Potholes Predict Chaos: Machine Learning Fuses Pavement Data with Traffic Risk

Poor pavement is usually treated as a maintenance problem, not a traffic-safety variable. A new study argues that this separation is a mistake.

The team, led by Mayssa Hamdani, Nafaa Jabeur and Ansar Yasar, combined segment-level Pavement Condition Index (PCI) data from the U.S. Government Open Data Portal with real-time traffic observations pulled from the TomTom Traffic API. Their target area covered 25 road segments in urban New York City, chosen to span the full PCI range from 0 to 10, grouped into poor (PCI ≤ 4), fair (5–7) and good (≥ 8) categories. For each segment they computed a speed ratio — actual speed divided by free-flow speed — where values near 1 indicate smooth flow and low values signal congestion.

Before any modeling, the researchers ran a Bayesian one-way analysis of variance using the PyMC framework to test whether congestion levels genuinely differ across pavement categories. They did: the 94% highest-density intervals for mean congestion ratios in poor segments showed no significant overlap with those of fair or good segments. That statistical check mattered, because it validated pavement condition as an informative feature rather than background noise before it entered the machine learning pipeline.

The core of the work is a supervised Random Forest classifier trained to estimate hourly “instability risk” — a proxy for risky driving behavior. Instability was defined operationally: a segment-hour was labeled unstable when it sat on a poor-pavement segment (PCI ≤ 4) and its instability score, an equal-weighted blend of mean congestion and short-term speed-ratio variability, exceeded the 85th-percentile threshold of roughly 0.204 computed from the training data. The resulting dataset was heavily imbalanced, with only about 4.7% of observations flagged as unstable, so the model used balanced class weighting during training.

Evaluation was deliberately conservative. The team applied a chronological split — the earliest 70% of segment-time observations for training, the most recent 30% for testing — with no temporal overlap, and all feature computation and threshold estimation confined to the training subset to prevent data leakage. On that held-out test set, the infrastructure-aware model achieved an ROC-AUC of 0.9804 and a precision-recall AUC of 0.9074. A traffic-only baseline that excluded pavement data managed an ROC-AUC of 0.8785 but collapsed to a PR-AUC of 0.5833. The gap in precision-recall performance is the more telling figure, since with roughly one unstable observation in twenty, rare-event detection is the real test.

Explainability analysis using SHAP added a second layer of insight. Pavement condition and congestion emerged as the two most influential features, with nearly identical mean absolute SHAP values of about 0.064. Low PCI values consistently pushed predicted instability upward, while high PCI values stabilized predictions. The authors read this as evidence that the model captured physically meaningful interactions — degraded pavement amplifying the destabilizing effects of congestion — rather than exploiting statistical artifacts.

Beyond the classifier, the study introduces a composite metric called PCI_T, which extends the traditional PCI by blending normalized pavement condition (weighted 0.3), the speed ratio (0.2) and a heuristic vehicle-induced stress factor (0.5) that approximates the relative road-degrading impact of different vehicle classes. PCI_T serves as an explanatory feature in the risk model rather than an optimization target. The researchers also embedded it in a broader proposed architecture, ORQCIAM — Optimized Routing with Quantified Composite Infrastructure-Aware Modeling — which layers multi-modal sensing, Bayesian preprocessing, an intelligent routing engine with a dynamic exponential-decay cost function, and predictive path-finding that forecasts future segment costs from historical data.

To probe how such routing might behave in practice, the team ran an agent-based simulation on real OpenStreetMap networks, dispatching 200 randomly generated origin-destination pairs along routes costed either by conventional PCI or by PCI_T. Each traversed edge accumulated an exponentially weighted stress score. The comparison showed that PCI_T-informed routing shifts traffic away from degraded, high-stress segments, reducing the network’s overall exposure to hazardous driving conditions. Separately, mapping predicted risk probabilities geospatially revealed hotspots clustered around major arterial routes and intersections, with residential side streets mostly rated low or very low risk — patterns the authors describe as persistent instability rather than isolated congestion events.

The practical payoff, the researchers suggest, lies in resource allocation. Their inspection-efficiency analysis showed that the infrastructure-aware approach captures most cumulative instability risk exposure while inspecting only about 10% to 15% of segments, outperforming both traffic-only and PCI-only prioritization strategies. In a budget-constrained maintenance world, that means a city could target the handful of road sections where crumbling pavement and volatile traffic intersect to produce the greatest risk, rather than spreading inspections thinly across the network.

The authors are careful about what the model does and does not show. The instability label is a proxy built from traffic and infrastructure features, not observed crash or incident data, so predicted risk is inferred rather than a direct measure of safety outcomes. The evaluation relied on a single Random Forest configuration without comparison to gradient boosting or neural networks, without cross-validation or confidence intervals, and the PCI_T weights were set heuristically with no sensitivity analysis. The SHAP work covered global attribution only, and the study drew on a single urban area, limiting generalizability. Still, the team argues the framework demonstrates a concrete path for future ITS: instead of treating road surface as static context, intelligent transportation systems could ingest pavement condition alongside live traffic dynamics to anticipate safety-critical situations, guide infrastructure-aware routing, and prioritize repairs — with validation across multiple cities and true safety-outcome metrics as the next step.

Subject of Research: Machine learning prediction of traffic instability risk by integrating road pavement condition with traffic data in New York City

Article Title: Integrating road infrastructure condition into intelligent transportation systems: a machine learning-based infrastructure-aware risk prediction approach

Article References: Hamdani, M., Jabeur, N., & Yasar, A. (2026). Integrating road infrastructure condition into intelligent transportation systems: a machine learning-based infrastructure-aware risk prediction approach. Autonomous Intelligent Systems, 6(1), Article 15. https://doi.org/10.1007/s43684-026-00137-0

Image Credits: AI Generated

DOI: 10.1007/s43684-026-00137-0

Keywords: intelligent transportation systems, machine learning, pavement condition index, traffic risk prediction, road infrastructure, random forest, SHAP explainability, New York City, routing optimization, traffic congestion, infrastructure-aware modeling, road safety