NEJM AI. 2025 May;2(5):AIoa2400421. doi: 10.1056/AIoa2400421. Epub 2025 Apr 24.
ABSTRACT
BACKGROUND: Despite optimism for the incorporation of deep-learning (DL) algorithms into clinical practice, there is a lack of prospective data on how U.S. Food and Drug Administration (FDA)-cleared algorithms function out of the box and clinicians' interpretations of model predictions. This study illustrates how model calibration, which rescales model outputs to represent probabilities of disease among those flagged, enables improved interpretability of the scores by end users.
METHODS: Using data from patients flagged as being at high risk for hypertrophic cardiomyopathy (HCM), patients 18 years of age and over were screened from a single health system using a DL electrocardiogram (ECG)-based algorithm for HCM detection. Cardiologists manually reviewed medical records and imaging of patients flagged as being suggestive of HCM to classify individuals as HCM positive (HCM+) or HCM negative (HCM-). Logistic regression was used for calibration to transform model outputs to probabilities in the flagged sample. We compare calibrated probabilities with the estimated probability of disease in the flagged sample and estimated positive predictive value (PPV) at different thresholds of the calibrated score.
RESULTS: Among 1522 patients were flagged as being high risk, 166 were classified as HCM+ and 1356 as HCM-. Being flagged as being at high risk on the initial screening has a PPV of 10.9% for HCM patients. Recalibrated model scores in the flagged sample closely approximated the probability of HCM among those who were initially flagged, lending improved interpretability. Setting a cutoff value that corresponds to the top 5 and 10 ranked values yields PPVs of 80% and 70%, respectively. As the ranks of the model outputs and the recalibrated scores are equivalent, PPVs derived from this approach do not depend on which score is used, but the calibrated scores have context-relevant interpretability as the probability of HCM.
CONCLUSIONS: Using the raw scores from an out-of-the-box-FDA-cleared algorithm for the identification of HCM from ECGs may have low PPV for identifying patients who truly have HCM. Calibration of scores for those flagged can improve the interpretability of the scores in terms of disease probability. (Funded by Viz.ai.).
PMID:42932535 | DOI:10.1056/AIoa2400421