Cost-Aware Distillation of a Commercial First-Trimester Preeclampsia Screening Engine in Northern Vietnam: Tiered Feature Sets and Explain Ability

Ta Van Thao, Vu Thi Minh Phuong, Nguyen Lien Huong, Duong Thanh Tam, Tran Sach Viet, Vu Ngoc Bac, Vu Ngoc Anh

Nội dung chính của bài viết

Tóm tắt

Commercial first-trimester preeclampsia (PE) screening engines (e.g., PerkinElmer/FMF) integrate maternal factors, biophysical measures, Doppler indices, and biochemical markers to generate continuous risk ratios, but their use in resource-constrained settings is limited by missing data, centralized assays, and limited transparency. We developed a tiered, interpretable machine-learning (ML) distillation pipeline to approximate the PerkinElmer risk ratio for PE <37 weeks, evaluate fidelity across cost-aware feature tiers, identify minimal deployable feature sets, and assess agreement at the clinical cut-off. A retrospective cohort of 1,051 singleton pregnancies from Northern Vietnam (2023–2025) was split into training (n=850) and test (n=201) sets. Complete-case tiers were defined as Tier 0 (maternal factors + MAP), Tier 1 (+ PlGF/PAPP-A), and Tier 2 (+ UtA-PI). Fidelity was assessed using ranking, regression, and calibration metrics, with interpretability via permutation importance, SHAP, and ablation, and threshold mimicry using the PerkinElmer cut-off. Despite declining tier availability, fidelity remained high (Spearman ρ=0.975–0.981), with strong top-rank agreement and excellent AUC-ROC (0.97–1.00). A minimal feature set (MAP MoM, BMI, parity, PlGF MoM, UtA-PI MoM) retained ≥95% fidelity, supporting scalable, explainable PE triage under real-world constraints.

Chi tiết bài viết

Tài liệu tham khảo

1. Steegers EAP, von Dadelszen P, Duvekot JJ, Pijnenborg R. Pre-eclampsia. Lancet. 2010; 376(9741): 631-644. doi:10.1016/S0140-6736(10)60279-6
2. Karumanchi SA, Granger JP. Preeclampsia and Pregnancy-Related Hypertensive Disorders. Hypertension. 2016; 67(2): 238-242. doi:10.1161/HYPERTENSIONAHA.115.05024.
3. Rolnik DL, Wright D, Poon LC, et al. Aspirin versus placebo in pregnancies at high risk for preterm preeclampsia. N Engl J Med. 2017; 377(7): 613-622. doi:10.1056/NEJMoa1704559.
4. Akolekar R, Syngelaki A, Poon L, Wright D, Nicolaides KH. Competing risks model in early screening for preeclampsia by biophysical and biochemical markers. Fetal Diagn Ther. 2013; 33(1): 8-15. doi:10.1159/000341264.
5. O’Gorman N, Wright D, Poon LC, et al. Multicenter screening for pre-eclampsia by maternal factors and biomarkers at 11-13 weeks’ gestation: comparison with NICE guidelines and ACOG recommendations. Ultrasound Obstet Gynecol. 2017; 49(6): 756-760. doi:10.1002/uog.17455.
6. Poon LC, Shennan A, Hyett JA, et al. The FIGO initiative on pre-eclampsia: a pragmatic guide for first-trimester screening and prevention. Int J Gynecol Obstet. 2019; 145(S1): 1-33. doi:10.1002/ijgo.12802.
7. Martins JG, Miller E, Aboukhater D, Bittner M, Rolnik DL, Kawakita T. Performance of a first-trimester combined screening for preterm preeclampsia in the United States population using the fetal medicine foundation competing risks model. Am J Obstet Gynecol MFM. 2025; 7(12): 101803. doi:10.1016/j.ajogmf.2025.101803.
8. Verlohren S, Herraiz I, Lapaire O, et al. The sFlt-1/PlGF ratio in different types of hypertensive pregnancy disorders and its prognostic potential in preeclamptic patients. Am J Obstet Gynecol. 2012; 206(1): 58.e1-58.e588. doi:10.1016/j.ajog.2011.07.037.
9. Breiman L. Random forests. Mach Learn. 2001; 45: 5-32. doi:10.1023/A:1010933404324.
10. von Dadelszen P, Magee LA, Roberts JM. Subclassification of preeclampsia. Hypertens Pregnancy. 2003; 22(2): 143-148. doi:10.1081/PRG-120021060.
11. Ribeiro MT, Singh S, Guestrin C. “Why should I trust you?” Explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. 2016: 1135-1144. doi:10.1145/2939672.2939778.
12. Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17). Curran Associates Inc., Red Hook, NY, USA, 2017; 4768-4777.
13. Bucur O, et al. Knowledge distillation in medical imaging: A survey. arXiv:2203.04742; 2022.
14. Perkins NJ, Cole SR, Harel O, et al. Principled Approaches to Missing Data in Epidemiologic Studies. Am J Epidemiol. 2018; 187(3): 568-575. doi:10.1093/aje/kwx348.
15. Järvelin K, Kekäläinen J. Cumulated gain-based evaluation of IR techniques. ACM Trans Inf Syst. 2002; 20(4): 422-446. doi:10.1145/582415.582418.
16. Collins GS, Reitsma JB, Altman DG. et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD Statement. BMC Med, 2015; 13:1. https://doi.org/10.1186/s12916-014-0241-z.
17. Sculley D, Holt G, Golovin D, Davydov E, Phillips T, Ebner D, et. al. Hidden technical debt in machine learning systems. In: Advances in Neural Information Processing Systems 28 (NeurIPS); 2015: 2503-2511.
18. Rudin C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell. 2019; 1: 206-215.
19. DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988; 44(3): 837-845.
20. Vickers AJ, Cronin AM, Elkin EB, Gonen M. Extensions to decision curve analysis. BMC Med Inform Decis Mak. 2008; 8: 53. doi:10.1186/1472-6947-8-53.
21. Vickers AJ, Cronin AM, Elkin EB et al. Extensions to decision curve analysis, a novel method for evaluating diagnostic tests, prediction models and molecular markers. BMC Med Inform Decis Mak. 2008; 8: 53. https://doi.org/10.1186/1472-6947-8-53.
22. Hinton G, Vinyals O, Dean J. Distilling the knowledge in a neural network. arXiv: 1503.02531; 2015.
23. Molnar C. Interpretable Machine Learning. 2nd ed. 2022.
24. Wiens J, et al. Do no harm: a roadmap for responsible machine learning for health care. Nat Med. 2019; 25: 1337-1340.