Advancing Choice Modelling for Time-Critical Decision-Making: Integrating Machine Learning to Improve Predictive Accuracy and Interpretability

Abstract

Aim: This thesis examines how time pressure affects decision-making in dynamic driving scenarios, integrating traditional discrete choice modeling with modern ma- chine learning approaches to improve predictive accuracy, interpretability, and un- derstanding of behavioral adaptations over time. Methods: An online experiment with 514 participants was conducted, where each participant made 10 simulated driving decisions under varying time pres- sures—from unlimited time to severe 5-second constraints. In each trial, participants were asked to choose whether to stay in their current lane or change lanes, making lane choice the primary dependent variable. Study 1 used mixed-effects logistic re- gression to explore how time constraints influenced optimal lane-change decisions, while also accounting for individual differences such as risk perception, education level, and gender. In Phase 2 of the study, these insights informed the develop- ment of a feature- engineered Long Short-Term Memory (LSTM) model, trained to detect temporal patterns in lane-choice behavior. Phase 3 extended this by training an LSTM directly on raw, sequential input data-removing preprocessing constraints to understand the intricate latent behavior which was not captured by traditional models. Results: Time pressure had a non-linear impact on decision performance, with par- ticipants performing worst under moderate constraints (10–30 seconds), supporting the hypothesis that cognitive conflict peaks when there is insufficient time to delib- erate but too much to act on intuition. Conversely, performance improved under both low and extreme time constraints, indicating shifts in cognitive strategy (e.g., deliberative vs. heuristic modes). In Phase 2, we transitioned from inference to pre- diction, using a feature-engineered LSTM model to address the limitations of static regression models. Although it improved predictive accuracy its performance was still constrained by the reliance on hand-crafted features. In Phase 3, a raw-sequence LSTM model captured temporal dependencies and adaptive learning more effec- tively, achieving 95.6% test accuracy and 97.5% AUC-ROC. SHAP analysis and Par- tial Dependence Plots (PDPs) further explained the non-linear feature interactions and adaptive cue weighting, enhancing model transparency. Conclusion: This multi-method thesis demonstrates that decisions under time pres- sure are not merely degraded, but adaptive and context-sensitive. The integration of behavioral theory with interpretable deep learning offers new pathways for build- ing neuroadaptive decision-support systems in high-stakes domains such as au- tonomous transport, space operations, and cognitive augmentation technologies.