Artificial intelligence for non-invasive hemodynamic profiling in pregnancy: Development of a novel scoring system

Hemodynamic profiling during pregnancy is critical for risk stratification and management of hypertensive disorders such as preeclampsia. Traditional assessment via echocardiography, while accurate, is costly and not universally accessible. This study aimed to develop and validate a simple, non-invasive, artificial intelligence-driven scoring system (AI-SCORE) that classifies maternal hemodynamic profiles using readily available clinical variables.

In this cross-sectional study, 90 low-risk pregnant women at a single tertiary center underwent standard echocardiographic evaluation as the reference. Mean arterial pressure (MAP) and cardiac output data extracted from echocardiography were subjected to principal component analysis followed by K means clustering to identify hemodynamic profiles. A general regression neural network was then trained on clinical predictors-trimester, age, weight, heart rate, and the MAP/heart rate index-to generate the AI-SCORE. Model performance was assessed by area under the receiver operating characteristic curve (AUC-ROC) for discriminating among profiles.

The median maternal age was 27 years (interquartile range [IQR] 24-31); 30 % of participants were classified as obese. Three hemodynamic profiles were identified: high-resistance (hypodynamic): 36.7 % of participants; normodynamic: 38.9 %; hyperdynamic: 24.4 %. Profiles were evenly distributed across trimesters. The AI-SCORE achieved AUC-ROC values of 0.706-0.811 for pairwise profile discrimination. Lower scores corresponded to the high-resistance/hypodynamic phenotype, while higher scores indicated a hyperdynamic state.

The AI-SCORE reliably identifies maternal hemodynamic phenotypes using non-invasive, low-cost clinical inputs. By enabling rapid bedside profiling, this tool has the potential to enhance screening, diagnosis, and tailored management of hypertensive disorders in pregnancy.