A federated learning framework for deep imputation of missing data in heterogeneous ICU time series

The proliferation of multimodal time-series healthcare data presents unprecedented opportunities for data-driven insights but also significant challenges due to pervasive missing values, especially in critical care environments. Traditional centralised imputation methods are often infeasible due to strict privacy regulations, while single-institution models suffer from poor generalizability. Federated learning offers a promising alternative but faces challenges, including statistical heterogeneity, temporal misalignment, and complex missingness patterns. This paper proposes Fed-HealthImp, a federated learning framework for deep imputation of missing values in multivariate, irregularly sampled clinical time-series. Our framework employs a self-attention-based imputation model with adaptive client weighting to handle non-IID data distributions across hospitals. We evaluate Fed-HealthImp on three real-world ICU datasets (eICU-CRD, MIMIC-IV, HiRID) under various missingness patterns. Results show that Fed-HealthImp achieves imputation quality within 3.5% of a privacy-violating centralised model, significantly outperforms local-only training, and improves downstream mortality prediction AUROC by up to 3.4%. Our work establishes a practical, privacy-preserving pathway for building robust imputation models from fragmented global ICU data.