Clinical Research · global
Clinical Scoring Followed by AI Echocardiography Analysis Improves HFpEF Case Detection in Study
Among 280 patients referred for suspected HFpEF, adding AI imaging analysis in stages increased the proportion of confirmed cases identified from 48% to 78%. The findings support further validation of the diagnostic pathway, but do not yet establish that it would perform similarly in general outpatient settings.
A normal proportion of blood ejected by the heart does not mean that heart function is entirely normal. The challenge in diagnosing heart failure with preserved ejection fraction (HFpEF) lies in identifying patients whose condition can be obscured by a single measure. A study in the November issue of the *Journal of Cardiac Failure* suggests that clinical scoring followed by AI echocardiography analysis in selected patients may address gaps in existing diagnostic pathways.
The study included 280 outpatients referred for suspected HFpEF, of whom 222, approximately 79%, were determined by the study to have HFpEF. The researchers applied the EchoGo HF model to resting echocardiography videos. The model analyzed the apical four-chamber view, which displays all four heart chambers simultaneously, and was then compared with the H₂FPEF clinical score to assess whether imaging analysis could provide additional clues.
This validation focused on cardiac filling pressures, rather than imaging appearances alone. Patients underwent exercise hemodynamic assessment; the disease criteria listed in the abstract encompassed elevated resting or exercise filling pressures measured by right heart catheterization, as well as resting or exercise echocardiographic E/e′ measures. The results therefore need to be interpreted with recognition that the reference standard included different examination methods; not all cases should be understood as having been confirmed by invasive testing.
The finding with the greatest practical relevance came from a two-step strategy: first calculate the H₂FPEF score, then selectively use AI in patients with low or intermediate scores. This pathway identified 78% of confirmed cases, compared with 48% using the score alone, a difference of 30 percentage points. The 78% represents the proportion of patients with the disease who were identified, rather than overall diagnostic accuracy across everyone tested or the performance of AI used alone.
For EchoGo HF alone, the study reported a sensitivity of 67.3% and specificity of 68.5%, indicating that the model still missed cases and classified some patients without the disease as positive. In addition, 15 patients, approximately 5% of the total, received an indeterminate classification. These results underscore that AI outputs still need to be interpreted alongside clinical information and subsequent testing, and are not yet sufficient on their own to confirm or rule out the disease.
Combining AI with H₂FPEF increased the AUROC, a measure of the ability to distinguish patients with and without the disease, from 0.656 to 0.724, a statistically significant difference (P = 0.021). Improvements occurred mainly in patients with low and intermediate scores. The study also found that an AI classification of HFpEF was independently associated with higher pulmonary capillary wedge pressure and lower peak oxygen uptake during exercise. This suggests that the imaging signals identified by the model may be related to impaired physiological function, but these associations do not establish the mechanism underlying its interpretation.
These patients had already been referred for suspected disease, and nearly 80% had the condition, making them markedly different from general outpatient populations or asymptomatic screening populations. The study supports further validation of AI's value within a specific diagnostic pathway. It has not yet demonstrated that adoption can reduce invasive testing, enable earlier effective treatment, or improve prognosis; the paper also disclosed that some authors were employed by the model's developer, Ultromics. Prospective studies are needed next to examine how this pathway affects testing decisions and patient outcomes across different healthcare settings.