Machine Learning Identifies High-Risk Paratransit Drivers in Bangladesh

Published in Discover Civil Engineering, the study demonstrates that a regularized logistic regression model outperforms more complex ensemble algorithms in distinguishing between drivers who have been involved in a single accident and those who have been involved in multiple incidents. The research addresses a critical gap in urban transport safety by providing a data-driven screening tool for informal transport sectors that serve as the backbone of mobility in many developing cities.

Paratransit systems, which include auto-rickshaws, easy bikes, and other intermediate public transport modes, account for approximately 60% of motorized trips in smaller cities in Bangladesh. These informal fleets are essential for urban resilience, particularly during disruptions to formal public transport networks such as monsoon flooding or fuel shortages. However, the informal nature of these operations means that drivers often enter the profession with minimal licensing, training, or regulatory oversight. This lack of formal supervision creates significant safety challenges that conventional statistical methods have struggled to address effectively.

The study collected primary data through structured face-to-face interviews with 507 active paratransit drivers at nine major paratransit stations across Gazipur between July and December 2021. The sample was purposively selected to capture the operational and geographic diversity of the district, covering all principal vehicle types including easy bikes, manual rickshaws, CNG auto-rickshaws, legunas, and tempos. The researchers categorized drivers based on their self-reported accident history, grouping 386 drivers (76.1%) as single-accident drivers and 121 drivers (23.9%) as repeated-accident drivers. Notably, the sample did not include any drivers who reported zero accidents, a limitation the authors acknowledge as a constraint on the generalizability of the baseline risk estimates.

This method generates synthetic examples of the minority class to balance the training data, allowing the models to learn more robust decision boundaries. The study compared five classification algorithms: Logistic Regression, Random Forest, XGBoost, Gradient Boosting, and a Stacking Ensemble. Model performance was evaluated using repeated stratified five-fold cross-validation, repeated 10 times, to ensure that the results were not artifacts of a single fortunate data partition.

Contrary to the common assumption that more complex models yield better results, the regularized logistic regression emerged as the best-performing and most stable classifier. It achieved a mean F1-score of 0.557 and an AUC-ROC of 0.751 across the cross-validation folds. The model’s simplicity provided a significant advantage in interpretability, as its coefficients yielded directly actionable odds ratios for policy makers. In contrast, while the Stacking Ensemble exhibited the best calibration with a Brier score of 0.204, its predictive accuracy was lower than that of the logistic regression model, suggesting that the added complexity was not justified for this specific dataset.

Feature importance analysis revealed that vehicle type was the dominant predictor of repeated-accident status. In the adjusted multivariable model, motorized three-wheeler operation was the only strongly significant predictor, with an adjusted odds ratio of 14.4. This finding suggests that the risk is driven more by the vehicle-and-exposure dimension, including higher operating speeds and greater exposure in mixed traffic, than by demographic characteristics such as age or education.

For instance, while repeated-accident drivers had significantly more driving experience on average, this association was largely absorbed by correlated operational variables in the multivariable model. The researchers caution against interpreting these bivariate associations as causal, noting that the cross-sectional design of the study precludes definitive causal inference.

When compared to a traditional Poisson regression benchmark, the machine learning classification approach offered a specific advantage in recall. While both models showed comparable overall discrimination ability, the logistic regression model recovered a substantially larger share of repeated-accident drivers, with a recall of 0.568 compared to 0.405 for the Poisson model. This higher recall is critical for a screening tool, as the primary goal is to identify drivers who warrant closer monitoring or intervention before further accidents occur.

The study highlights the potential for integrating predictive screening into existing administrative processes. Because the strongest predictors, such as vehicle type and income, are already recorded in licensing files, authorities could generate risk scores administratively without the need for additional telematics or roadside sensors.

Several limitations warrant consideration. The reliance on self-reported accident history may introduce recall bias or social desirability effects, and the absence of accident-free drivers in the sample limits the ability to model the transition from zero to first accident. Additionally, the study was conducted in a single district, and the surveyed workforce was entirely male, which may constrain the generalizability of the findings to other geographic contexts or gender-diverse paratransit settings. Future research should incorporate objective crash records and longitudinal designs to strengthen causal inference and external validity.

The study was conducted in Gazipur, a rapidly urbanizing district adjacent to Dhaka with a population exceeding 3.4 million. As one of Bangladesh’s fastest-growing industrial zones, the district hosts dense concentrations of garment and manufacturing workers whose daily commutes are heavily dependent on an informal fleet of easy bikes, rickshaws, CNG auto-rickshaws, legunas, and tempos. It is estimated that between 8,000 and 10,000 registered paratransit vehicles operate within the district, serving a population that often lacks access to formal public transportation. According to data from the Gazipur Metropolitan Police, these paratransit vehicles are involved in roughly 35–40% of all reported road accidents. The informal nature of these operations means that drivers often enter the field with minimal licensing, training, or route regulation, creating a safety environment where conventional regulatory systems struggle to maintain oversight. This context makes the district a critical site for analyzing driver-level safety, as the reliability of the informal fleet is a practical component of urban transport resilience, particularly when formal networks are interrupted by monsoon flooding, transport strikes, or fuel shortages.

The research team selected five classification algorithms to evaluate their effectiveness in distinguishing between single-accident and repeated-accident drivers. These included Logistic Regression, Random Forest, XGBoost, Gradient Boosting, and a Stacking Ensemble. The selection was designed to span distinct modeling paradigms, ranging from interpretable linear models to complex ensemble methods. Logistic Regression was included as a baseline because its coefficients provide directly actionable odds ratios for policy. Random Forest and Gradient Boosting represent bagging- and boosting-based tree ensembles, respectively, which are capable of capturing non-linear interactions among driver, vehicle, and exposure variables. XGBoost was added as a regularized, high-performance implementation of gradient boosting, while the Stacking Ensemble combined the complementary strengths of the base learners to test whether further gains in accuracy were possible. This range allowed the researchers to determine whether added complexity was justified for this specific problem and to assess the trade-offs between accuracy and explainability.

Model evaluation was based on repeated stratified five-fold cross-validation, repeated 10 times for a total of 50 folds. This approach was chosen to ensure that the reported performance was not an artifact of a single fortunate data partition. Within each fold, feature standardization was fitted on the training portion only and then applied to the held-out portion, and SMOTE was applied exclusively to the training portion to prevent information leakage. The models were evaluated using Accuracy, Precision, Recall, F1-Score, AUC-ROC, and Matthews Correlation Coefficient. The F1-score was selected as the primary metric because it balances false positives and false negatives, which is critical for imbalanced classification tasks where safe drivers significantly outnumber high-risk ones. The researchers also assessed model calibration using Brier scores and reliability diagrams to determine whether predicted probabilities aligned with actual outcome frequencies.

Feature importance analysis revealed that vehicle type was the single most important predictor of repeated-accident status, with an average importance score of 0.230 across tree-based models. This dominance indicates that the mode of operation separates high-risk from low-risk operators more effectively than any driver attribute. Daily Income ranked second with an importance score of 0.114, followed by the Experience-to-Age Ratio and Driving Experience. Vehicle Speed and Income Efficiency also emerged as leading predictors, confirming that operating mode, economic pressure, and exposure together shape crash risk. In the logistic regression model, motorized vehicle operation carried the largest positive coefficient, confirming its role as the dominant risk-increasing factor. Conversely, easy bike operation and higher education levels were identified as the strongest protective factors, suggesting that lower literacy may compound operational risk by limiting access to formal training or traffic-rule comprehension.

The study’s findings have direct implications for how transportation authorities approach safety regulation in informal sectors. The results suggest that targeted safety interventions should prioritize the vehicle-and-exposure dimension over demographic screening. Because the strongest predictors, such as vehicle type and income, are already recorded in licensing files, authorities could generate risk scores administratively without the need for additional telematics or roadside sensors. The researchers emphasize that framing this screening as a diagnostic tool that directs inspection and training, rather than as a sanction, may be more effective in a sector where operators are politically organized and enforcement is often viewed with suspicion. This strategy aligns with broader goals of resilient and inclusive urban mobility by providing an affordable pathway to improve the reliability of the informal transport system.

Subject of Research: Machine learning classification of high-risk paratransit drivers in Gazipur, Bangladesh

Article Title: Classification of high-risk paratransit drivers using ensemble machine learning

Article References: Siraj, M. S. B., Noor, S., Miah, M. E., Islam, M. J., Mahmud, S., Ferdous, J., & Ahaduzzaman, S. M. (2026). Classification of high-risk paratransit drivers using ensemble machine learning. Discover Civil Engineering, 3(1), Article 252. https://doi.org/10.1007/s44290-026-00652-2

Image Credits: AI Generated

DOI: 10.1007/s44290-026-00652-2

Keywords: paratransit safety, machine learning, logistic regression, urban transport, Bangladesh, driver risk assessment, ensemble methods, traffic accidents, informal transport, predictive modeling, Classification, high-risk