Machine Learning Algorithm Enhances Detection of Critical Congenital Heart Disease: Study
Researchers have found that incorporating machine learning (ML) algorithms into pulse oximetry screenings significantly improves the detection of critical congenital heart disease (CCHD) in newborns. This innovative approach addresses the limitations of traditional oxygen saturation (Spo2) screening, particularly in identifying conditions like coarctation of the aorta (CoA). This study was published in the Journal of American Heart Association by Heather S. and colleagues.
Traditional Spo2 screening has not been effective in early detection of CCHD, often missing crucial diagnoses such as CoA. To enhance detection rates, researchers developed an ML pulse oximetry algorithm that integrates perfusion data and radiofemoral pulse delay. This study aimed to evaluate the effectiveness of this advanced screening method.
The study enrolled 523 newborns from six different sites, including those with and without CCHD. Researchers recorded simultaneous pre- and postductal pulse oximetry readings. The ML algorithms were tested under various conditions: with one versus two time points and with or without pulse delay data. The sensitivity, specificity, and area under the receiver operating characteristic (ROC) curve were compared between the traditional Spo2-alone method and the new ML algorithms.
• Out of the 523 newborns, 317 had no congenital heart disease (CHD), 74 had CHD, and 132 had CCHD, including 21 cases of isolated CoA.
• The Spo2-alone method missed 26.2% of CCHD cases. When focusing on patients with both two time-point measurements and pulse-delay data (65 without CHD, 14 with CCHD), the ML algorithm significantly improved detection rates.
• The ML model incorporating two time points and pulse delay data achieved a sensitivity of 92.86% for CCHD detection, compared to 71.43% with Spo2 alone.
• For CoA, the sensitivity improved from 0% to 66.67%.
• All ML models maintained 100% specificity.
• The area under the ROC curve for CCHD detection increased from 0.83 to 0.96, and for CoA detection from 0.48 to 0.83, both showing significant improvements (P=0.03).
The study highlights the substantial benefits of using an ML algorithm that incorporates oxygenation, perfusion data, and pulse delay at two time points for detecting CCHD and CoA in newborns. The significant improvement in sensitivity and specificity indicates that this approach can greatly enhance early diagnosis and intervention, potentially saving lives and reducing long-term health complications associated with delayed diagnosis.
In conclusion, ML-enhanced pulse oximetry that combines multiple data points and time measurements significantly improves the detection of CCHD and CoA in newborns within the first 48 hours of life. This advanced screening method holds promise for better early diagnosis and improved outcomes for infants with these critical conditions.
Reference:
Siefkes, H., Oliveira, L. C., Koppel, R., Hogan, W., Garg, M., Manalo, E., Cresalia, N., Lai, Z., Tancredi, D., Lakshminrusimha, S., & Chuah, C.-N. (2024). Machine learning–based critical congenital heart disease screening using dual‐site pulse oximetry measurements. Journal of the American Heart Association, 13(12). https://doi.org/10.1161/jaha.123.033786
