Heart failure with preserved ejection fraction remains one of cardiology's most vexing diagnostic challenges — a condition affecting half of all heart failure patients yet requiring expensive, sometimes invasive workups to confirm. The prospect of a standard 12-lead ECG, reinterpreted by machine learning, serving as a frontline screening tool could fundamentally shift how this condition is identified at scale, particularly in resource-limited or primary care settings.
This PRISMA-registered systematic review and meta-analysis synthesized ten studies published between 2021 and 2025, collectively enrolling more than 270,000 participants across varied clinical populations. Seven studies supplied sufficient data for statistical pooling across 11 independent cohorts. Using a logit-transformed random-effects model, the pooled area under the receiver operating characteristic curve (AUROC) reached 0.84 (95% CI: 0.78–0.88), a threshold generally considered indicative of good discriminatory performance for a binary diagnostic tool. The analysis applied QUADAS-AI for bias assessment, a methodological refinement specifically designed for AI-based diagnostic studies.
The result is notable, but the critical caveat is a measured one: heterogeneity across included studies was extreme, with I² = 100%, meaning the performance estimates varied so substantially between cohorts that the pooled figure must be interpreted cautiously rather than as a reliable universal benchmark. This degree of heterogeneity likely reflects meaningful differences in patient populations, reference standards used (echocardiographic versus invasive hemodynamic), AI model architectures, and ECG acquisition protocols. HFpEF itself is a heterogeneous syndrome with multiple phenotypes, which compounds the challenge of building a single high-performing classifier.
For the broader research landscape, this meta-analysis represents an important early signal rather than a green light for clinical deployment. AI-ECG tools have already demonstrated strong performance in detecting reduced ejection fraction, atrial fibrillation, and hypertrophic cardiomyopathy; extending that capability to the diastolic dysfunction spectrum is the logical next frontier. Prospective validation in unselected, real-world primary care cohorts — rather than retrospective hospital datasets — will be essential before these tools can meaningfully change screening pathways.