Deep learning on chest radiographs detects reduced ejection fraction, may cut echo volume

A deep learning model detected LVEF ≤40% on chest X-rays with AUC 0.92-0.97 in checkup cohorts, enriching baseline prevalence 37-87-fold, and could reduce echocardiography volume up to 85% without missing cases.

• In health checkup cohorts, AUCs for LVEF ≤40% were 0.92–0.97 with low AUPRCs (0.04) but 37–87-fold enrichment over baseline prevalence.
• Counterfactual reinforcement learning improved heatmap localization precision (P<.001) and AUC for LVEF <50% (P=.04).
• Model-based prescreening estimated TTE volume reductions up to 85% while identifying all 12 and 7 individuals with LVEF ≤40% in the two external checkup cohorts.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.