AI-Enhanced ECG Analysis Shows Strong Accuracy in Detecting Hypertrophic Cardiomyopathy

Use of artificial intelligence-enhanced 12-lead echocardiography (AI-ECG) was associated with a high degree of accuracy in detection of hypertrophic cardiomyopathy (HCM) in real-world clinical use, enabling identification of HCM in 5% of patients not previously diagnosed with HCM. So finds a prospective observational study reported by Cleveland Clinic researchers at the American College of Cardiology Scientific Session and published online in JACC Clinical Electrophysiology.

“The findings of our real-world study suggest that clinical application of AI-ECG algorithms offers incremental diagnostic value in evaluating patients who might have HCM, enabling earlier disease detection and intervention, which improves the likelihood of successful treatment outcomes,” says principal investigator Milind Desai, MD, MBA, Director of Cleveland Clinic’s Hypertrophic Cardiomyopathy Center. “The algorithm we used performed well across patient subgroups differing by age, sex and whether or not patients had prior septal myectomy.”

The need for enhanced HCM diagnosis

HCM usually is diagnosed based on some combination of clinical evaluation, cardiac imaging methods such as echocardiography and cardiac magnetic resonance imaging (CMR), and genetic testing. Yet HCM patients can have no symptoms or nonspecific symptoms for considerable periods, and evidence indicates that nearly 85% of HCM cases are undiagnosed, underdiagnosed or misdiagnosed.

“In view of the considerable clinical implications of a delayed diagnosis or misdiagnosis of HCM for both patients and their at-risk family members, early and accurate diagnosis is a top priority,” Dr. Desai says.

More than 90% of patients with HCM have abnormal ECG findings, with the most common abnormalities being left ventricular hypertrophy, Q waves and repolarization abnormalities. Because these ECG findings are nonspecific and overlap with those of many other conditions, ECG screening alone has limited utility for detecting HCM, despite its low expense, noninvasive nature and wide availability.

To overcome the limits of standard ECG-based screening, convolutional neural networks have been used in recent years to develop AI-powered ECG algorithms for detecting HCM from 12-lead ECGs. Cleveland Clinic implemented one such algorithm (from Viz.ai, San Francisco) in February 2024 to assist in the diagnosis of HCM. “We undertook our study to evaluate the real-world clinical application of this algorithm in a large, undifferentiated patient population at our tertiary care center,” Dr. Desai notes.

Study design and key findings

Following implementation of the AI-ECG algorithm for routine clinical use, all ECGs performed in adult patients evaluated at Cleveland Clinic’s Main Campus underwent analysis by the algorithm in addition to standard ECG interpretation. The study included all first ECGs for the 45,873 consecutive adults evaluated during the first 8.5 months of the algorithm’s use. Other than exclusion of patients with pacemakers, there were no restrictions for inclusion by the presence or absence of a cardiovascular diagnosis.

The investigators assessed the AI-ECG algorithm’s diagnostic performance against clinical diagnoses using three different HCM probability thresholds: ≥0.95, ≥0.90 and ≥0.85 (within a range from 0 to 1).

Patients were divided into the following groups:

• Those with a confirmed prior diagnosis of HCM
• Those with no formal HCM diagnosis in whom likely HCM was flagged by the AI-ECG algorithm (at or above the probability threshold of 0.85) and then confirmed by clinical assessment (including advanced imaging)
• Those with a prior confirmed alternate diagnosis without clinical or imaging evidence of HCM

Clinical diagnoses were made by experienced cardiologists, whether for patients with a diagnosis of HCM (in both each of the first two groups above) or those with an alternative diagnosis excluding HCM.

Among the first ECGs analyzed for the study’s 45,873 patients, 1,265 (2.76%) were flagged by the algorithm for suspected HCM using the ≥0.85 probability threshold (i.e., the most inclusive threshold). The 1,265 patients with these flagged ECGs broke down as follows:

• 511 (40.4%) had previously documented clinical HCM
• 63 (5.0%) had a new HCM diagnosis prompted by AI-ECG detection
• 691 (54.6%) had no clinical evidence of HCM and had an alternate diagnosis

As expected, the algorithm’s accuracy varied by the chosen HCM probability threshold. Key observations included the following:

• The ≥0.95 threshold offered the lowest sensitivity for HCM (50%), which improved to 80% at the ≥0.90 threshold and to 100% (based on the study design) at the ≥0.85 threshold.
• The algorithm had a specificity of 99.6% at the ≥0.95 threshold, 99.1% at the ≥0.90 threshold and 98.4% at the ≥0.85 threshold.
• The algorithm’s accuracy and negative predictive value were greater than 98% across all three probability thresholds.

The algorithm performed similarly well in men and women, and it was more sensitive in patients under age 50 but more specific in those 50 or older. Among patients with previously diagnosed HCM, the algorithm performed similarly across all probability thresholds in patients regardless of whether or not they had undergone myectomy and whether or not they had received mavacamten therapy.

The 691 false-positive patients flagged for HCM had various alternate diagnoses, including aortic stenosis and hypertension. This suggests that concomitant secondary left ventricular hypertrophy can potentially confound the AI-ECG algorithm’s diagnosis of HCM, the researchers note.

A promising addition to HCM screening tools

“These findings suggest that AI-ECG can potentially offer a less costly and faster alternative to traditional screening tools like echocardiography and CMR in the initial evaluation for HCM,” Dr. Desai notes. He adds that the algorithm’s ability to flag 63 cases of undiagnosed HCM during this short study period — 5% of the total ECGs flagged — underscores its potential to improve patient outcomes through earlier HCM detection and treatment.

For example, Dr. Desai and colleagues recently reported the case of a patient who was diagnosed with obstructive HCM through the use of AI-ECG who has since been initiated on therapy with the novel agent mavacamten and has shown excellent clinical response (JACC Case Rep. 2025 Epub 26 Feb).

In their study report, the authors note that the ideal probability threshold to use when employing an AI-ECG algorithm may vary based on specific clinical context. For instance, a threshold of ≥0.95 has a high positive predictive value, suggesting particular utility when screening high-risk patients, such as those with a family history of HCM. In contrast, a threshold of ≥0.85-0.90 offers a low risk of missing actual HCM cases, but at the expense of excess false-positives. Such a threshold could be useful when screening a lower-risk population like young athletes, in whom HCM is the most common cause of sudden cardiac death.

“The symptoms of HCM can be nonspecific, and often other diagnoses — both cardiac and noncardiac — are entertained,” observes study co-author Maran Thamilarasan, MD, staff cardiologist in Cleveland Clinic’s Section of Cardiovascular Imaging. “If AI-enhanced ECG can alert to the possibility of this diagnosis, as this study suggests is the case, this can lead to earlier appropriate diagnostic testing and treatment.”

Dr. Desai is a consultant to and has research agreements with Viz.ai as well as several pharmaceutical companies. No external funding or industry support was used in the conduct of the study reported here, which was fully funded by philanthropic support for Dr. Desai’s research. Dr. Thamilarasan reported no potential conflicts of interest.

Image at top is reprinted from Desai et al., JACC: Case Reports (2025 Epub 26 Feb), under the Creative Commons CC-by-NC-ND license.