Understanding Miscarriage Risk in PCOS During Assisted Reproduction

Polycystic ovary syndrome remains one of the most prevalent endocrine disorders affecting women of reproductive age. Clinicians frequently encounter substantial reproductive challenges when managing these patients during assisted reproductive technology cycles. Although controlled ovarian stimulation and fresh embryo transfer often achieve clinical pregnancy, managing miscarriage risk in PCOS remains an urgent priority. Women with this condition experience elevated rates of pregnancy loss compared to normo-ovulatory peers. This heightened vulnerability stems from complex mechanisms, including chronic systemic inflammation, altered endometrial receptivity, and hyperandrogenemia. Furthermore, elevated luteinizing hormone and disrupted insulin signaling impair the peri-implantation microenvironment. Consequently, early ultrasound confirmation does not eliminate clinician anxiety regarding ongoing fetal viability. Historically, reproductive specialists relied on broad demographic parameters to estimate prognosis. However, individualized post-confirmation risk stratification has remained challenging in daily clinical practice. Clinicians require reliable prognostic tools to guide patient counseling and tailor post-conception care. Therefore, researchers developed machine learning architectures to synthesize multi-parametric clinical data into precise prognostic insights.

Multicenter Machine Learning Model Development and Algorithm Selection

To address this dilemma, investigators launched a multicenter study analyzing 8,911 clinical pregnancies following fresh in vitro fertilization. Specifically, the researchers partitioned cycles across a development cohort of 7,241 cycles, an external geographic cohort of 1,170 cycles, and a prospective temporal cohort of 500 cycles. Observed miscarriage rates were 13.4%, 12.5%, and 12.4% across the three cohorts, demonstrating consistent baseline outcomes. Furthermore, the investigators prespecified predictors without missing data, utilizing training-fold median imputation strictly as a leakage-safe fallback. Within the primary development cohort, the authors evaluated 13 distinct machine learning algorithms alongside conventional logistic regression. They applied repeated stratified 5-fold cross-validation with five repeats to guarantee unbiased model comparisons. Importantly, the team executed all data preprocessing and hyperparameter tuning exclusively inside isolated training folds. Among candidate algorithms, Extreme Gradient Boosting, or XGBoost, achieved superior discriminative accuracy with an area under the curve of 0.885. Similarly, LightGBM showed competitive performance with an area under the curve of 0.882. Because the paired DeLong test confirmed comparable performance, researchers selected and prelocked the XGBoost model for downstream external validation.

External Validation Findings and Cross-Cohort Performance Attenuation

Clinical adoption requires that predictive algorithms maintain diagnostic integrity across diverse patient settings. Therefore, the investigators tested the locked XGBoost model on external cohorts without modifying underlying algorithm parameters. In geographic external validation, the model attained an area under the curve of 0.787 and a Brier score of 0.099. Similarly, prospective temporal evaluation demonstrated an area under the curve of 0.763 with a Brier score of 0.104. These results confirm meaningful discriminative utility, yet they illustrate noticeable performance attenuation compared to internal derivation results. Furthermore, unadjusted raw XGBoost probabilities substantially overestimated absolute risk within the development population. Specifically, the algorithm generated a mean predicted probability of 30.3%, whereas the true observed miscarriage rate was only 13.4%. Logistic recalibration effectively resolved this probability skew within the development cohort. However, the study team deliberately avoided applying recalibration adjustments to external test sets. When assessing risk-quartile stratification using fixed development cutoffs, the model successfully enriched events across validation cohorts. Nevertheless, positive predictive values remained modest, proving that algorithmic classifications cannot serve as definitive diagnostic endpoints.

Key Biological Predictors Identified Through Interpretability Analyses

Clinicians frequently express skepticism regarding predictive models because complex algorithms function as opaque decision matrices. To overcome this challenge, the investigators leveraged Shapley additive explanations to reveal internal predictor interactions. Notably, the analysis identified serum anti-Müllerian hormone concentration, endometrial thickness on the day of trigger administration, and basal luteinizing hormone as principal contributors. In addition, maternal age and baseline metabolic indices added valuable predictive weight. These statistical findings closely mirror established gynecological pathophysiology. For example, excessive anti-Müllerian hormone levels reflect profound follicular arrest and severe metabolic dysregulation in polycystic ovary syndrome. Meanwhile, diminished endometrial thickness on trigger day indicates sub-optimal uterine perfusion and impaired endometrial receptivity. Furthermore, elevated basal luteinizing hormone disrupts oocyte nuclear maturation and impairs early embryonic competence. Thus, the machine learning model successfully captures multi-factorial reproductive mechanisms rather than arbitrary numerical patterns. By combining baseline endocrine biomarkers with sonographic endometrial features, the algorithm provides actionable physiological context. Consequently, reproductive specialists can interpret individual patient risk scores through recognized clinical biology rather than relying on uninterpretable algorithmic calculations.

Clinical Implementation Considerations and the Imperative for Recalibration

The potential integration of artificial intelligence into reproductive medicine offers powerful opportunities for clinical workflows. However, translating machine learning algorithms into routine assisted reproduction requires extreme caution and clinical oversight. Because the locked model exhibited performance attenuation and calibration drift in independent validation, fertility specialists cannot utilize it as an autonomous diagnostic device. Instead, providers should employ the algorithm strictly as a risk-ranking instrument to structure empathetic patient consultations. For example, identifying individuals in higher risk quartiles allows clinicians to deliver targeted counseling, schedule timely early sonograms, and reinforce treatment adherence. Moreover, healthcare centers must mandate local algorithm recalibration before integrating predictive software into everyday clinical practice. Patient populations across different geographic territories, including clinics across India, exhibit diverse demographic profiles, body mass distributions, and laboratory testing platforms. Therefore, deploying uncalibrated models risks generating misleading probability estimates that could cause unnecessary psychological distress. Moving forward, prospective implementation studies must evaluate whether algorithm-guided monitoring protocols truly enhance reproductive outcomes and improve patient satisfaction.

Why is miscarriage risk elevated in women with PCOS after fresh embryo transfer?

Women with polycystic ovary syndrome experience elevated miscarriage risk due to complex endocrine and metabolic disturbances. Specifically, chronic hyperandrogenemia, insulin resistance, and abnormal luteinizing hormone surges alter oocyte competence and disrupt early endometrial receptivity. Furthermore, supra-physiological estradiol levels during fresh ovarian stimulation impair decidual vascular remodeling. Consequently, these combined physiological factors weaken the delicate peri-implantation microenvironment, thereby predisposing affected patients to early pregnancy loss before completion of the first trimester.

Which clinical predictors drove the machine learning model predictions?

Interpretability analysis using Shapley additive explanations revealed that serum anti-Müllerian hormone concentration, endometrial thickness on trigger administration day, and basal luteinizing hormone served as the three leading predictive drivers. In addition, maternal age and baseline metabolic indices added valuable predictive weight. Together, these routine clinical measurements capture ovarian reserve magnitude, neuroendocrine disruption, and uterine receptivity. By combining these variables, the model accurately mirrors the underlying biological dynamics that influence clinical pregnancy maintenance.

Can the XGBoost model guide clinical treatment or luteal support modifications?

The locked XGBoost model should not dictate therapeutic interventions or guide luteal support modifications because it functions strictly as a risk-ranking instrument. Furthermore, external validation demonstrated noticeable calibration drift, with raw probabilities significantly overestimating actual miscarriage incidence. Consequently, fertility specialists should utilize algorithmic outputs exclusively for personalized risk counseling and scheduling closer early pregnancy monitoring, while adhering strictly to validated reproductive endocrinology guidelines for clinical medical management.

Disclaimer: This content is for informational and educational purposes only. It is not intended to provide medical advice or to be a substitute for professional medical assessment, diagnosis, or treatment. Always seek the advice of a qualified healthcare professional with any questions you may have regarding a medical condition or treatment options. Refer to the latest local and national guidelines for clinical practice.

References

1. Tong J et al. Miscarriage Risk Prediction for Women With Polycystic Ovary Syndrome Who Achieve Clinical Pregnancy After Fresh Embryo Transfer: Multicenter Development, Geographic External Validation, and Prospective Temporal Validation Study. J Med Internet Res. 2026 Nov 05. doi: 10.2196/101829. PMID: 42987309.


2. Teede HJ, Tay CT, Laven JSE, et al. Recommendations from the 2023 international evidence-based guideline for the assessment and management of polycystic ovary syndrome. Eur J Endocrinol. 2023;189(2):G43-G64. doi: 10.1093/ejendo/lvad096.


3. Chi H, Wang S, Zou Y, et al. Polycystic ovary syndrome and miscarriage risk: dual evidence from cohort study and Mendelian randomization. Front Endocrinol. 2026;17:1854797. doi: 10.3389/fendo.2026.1854797.