A Novel Explainable Deep Learning STING Kernel Approach for Multivariate Time Series Imputation in Healthcare

Multivariate time series (MTS) data is crucial in fields such as healthcare, finance, and traffic management, particularly for forecasting clinical outcomes like mortality rates, disease risks, and hospital stay durations.In healthcare, imputing missing values from complex MTS datasets can enhance critical care management and enable personalized treatments.This study focuses on imputing missing values in health-related data, especially intravenous vital signs and data essential to physicians' decision-making.To address limitations of existing imputation methods, we introduce a novel approach: the STING Kernel Deep Level (SKDL), coupled with an explainable framework.SKDL is designed to improve accuracy and effectively handle categorical outputs.Our evaluation using the MIMIC-IV dataset shows that SKDL outperforms traditional imputation methods such as Mean Imputation, k-Nearest Neighbors (KNN), and standard GAN-based approaches, based on performance metrics including Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and R Squared.SKDL achieved an MAE of 0.0870, MSE of 0.0175, RMSE of 0.0040, and R Squared of 0.3367, indicating strong accuracy.Furthermore, the integration of Explainable AI (XAI) enables interpretability by visualizing imputation rationale, helping clinicians verify that the predicted values align with physiological expectations, thereby reinforcing trust in the imputation process.These results suggest that SKDL, with its interpretable design, provides a reliable solution for missing data imputation and supports more consistent and transparent clinical decisionmaking.