Limits of Sociodemographic Data for Predicting Postpartum Readmission: Conventional, Ensemble and Automated Machine Learning Compared at Equal Workload, With a Decision-Economic Analysis

A tuned logistic regression using nuMoM2b first-pregnancy sociodemographic data achieved AUC 0.614 (95% CI 0.542–0.685) for 14-day postpartum readmission, while automated models performed worse.

• Why it matters: Postpartum readmission is rare and costly, yet risk tools often depend on routinely collected sociodemographic variables whose predictive value and economic payoff are unclear.
• What they did: Using 8,774 nuMoM2b participants (154 readmissions), conventional, ensemble, and automated ML models were trained on 70% and evaluated once on 30% (46 readmissions), with equal screening workload and post hoc recalibration.
• The result: After recalibration, no model improved beyond the base rate; adding delivery/discharge information yielded AUC 0.611–0.631, and decision-economic analysis showed contacting everyone is optimal at low outreach cost, with no model gaining more than running cost near break-even.

AI summary of the abstract · may contain errors · not medical advice · read the paper.