(CCSP026) MACHINE LEARNING ALGORITHMS TO PREDICT OUTCOME OF FETAL CARDIAC DISEASE

Background: Machine learning algorithms are evolving to be used by healthcare providers to reduce clinical uncertainty, improve prognostic accuracy, and ideally improve patient outcomes. This is particularly relevant in the field of fetal medicine, where the probability of adverse outcomes may be difficult to predict based on a fetal echocardiogram.

METHODS AND RESULTS: We utilized a machine learning algorithm to facilitate prediction of outcomes, based on retrospective data of fetal cardiac disease and known outcomes. Specific objectives were to predict the following outcomes: 1) in utero demise/stillbirth, 2) high-acuity care needed postnatally (intensive care admission, ventilation, prostaglandin or cardiac intervention < 30 days of age) and 3) favourable medium-term prognosis (medium-term survival with developmental delay no greater than mild severity). Retrospective study including all fetuses with cardiac disease evaluated with a fetal echocardiogram at Sunnybrook Health Sciences Centre, Toronto, Canada between January 2010 and December 2021. Outcomes were acquired from 3 other Toronto Hospitals (Sick Kids, Mount Sinai and Michael Garron). Cases with termination of pregnancy were excluded. Prediction models were created using the XgBoost algorithm (tree-based) with 5-fold cross validation.

Among 211 cases of fetal cardiac disease, 61 were excluded (39 terminations, 21 lost to follow up, 1 arrhythmia), leaving a cohort of 150 fetuses. Of those, 15 (10%) had fetal demise/stillbirth and 70 (52%) of live births resulted in the need for high acuity neonatal care. Of those patients with medium-term follow-up, 57/82 (70%) had a favourable prognosis. At the diagnostic fetal echocardiogram (mean gestational age of 24±6 weeks): 63 (42%) had minor cardiac abnormalities, 46 (31%) had major abnormalities and 41 (37%) had not yet developed cardiac abnormalities. Prediction models for live birth, high acuity neonatal care and positive prognosis had AUCs of 0.74, 0.80 and 0.67 respectively with a good balance between sensitivity and specificity. Most important features for the livebirth prediction models were: presence of non-cardiac or genetic abnormalities and more severe structural heart disease. High acuity of postnatal care was predicted by nuchal thickness, gestational age and weight at birth. Medium-term prognosis was predicted by abnormal fetal right ventricular function, tricuspid valve abnormalities, and gestational age and weight at birth.

Conclusion: Prediction models using machine learning provide reasonably good discrimination of key prenatal and postnatal outcomes among fetuses with congenital heart disease.