Predictive Failure Detection in Healthcare Integration Middleware Using Hybrid Ensemble Time-Series Machine Learning

Healthcare integration engines process millions of clinical messages daily, yet operational failures including queue saturation, memory exhaustion, thread starvation, and connection pool depletion are detected only after disrupting clinical workflows. This paper presents a Predictive Failure Detection System (PFDS) applying time-series machine learning to integration engine telemetry for proactive failure identification. Three model architectures are evaluated: Long Short-Term Memory (LSTM) networks, Isolation Forest, and a hybrid ensemble combining both with a gradient-boosted meta-classifier. Evaluation across 180 days of simulated enterprise telemetry (200+ channels, 500 messages/second, 847 injected failure events) demonstrates the hybrid ensemble achieves an F1-score of 0.91, median predictive lead-time of 22 minutes, and false positive rate of 4.2%. Detection rates reach 93% for queue saturation and thread starvation, 87–88% for memory exhaustion and connection pool depletion, with the longest observed lead-time at 47 minutes. Aggregate detection of gradual-onset failures (F1–F4) reaches 90.4%. PFDS enables a paradigm shift from reactive incident response to proactive failure