Agents of Prediction: AI Framework Boosts Traffic Crash Severity Analysis

by

Author: Denis Avetisyan

A new multi-agent system leverages the power of large language models and traditional machine learning to dramatically improve the accuracy and interpretability of traffic accident severity predictions.

The Transport Agent framework operates on the principle that severity assessment isn’t a calculation, but an emergent property-data preprocessing selects features, a dedicated team evaluates category-specific impacts, and an MLP module integrates these disparate evaluations, not to <i>determine</i> severity, but to <i>reveal</i> the system’s inherent understanding of it-a process destined to amplify existing biases and propagate unforeseen errors.
The Transport Agent framework operates on the principle that severity assessment isn’t a calculation, but an emergent property-data preprocessing selects features, a dedicated team evaluates category-specific impacts, and an MLP module integrates these disparate evaluations, not to determine severity, but to reveal the system’s inherent understanding of it-a process destined to amplify existing biases and propagate unforeseen errors.

TransportAgents combines hybrid reasoning and data fusion to provide a robust framework for severity assessment.

Accurate prediction of traffic accident severity remains a challenge due to the complexity and heterogeneity of crash data. This paper introduces ‘TransportAgents: a multi-agents LLM framework for traffic accident severity prediction’, a novel hybrid approach that integrates the reasoning capabilities of large language models with traditional machine learning techniques. Through a multi-agent system specializing in distinct data subsets-demographics, environment, and incident details-TransportAgents consistently outperforms both conventional machine learning and single-agent LLM baselines across multiple U.S. datasets. Can this framework’s improved accuracy and calibrated predictions unlock more effective emergency response and proactive safety planning?

The Inevitable Complexity of Crash Prediction

Conventional approaches to predicting traffic accident severity frequently fall short due to the inherent complexity of real-world crashes and the subtle variations in resulting injuries. Single predictive models often oversimplify the multitude of interacting factors – encompassing environmental conditions, vehicle dynamics, driver behavior, and the specific characteristics of the collision itself – leading to inaccurate assessments. The nuanced nature of injury patterns, where seemingly minor differences in impact can drastically alter the severity of harm, further challenges these models. Consequently, predictions may fail to capture the full spectrum of potential outcomes, hindering effective risk analysis and the development of targeted safety measures. This limitation underscores the need for more sophisticated methodologies capable of accommodating the multifaceted nature of traffic accidents and their consequences.

The inability to precisely forecast crash severity significantly impedes proactive safety measures and efficient allocation of resources. Current risk assessment tools, frequently based on limited data or simplistic algorithms, often fail to capture the intricate interplay of factors contributing to injury outcomes. This imprecision leads to interventions that may not address the most critical risks, and hinders the development of targeted preventative strategies. Consequently, a pressing need exists for analytical methods that not only improve predictive accuracy, but also offer clear insights into the key drivers of crash severity – allowing safety professionals to understand why certain crashes result in specific injury levels and, crucially, to design more effective countermeasures based on these understandings.

Using the CPSRMS dataset, the TransportAgent framework demonstrates superior performance in predicting severity levels by generating distributions more closely aligned with ground truth compared to baseline methods like Vanilla LLMs, CoT prompting, and MLPs, which exhibit a bias towards under-representing higher severity levels.
Using the CPSRMS dataset, the TransportAgent framework demonstrates superior performance in predicting severity levels by generating distributions more closely aligned with ground truth compared to baseline methods like Vanilla LLMs, CoT prompting, and MLPs, which exhibit a bias towards under-representing higher severity levels.

Architecting for Adaptation: The TransportAgent Framework

TransportAgent utilizes a hybrid architecture to enhance crash severity prediction by integrating Large Language Models (LLMs) with established machine learning techniques. This approach addresses limitations inherent in either system operating independently; LLMs provide contextual understanding from unstructured data, while structured machine learning components offer precision and reliability in numerical analysis. The framework processes crash data, incorporating textual reports and structured features such as vehicle speed and road conditions. By combining these data sources, TransportAgent aims to improve prediction accuracy and provide more nuanced risk assessments compared to traditional methods relying solely on statistical modeling or natural language processing. This integration allows for a more comprehensive analysis, leveraging the strengths of both LLMs and structured machine learning to identify critical factors influencing crash severity.

The TransportAgent framework utilizes a multi-agent system architecture, distributing processing tasks among specialized agents that operate collaboratively. This approach contrasts with monolithic models by allowing for modularity and independent refinement of individual components. Each agent is designed with a specific function – such as feature selection, data categorization, or severity assessment – and communicates with others to achieve a comprehensive risk evaluation. This distributed design enhances both the flexibility of the system, allowing for easy adaptation to new data sources or evolving requirements, and its interpretability, as the contribution of each agent to the final prediction can be individually examined and validated.

The TransportAgent framework utilizes a multi-agent system comprised of three primary components: the Feature Selection Agent, the Conceptual Category Organizer, and the Severity Assessment Team. The Feature Selection Agent operates on raw input data to identify and prioritize the most relevant features for crash severity prediction, reducing dimensionality and improving model efficiency. Following feature selection, the Conceptual Category Organizer structures the refined data into predefined conceptual categories – such as vehicle dynamics, environmental conditions, and driver behavior – to facilitate standardized information processing. Finally, the Severity Assessment Team leverages the categorized features to evaluate crash risk levels, providing a structured assessment of potential crash severity based on the processed inputs.

The Transport Agent utilizes a feature selection process to identify relevant variables, links them to conceptual categories, organizes those variables, and then evaluates category-specific severity scores to determine overall risk <span class="katex-eq" data-katex-display="false"> ightarrow</span>.
The Transport Agent utilizes a feature selection process to identify relevant variables, links them to conceptual categories, organizes those variables, and then evaluates category-specific severity scores to determine overall risk ightarrow.

The Illusion of Accuracy: Modeling Injury Dynamics

TransportAgent employs a diverse set of predictive models to analyze injury data, incorporating both econometric and machine learning techniques. Specifically, the framework utilizes Binary Logit, Multinomial Logit, and Ordered Logit models, alongside tree-based methods including Random Forests, Gradient Boosting, and XGBoost. This integration allows the system to capture a wider range of patterns and relationships within the data than would be possible with a single modeling approach, improving the overall predictive capability and robustness of the injury risk assessment.

The TransportAgent framework utilizes a Multilayer Perceptron (MLP) module to consolidate predictions from diverse modeling techniques, including Binary Logit, Multinomial Logit, Ordered Logit, Random Forests, Gradient Boosting, and XGBoost. This MLP functions as a meta-learner, accepting the outputs of these individual models as inputs and synthesizing them into a single, unified prediction. By considering the combined insights from multiple algorithms, the MLP is designed to leverage complementary strengths and mitigate individual model weaknesses, resulting in a more robust and accurate injury risk assessment based on comprehensive feature analysis.

The TransportAgent framework demonstrates state-of-the-art predictive accuracy when utilizing the LLaMA-3.3-70B-Instruct backbone. Specifically, the model achieves an accuracy of 73.31% on the Consumer Product Safety Risk Management System (CPSRMS) dataset and 76.9% on the National Electronic Injury Surveillance System (NEISS) dataset. These results indicate a significant advancement in injury prediction capabilities compared to existing methods, as validated by performance on these established datasets.

Evaluation of the TransportAgent model across ten distinct train-test splits yielded a consistent average accuracy of 70.69%, with a standard deviation of ± 1.76%. This low standard deviation indicates the model’s robustness and its ability to generalize effectively to unseen data, as performance does not significantly fluctuate with different data partitions. The consistent performance across multiple splits strengthens confidence in the model’s predictive capabilities and reduces the risk of overfitting to a specific training set.

The Inevitable Future: Scaling Adaptation and Understanding Failure

The TransportAgent framework distinguishes itself through a deliberately modular architecture, facilitating seamless incorporation of evolving data streams, analytical methods, and agent capabilities. This design philosophy permits researchers and practitioners to readily update the system with novel crash data – such as high-resolution sensor information or real-time traffic patterns – without disrupting core functionality. Similarly, advanced modeling techniques, including machine learning algorithms or econometric approaches, can be integrated as plug-and-play components. The framework’s open structure also extends to agent functionalities; new behaviors, decision-making processes, or interaction protocols can be added to simulate a wider range of driver behaviors and traffic scenarios, ultimately fostering a highly adaptable and extensible platform for traffic safety analysis and intervention.

The potential of the TransportAgent framework extends beyond mere prediction; it offers a pathway towards proactive traffic safety management. By generating more accurate assessments of crash severity, the system facilitates targeted interventions – allowing resources to be allocated precisely where they are most needed to mitigate risk. This capability moves beyond reactive responses to accidents, enabling preventative measures such as optimized road design, focused driver education programs, and the strategic deployment of safety technologies. Ultimately, enhanced predictive power, coupled with improved interpretability of the factors influencing crash outcomes, promises not only to lessen the severity of accidents when they do occur, but also to contribute to a demonstrable reduction in their overall incidence, fostering safer roadways for all.

Investigations are planned to integrate advanced reasoning strategies-specifically, Chain-of-Thought prompting-into the TransportAgent framework. This technique encourages the agent to articulate its decision-making process step-by-step, moving beyond simple predictions to provide a transparent rationale for crash severity assessments. Concurrently, researchers will refine prompt engineering methodologies, carefully crafting input instructions to elicit more nuanced and accurate responses from the underlying models. These combined efforts aim to not only boost the agent’s predictive performance but also to make its internal logic more interpretable, fostering trust and enabling stakeholders to better understand the factors driving risk assessments and, ultimately, to implement more effective safety measures.

The pursuit of predictive accuracy, as demonstrated by TransportAgent, often neglects the inherent instability within complex systems. This framework, while striving for superior crash severity prediction through data fusion and hybrid reasoning, operates within a probabilistic landscape. Ada Lovelace observed that “the Analytical Engine has no pretensions whatever to originate anything.” Similarly, TransportAgent doesn’t create foresight; it extrapolates patterns from existing data, acknowledging that even the most refined model is susceptible to unforeseen variables. Stability, in this context, is merely an illusion that caches well – a temporary alignment with observed data, easily disrupted by the inevitable chaos of real-world events. The system’s architecture, therefore, isn’t a solution, but a carefully constructed prophecy of potential failure, a point readily acknowledged by those who understand systems aren’t tools, but ecosystems.

The Road Ahead

TransportAgent, like any attempt to map the chaotic dance of cause and effect, offers a snapshot, not a solution. It’s a clever arrangement of mirrors, reflecting the problem with increasing fidelity, but still fundamentally reliant on the ghosts of accidents past. Every feature engineered is a promise made to the past, a belief that history will reliably repeat itself. The framework’s success hints not at mastery over risk, but at an improved capacity to describe it – a subtle, yet critical distinction.

The true challenge lies not in incremental improvements to prediction accuracy, but in acknowledging the inherent limitations of such models. Systems built to anticipate failure will, inevitably, fail in novel ways. The focus must shift towards resilience – towards architectures that anticipate their own obsolescence and gracefully accommodate the unexpected. One imagines a future where these agents aren’t merely predicting severity, but actively participating in the negotiation of safety – a complex interplay of foresight and adaptation.

Ultimately, this work points toward a broader truth: control is an illusion that demands SLAs. Every system, even one as elegantly constructed as TransportAgent, is an ecosystem, not a tool. It will grow, evolve, and eventually, begin fixing itself-in ways its creators never intended, and perhaps, never could have foreseen.

Original article: https://arxiv.org/pdf/2601.15519.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2026-01-23 12:50