What is it about?
Chagas disease is a parasite-borne illness, common in Latin America that can silently damage the heart over many years, but the blood tests used to diagnose it are expensive and often unavailable. We explored whether artificial intelligence could instead spot the disease from a simple, low-cost ECG recording. We tested a two-step training strategy, first teaching the AI on a large but loosely-labeled dataset, then fine-tuning it on smaller, more reliable data, and compared it to simply training the AI directly on the reliable data. Surprisingly, the simpler, direct approach worked better. Our results suggest that when AI is trained on heart recordings collected from different countries and time periods, it can end up learning to tell the datasets apart rather than the disease itself, a pitfall future researchers should watch for when combining medical data from different sources.
Featured Image
Photo by Towfiqu barbhuiya on Unsplash
Why is it important?
This work was conducted as part of the 2025 George B. Moody PhysioNet Challenge, a timely, actively-running international benchmark for AI-based Chagas disease detection from ECGs, a disease that affects millions worldwide yet remains chronically underdiagnosed due to costly serological testing. While combining large weakly-labeled datasets with small, high-confidence ones is a common strategy for boosting AI performance in medicine, our findings show this can backfire: when positive and negative examples come from different countries, eras, or recording setups, the model risks learning to distinguish datasets rather than disease. This is a practical, generalizable caution for researchers building AI screening tools from patchwork medical datasets, a scenario increasingly common as global health AI relies on pooling data across institutions and regions with unequal access to gold-standard diagnostics.
Perspectives
I keep coming back to these PhysioNet/CinC challenges because I genuinely enjoy the format, an open problem, a hidden test set, and the chance to see how different teams tackle the same messy real-world data. This one pulled me in further as Chagas disease is a largely neglected disease due to socioeconomic marginalization and an absence of early symptoms. I was curious whether clever pretraining tricks could squeeze more signal out of imperfect labels. It was a little frustrating to watch our more sophisticated approach lose to the simpler baseline, but this is one of the first times I've deliberately written up a negative result, but I think that's actuall the kind of finding the field needs more of. My hope is that it nudges other researchers to scrutinize their datasets a bit harder before trusting an AI model's results, and in a keeps Chagas disease on more people's radar.
Bjørn-Jostein Singstad
Akershus University Hospital
Read the Original
This page is a summary of: Auxiliary Pretraining and Fine-Tuning Across Heterogeneous Datasets for ECG-Based Chagas Disease Detection, December 2025, Computing in Cardiology,
DOI: 10.22489/cinc.2025.179.
You can read the full text:
Contributors
The following have contributed to this page
