(881) A personalized risk stratification tool for perinatal morbidity and mortality using explainable artificial intelligence (AI)

Late preterm delivery is frequently considered when competing maternal risks begin to outweigh perinatal risks, but tools for precise, personalized perinatal risk estimation are lacking. Traditional regression methods do not capture the complex interactions specific to each possible combination of risk factors, which can be synergistic rather than simply additive. AI-based Probabilistic Graphical Models (PGMs) can quantify these synergies transparently. We aimed to develop an ‘explainable AI’ machinery for personalized perinatal morbidity risk estimation using variables from clinical and social determinants of health domains.  

Study Design:

Prospectively-collected longitudinal data from a cohort of 7,646 nulliparas were analyzed. The primary outcome was a composite of perinatal morbidity and mortality (Fig footnote). Data were randomly sectioned into training and testing sets (75-25 split). Chi-squared and mutual information feature selection were performed on the training data. The top 12 variables identified by both methods were selected as features for logistic regression (LR) and PGM models, which were used to plot receiver-operating characteristic (ROC) curves and estimate the relative risk (RR) of the outcome conferred by late preterm birth (PTB) in various scenarios.

Results:

LR and PGM models were trained to predict perinatal outcome (n=5,734) and tested on the remaining cohort (n=1,912). Predictive features included preeclampsia, metrics of fetal growth, uterine artery Doppler, maternal stress and education, and gestational age (GA) at delivery. The PGM (Fig) performed better than the LR model (AUC 0.72 vs 0.64). The PGM-estimated RR conferred by late PTB with other factors ranged from 1.3±0.4 to 6.5±1.1, demonstrating the variability in risk given different factor combinations (Table).

Conclusion:

PGMs outperform LR models for morbidity prediction, provide clinical interpretability of risk relationships, and identify synergies arising from specific variable combinations.