AI Model Spots Weak Heart Function on Chest X‑Rays, Cuts Need for Extra Scans

The study set out to build a deep learning tool that could flag a weakened heart on ordinary chest X‑rays. Researchers gathered pairs of X‑ray and echocardiogram scans taken on the same day between 2016 and 2021. They used a technique called counterfactual reinforcement learning to make the model easier to understand and to boost its accuracy. The final system was named DLLV and was trained on a large collection of 24,674 patients, which included 32,135 paired scans. The goal was to use DLLV as a first‑pass filter before ordering a more detailed echocardiogram in routine health checks.

When DLLV was tested on three separate groups of patients, it showed strong performance. In the main referral center, the model reached an AUC of 0.86, meaning it could correctly rank patients with low ejection fraction above those with normal function. In two health‑check cohorts, the AUC rose to between 0.92 and 0.97. Although precision‑recall scores stayed low at 0.04, the enrichment over the usual prevalence was dramatic‑87 times higher in one group and 37 times higher in the other. The improved heatmaps helped the model focus on the heart area, and the AUC for a broader cut‑off (LVEF < 50%) also increased (p = .04). Using DLLV as a prescreen could cut the number of full echocardiograms by as much as 85 % while still catching every patient with an LVEF at or below 40 % (12 cases in one set, 7 in another).

The findings suggest that DLLV could become a useful first step in large‑scale screening programs. By flagging likely cases, it would let clinics reserve detailed echo exams for those who truly need them. This could save time, lower costs, and keep the workflow smooth for busy health‑check centers. More real‑world testing will be needed before the tool is rolled out widely, but the early results are promising.

actions