SeqBoost: a sequential explainable model for predicting ED revisits within 72 hours.

PURPOSE: Accurately predicting emergency department (ED) revisits within 72 hours remains challenging due to irregular and short patient visit histories. This study investigates how temporal representations of historical ED utilization and sequential modeling can improve predictive performance and model explainability. METHODS: We evaluate two complementary approaches using the MIMIC-IV-ED dataset. First, we introduce interval-based temporal features that summarize the recency and spacing of prior ED visits, and compare them with conventional visit-count summaries. Second, we examine longitudinal visit-level representations and propose Sequential Boosting (SeqBoost), a gradient-boosting framework that incorporates historical visits sequentially without padding missing values. Models are evaluated using patient-level cross-validation and AUROC, and feature attributions are analyzed using SHAP. RESULTS: Interval-based temporal features consistently outperformed visit-count summaries, particularly for short histories. For patients with a single prior visit, AUROC increased from 0.590 to 0.641, and further to 0.657 when combined with visit counts. Incorporating longitudinal visit-level features improved performance to an AUROC of 0.691 at history length two. SeqBoost achieved predictive performance comparable to that of non-sequential longitudinal models while avoiding padded missing values. SHAP analyses showed that SeqBoost generated feature attributions based solely on observed visits, whereas non-sequential models relied heavily on missing historical features. CONCLUSION: Temporal-interval statistics and longitudinal visit-level representations improve 72-hour ED revisit prediction in MIMIC-IV-ED. SeqBoost achieves competitive predictive performance while providing more stable, clinically interpretable feature attributions by eliminating artifacts caused by missing values. These findings support more reliable and interpretable clinical decision support for identifying patients at risk of return to the ED.