Heterogeneous Effects of Cesarean Delivery Across Maternal and Infant Outcomes Using Population-Based Causal Machine Learning
Faculty Mentor
Lili Yu
Location
Russell Union Ballroom
Type of Research
On-going
Session Format
Poster Presentation
College
Jiann-Ping Hsu College of Public Health
Department
Biostatistics, Epidemiology and Environmm
Abstract
Title: Heterogeneous Effects of Cesarean Delivery Across Maternal and Infant Outcomes Using Population-Based Causal Machine Learning
Background: Cesarean delivery rates remain high and unevenly distributed, yet most studies estimate average effects that may mask clinically meaningful heterogeneity. We examined individualized causal effects of cesarean versus vaginal delivery across multiple maternal and infant outcomes.
Methods: We analyzed Pregnancy Risk Assessment Monitoring System (PRAMS) data from 197,942 mothers. High-dimensional covariates were encoded and reduced using a double-selection framework combining LASSO logistic regression and gradient-boosted variable importance screening to identify predictors of treatment assignment and outcomes while excluding post-treatment variables. Propensity scores were estimated using gradient-boosted trees and applied via inverse probability of treatment weighting (IPTW). Covariate balance improved substantially after weighting (99.39% of covariates with |SMD| < 0.10). Heterogeneous treatment effects were estimated using an X-learner framework with gradient-boosted outcome models, yielding individualized probability differences for each outcome.
Results: Propensity score performance in the test set demonstrated adequate discrimination (log-loss 0.4638; Brier score 0.1516; sensitivity 0.565; specificity 0.873). IPTW weights were stable (median 1.362; 99th percentile 9.083). Substantial heterogeneity was observed, particularly for abnormal labor, where predicted individual effects ranged from marked reductions to substantial increases in risk. Rare outcomes showed smaller but structured variability. Clustered treatment-effect profiles revealed multidimensional trade-offs rather than uniformly beneficial or harmful patterns. Key drivers of heterogeneity included gestational age, maternal hypertension, prenatal care utilization, smoking, and state-level indicators.
Conclusions: Cesarean delivery effects vary meaningfully across individuals. Individualized probability differences may inform shared decision making and support development of clinician decision support tools in obstetric care.
Program Description
.
Start Date
4-23-2026 2:00 PM
End Date
4-23-2026 4:00 PM
Recommended Citation
Azu, Emmanuel Unimke; Oloyede, Tobi F.; Asifat, Olamide; Kizza, Timothy; Shah, Gulza; and Yu, Lili, "Heterogeneous Effects of Cesarean Delivery Across Maternal and Infant Outcomes Using Population-Based Causal Machine Learning" (2026). GS4 Student Scholars Symposium. 200.
https://digitalcommons.georgiasouthern.edu/research_symposium/2026/2026/200
Heterogeneous Effects of Cesarean Delivery Across Maternal and Infant Outcomes Using Population-Based Causal Machine Learning
Russell Union Ballroom
Title: Heterogeneous Effects of Cesarean Delivery Across Maternal and Infant Outcomes Using Population-Based Causal Machine Learning
Background: Cesarean delivery rates remain high and unevenly distributed, yet most studies estimate average effects that may mask clinically meaningful heterogeneity. We examined individualized causal effects of cesarean versus vaginal delivery across multiple maternal and infant outcomes.
Methods: We analyzed Pregnancy Risk Assessment Monitoring System (PRAMS) data from 197,942 mothers. High-dimensional covariates were encoded and reduced using a double-selection framework combining LASSO logistic regression and gradient-boosted variable importance screening to identify predictors of treatment assignment and outcomes while excluding post-treatment variables. Propensity scores were estimated using gradient-boosted trees and applied via inverse probability of treatment weighting (IPTW). Covariate balance improved substantially after weighting (99.39% of covariates with |SMD| < 0.10). Heterogeneous treatment effects were estimated using an X-learner framework with gradient-boosted outcome models, yielding individualized probability differences for each outcome.
Results: Propensity score performance in the test set demonstrated adequate discrimination (log-loss 0.4638; Brier score 0.1516; sensitivity 0.565; specificity 0.873). IPTW weights were stable (median 1.362; 99th percentile 9.083). Substantial heterogeneity was observed, particularly for abnormal labor, where predicted individual effects ranged from marked reductions to substantial increases in risk. Rare outcomes showed smaller but structured variability. Clustered treatment-effect profiles revealed multidimensional trade-offs rather than uniformly beneficial or harmful patterns. Key drivers of heterogeneity included gestational age, maternal hypertension, prenatal care utilization, smoking, and state-level indicators.
Conclusions: Cesarean delivery effects vary meaningfully across individuals. Individualized probability differences may inform shared decision making and support development of clinician decision support tools in obstetric care.
