusing on two distinct cultural contexts, this

research aims to bridge methodological gaps and contribute to global road safety efforts.

2. Literature Review

Traditional data-driven regression models have been widely used to model crash sever-

ity, offering valuable mathematical interpretations and insights into individual predictor

variables. However, these models are limited by their reliance on underlying assumptions,

such as linear link functions and error distribution terms, which, if violated, can lead

to biased estimates and reduced predictive accuracy [

]. This limitation is particularly

critical when modeling complex crash dynamics, where interactions between factors are

non-linear and highly context-dependent. Advanced data collection techniques, such as

mobile LiDAR systems and point cloud applications, have signiﬁcantly improved trafﬁc

safety analysis by capturing detailed spatial and environmental data [

]. Despite their

contributions, these technologies often fall short of providing contextual insights, such as

driver behavior and cultural factors that signiﬁcantly inﬂuence crash dynamics.

In contrast, text-based Natural Language Processing (NLP) methods like BERT enable

the extraction of nuanced crash factors from narrative reports, offering insights into the cul-

tural and behavioral elements surrounding crash incidents. A recent study highlighted the

potential of deep NLP approaches, including ensemble learning with pre-trained transform-

Electronics 2025,14, 272 3 of 35

ers, for improving crash severity classiﬁcation, demonstrating their capability in harnessing

unconventional data sources for trafﬁc safety analysis [

]. Integrating narrative analysis

with spatial data facilitates a holistic understanding of road safety factors, addressing gaps

left by traditional approaches.

Various methodologies, including association rule mining, spatial statistical analysis,

kernel density estimation (KDE), and sequence analysis, to investigate crash patterns and

understand crash patterns [

–

]. These methods have proven effective in identifying spatial

and temporal crash trends, crash-prone areas, and crash sequences, forming a foundation

for targeted safety measures [

–

]. Machine learning-based models, including support

vector machines, decision trees, and deep learning, have emerged as promising tools in

road safety research, particularly in addressing the limitations of traditional statistical

approaches [

–

]. These models leverage real-time trafﬁc data, such as trafﬁc ﬂow,

speed, and volume, to identify crash-related trends and circumstances [

–

]. Resampling

approaches have been recommended to enhance the predictive performance of machine

learning algorithms in handling imbalanced crash datasets [

], thereby improving the

reliability of predictions. A recent study by Jaradat et al. (2024) introduced a multitask

learning framework that utilizes social media data, speciﬁcally Twitter, to analyze and

detect real-time crash patterns, enabling faster and more targeted trafﬁc management

responses in dynamic conditions [

]. This innovation demonstrates the potential of AI in

harnessing unconventional data sources for road safety improvements.

NLP techniques, such as BERT, facilitate the extraction of nuanced crash factors,

allowing for a richer understanding of the elements inﬂuencing road safety in diverse

cultural contexts. The ﬂexibility and scalability of BERT in processing unstructured text

have expanded the horizons of crash data analysis, enabling researchers to explore complex

relationships within narrative data [21].

Numerous studies have investigated the factors affecting road crashes. Researchers

investigated the effectiveness of artiﬁcial neural network models in simulating motorway

crashes and established that the core reason for the occurrence of crashes is the average

daily trafﬁc volume and average vehicle speed [

]. Text-mining analyses have been em-

ployed to classify road trafﬁc injury collision features, demonstrating their utility in road

injury prevention [

]. Giummarra et al. (2022) performed text mining to classify crash

circumstances by road user group and found that it can be used to uncover the features of

road trafﬁc injury [

]. These studies emphasize the value of text analysis in understanding

crash dynamics across various contexts. Wang et al. (2017) analyzed the severity of trafﬁc

crash injuries at intersections and different sections of roads and showed the importance of

understanding the differences in collision injury severity and their contributing factors [

This is complemented by studies like Darus et al. (2022), which highlight road charac-

teristics and trafﬁc conditions as critical contributors to collision severity and emphasize

the necessity of incorporating multiple parameters into injury classiﬁcation models [

Donnelly-Swift and Kelly (2015) used generalized linear regression models to identify

variables related to fatal or serious injuries in single-vehicle road trafﬁc crashes. Their study

was important because it provided insight into the multidimensional aspects that result in

injury severity [25].

Text mining methods, including thematic analysis, content analysis, and NLP, have

been used to mine and analyze crash narrative textual data [

–

]. These methods can

be employed to identify the contributing components that characterize crash patterns and

explore the causation of crashes within unstructured narrative reports [

]. Association

rule mining has played an instrumental role in identifying parameters associated with

crash types, such as senior-driver crossing crashes [

]. Combining real-time trafﬁc data

with insights gleaned from crash narratives yields a more complete understanding of crash

Electronics 2025,14, 272 4 of 35

risk and causation. This holistic mechanism establishes the foundation for developing

forecasting models for crashes in real-time, which are reliable, effective, and fundamental

for proactive trafﬁc management and safety enhancement [6,30].

Topic modeling is an unsupervised machine learning technique aimed at identifying

abstract “topics” by clustering groups of words within a set of documents. The evolution

of topic modeling techniques has signiﬁcantly advanced the ﬁeld of NLP, beginning with

the introduction of Latent Semantic Analysis (LSA) and culminating in the development of

Bidirectional Encoder Representations from Transformers (BERT) [

]. LSA, introduced by

Deerwester et al. [

], marked the inception of extracting latent topics from text by decom-

posing term-document matrices, thereby uncovering the underlying semantic structure of

the corpus. Following LSA, Blei et al. [

] introduced Latent Dirichlet Allocation (LDA).

This generative probabilistic model improved upon LSA by allowing documents to be

represented as mixtures of multiple topics, thus providing a more ﬂexible and detailed

method for topic discovery. Despite their effectiveness, both LSA and LDA are limited by

their reliance on bag-of-words representations, which ignore the order of words and the

contextual nuances of language. The advent of deep learning brought about signiﬁcant ad-

vancements in topic modeling, with the introduction of models that could understand the

context and semantics of words in text. BERT [34] represents a paradigm shift, leveraging

a transformer architecture to generate deep contextualized word embeddings. Although

LDA and BERT are established techniques in natural language processing, the present study

distinguishes itself by applying these methods to a novel cross-cultural context in trafﬁc

safety analysis. The integration of BERT for topic modeling and text classiﬁcation with

the addition of SHAP for model interpretability offers a fresh perspective that enhances

the understanding of crash patterns in different cultural settings. This approach applies

advanced NLP techniques to a new domain and introduces a comprehensive methodology

that can be adapted for broader applications in road safety research. This work, therefore,

contributes meaningfully to the ﬁeld by bridging the gap between methodological rigor

and practical application in a critical public safety domain.

Trafﬁc crashes have been studied from a cross-cultural perspective in numerous

research. These studies have assessed risks in road behavior, the incidence of aggressive

driving, and driving behaviors across various nations and cultures [

–

]. The results

reveal that driver anger and road safety behaviors are signiﬁcantly inﬂuenced by cultural

factors, underscoring the interplay between societal norms and driving practices [39–41].

Given the global nature of road safety challenges, there is a critical need for research

that transcends national boundaries and incorporates cross-cultural comparisons. Existing

studies have left signiﬁcant gaps in understanding how diverse socio-economic, infras-

tructural, and cultural contexts inﬂuence crash patterns and outcomes. Moreover, while

advancements in analytical methods, such as Natural Language Processing (NLP) and

machine learning, have greatly enhanced crash analysis, their application to cross-cultural

comparisons remains limited. Despite progress in leveraging advanced models, many stud-

ies predominantly focus on single-country analyses or speciﬁc datasets, failing to account

for the complex interplay of cultural, legal, and environmental contexts in crash causation.

This gap is particularly evident in the comparison of crash narratives between coun-

tries with vastly different socio-economic and cultural contexts. For instance, the United

States, a highly motorized and developed nation, features extensive roadway networks

and advanced trafﬁc safety measures. In contrast, a developing nation, Jordan presents

challenges such as varied trafﬁc law enforcement and distinct driving behaviors. This study

addresses these gaps by integrating BERT-based topic modeling and text classiﬁcation with

SHAP interpretability to analyze crash data from the USA and Jordan. Unlike prior studies

focusing solely on infrastructural or behavioral factors, this research incorporates cultural,

Electronics 2025,14, 272 5 of 35

legal, and environmental elements to inform region-speciﬁc road safety interventions. The

contributions of this study are as follows:

•

Evaluate state-of-the-art AI models, such as BERT, for crash severity classiﬁcation,

thereby enhancing predictive accuracy.

•

Leverage SHAP for AI model transparency, offering actionable insights into crash

severity factors for informed policymaking.

•

Analyze crash narratives to uncover nuanced safety risks and support comprehensive

safety measures.

•

Provide cross-cultural safety insights, facilitating the adaptation of successful inter-

ventions globally.

•

Identify unique and shared crash severity factors in the USA and Jordan, guiding

targeted safety interventions.

•

Safety countermeasures are recommended to prevent crashes and reduce their severity

in both countries.

3. Methodology

3.1. Proposed Framework

This study utilizes state-of-the-art AI techniques, including BERT and BERT/Bi-LSTM,

to analyze factors contributing to fatal and non-fatal crashes using datasets from the

USA and Jordan. The proposed research framework, illustrated in Figure 1, involves

multiple stages:

Data Collection from USA and Jordan crash datasets, encompassing both tabular and

narrative data.

2. Preprocessing to standardize and prepare the data for analysis.

3. Exploratory and Textual Visualizations, including Exploratory Data Analysis (EDA),

Topic Modeling, unigrams, bigrams, similarity matrices, and hierarchical structuring.

4. Text Classiﬁcation using BERT/Bi-LSTM with SHAP interpretability.

5. Generating actionable insights and policy recommendations based on the analysis.

This comprehensive framework thoroughly compares crash patterns and highlights

region-speciﬁc factors inﬂuencing road safety.

3.2. Dataset

This study utilizes crash data from two distinct regions: Jordan and the USA. Each

dataset contains both tabular and narrative crash data. The tabular data includes core vari-

ables such as Severity, Crash Type, Light Condition, and Number of Vehicles. Additional

contextual features like Weekday, Month, Season, Car Type, Accident Time, and Driver

Age were extracted for deeper analysis.

3.2.1. Jordan Dataset

The Jordan dataset consists of 6359 trafﬁc narrative reports from ﬁve major freeways:

Airport Roads, Desert Highways, Jordan Highway, Route 30, and Route 35. These reports,

obtained from the Jordan Trafﬁc Institute (JTI), include both tabular and narrative data.

Crashes were categorized into four levels: Fatal, Severe Injury, Moderate Injury, and

Minor Injury.

3.2.2. USA Dataset

The USA dataset was obtained from the Missouri State Highway Patrol, covering

reports from 2019–2020. Crashes were classiﬁed into three categories: Fatal, Property

Damage, and Personal Injury. An equal sample size of 6359 crashes was selected from the

USA dataset to ensure comparability, maintaining balance across categories.

Electronics 2025,14, 272 6 of 35

Electronics 2025, 14, x FOR PEER REVIEW 6 of 34

Figure 1. Proposed framework ﬂowchart for cross-cultural road crash analysis.

3.2. Dataset

This study utilizes crash data from two distinct regions: Jordan and the USA. Each

dataset contains both tabular and narrative crash data. The tabular data includes core var-

iables such as Severity, Crash Type, Light Condition, and Number of Vehicles. Additional

contextual features like Weekday, Month, Season, Car Type, Accident Time, and Driver

Age were extracted for deeper analysis.

3.2.1. Jordan Dataset

The Jordan dataset consists of 6359 traﬃc narrative reports from ﬁve major freeways:

Airport Roads, Desert Highways, Jordan Highway, Route 30, and Route 35. These reports,

obtained from the Jordan Traﬃc Institute (JTI), include both tabular and narrative data.

Crashes were categorized into four levels: Fatal, Severe Injury, Moderate Injury, and Mi-

nor Injury.

3.2.2. USA Dataset

The USA dataset was obtained from the Missouri State Highway Patrol, covering

reports from 2019–2020. Crashes were classiﬁed into three categories: Fatal, Property

Damage, and Personal Injury. An equal sample size of 6359 crashes was selected from the

USA dataset to ensure comparability, maintaining balance across categories.

Figure 1. Proposed framework ﬂowchart for cross-cultural road crash analysis.

3.2.3. Narrative Data and NLP Analysis

Crash narratives provided a rich source of contextual information. Given structural

differences between datasets, these narratives became the primary focus for comparison.

Using BERT and BERT/Bi-LSTM, the study analyzed key factors contributing to crash

severity. This approach enhanced understanding of underlying crash dynamics in Jordan

and the USA, as illustrated by sample narratives in Tables 1and 2.

Electronics 2025,14, 272 7 of 35

Table 1. USA dataset of fatal and non-fatal crash samples.

Crash Narrative Severity

VEHICLE 1 WAS WESTBOUND ON HIGHWAY D, TRAVELING

TOO FAST FOR CONDITIONS. DRIVER 1 FAILED TO NEGOTIATE

A CURVE TO THE LEFT. VEHICLE 1 BEGAN TO SLIDE AND

TRAVELED OFF THE LEFT SIDE OF THE ROADWAY, THEN

OVERTURNED.

Non-fatal

ATAL CRASH NEXT OF KIN NOTIFIED TROOP A FATAL CRASH

#24 AND FATALITY #27 FOR 2019. THE CRASH OCCURRED AS

DRIVER 1 FAILED TO NEGOTIATE A CURVE; VEHICLE 1

TRAVELED OFF THE ROADWAY AND OVERTURNED. DRIVER 1

WAS EJECTED AND STRUCK A TREE. VEHICLE 1 AND DRIVER 1

CAME TO REST IN A DITCH OFF THE SOUTH SIDE OF US-50.

DRIVER 1 PRONOUNCED ON SCENE AT 1346 BY DEPUTY

CORONER TODD ASBURY. ASSISTED BY CPL R. W.

SHAUL/528/AND PETTIS COUNTY SHERIFF’S DEPARTMENT.

Fatal

Table 2. Jordan dataset of fatal and non-fatal crash samples.

Crash Narrative Severity

As a result of a sudden leaking change violation by the driver of

vehicle two, he collided with vehicle number one in N, and the crash

resulted in material damage estimated by a specialized technician

Non-fatal

While vehicle no. 1 was traveling from Ma’an Governorate to Aqaba

Governorate, and in the al-humaima area, we were charged with a

violation of failure to take necessary trafﬁc safety precautions

Fatal

3.3. Data Preprocessing

3.3.1. Standardization for Cross-Cultural Analysis

To enable meaningful comparisons between the Jordan and USA crash datasets, crash

severity levels were standardized into two categories: Fatal and Non-fatal. Shared variables,

including Severity, Crash Type, Light Condition, and Number of Vehicles, were retained for

core analysis. Additional contextual features, such as Weekday, Month, Season, Car Type,

Accident Time, and Driver Age, were extracted for deeper investigation. Unique variables

were excluded to maintain cross-dataset comparability.

3.3.2. Text Preprocessing

A systematic preprocessing approach was implemented to ensure consistency and

data quality. Key steps included:

Text Normalization: Text was converted to lowercase, and non-textual charac-

ters, URLs, and commonly repeated irrelevant terms (e.g., “VEHICLE”, “KIN”, “PRO-

NOUNCED”, “FATALITY”) were removed to reduce noise while preserving critical crash-

related terms, such as “death”, “killing”, and “notiﬁed”.

Minimal Preprocessing: Techniques such as tokenization, stop-word removal, and

stemming were deliberately avoided to retain important contextual information, particu-

larly region-speciﬁc nuances such as rollover dynamics in the USA or vehicle-related terms

in Jordan.

3.3.3. Advanced Encoding

Advanced transformer models like BERT were employed to process crash narratives

effectively. Raw text inputs were encoded using DistilBERT, with each narrative truncated

Electronics 2025,14, 272 8 of 35

or padded to a maximum length of 200 tokens to standardize input dimensions. This ap-

proach aligns with modern research emphasizing minimal preprocessing for deep learning

models, ensuring that nuanced meanings and contextual details are preserved.

Following preprocessing, exploratory data analysis (EDA) was conducted to uncover

initial insights and patterns. This included descriptive statistics and visualization tech-

niques to highlight prevalent terms and relationships in the data. The standardized and

reﬁned datasets provided a robust foundation for downstream analyses, including the

application of BERT/Bi-LSTM models and SHAP-based interpretations to uncover key

crash factors.

3.4. Exploratory Data Analysis (EDA)

Following preprocessing, we conducted Exploratory Data Analysis (EDA) to uncover

patterns, trends, and potential outliers in the data. Descriptive statistics and graphical rep-

resentations, such as unigram and bigram frequency visualizations, highlighted prevalent

terms and relationships, providing valuable insights into crash dynamics across the two

datasets. However, we acknowledge that aligning crash severity categories during data

standardization may introduce potential biases, as cultural and contextual differences in re-

porting practices could affect the comparability of datasets. For example, terms describing

technical defects in Jordan may differ in granularity from terminology used in the USA,

necessitating careful interpretation of results.

This robustly reﬁned dataset, which retained its contextual integrity, provided a solid

foundation for subsequent analysis, as summarized in Tables 3and 4.

Table 3. USA data descriptive statistics.

Feature Original/Extracted Value # %

Severity Original Non-fatal 5883 92.515

Fatal 476 7.485

Crash Type Original

Fixed Object 2654 41.736

Overturn 2553 40.148

Motor Vehicle in Transport 699 10.992

Animal 185 2.909

Other Object 97 1.525

Immersion 35 0.550

Other Non-Collision 34 0.535

Pedestrian 31 0.487

Working Motor Vehicle 24 0.377

Pedalcycle 17 0.267

Parked Motor Vehicle 10 0.157

Fell/Jumped from MV 8 0.126

Jackknife 6 0.094

Railway Vehicle 3 0.047

Animal Drawn

Veh/Animal Ridden Trans 2 0.031

Fire/Explosion 1 0.016

Electronics 2025,14, 272 9 of 35

Table 3. Cont.

Feature Original/Extracted Value # %

Light Condition Original

Dark-Unlighted 4286 67.401

Daylight 1822 28.652

Dark-Lighted 241 3.790

Unknown 7 0.110

Other 2 0.031

Dark-Unknown Lighting 1 0.016

No of Vehicles Original

1 3488 54.851

2 2452 38.560

3 316 4.969

4 67 1.054

5 27 0.425

6 7 0.110

11 1 0.016

8 1 0.016

Weekday Extracted

Monday 995 15.647

Tuesday 960 15.097

Wednesday 958 15.065

Thursday 901 14.169

Friday 880 13.839

Saturday 835 13.131

Sunday 830 13.052

Month Extracted

4 698 10.977

5 692 10.882

6 682 10.725

7 655 10.300

8 649 10.206

9 640 10.064

10 611 9.608

11 506 7.957

12 418 6.573

1 349 5.488

2 339 5.331

3 120 1.887

Season Extracted

Autumn 2039 32.065

Winter 1948 30.634

Spring 1263 19.862

Summer 1109 17.440

Electronics 2025,14, 272 10 of 35

Table 3. Cont.

Feature Original/Extracted Value # %

Car Type Extracted

Motor Vehicle 3292 51.769

Truck 2085 32.788

Commercial Vehicle 416 6.542

Motorcycle 413 6.495

Emergency Vehicle 113 1.777

Electric vehicle 39 0.613

School Bus 1 0.016

Late_night 1595 25.083

Acc Time Extracted

Morning 1492 23.463

Midday 925 14.546

Afternoon 888 13.964

Evening 801 12.596

Night 446 7.014

Early_morning 212 3.334

Driver Age Extracted

<25 3297 51.848

25–54 1298 20.412

>55 985 15.490

>64 779 12.250

Table 4. Jordan data descriptive statistics.

Variable Original/Extracted Value # %

Severity Original Fatal 123 1.934

Non-fatal 6236 98.066

Crash Type Original

Deterioration 308 4.844

Collision 5928 93.222

Run over 123 1.934

Light Conditions Original

day 4421 69.524

Night and insufﬁcient

lighting 525 8.256

Night and road with

adequate lighting 1050 16.512

sunset 197 3.098

sunrise 35 0.550

darkness 131 2.060

Electronics 2025,14, 272 11 of 35

Table 4. Cont.

Variable Original/Extracted Value # %

No of Vehicles Original

1 1358 21.356

2 4554 71.615

3 366 5.756

4 52 0.818

6 6 0.094

5 21 0.330

7 2 0.031

Weekday Extracted

Thursday 363 5.708

Tuesday 396 6.227

Sunday 361 5.677

Wednesday 308 4.844

Saturday 342 5.378

Monday 337 5.300

Friday 406 6.385

Month Extracted

9 205 3.224

11 197 3.098

1 190 2.988

2 206 3.240

3 258 4.057

4 209 3.287

5 202 3.177

6 219 3.444

7 210 3.302

8 209 3.287

10 218 3.428

12 190 2.988

Season Extracted

Spring 620 9.750

Autumn 4515 71.002

Summer 586 9.215

Winter 638 10.033

Electronics 2025,14, 272 12 of 35

Table 4. Cont.

Variable Original/Extracted Value # %

Car Type Original

Small ride-on car 4076 64.082

The tug had not prepared

for shipment 415 6.526

shipping 801 12.596

Shared transfer 763 11.999

bus 63 0.991

Works/construction vehicle

12 0.189

Medium ride 161 2.532

Vehicle category 5 0.079

Special use vehicle 35 0.550

Shipping locomotive and

trailer 5 0.079

Motorized bicycle 21 0.330

Shipping locomotives and

semi-trailers 1 0.016

Agricultural vehicle 1 0.016

Crash Time Extracted

Evening 677 10.646

Midday 680 10.694

Afternoon 393 6.180

Late_night 155 2.437

Night 362 5.693

Morning 220 3.460

Early_morning 26 0.409

Driver Age Original

<25 1044 16.418

25–54 4700 73.911

>55 390 6.133

>64 204 3.208

Based on the descriptive statistics provided in Tables 3and 4, we conducted a com-

parative analysis of trafﬁc crash patterns between Jordan and the USA, revealing notable

differences and similarities across various dimensions. In terms of crash severity, both

countries predominantly experience personal injury outcomes, yet Jordan has a signiﬁ-

cantly lower fatality rate (1.93%) compared to the USA (7.49%). Crash types in Jordan are

overwhelmingly collisions (93.22%), in stark contrast to the USA’s diverse crash types, in-

cluding ﬁxed object collisions and overturns. Light conditions further differentiate the two,

with most crashes in Jordan occurring during daylight (69.52%) and the USA experiencing

a majority under dark-unlighted conditions (67.40%). Jordan exhibits a higher incidence of

multi-vehicle collisions (71.62%), whereas single-vehicle crashes are more prevalent in the

USA (54.85%).

Seasonal trends show Jordan peaked in crashes during Autumn (71%), aligning more

closely with the USA’s increased crashes in colder months. Vehicle types involved in crashes

also vary, with Jordan dominated by small ride-on cars (64.08%) and the USA showing a

broader distribution, including motor vehicles and trucks. Crash timing indicates Jordan’s

Electronics 2025,14, 272 13 of 35

crashes peak during evening and midday, contrasting with the USA’s pattern of nighttime

crashes. Lastly, the driver age group involved in crashes in Jordan is predominantly the

25–54 age group (73.91%), whereas the USA sees a broader age distribution, including a

notable involvement of younger drivers. These ﬁndings underscore the complex interplay

of environmental, societal, and regulatory factors inﬂuencing road safety in each country,

highlighting the need for tailored road safety measures and policies. In the USA, a vast

majority of crashes occur during the day, with a signiﬁcant drop-off as lighting conditions

worsen, suggesting that visibility plays a critical role in trafﬁc safety.

Conversely, in Jordan, crashes are more evenly distributed between daylight and dark

but lighted conditions, with fewer incidents occurring in the dark unlighted, which may

indicate better adaptation or less trafﬁc during these times. This comparison reveals stark

differences in how light conditions affect driving safety in the two countries. While the USA

clearly prefers daytime driving safety, Jordan displays a higher tolerance for less optimal

lighting conditions. This insight could inform targeted strategies for improving road safety

tailored to the unique driving environments of each country.

3.5. Topic Modeling Using BERT

BERT’s ability to capture bidirectional contexts has advanced the development of

sophisticated topic modeling approaches that go beyond simple word co-occurrence, al-

lowing for the extraction of semantically coherent and contextually relevant topics. This

evolution from earlier models like Latent Semantic Analysis (LSA) to BERT highlights the

shift from linear algebra-based and probabilistic methods to deep learning, marking signiﬁ-

cant progress toward richer, more nuanced text representations for topic modeling. The

use of contextual embeddings has made BERT a valuable tool for identifying underlying

themes in crash narratives. A critical component of topic modeling involves selecting a

hyperparameter, denoted as

, which represents the number of latent topics. However,

determining the optimal number of topics is often challenging, as it affects the quality and

meaningfulness of the extracted topics [34].

BERT continues to be widely employed for natural language processing (NLP) tasks

due to its ability to capture deep contextual information. This strength lies in its foundation

on the transformer architecture introduced by Vaswani et al., which employs protocol-

deﬁned encoder blocks to process text in parallel rather than sequentially [

]. BERT

leverages its bidirectional and context-sensitive nature to improve language understand-

ing and prediction accuracy. Since textual data can often include symbols and numbers

that introduce noise during data processing, standard preprocessing techniques like text

cleaning, stop-word removal, tokenization, stemming, and word embedding are applied.

However, these preprocessing steps can sometimes remove valuable contextual or semantic

information from sentences or phrases. BERT addresses this limitation by being ﬁne-tuned

with an additional output layer, making it adaptable for a wide range of NLP tasks—such as

question answering and language inference—without signiﬁcant task-speciﬁc architecture

changes.

Pre-trained on large datasets like English Wikipedia and the Book Corpus (800 M

words), BERT is capable of handling tasks in over 100 languages. Its training in next-

sentence prediction and masked language modeling further enhances its versatility. BERT

incorporates three embedding layers: token, segment, and position embeddings [

In classiﬁcation tasks, the Bi-Directional Long Short-Term Memory (Bi-LSTM) network

processes the 768 hidden states of the [CLS] token representation produced by BERT.

Models like TK-BERT and BERT-LDA have been shown to outperform traditional

methods such as Latent Dirichlet Allocation (LDA) and Non-negative Matrix Factorization

(NMF) by integrating contextual semantics and thematic narratives [

]. Coherence was

Electronics 2025,14, 272 14 of 35

the primary metric in this study, as it reﬂects the logical association between words within

a topic, ensuring the validity and relevance of the generated topics. Hyperparameter

selection for k was performed in collaboration with domain expert to maximize coherence

and interpretability. While Coherence was central to our evaluation, future studies may

beneﬁt from incorporating additional metrics, such as exclusivity, to further enhance the

uniqueness and interpretability of topics.

3.6. Text Classiﬁcation Using BERT/Bi-LSTM

Long Short-Term Memory (LSTM), a form of recurrent neural network (RNN), has

been extensively used in natural language processing and sequential data analysis, due

to its ability to capture long-term dependencies [

]. Unlike RNNs, LSTM networks

utilize memory cells and gating techniques to store and retrieve information over long

sequences in an elegant manner. This architectural design resolves the vanishing and

exploding gradient issues when training deep networks on sequential data, making LSTMs

particularly suitable for language modeling, text classiﬁcation, and time series prediction,

among other applications. Their ability to model sequential relationships is critical in

analyzing crash narratives, where temporal context plays a vital role.

The LSTM network, a specialized variant of recurrent neural networks (RNNs), ex-

hibits unique capabilities in handling short- and long-term correlations within time series

data. The network’s architecture includes a memory unit, where a central memory cell,

denoted by the red circle, plays a pivotal role. This structure enables the LSTM network

to effectively capture dependencies and patterns in sequential data, making it particu-

larly advantageous for various applications where understanding short- and long-term

relationships is essential [44].

3.7. Evaluation Metrics

The BERT/Bi-LSTM model was validated on the test dataset. Key metrics included

accuracy, precision, recall, F1-score, SHAP, and coherence score.

•

Accuracy measures the overall correctness of the model in classifying fatal and non-

fatal crashes. The BERT/Bi-LSTM model achieved 99.5% accuracy for the USA dataset

and 99% accuracy for the Jordan dataset, highlighting its effectiveness in classifying

crash severity.

•

Precision indicates the proportion of correctly predicted fatal crashes out of all pre-

dicted fatal crashes. High precision reduces the occurrence of false positives, which is

critical for reliable crash severity prediction.

•

Recall evaluates the model’s ability to capture all actual fatal crashes out of the total

true fatal crashes. This metric is vital for ensuring that no fatal crashes are overlooked.

•

The F1-score balances precision and recall, ensuring the model performs well in both

aspects. It is particularly important in this study to ensure both false positives and

false negatives are minimized. The F1-scores for BERT/Bi-LSTM models exceeded

98% for both datasets.

SHAP interpretability values were employed to interpret the model predictions, iden-

tifying key features such as “overturned” and “attempted” for fatal crashes in the USA,

and “damage” and “technical” issues for Jordan. This interpretability is crucial for under-

standing the underlying factors contributing to crashes in different cultural contexts.

For topic modeling, the coherence score was used to evaluate the semantic association

of words within each topic. Higher coherence indicates better quality topics. The USA

dataset achieved the highest coherence with 25 topics, while Jordan peaked with 10 topics,

reﬂecting the differences in crash narratives between the two regions.

Electronics 2025,14, 272 15 of 35

The coherence score

was computed using four different statistics: the segment

of conﬁrmed co-occurrences of two words, the probability of these two words appearing

together, and their probabilities. The Cvfor a single topic is as follows [45]:

Cv=

∑

m=2

m−1

∑

l=1

score(wm,wl)

M(M−1)/2

where

is the coherence score for a single topic, measuring the logical association between

words within that topic;

is the number of words in a topic;

is the positions of words in

a topic; and

score(wm,wl)

is a function that calculates a score for the relationship between

two words, wmand wl, at positions mand lwithin the topic.

4. Experimental Setup

4.1. Model Selection

The BERT and BERT/Bi-LSTM models were selected for this study due to their proven

superiority in handling complex natural language processing (NLP) tasks, especially when

context and sequence play a crucial role, as in the analysis of crash narratives. These models

were chosen over traditional approaches such as Naive Bayes, Support Vector Machines

(SVM), and Random Forests for several reasons:

1. Contextual Understanding

BERT’s bidirectional nature captures the full context of words by analyzing both

preceding and following text. This makes it well-suited for interpreting nuanced

crash narratives, where the sequence of events and relationships between terms can

signiﬁcantly affect crash severity analysis [46].

2. Handling Sequential Data

The integration of Bi-LSTM (Bidirectional Long Short-Term Memory) enhances BERT’s

ability to process sequential data in both forward and backward directions. This is

critical for understanding the progression of events leading up to crashes, as it captures

long-range dependencies within crash narratives [

]. This feature is especially

important for temporal event analysis, where the order of events directly impacts the

interpretation of severity.

3. Interpretability with SHAP

The inclusion of SHAP (Shapley Additive Explanations) provides transparency into

the model’s predictions, offering insights into the most inﬂuential factors driving

crash severity. This interpretability ensures that complex models, like BERT/Bi-LSTM,

do not function as “black boxes”, but instead provide clear explanations for decision-

making processes [49].

4. Generalization Across Datasets

BERT’s pre-trained architecture and its ability to be ﬁne-tuned on large datasets

provide excellent generalization capabilities, making it suitable for cross-cultural

crash analysis. This generalization is particularly useful when analyzing datasets

from different regions, such as the USA and Jordan [50].

4.2. Classiﬁcation Models

In this experiment, three models: Naive Bayes, AdaBoost, and BERT/Bi-LSTM were

evaluated. These models were selected for their relevance to text classiﬁcation tasks

and their ability to capture both simple and complex relationships within crash narra-

tives

[51,52]

. Based on the initial results, BERT/Bi-LSTM outperformed the other baseline

classiﬁers and was subsequently used for the remainder of the experiment.

Electronics 2025,14, 272 16 of 35

We selected the BERT/Bi-LSTM model because it combines BERT’s contextual un-

derstanding with Bi-LSTM’s sequential processing capabilities. Compared to GPT-based

models, which are primarily designed for generative tasks, BERT/Bi-LSTM is better suited

for classiﬁcation and interpretability in analyzing structured crash data.

The conceptual framework integrated an LSTM-based language model (i.e., BERT/Bi-

LSTM), as illustrated in Figure 2. To enhance the model’s performance, we divided the

dataset into three subsets: training, validation, and testing, constituting 70%, 15%, and 15%

of the data, respectively. The last layer weights generated by the pre-trained models were

fed into the Bi-LSTM layer to perform the ﬁne-tuning classiﬁcation task. The optimization

of the LSTM model involved experimenting with conﬁgurations, resulting in three types

of dense layers. The ﬁrst layer has 64 units with ReLU activation and a 0.2 dropout

layer, followed by another layer with 32 units, ReLU activation, and a 0.2 dropout layer.

The ﬁnal layer is an output layer with two units using the SoftMax activation function.

Furthermore, our conﬁguration includes 128 LSTM cells with a learning rate of 10

−4

, a

batch size of 32, Adam optimizer, and Cross Entropy Loss as the chosen loss function [

The hyperparameters used in the BERT/Bi-LSTM conﬁguration are presented in Table 5.

Electronics 2025, 14, x FOR PEER REVIEW 15 of 34

Figure 2. Conceptual LSTM-based language model classiﬁer (i.e., BERT/Bi-LSTM).

Table 5. BERT/Bi-LSTM model conﬁgurations.

Parameter/Hyperparameter Value

MAX_LENGTH 200

Pre-trained Model DistilBERT

Tokenizer DistilBertTokenizer

Number of LSTM Units 128

Number of Dense Layers 3

Hidden Units in Dense Layers [32, 64]

Dropout Rate 0.2

Batch Size 32

Activation Function Softmax

Loss Function Cross entropy

Optimizer Adam

Learning Rate Scheduler ReduceLROnPlateau

Learning Rate 10

Minimum Learning Rate 0.000001

Patience 3 (Early Stopping), 1 (Reduce LR)

Training Batch Size 16

Number of Training Epochs 6

4.3. SHAP Interpretability

Shapley Additive Explanations (SHAP), a method grounded in Shapley values, is

implemented to interpret the predictions of the classiﬁcation models globally [53]. SHAP

assigns each feature a value that reﬂects its contribution to a prediction, providing a com-

prehensive understanding of feature importance. This technique enhances transparency

and trust by oﬀering insights into the decision-making process of complex models.

By elucidating how speciﬁc features, such as crash type or driver-related factors, in-

ﬂuence fatal and non-fatal crash outcomes, SHAP facilitates both global and local

Figure 2. Conceptual LSTM-based language model classiﬁer (i.e., BERT/Bi-LSTM).

Table 5. BERT/Bi-LSTM model conﬁgurations.

Parameter/Hyperparameter Value

MAX_LENGTH 200

Pre-trained Model DistilBERT

Tokenizer DistilBertTokenizer

Number of LSTM Units 128

Number of Dense Layers 3

Hidden Units in Dense Layers [32, 64]

Electronics 2025,14, 272 17 of 35

Table 5. Cont.

Parameter/Hyperparameter Value

Dropout Rate 0.2

Batch Size 32

Activation Function Softmax

Loss Function Cross entropy

Optimizer Adam

Learning Rate Scheduler ReduceLROnPlateau

Learning Rate 10−4

Minimum Learning Rate 0.000001

Patience 3 (Early Stopping), 1 (Reduce LR)

Training Batch Size 16

Number of Training Epochs 6

4.3. SHAP Interpretability

Shapley Additive Explanations (SHAP), a method grounded in Shapley values, is

implemented to interpret the predictions of the classiﬁcation models globally [

]. SHAP

assigns each feature a value that reﬂects its contribution to a prediction, providing a com-

prehensive understanding of feature importance. This technique enhances transparency

and trust by offering insights into the decision-making process of complex models.

By elucidating how speciﬁc features, such as crash type or driver-related factors,

inﬂuence fatal and non-fatal crash outcomes, SHAP facilitates both global and local inter-

pretability. This enables the identiﬁcation of the most inﬂuential factors across the dataset,

ensuring a deeper understanding of model behavior and its alignment with real-world

crash dynamics.

SHAP was chosen over other interpretability frameworks, such as LIME, because it

provides consistent, additive feature attributions for global and local explanations. Unlike

LIME, which relies on perturbation and is model-agnostic, SHAP’s game-theoretic founda-

tion ensures stability and robustness when applied to complex deep learning models like

BERT/Bi-LSTM.

In the ﬁnal stage, we integrated the outcomes of topic modeling and text classiﬁcation

by combining the identiﬁed themes and topics with the interpretability provided by SHAP.

This holistic approach allows us to comprehensively understand the datasets, revealing

nuanced insights into the factors that inﬂuence crash occurrence and severity. By integrat-

ing these analyses, this study aims to provide actionable insights for policymakers and

practitioners, contributing to advancing road safety research and interventions.

5. Analysis and Results

5.1. Modeling Results

Tables 6and 7display the experimental outcomes for the USA and Jordan datasets.

The USA classiﬁcation models, including NB-TFIDF and AdaBoost-TFIDF, showed high

accuracy (96–98%). The BERT/Bi-LSTM model achieved superior performance with 99.5%

accuracy and macro metrics (precision, recall, and F1-score) exceeding 98%. In the Jordan

dataset, NB-TFIDF and AdaBoost-TFIDF exhibited 99% accuracy, while BERT/Bi-LSTM

maintained 99% accuracy with balanced macro/weighted metrics (precision, recall, and F1-

score exceeding 0.87). This suggests BERT/Bi-LSTM’s efﬁcacy in capturing crash narrative

nuances in diverse contexts. Jordan’s models showed slightly lower recall, indicating

Electronics 2025,14, 272 18 of 35

the complexity of classifying crash narratives in this context. Further insights into model

convergence and generalization are provided by the visual representation of training and

validation accuracy in Figures 3and 4.

Table 6. Experiment results: USA dataset.

Embedding Accuracy Macro

Precision

Macro

Recall

Macro

F1-Score

NB-TFIDF 0.96 0.98 0.80 0.86

AdaBoost-TFIDF 0.98 0.98 0.98 0.98

BERT/Bi-LSTM 99.5 0.99 0.98 0.98

Table 7. Experiment results: Jordan dataset.

Embedding Accuracy Macro

Precision

Macro

Recall

Macro

F1-Score

NB-TFIDF 0.99 0.99 0.65 0.73

AdaBoost-TFIDF 0.99 0.88 0.84 0.86

BERT/Bi-LSTM 0.99 0.99 0.99 0.99

Electronics 2025, 14, x FOR PEER REVIEW 16 of 34

interpretability. This enables the identiﬁcation of the most inﬂuential factors across the

dataset, ensuring a deeper understanding of model behavior and its alignment with real-

world crash dynamics.

SHAP was chosen over other interpretability frameworks, such as LIME, because it

provides consistent, additive feature aributions for global and local explanations. Unlike

LIME, which relies on perturbation and is model-agnostic, SHAPs game-theoretic foun-

dation ensures stability and robustness when applied to complex deep learning models

like BERT/Bi-LSTM.

In the ﬁnal stage, we integrated the outcomes of topic modeling and text classiﬁcation

by combining the identiﬁed themes and topics with the interpretability provided by

SHAP. This holistic approach allows us to comprehensively understand the datasets, re-

vealing nuanced insights into the factors that inﬂuence crash occurrence and severity. By

integrating these analyses, this study aims to provide actionable insights for policymakers

and practitioners, contributing to advancing road safety research and interventions.

5. Analysis and Results

5.1. Modeling Results

Tables 6 and 7 display the experimental outcomes for the USA and Jordan datasets.

The USA classiﬁcation models, including NB-TFIDF and AdaBoost-TFIDF, showed high

accuracy (96–98%). The BERT/Bi-LSTM model achieved superior performance with 99.5%

accuracy and macro metrics (precision, recall, and F1-score) exceeding 98%. In the Jordan

dataset, NB-TFIDF and AdaBoost-TFIDF exhibited 99% accuracy, while BERT/Bi-LSTM

maintained 99% accuracy with balanced macro/weighted metrics (precision, recall, and

F1-score exceeding 0.87). This suggests BERT/Bi-LSTMs eﬃcacy in capturing crash nar-

rative nuances in diverse contexts. Jordans models showed slightly lower recall, indicat-

ing the complexity of classifying crash narratives in this context. Further insights into

model convergence and generalization are provided by the visual representation of train-

ing and validation accuracy in Figures 3 and 4.

Table 6. Experiment results: USA dataset.

Embedding Accuracy Macro Precision Macro Recall Macro F1-Score

NB-TFIDF 0.96 0.98 0.80 0.86

AdaBoost-TFIDF 0.98 0.98 0.98 0.98

BERT/Bi-LSTM 99.5 0.99 0.98 0.98

Figure 3. Training and validation (accuracy vs. loss) for the USA dataset.

Electronics 2025, 14, x FOR PEER REVIEW 17 of 34

Table 7. Experiment results: Jordan dataset.

Embedding Accuracy Macro Precision Macro Recall Macro F1-Score

NB-TFIDF 0.99 0.99 0.65 0.73

AdaBoost-TFIDF 0.99 0.88 0.84 0.86

BERT/Bi-LSTM 0.99 0.99 0.99 0.99

Figure 4. Training and validation (accuracy vs. loss) for the Jordan dataset.

5.2. Topic Modeling Results

The coherence score graphs for topic modeling in the USA and Jordan datasets show

distinct peaks, indicating each datasets optimal number of topics. For the USA, the peak

coherence is achieved with about 25 topics, suggesting a wider variety of distinct discus-

sion points within crash narratives. In contrast, Jordans dataset reaches peak coherence

with approximately ten topics, indicating a more concentrated range of themes that best

capture the discussions within the crash data. This comparison highlights the complexity

of the USAs crash narratives and a more focused thematic structure within Jordans da-

taset. The larger number of topics in the USA dataset reﬂects its more diverse crash cir-

cumstances, potentially linked to diﬀerences in infrastructure, driving behaviors, and re-

porting practices.

Figure 5 shows the USA and Jordan Topics per dataset. The following insights are

extracted from the ﬁgure. In the USA dataset, topics like “overturned” and “roadway”

suggest a concentration on accidents and road conditions, which may be indicative of the

countrys extensive road networks and the higher incidence of vehicle rollovers on high-

ways. Conversely, Jordans topics, such as “lane” and “change”, reﬂect concerns more

closely related to driver behavior and road usage, possibly due to denser traﬃc conditions

and the importance of maneuvering in more constrained driving environments. The key-

words presented in Figure 5 further emphasize these regional diﬀerences, with the USA

dataset reﬂecting challenges associated with high-speed driving and wildlife interactions,

while the Jordan dataset highlights technical issues and traﬃc safety measures.

These topic insights reveal key diﬀerences in driving culture and infrastructure be-

tween the two countries. For the USA, terms like “deer” and “guardrail” appear, high-

lighting interactions with wildlife and road safety features unique to sprawling, diverse

landscapes. In Jordan, the recurrence of “precautions” and “safety” indicates a heightened

awareness and perhaps a more precautionary approach to driving, inﬂuenced by regional

driving norms and infrastructure development stages. The ﬁgure underscores how re-

gional themes drive distinct safety priorities, with the USA focusing on mitigating high-

speed risks and Jordan emphasizing adherence to safety precautions in urban contexts.

Figure 4. Training and validation (accuracy vs. loss) for the Jordan dataset.

Electronics 2025,14, 272 19 of 35

5.2. Topic Modeling Results

The coherence score graphs for topic modeling in the USA and Jordan datasets show

distinct peaks, indicating each dataset’s optimal number of topics. For the USA, the peak

coherence is achieved with about 25 topics, suggesting a wider variety of distinct discussion

points within crash narratives. In contrast, Jordan’s dataset reaches peak coherence with

approximately ten topics, indicating a more concentrated range of themes that best capture

the discussions within the crash data. This comparison highlights the complexity of the

USA’s crash narratives and a more focused thematic structure within Jordan’s dataset.

The larger number of topics in the USA dataset reﬂects its more diverse crash circum-

stances, potentially linked to differences in infrastructure, driving behaviors, and reporting

practices.

Figure 5shows the USA and Jordan Topics per dataset. The following insights are

extracted from the ﬁgure. In the USA dataset, topics like “overturned” and “roadway”

suggest a concentration on accidents and road conditions, which may be indicative of

the country’s extensive road networks and the higher incidence of vehicle rollovers on

highways. Conversely, Jordan’s topics, such as “lane” and “change”, reﬂect concerns more

closely related to driver behavior and road usage, possibly due to denser trafﬁc conditions

and the importance of maneuvering in more constrained driving environments. The

keywords presented in Figure 5further emphasize these regional differences, with the USA

dataset reﬂecting challenges associated with high-speed driving and wildlife interactions,

while the Jordan dataset highlights technical issues and trafﬁc safety measures.

Electronics 2025, 14, x FOR PEER REVIEW 18 of 34

Figure 5. Comparison of topic results: USA (left) vs. Jordan (right).

In this experiment, the coherence score is systematically computed for topics ranging

from 2 to 50. The results in Figure 5 revealed that the highest coherence score is achieved

at 𝑘 = 25 and 10 for the US and Jordan datasets, respectively. Table 8 provides an il-

lustrative example of some topics identiﬁed in the process. This analysis highlights the

adaptability of the topic modeling approach to identify region-speciﬁc issues, which can

inform tailored policy interventions.

Table 8. Top 10 words across the ﬁrst three topics per dataset.

Dataset Topic # Keywords

USA dataset

Topic 1 [(overturned, 0.061), (roadway, 0.059), (curve, 0.050), (negotiate, 0.050), (right,

0.049), (ditch, 0.046), (occurred, 0.044), (ran, 0.044), (traveled, 0.043), (crash, 0.039)]

Topic 2 [(rear, 0.093), (vehicle, 0.078), (turn, 0.050), (struck, 0.049), (traffic, 0.041), (stopped

0.039), (left, 0.036), (make, 0.034), (slowed, 0.032), (lane, 0.028)]

Topic 3 [(crash, 0.040), (vehicle, 0.037), (assisted, 0.036), (occurred, 0.034), (county, 0.032),

(sheriffs, 0.028), (tpr, 0.028), (department, 0.027), (rest, 0.026), (came, 0.025)]

Jordan dataset

Topic 1

[(lane, 0.052), (change, 0.044), (sudden, 0.042), (vehicle, 0.019), (damage, 0.019),

(estimated, 0.017), (material, 0.017), (specialized, 0.017), (result, 0.017), (expert,

0.017)]

Topic 2

[(precautions, 0.048), (failure, 0.047), (necessary, 0.043), (driving, 0.028), (violation

0.025), (safety, 0.025), (vehicle, 0.022), (technical, 0.019), (driver, 0.019), (result,

0.019)]

Figure 5. Comparison of topic results: USA (left) vs. Jordan (right).

These topic insights reveal key differences in driving culture and infrastructure be-

tween the two countries. For the USA, terms like “deer” and “guardrail” appear, high-

Electronics 2025,14, 272 20 of 35

lighting interactions with wildlife and road safety features unique to sprawling, diverse

landscapes. In Jordan, the recurrence of “precautions” and “safety” indicates a heightened

awareness and perhaps a more precautionary approach to driving, inﬂuenced by regional

driving norms and infrastructure development stages. The ﬁgure underscores how regional

themes drive distinct safety priorities, with the USA focusing on mitigating high-speed

risks and Jordan emphasizing adherence to safety precautions in urban contexts.

In this experiment, the coherence score is systematically computed for topics ranging

from 2 to 50. The results in Figure 5revealed that the highest coherence score is achieved at

25 and 10 for the US and Jordan datasets, respectively. Table 8provides an illustrative

example of some topics identiﬁed in the process. This analysis highlights the adaptability

of the topic modeling approach to identify region-speciﬁc issues, which can inform tailored

policy interventions.

Table 8. Top 10 words across the ﬁrst three topics per dataset.

Dataset Topic # Keywords

USA dataset

Topic 1

[(‘overturned’, 0.061), (‘roadway’, 0.059), (‘curve’, 0.050),

(‘negotiate’, 0.050), (‘right’, 0.049), (‘ditch’, 0.046), (‘occurred’,

0.044), (‘ran’, 0.044), (‘traveled’, 0.043), (‘crash’, 0.039)]

Topic 2

[(‘rear’, 0.093), (‘vehicle’, 0.078), (‘turn’, 0.050), (‘struck’, 0.049),

(‘trafﬁc’, 0.041), (‘stopped’, 0.039), (‘left’, 0.036), (‘make’, 0.034),

(‘slowed’, 0.032), (‘lane’, 0.028)]

Topic 3

[(‘crash’, 0.040), (‘vehicle’, 0.037), (‘assisted’, 0.036), (‘occurred’,

0.034), (‘county’, 0.032), (‘sheriffs’, 0.028), (‘tpr’, 0.028),

(‘department’, 0.027), (‘rest’, 0.026), (‘came’, 0.025)]

Jordan dataset

Topic 1

[(‘lane’, 0.052), (‘change’, 0.044), (‘sudden’, 0.042), (‘vehicle’,

0.019), (‘damage’, 0.019), (‘estimated’, 0.017), (‘material’, 0.017),

(‘specialized’, 0.017), (‘result’, 0.017), (‘expert’, 0.017)]

Topic 2

[(‘precautions’, 0.048), (‘failure’, 0.047), (‘necessary’, 0.043),

(‘driving’, 0.028), (‘violation’, 0.025), (‘safety’, 0.025), (‘vehicle’,

0.022), (‘technical’, 0.019), (‘driver’, 0.019), (‘result’, 0.019)]

Topic 3

[(‘sequence’, 0.072), (‘close’, 0.071), (‘vehicle’, 0.027), (‘number’,

0.027), (‘collided’, 0.027), (‘driver’, 0.027), (‘followup’, 0.025),

(‘violation’, 0.025), (‘expert’, 0.024), (‘technical’, 0.023)]

5.3. SHAP Interpretation of Crash Outcomes in the USA and Jordan

The SHAP analysis provides insights into the inﬂuential factors associated with fatal

and non-fatal crashes in the USA and Jordan, revealing distinct and shared contributors

across both regions. While SHAP values offer valuable interpretability by indicating which

factors have a strong association with crash outcomes, it is essential to recognize that

SHAP relies on observed data and does not establish causation. SHAP’s correlation-based

approach cannot account for unobserved variables, which may also play critical roles in

crash dynamics.

5.3.1. Causality Limitations and Unobserved Factors

One major limitation of the SHAP framework, as noted, is its inability to identify root

causes or account for unobserved factors inﬂuencing crash outcomes. This limitation is

particularly important in real-world crash analysis, where certain inﬂuential factors, such

as driver distraction, road conditions, or unseen vehicle malfunctions, may not be explicitly

captured in the dataset. While SHAP highlights important associations, a well-posed causal

Electronics 2025,14, 272 21 of 35

discovery framework could provide a more comprehensive understanding by directly

identifying relationships beyond those observed.

Future research could incorporate causal inference methods alongside SHAP to better

address unobserved variables. Techniques such as structural equation modeling, propensity

score matching, or latent variable models could help capture hidden factors inﬂuencing

crashes. By expanding the analytical approach in this way, researchers could develop a

more nuanced understanding of crash causality, enhancing the insights provided by SHAP.

5.3.2. Key Findings in the USA

In the USA dataset, SHAP analysis identiﬁes prominent factors for fatal crashes, with

terms such as “overturned” and “attempted” indicating rollover risks and failed evasive

maneuvers as signiﬁcant dangers. Other factors, including “hours”, “scene”, “hospital”,

and “alcohol”, highlight the role of delayed emergency response, serious injuries, and

substance impairment in fatal outcomes. Environmental factors like “tree” and “ﬁre” also

appear in fatal cases, underscoring the impact of roadside obstacles on crash severity (see

Figure 6: USA dataset SHAP values for fatal and non-fatal).

Electronics 2025, 14, x FOR PEER REVIEW 19 of 34

Topic 3 [(sequence, 0.072), (close, 0.071), (vehicle, 0.027), (number, 0.027), (collided

0.027),

(driver, 0.027), (followup, 0.025), (violation, 0.025), (expert, 0.024), (technical, 0.023)]

5.3. SHAP Interpretation of Crash Outcomes in the USA and Jordan

The SHAP analysis provides insights into the inﬂuential factors associated with fatal

and non-fatal crashes in the USA and Jordan, revealing distinct and shared contributors

across both regions. While SHAP values oﬀer valuable interpretability by indicating

which factors have a strong association with crash outcomes, it is essential to recognize

that SHAP relies on observed data and does not establish causation. SHAPs correlation-

based approach cannot account for unobserved variables, which may also play critical

roles in crash dynamics.

5.3.1. Causality Limitations and Unobserved Factors

One major limitation of the SHAP framework, as noted, is its inability to identify root

causes or account for unobserved factors inﬂuencing crash outcomes. This limitation is

particularly important in real-world crash analysis, where certain inﬂuential factors, such

as driver distraction, road conditions, or unseen vehicle malfunctions, may not be explic-

itly captured in the dataset. While SHAP highlights important associations, a well-posed

causal discovery framework could provide a more comprehensive understanding by di-

rectly identifying relationships beyond those observed.

Future research could incorporate causal inference methods alongside SHAP to bet-

ter address unobserved variables. Techniques such as structural equation modeling, pro-

pensity score matching, or latent variable models could help capture hidden factors inﬂu-

encing crashes. By expanding the analytical approach in this way, researchers could de-

velop a more nuanced understanding of crash causality, enhancing the insights provided

by SHAP.

5.3.2. Key Findings in the USA

In the USA dataset, SHAP analysis identiﬁes prominent factors for fatal crashes, with

terms such as “overturned” and “aempted” indicating rollover risks and failed evasive

maneuvers as signiﬁcant dangers. Other factors, including “hours”, “scene”, “hospital”,

and “alcohol”, highlight the role of delayed emergency response, serious injuries, and

substance impairment in fatal outcomes. Environmental factors like “tree” and “ﬁre” also

appear in fatal cases, underscoring the impact of roadside obstacles on crash severity (see

Figure 6: USA dataset SHAP values for fatal and non-fatal).

Figure 6. USA dataset SHAP values for fatal and non-fatal.

For non-fatal crashes, terms such as “control”, “wheels”, “fence”, and “facing” suggest

that vehicle stability issues and interactions with roadside structures are common inﬂuences

on injury outcomes.

5.3.3. Key Findings in Jordan

In Jordan, the SHAP analysis indicates a distinct pattern in crash outcomes, with

keywords like “damage” and “resulted” serving as primary factors in both fatal and non-

fatal crashes, suggesting a direct correlation between the extent of damage and crash

severity (see Figure 7: Jordan dataset SHAP values for fatal and non-fatal).

Additional factors in fatal incidents include “towards”, “highway”, “bridge”, and “vi-

olation”, which emphasize the inﬂuence of high-speed roadways, infrastructure elements,

and trafﬁc violations. The presence of “technical” issues in fatal cases suggests that vehicle

maintenance and technical faults are common concerns in Jordan, where older vehicle

ﬂeets may increase crash risks. For non-fatal crashes in Jordan, terms such as “island”,

“pedestrian”, “defect”, and “device” point to pedestrian involvement and technical vehicle

issues as prominent factors.

Electronics 2025,14, 272 22 of 35

Electronics 2025, 14, x FOR PEER REVIEW 20 of 34

For non-fatal crashes, terms such as “control”, “wheels”, “fence”, and “facing” sug-

gest that vehicle stability issues and interactions with roadside structures are common

inﬂuences on injury outcomes.

5.3.3. Key Findings in Jordan

In Jordan, the SHAP analysis indicates a distinct paern in crash outcomes, with key-

words like “damage” and “resulted” serving as primary factors in both fatal and non-fatal

crashes, suggesting a direct correlation between the extent of damage and crash severity

(see Figure 7: Jordan dataset SHAP values for fatal and non-fatal)

Figure 7. Jordan dataset SHAP values for fatal and non-fatal.

Additional factors in fatal incidents include “towards”, “highway”, “bridge”, and

“violation”, which emphasize the inﬂuence of high-speed roadways, infrastructure ele-

ments, and traﬃc violations. The presence of “technical” issues in fatal cases suggests that

vehicle maintenance and technical faults are common concerns in Jordan, where older

vehicle ﬂeets may increase crash risks. For non-fatal crashes in Jordan, terms such as “is-

land”, “pedestrian”, “defect”, and “device” point to pedestrian involvement and technical

vehicle issues as prominent factors.

5.4. Hierarchical Clustering

The clustering dendrograms for Jordan and the USA depict the thematic structures

of each countrys crash reports. Jordans data suggests focusing on speciﬁc issues like ve-

hicle damage and driver skills, along with safety and prevention measures. It also shows

distinct discussions on fuel and animal interactions. The USAs dendrogram, in contrast,

reveals a more complex array of topics that intertwine various aspects of road incidents,

including animal crossings, vehicular dynamics, and legal responses, suggesting richer

narrative detail and a broader spectrum of crash-related factors. Figures 8 and 9 show the

topics for the USA and Jordanian datasets, respectively.

Figure 7. Jordan dataset SHAP values for fatal and non-fatal.

5.4. Hierarchical Clustering

The clustering dendrograms for Jordan and the USA depict the thematic structures of

each country’s crash reports. Jordan’s data suggests focusing on speciﬁc issues like vehicle

damage and driver skills, along with safety and prevention measures. It also shows distinct

discussions on fuel and animal interactions. The USA’s dendrogram, in contrast, reveals a

more complex array of topics that intertwine various aspects of road incidents, including

animal crossings, vehicular dynamics, and legal responses, suggesting richer narrative

detail and a broader spectrum of crash-related factors. Figures 8and 9show the topics for

the USA and Jordanian datasets, respectively.

Electronics 2025, 14, x FOR PEER REVIEW 21 of 34

Figure 8. USA hierarchical clustering.

Figure 9. Jordan hierarchical clustering.

5.5. Similarity Matrix

A similarity matrix is a mathematical representation used in text analysis to measure

the pairwise similarity between data points, such as documents, words, or topics. In crash

data analysis, similarity matrices are employed to identify thematic overlaps and clusters,

oﬀering insights into paerns within the dataset. For instance, the USAs crash data simi-

larity matrix (Figure 10) highlights thematic overlaps in vehicle-related damage and hu-

man factors, forming distinct clusters for specialized incidents such as pedestrian injuries

and animal-related crashes. This clustering reﬂects a clear separation of themes, capturing

speciﬁc regional crash characteristics in the USA. Conversely, Jordans similarity matrix

(Figure 11) reveals a more intricate web of interrelated topics, often emphasizing infra-

structure and vehicle dynamics. Distinct intersections in crash narratives, such as those

linking road conditions, vehicle types, and crash scenarios, suggest a high degree of the-

matic connectivity. Both datasets exhibit unique paerns that mirror their regional traﬃc

incident characteristics, with the USA showing greater thematic integration and Jordan

displaying more deﬁned separations in speciﬁc themes. Such matrices provide a nuanced

understanding of the structural and thematic diﬀerences in crash data reporting, revealing

how regional factors inﬂuence traﬃc incident analysis [54–56].