Modelling an hCG Immunoassay for Predicting Treatment in Low-Risk Gestational Trophoblastic Disease

Gestational trophoblastic disease (GTD) encompasses disorders of abnormal trophoblast proliferation, including hydatidiform mole and choriocarcinoma, which are clinically monitored by serial measurements of human chorionic gonadotropin (hCG). Early identification of patients whose hCG levels will normalise spontaneously versus those requiring chemotherapy is critical for optimising care and avoiding overtreatment. This study develops predictive machine learning tools to stratify GTD patients based on early hCG trajectory patterns. Two large cohorts were analysed: a post-molar pregnancy cohort with longitudinal urinary hCG measurements, and a cohort of GTD patients treated with first-line methotrexate (MTX) with serial serum hCG monitoring. Time-series feature engineering was applied to capture kinetic characteristics of hCG decline (such as rates of change and pattern metrics), which were used as inputs to classifiers including logistic regression, gradient boosting machines, and random forests. Ensemble models (such as gradient boosting and random forest, which combine multiple decision trees to improve predictive performance) outperformed the linear logistic regression model by capturing nonlinear trends in hCG dynamics. In the urine cohort, the ensemble classifier achieved high specificity and strong sensitivity for predicting cases that ultimately required chemotherapy, indicating robust early recognition of high-risk patients. In the MTX treated serum cohort, the models detected early deviations in hCG decline indicative of emerging methotrexate resistance, enabling identification of patients likely to need second-line therapy. These findings demonstrate that early hCG trajectories contain informative patterns correlated with treatment outcome. Integrating these predictive models into clinical practice could support decision-making by forecasting individual patient courses, enabling prompt intervention for non responders while reducing unnecessary chemotherapy. Together, this work supports the use of early hCG trajectories for personalised risk stratification, enabling timely intervention while minimising unnecessary chemotherapy.