Interpretable AI methods for prompt classification of temporal data

This study presents two novel interpretable models for the prompt classification of temporal data in both event log (el) and time series (TS) domains, with a principal focus on applications in healthcare and wellness. First, we introduce the Sequence-Walking Decision Tree (SWDT), a tree-based model that unifies el data---discrete events annotated with times and values---to enable both supervised and unsupervised tasks. SWDT integrates multivariate temporal relations in a highly interpretable structure, offering mechanisms for predictive modeling, distance-based clustering, and associate rule mining. Second, we propose the Time Series Step Tree (TSST), a decision-tree-based approach tailored to the progressive analysis of univariate and multivariate TS. TSST adopts a step-wise evaluation strategy, making it capable of classifying partial time series promptly (''early'') without awaiting the entire sequence. This design enables reduced latency in mission-critical scenarios, like real-time patient monitoring, while preserving interpretability by generating human-readable decision paths. Experimental evaluations on benchmark datasets demonstrate the effectiveness of both SWDT and TSST in delivering robust predictive performance. Additionally, the proposed models exhibit strong potential for healthcare applications such as prompt detection of critical conditions, where timely, transparent, and reliable decisions are essential.