Presentation (CTI)

OBJECTIVES: Large variance in pregnancy costs driven by NICU and preterm birth can significantly impact healthcare budgets. Dynamically predicting pregnancy-related costs before delivery can help value-based medical practices intervene at the right time before costs rise. This study aimed to develop and assess the performance of a predictive model estimating remaining total cost of care for in-progress maternity episodes using data from a 24/7 virtual maternity care program.
METHODS: This retrospective study utilized medical claims, patient-reported data, patient messages, EHR data, and program-specific completion data across 822 patients with completed pregnancies. A Gradient Boosted Trees (GBT) model was developed using lagged, rolled-up, and static features depending on the input variable. Data were aggregated into chunked groups in weekly segments to predict going-forward cost for the remainder of the pregnancy at each given gestational week. Singular value decomposition (SVD) was used to reduce dimensionality of diagnosis code features. SHAP values were computed to interpret feature importance in the GBT. Coefficient of determination (R^2) was calculated for log(costs) of the episode overall and for the episode each week of gestational age relative to remaining episode costs.
RESULTS: The model achieved an overall R^2 of 0.70 for logged costs, with variability in performance depending on the week of maternal gestational age. Earlier weeks in pregnancy, such as those in the first trimester, showed an R^2 range of 0.23-0.29. Later weeks had higher performance, with R^2 ranging from 0.30-0.42. Average SHAP values decreased substantially around the time of delivery for patients, reflecting the reduced uncertainty in future costs since the majority of episode costs were incurred by that point.
CONCLUSIONS: The predictive model demonstrated robust performance in estimating pregnancy-related costs, which can guide resource allocation and patient interventions. Future research should explore additional optimization techniques and consider real-time applications to enhance perinatal healthcare efficiency and outcomes.