(1181) External validation of predictive models for postpartum readmission for hypertension management

Objective:
Prediction models for the risk of postpartum readmission for hypertension management (PRHM) have been developed but require external validation. We externally validated and compared the performance of two published risk prediction models by Venkatesh et al. (Model #1) and McLaren et al. (Model #2) in a large cohort.

Study Design:
Using a large observational cohort of all deliveries at a New York City health system between 1/2020-6/2024, we externally validated Model #1 and Model #2. We defined PRHM as a readmission within 42 days of delivery in which rapid acting antihypertensive medications (i.e. IV labetalol, IV hydralazine, or immediate release nifedipine) were administered. Both models include maternal age, delivery encounter blood pressure data points, and preeclampsia diagnosis; Model #1 further includes parity, birthweight, and delivery mode (race neutral) while Model #2 includes race/ethnicity (race-conscious) beyond the initial set. We evaluated each model through recalibration in the large, logistic recalibration, and model estimate revision. Model discrimination was assessed with the c-statistic and calibration with the Brier statistic and E-average.

Results:
Of 41,216 deliveries, 535 (1.3%) resulted in PRHM. Those readmitted were more often older, had higher diastolic blood pressure prior to discharge, had a diagnosis of severe preeclampsia, and delivered by cesarean section (Table 1). We found moderate performance of both models for predicting PRHM, with c-statistics ranging from 0.75-0.77 (Table 2) and acceptable calibration when the prediction model was allowed to estimate a model intercept slope for our cohort.

Conclusion:
In this large, diverse cohort, two published models for predicting PRHM demonstrated moderate discrimination and calibration. Performance improved with model updating, underscoring the need for tailoring prediction tools to the target population before clinical implementation. In line with other models and calculators that have excluded race without degrading performance, the race-neutral model performed well.