MLOps-BASED FRAMEWORK FOR PRODUCTION MACHINE LEARNING IN LARGE-SCALE HEALTHCARE CLAIMS PROCESSING

Abstract

Large-scale healthcare claims processing environments manage highly heterogeneous, high-volume, and operationally sensitive information associated with member eligibility, provider services, diagnosis codes, procedure codes, pharmacy transactions, authorization records, claim submissions, adjudication outcomes, payment decisions, denials, appeals, and revenue cycle operations. The increasing adoption of machine learning in healthcare payer organizations has created opportunities for automated claim classification, denial prediction, anomaly detection, payment integrity, fraud risk prioritization, processing optimization, and operational decision support. However, many machine learning initiatives remain limited to experimental environments because production deployment introduces challenges involving data quality, model reproducibility, feature consistency, infrastructure scalability, security, drift, monitoring, governance, explainability, and controlled retraining. This paper proposes an MLOps-Based Framework for Production Machine Learning in Large-Scale Healthcare Claims Processing. The framework integrates healthcare claims ingestion, data validation, privacy-aware preprocessing, feature engineering, feature management, model development, experiment tracking, automated testing, model registry, containerized deployment, scalable inference, continuous monitoring, drift detection, controlled retraining, secure API lifecycle management, enterprise data integration, and governance. The proposed methodology supports both batch and near-realtime claim analytics while maintaining traceability across data, model, code, configuration, and deployment versions. Machine learning services are continuously monitored for predictive quality, operational latency, data drift, concept drift, fairness indicators, infrastructure utilization, and business outcomes. A human-governed promotion process ensures that candidate models are validated before production release. The conceptual evaluation compares the proposed MLOps framework with a conventional manually managed machine learning lifecycle. Results indicate improvements in deployment frequency, model reproducibility, inference scalability, drift response time, claim-processing throughput, prediction consistency, rollback capability, and operational governance. The study demonstrates that MLOps provides a systematic foundation for transforming isolated healthcare machine learning models into reliable, secure, monitored, scalable, and continuously improving production services.