Electrocardiographic artificial intelligence model for timely detection of preeclampsia

Abstract Background/Introduction Preeclampsia (PE) is a major concern for maternal and fetal health, affecting about 5%-8% of women worldwide. Assessing PE early remains an obstetric challenge. PE increases risk of cardiovascular diseases such as heart failure and ischemic and hypertensive heart disease. This is more important when the disorder persists or newly occurs within the postpartum period (postpartum preeclampsia; PPPE). Timely detection and intervention from low cost and accessible data modalities is the key to reduce the burden of preeclampsia. Purpose The aim of this research was to develop ECG-based models that can predict risk of PE and PPPE before onset by using raw 12-lead ECGs, and a comparative model that includes simple demographic features. Methods 10s 12-lead digital supine ECG and demographic data was obtained from the electronic health records at a university health science center. Case data was extracted using ICD9/10 codes. Controls were matched by age and race. The data was split 80%-20% with the 20% retained as holdout data used in any way during training, while five-fold cross validation was performed on the remaining 80%. The ECG-AI model, a modified ResNet convolutional neural network (CNN) with 1D ECG signals as inputs, was used to predict incidence of PE/PPPE. The ECG-AI output, together with age, race and gestational age at time of delivery were inputted into a Cox Proportional Hazard Model (ECG-AI-Cox). Models were evaluated using the area under the receiver operator characteristics curve (ROCAUC). We also assessed time-dependent AUC at time intervals at 10 days before PE onset and 7 days into PPPE. Results A total of 316 (Controls=190, Cases=126) women and a total of 374 ECGs were included in the study. The patient cohort had an average age of 27.76±4.76 of which 71% were African American, 24% were of White race. The mean gestational age at time of delivery was 32.76±4.45 weeks. From the 126 cases, 114 were diagnosed with PE while 12 developed de novo PPPE from 24 hours to a maximum of 14 days after delivery. AUC(95%CI) for ECG-AI was 81.44(0.71-0.92) with Sensitivity/Specificity=0.86/074 and time-dependent AUC>0.85 for ECG-AI-Cox (Table 1). Conclusion Results from this research show that PE and PPPE can be predicted at moderate accuracy at least 10 days before childbirth and PE onset while PPPE can be predicted with high accuracy up to 1 week after delivery. The ECG-AI and ECG-AI-Cox models are easy to implement within clinical workflows due to their simple data requirement. There is also potential for the models’ integration within smart devices to help early detection for women that might not have access to healthcare.