MedImputeX: A Hybrid Generative and Explainable Framework for Missing Data Imputation in Healthcare Time-Series

Abstract

Data imputation is essential for managing missing values in multivariate time series data across several disciplines, including healthcare, transportation, and economics. Diffusionbased and deep learning approaches that are already out there work better, but they frequently make the borders between observed and missing areas uneven. They also have trouble capturing long-range temporal connections, which makes reconstruction accuracy less than ideal. To solve these problems, we suggest MedImputeX, a hybrid imputation system. The model uses BRITS-based bidirectional temporal reasoning to learn from both past and future situations. It uses SAITS multi-head selfattention to do a good job at modeling long-range dependencies. An adversarial refinement module that uses GAIN improves the realism of imputed values at the distribution level. A SHAP-driven interpretability block also makes sure that everything is clear and that each feature is explained. Proposed outcome on the MIMIC-IV dataset show that it works better than BRITS, SAITS, and GAN baselines. MedImputeX gives a correct and easy-tounderstand answer for time series imputation by lowering RMSE and MAE and raising $\mathrm{R}^{2}$.