Transfer Learning for ECG-Based Age Estimation from Adult to Pediatric Populations

Sara Battiston1, Niccolò Gonzato1, Md Moklesur Rahman2, Massimo W Rivolta2, Antonio Sanzo3, Irene Raso4, Sara Santacesaria5, Gianvincenzo Zuccotti6, Savina Mannarino5, Roberto Sassi2
1UniversitĂ  degli studi di Milano, 2Dipartimento di Informatica, UniversitĂ  degli Studi di Milano, 3Fondazione IRCCS Policlinico San Matteo, 4Pediatric Cardiology Unit, Buzzi Children's Hospital, Milan, Italy, 5Pediatric Cardiology Unit, Buzzi Children's Hospital, 6Department of Biomedical and Clinical Science, University of Milan


Abstract

Electrocardiogram (ECG)-based age prediction has emerged as a promising tool in medical AI, providing insights into physiological aging and potential health risks. While existing deep learning models have shown strong performance on adult populations using 10-second ECG recordings, their applicability to pediatric subjects remains largely unexplored. In this study, we investigated a transfer learning strategy for pediatric age estimation using ECGs, starting from a model pre-trained on adult data. Specifically, we first trained a convolutional neural network on single heartbeats from adult ECGs taken from the PTB-XL database, and then, we fine-tuned it on pediatric ECGs collected at the Buzzi Children Hospital, Milan, Italy. Our model achieved a RMSE of 10.32 years and a MAE of 8.03 years on adult data, which were found comparable to prior works trained on longer segments of ECG signals. In the pediatric dataset, the model achieved a RMSE of 2.67 years and a MAE of 1.88 years. These results suggest that meaningful age-related features can be extracted even from single heartbeats and that transfer learning enables effective adaptation across age groups, offering a practical solution for pediatric age estimation or in other contexts where available data might be typically more scarce.