Uncertainty Quantification in Ambulance TravelTime Prediction : Evaluating Probabilistic DeepLearning under Spatial Anonymization

Sammanfattning

Predictive modeling for ambulance travel time is critical for emergency services.Traditional deterministic models fail to quantify predictive uncertainty duringinference, especially when encountering situations for which the model has notbeen trained. When using real-world data from historical ambulance transports,the data are required to go through an anonymization process, to ensure thatpatient privacy is preserved. This introduces significant geographical noise intothe dataset. This thesis evaluates the reliability of several probabilistic deeplearning models: Heteroscedastic Regression, Monte Carlo Dropout, DeepEnsembles, and Evidential Deep Learning. The models are evaluated on howwell they capture both epistemic and aleatoric uncertainty for ambulance traveltime predictions. To ensure a fair comparison and maximize modelperformance, the Optuna framework was used to identify high-performing modelarchitectures and hyperparameters. Using a real-world dataset provided by Region Skåne, the models are evaluatedboth under standard baseline conditions and under simulated distribution shifts.The results show that the Deep Ensemble architecture successfullydisentangles the uncertainty components and utilizes them to detect and signal when the model is faced with out-of-distribution data without generatingoperationally impractical prediction intervals. Additionally, the study estimates the effects of spatial anonymization, showingan irreducible lower bound of aleatoric uncertainty that correlates directly withthe geographical size of the origin municipality. These findings illustrate thefundamental dilemma that the spatial masking that protects patient privacy limitsthe predictive sharpness of the models. Furthermore, the evaluation confirmsthat integrating uncertainty quantification does not compromise model accuracy,the probabilistic models achieved comparable, and for some evaluation metrics,even superior point-prediction performance compared to the deterministicbaseline.