Evaluating the Role of Order and Time Information for Classifying Sequences of Healthcare Events Derived from Claims Data.

Sequences of healthcare events from claims data are increasingly used for predictive modeling. Despite the rise in popularity of neural networks, the benefit of modeling event order and timing remains underexplored. We simulated patient trajectories using parameters estimated from claims data including 22,000 amyotrophic lateral sclerosis (ALS) cases and 22,000 age and gender-matched controls. The simulation reproduced irregular event timing and condition-related activity that increased near an incident diagnosis. We compared three model families for population classification: frequency-based models, sequential models (long short-term memory and Transformer architectures) capturing event order, and sequential models incorporating temporal information. Performance was evaluated using the area under the ROC curve (AUC) at multiple time points before diagnosis. Sequential models outperformed frequency-based baselines by about +0.06 AUC on average, confirming the benefit of modeling event order. Adding time information provided no noticeable improvement, and results were stable across different time encoding and scaling methods. These findings suggest that, in claims-based classification tasks, event order captures most of the useful temporal signal, while explicit time encoding offers limited additional benefit under similar conditions.