A Genetic Algorithm–Optimized EWMA Control Chart Integrating Machine Learning–Based Risk Prediction for Multistage Thyroid Cancer Treatment Monitoring (Preprint)

BACKGROUND Background: Effective monitoring of thyroid cancer treatment requires timely detection of clinically meaningful changes across complex, multistage care pathways. Traditional monitoring approaches often rely on static thresholds and periodic assessments, limiting their sensitivity to subtle yet important process deviations and potentially delaying clinical intervention. OBJECTIVE Objectives: This study aimed to develop and evaluate an integrated, data-driven framework that combines deep learning–based risk prediction with genetic algorithm–optimized statistical process monitoring to enhance thyroid cancer treatment surveillance. METHODS Methods: Informative clinical variables were first identified using an information-theoretic feature selection approach. A Multilayer Perceptron (MLP) neural network was then employed to predict patient-specific treatment risk, capturing nonlinear relationships within longitudinal electronic medical record data. The resulting risk estimates were sequentially incorporated into Exponentially Weighted Moving Average (EWMA) control charts to enable continuous process monitoring. To further improve chart performance, a genetic algorithm was used to optimize EWMA parameters, enhancing accuracy and sensitivity while maintaining chart stability. RESULTS Results: The results demonstrate that the MLP model outperforms conventional machine learning approaches in risk prediction across multiple evaluation metrics. Genetic algorithm–optimized EWMA charts exhibit smoother trajectories, improved stability, and more reliable detection of clinically meaningful deviations compared with non-optimized charts. The consistency of these improvements across multiple features suggests that GA-based optimization enhances overall control chart performance rather than feature-specific behavior. CONCLUSIONS Conclusions: This hybrid framework demonstrates the feasibility of integrating machine learning–derived risk prediction with optimized statistical process control for longitudinal treatment monitoring. The proposed approach supports proactive treatment surveillance, risk-stratified follow-up, timely clinical intervention, and data-driven decision-making in thyroid cancer care. Furthermore, the framework has the potential to enable continuous and adaptive monitoring across complex multi-stage treatment pathways if integrated with electronic medical record (EMR) systems in future implementations.