Early In-Hospital Mortality Prediction Based on xTimesNet and Time Series Interpretable Methods

Accurate prediction of in-hospital mortality is critical for healthcare risk management and clinical decision-making. However, existing time-series models struggle to model the non-stationarity and cross-periodic patterns in clinical data effectively, and their interpretability remains limited, particularly in addressing the dynamic importance of time-varying features and actionable counterfactual interventions. To address these challenges, this paper proposes the xTimesNet-TSR-CoMTE framework, which integrates a hybrid spectral-temporal model, xTimesNet, with two interpretability methods: Two-Step temporal saliency Rescaling (TSR) and Counterfactual Multivariate Time series Explainability (CoMTE). xTimesNet leverages adaptive Wavelet-FFT mixture and Inception modules to capture cross-period temporal interactions in clinical time series. Experiments on the MIMIC-III dataset, involving 27 features over 48 timesteps, demonstrate the framework's superior performance, achieving a ROC_AUC of 0.785 and PR_AUC of 0.403. Comparative analysis shows that CoMTE outperforms existing methods by generating clinically plausible counterfactuals with minimal modifications, as evidenced by the lowest L1 distance of 0.024 and zero violent feature modifications. TSR identifies critical time-varying features and their temporal importance, pinpointing systolic blood pressure, diastolic blood pressure, and body temperature as the most impactful predictors of mortality. CoMTE generates actionable counterfactual explanations by suggesting minimal interventions on these key features, such as maintaining systolic blood pressure within a stable range during critical early time points. This study advances the application of time-series models in clinical risk analysis and offers a robust solution for interpretable and actionable mortality prediction.