Flexible decay metrics for early prediction of gestational trophoblastic neoplasia: A comparative modeling approach

Alvin Duke R. Sy; Fernando B. Garcia, Jr.; Clarissa L. Velayo; Michael Daniel C. Lucagbo; Abubakar S. Asaad; Maria Stephanie Fay S. Cagayan

VOLUME 18 (Supplement)

PSL%202021 vol14-no01-p12-28-Mikita%20and%20Padlan

SciEnggJ 18 (Supplement) 469-482
available online: 11 December 2025
DOI: https://doi.org/10.54645/202518SupIFI-95

*Corresponding author
Email Address: arsy3@outlook.up.edu.ph
Date received: 02 September 2025
Dates revised: 16 November 2025
Date accepted: 24 November 2025

Read Full PDF

ARTICLE

Flexible decay metrics for early prediction of gestational trophoblastic neoplasia: A comparative modeling approach

Alvin Duke R. Sy*^1,2, Fernando B. Garcia, Jr.³, Clarissa L. Velayo⁴, Michael Daniel C. Lucagbo⁵, Abubakar S. Asaad¹, and Maria Stephanie Fay S. Cagayan^2,6

¹Department of Epidemiology and Biostatistics, College of
     University of the Philippines Manila, Manila, Philippines 1000
²Department of Pharmacology and Toxicology, College of Medicine,
     University of the Philippines Manila, Manila, Philippines 1000
³Department of Health Policy and Administration, College of Public
     Health, University of the Philippines Manila, Manila, Philippines
     1000
⁴Department of Physiology, College of Medicine, University of the
     Philippines Manila, Manila, Philippines 1000
⁵School of Statistics, University of the Philippines Diliman, Quezon
     City, Philippines 1101
⁶Section of Trophoblastic Diseases, Department of Obstetrics and
     Gynecology, University of the Philippines Philippine General
     Hospital, Taft Avenue, Manila, Philippines 1000

KEYWORDS: gestational trophoblastic disease, prognosis, retrospective studies, area under curve, chorionic gonadotropin, logistic models

Background: Molar pregnancy is a rare condition carrying a 15-22% risk of progression to gestational trophoblastic neoplasia (GTN). Although serial β-hCG monitoring is standard, its ability to predict early malignant change early remains uncertain.

This study examined whether flexible β-hCG decay metrics could improve early GTN prediction compared with conventional thresholds, and whether machine-learning classifiers provide meaningful gains beyond interpretable statistical models.

Methods: This retrospective cohort analyzed 413 post-molar patients with longitudinal β-hCG data. Seven decay metrics were derived from early follow-up measurements and evaluated using logistic and gradient boosting machine (GBM) models.

Model performance was assessed through cross-validated discrimination (area under the curve (AUC)), calibration, and decision-curve analysis (DCA).

Results: The GBM model achieved higher apparent discrimination (AUC: 0.96) but negligible net clinical benefit (NCB) across thresholds, indicating probable overfitting and limited bedside utility. A parsimonious logistic model (AUC: 0.77; calibration slope: 0.85) showed stable calibration and consistent net benefit for identifying low-risk patients. Among the decay metrics, time-to-75% β-hCG decline (~25 days) emerged as the most robust and interpretable predictor (sensitivity 92%, negative predictive value 88%), offering a simple signal of malignant persistence.

Conclusion: Early β-hCG decay dynamics, particularly time-to-75% decline, can guide risk-adaptive follow-up after molar evacuation. Complex machine-learning models contributed little beyond traditional approaches in this moderate-sized cohort. These findings support prospective validation and exploration of cost-effective surveillance models tailored to resource-limited settings.