AI Detects Heart Failure From an ECG With AUC Up to 0.96

The 10 second test that keeps missing heart failure

Margaret is 68. She gets breathless on a hill, sleeps on two pillows, and her ECG looks “normal.” Six months later she is in the ED with fluid filled lungs. Her echo shows HFpEF, the stiff heart type that makes up half of all heart failure and is missed most often.

What if her first ECG already held the answer?

I broke down a new preprint from Norway where researchers trained an open source AI on 284,000 ECGs using a clever fix they call pragmatic labelling. Instead of trusting noisy ICD codes alone, they paired codes with NT-proBNP. The result is a model that reads raw 12 lead voltage and spots heart failure across the full EF spectrum.

In prospective testing on 43,109 patients it hit AUC 0.84 overall, 0.91 for HFrEF, and up to 0.96 with strict labelling. It even outperformed NT-proBNP head to head, and flagged HFpEF in patients with normal biomarkers.

No new hardware. Just better eyes on the ECG you already order.

I wrote a simple walkthrough of how it works, where it fits in primary care, and what it gets wrong.

Read the full breakdown here:
https://sharetxt.live/blog/heart-failure-detection-in-ecg-using-ai