Automated Anomaly Detection and Correction in High-Dimensional Clinical Time-Series Data Using Robust Dimensionality Reduction and Bayesian Filtering

**Abstract:** The increasing complexity and volume of clinical time-series data, such as continuous glucose monitoring (CGM) and electrocardiogram (ECG) recordings, presents significant challenges for anomaly detection and correction. Existing methods often struggle with high dimensionality, non-stationarity, and the presence of noise. This paper introduces a novel framework, Robust Dimensionality Reduction and Bayesian Filtering (RDBF), for automated anomaly detection and correction in high-dimensional clinical time-series data. RDBF combines sparse principal component analysis (sPCA) with a Bayesian filtering approach to effectively reduce dimensionality, identify anomalous patterns, and dynamically correct for measurement errors. The approach exploits sparsity to efficiently represent the data while the Bayesian filter incorporates prior knowledge about physiological processes to improve anomaly detection accuracy and robustness. Demonstrations using simulated CGM data and a small dataset of real ECG recordings show promising results, achieving a >95% detection rate for simulated anomalies and a 15% improvement in accuracy compared to standard Kalman filtering on the ECG data. The framework is readily deployable and adaptable to various clinical time-series applications.