CSAI: Conditional Self-Attention Imputation for Healthcare Time-series

The introduction of the Conditional Self-Attention Imputation (CSAI) model represents a significant leap in addressing the challenges posed by missing data in multivariate time series derived from hospital electronic health records (EHRs). By employing attention-based hidden state initialization and a non-uniform masking strategy, CSAI effectively captures both long- and short-range temporal dependencies, which are critical in clinical settings. Comprehensive evaluations across four benchmark EHR datasets have shown that CSAI outperforms existing state-of-the-art architectures in data restoration and related tasks, underscoring its potential impact on healthcare analytics. Furthermore, CSAI's integration into the open-source PyPOTS toolbox facilitates broader access to advanced machine learning techniques for researchers and practitioners working with partially observed time series data, paving the way for improved patient care and data management in healthcare.