Moderate evidence
More
AI-read ECGs matched heart ultrasound at predicting cardiomyopathy in a study of 343 at-risk relatives
Study title: Artificial intelligence-enhanced electrocardiography for detection and prediction of dilated cardiomyopathy
- Evidence type
- Controlled or validated comparison
- Sample
- n = 343
- Sources
- 1 sourcePeer-reviewed
- Published
- 2 min read
- Truvace record
- TRV-2026-2934cite or check it
Study title: Artificial intelligence-enhanced electrocardiography for detection and prediction of dilated cardiomyopathy
Researchers trained a deep learning model on more than 330,000 ECG-echocardiogram pairs from Beth Israel Deaconess Medical Center in Boston to spot the echocardiographic signature of dilated cardiomyopathy, a heart muscle disease marked by an enlarged left ventricle that pumps poorly. It diagnosed the current phenotype with areas under the curve of .87 to .97 across outside cohorts and beat models built on traditional ECG features.
Read the study abstract
Background and aims A major challenge in dilated cardiomyopathy (DCM) care is the identification and management of individuals at risk before disease manifestation. An artificial intelligence (AI) model applied to electrocardiography (ECG) was developed to detect and predict the echocardiographic DCM phenotype (left ventricular dilatation with systolic dysfunction) and was evaluated for its ability to stratify the risk of incident DCM in familial cohorts. Methods A deep learning model was trained to detect the e…
The gain
In a multi-cohort study, an AI model reading ECGs diagnosed the dilated cardiomyopathy phenotype with AUCs of .87-.97 and predicted future disease about as well as echocardiography (C-index .80-.87 versus .80-.90).
In brief
The model also predicted who would develop the disease later, with accuracy comparable to echocardiography. In two groups of genetically predisposed relatives found through cascade screening, 5-year negative predictive values were 95.5% in Antwerp and 91.5% in Maastricht, so a low score largely ruled out near-term disease.
Positive predictive values were lower, at 42.4% and 39.4%. High-risk scores were linked to a 5.1 to 6.4-fold higher risk of incident DCM. The authors say the low-cost tool could reduce surveillance burden and allow earlier intervention.
Main points
- Training used more than 330,000 ECG-echocardiogram pairs from Beth Israel Deaconess Medical Center in Boston.
- External validation came from the UK Biobank, Zhongshan Hospital and Antwerp University Hospital.
- Diagnostic AUCs for the prevalent DCM phenotype ranged from .87 to .97, outperforming traditional ECG parameters.
- For future DCM, AI-ECG had a C-index of .80-.87, against .80-.90 for echocardiography.
- In the familial cohorts, Antwerp (n = 166) and Maastricht (n = 177) each had 18% DCM-positive individuals.
- Five-year negative predictive values were 95.5% in Antwerp and 91.5% in Maastricht.
- Positive predictive values were 42.4% and 39.4%, and high-risk scores carried a 5.1 to 6.4-fold greater risk of incident DCM.
The full story
The model was built to address a core problem in dilated cardiomyopathy care: finding and managing at-risk people before disease appears. It was trained to detect the echocardiographic phenotype (left ventricular dilatation with systolic dysfunction) and to predict its future onset from a standard ECG.
Performance was compared against models using traditional ECG features and echocardiographic traits. Across external cohorts, the AI-ECG model outperformed traditional ECG parameters for diagnosis and matched echocardiography for predicting incident disease.
In familial testing, genetically predisposed people were identified through cascade screening at cardiogenetics clinics in Antwerp (n = 166) and Maastricht (n = 177); 18% in each group had DCM. High negative predictive values mean the tool best supports excluding short- to mid-term disease, which the authors say could lighten surveillance for relatives.
Sources
1 on file- European Heart Journaldoi:10.1093/eurheartj/ehag917
Talk it over
The 5-year negative predictive values were 95.5% in Antwerp and 91.5% in Maastricht, but positive predictive values were 42.4% and 39.4%. How would you weigh the value of ruling out disease against the many false alarms?
AI-read ECGs matched heart ultrasound at predicting cardiomyopathy in a study of 343 at-risk relativeshttps://truvace.com/item/artificial-intelligence-enhanced-electrocardiography-for-detection-and-predictio
The 5-year negative predictive values were 95.5% in Antwerp and 91.5% in Maastricht, but positive predictive values were 42.4% and 39.4%. How would you weigh the value of ruling out disease against the many false alarms?
Answer in the threadIf you were one of the genetically predisposed relatives found through cascade screening in Antwerp or Maastricht, how would you want a low or high AI-ECG score used in your care?
Answer in the threadThe model was trained on ECG-echocardiogram pairs from Beth Israel Deaconess Medical Center in Boston. Which people might be missing from that training group, and how could that change how it performs elsewhere?
Answer in the thread
If you have more time
Antwerp (n = 166) and Maastricht (n = 177) each had 18% DCM-positive individuals. What would you want to ask about how well groups this size can show the tool's accuracy?
Answer in the threadHow might a cardiogenetics clinic you know change its surveillance schedule if a low score largely ruled out near-term disease, and what evidence would settle that?
Answer in the thread
ace
What readers say