I built a predictive analytics system that processes continuous biosignal streams — including ECG, SpO2, and heart rate variability — to identify patients at elevated risk of hospitalization. The system ingests real-time physiological data and transforms it into actionable risk scores that clinicians can use during triage and ongoing patient monitoring.

A core component of the system is its real-time alerting mechanism, which generates clinician-facing notifications when a patient's risk profile crosses critical thresholds. These alerts integrate directly into existing clinical workflows, ensuring that high-risk patients receive timely attention without overwhelming care teams with false positives.

The system was designed to support triage workflows in a healthcare environment serving thousands of patients under continuous monitoring, enabling care teams to allocate resources more effectively based on data-driven risk assessments.

The foundation of the system is a robust time-series feature engineering pipeline that extracts clinically meaningful features from raw physiological signals. This includes statistical summaries, frequency-domain features, and temporal patterns derived from sliding windows over continuous biosignal streams.

For risk scoring, I developed an ensemble approach combining gradient-boosted models with deep learning architectures. The gradient-boosted models provided strong baseline performance with interpretable feature importances, while the deep learning components captured complex temporal dependencies in the biosignal data that traditional models could not.

The inference layer was built as a FastAPI-based service optimized for low-latency predictions in a clinical decision support context. The service handles real-time data ingestion, feature computation, and model inference within tight latency requirements to ensure that risk scores are always current and clinically relevant.

The system achieved a 92% F1-score on the hospitalization risk prediction task, demonstrating strong performance across both precision and recall. This balance was critical in a clinical setting where both missed detections and false alarms carry significant consequences.

Through iterative model optimization and threshold tuning, false positives were reduced by 30% compared to the baseline approach. This reduction enabled clinicians to focus their attention on genuinely at-risk patients, improving both workflow efficiency and the quality of clinical decision-making.

The system was deployed in a production healthcare environment serving thousands of patients under continuous monitoring, providing round-the-clock risk assessment and contributing to more proactive and data-informed patient care.