This study finds that an out-of-the-box risk classification by a U.S. Food and Drug Administration–cleared artificial intelligence–enabled algorithm for identifying hypertrophic cardiomyopathy from electrocardiograms can have a low predictive value when used for diagnosis in a screened population. Calibrated model scores approximate the probability of disease among those flagged and can both improve the interpretability of model outputs and inform decision-making about subsequent follow-up.
Unpacking an important application of model calibration for an out-of-the-box artificial intelligence algorithm used to flag patients who are at high risk for hypertrophic cardiomyopathy, this article explains the use of calibrated scores for patient-level decision-making.
This perspective draws parallels between artificial intelligence (AI) regulation and competency-based medical education (CBME), an existing framework that addresses similar complexities in training human clinicians. The authors propose a CBME framework as a life-cycle approach to regulating AI in health care that offers a dynamic, continuous, and outcomes-based regulatory model to address the unique challenges posed by AI technologies.
This paper discusses the challenges of postdeployment assessment of predictive artificial intelligence (AI) models in health care, particularly how confounding medical interventions can distort performance evaluation. It explores current limitations in monitoring strategies and suggests advanced causal modeling as a potential solution to ensure that AI models remain effective in dynamic clinical settings.
PODCAST
AI Grand Rounds features informal conversations with a variety of unique experts exploring the deep issues at the intersection of artificial intelligence, machine learning, and medicine.
This perspective explains how the unique nature of primary care translates to important differences in machine learning model development and adoption for clinical use. Four key considerations are presented to address the use of ML in primary care settings.
This perspective examines the ongoing challenges in integrating artificial intelligence into health care, emphasizing the persistent issues of algorithmic bias, privacy concerns, and the necessity for rigorous validation. It highlights the balance between automation and human oversight, particularly in ethical decision-making and informed consent processes.
This study describes the development and application of a self-supervised deep-learning approach called temporal learning for analyzing longitudinal medical imaging. It demonstrates that temporal learning improves longitudinal, imaging-based recurrence prediction in pediatric gliomas, and that performance increases incrementally with the number of historical scans available to the algorithm.
This study evaluates the CathEF deep-learning algorithm, which estimates left ventricular ejection fraction directly from routine left coronary angiograms in acute coronary syndrome patients, eliminating the need for additional contrast or catheterization. The study found that CathEF achieved diagnostic performance comparable to ventriculography, although its accuracy was lower in patients with ST-elevation myocardial infarction or right coronary artery involvement.
This review examines the evolving role of artificial intelligence (AI) in oncology, particularly in radiology and pathology, and treatment selection using large language models. It highlights key challenges — such as data quality, model validation, and clinical integration — and proposes strategies to overcome them to support the effective adoption of AI to improve cancer care and patient outcomes.
Your trusted guide to rigorous evaluation of clinical applications of AI.
Monthly & Annual Options