REINFORCEMENT LEARNING for coronary artery disease treatment revealed significant variations in clinical practice across patient sexes and hospital sites, according to a new artificial intelligence study.
Meanwhile transfer learning helped reduce performance gaps when models were applied to new patient groups.
Researchers analysed data from more than 41,000 adults with obstructive coronary artery disease who underwent diagnostic coronary angiography between 2009 and 2019.
The study investigated whether reinforcement learning could detect differences in treatment decision-making and whether transfer learning could address distribution shifts between patient populations and care settings.
Evaluation of physician treatment policies showed notable variation between groups. Female patients received lower expected rewards under physician policies derived from female patient data compared with male patients under male physician policies.
Differences were also identified between hospital sites. Policies derived from one site frequently performed differently when applied to patients treated at another site, suggesting variation in historical treatment approaches.
The investigators developed reinforcement learning policies and compared their performance with physician behaviour policies. Across sex and site analyses, reinforcement learning generally delivered higher expected rewards than physician policies.
The findings showed that reinforcement learning models trained on group specific data consistently outperformed the corresponding physician policies when evaluated within the same group.
Researchers also observed differences in treatment recommendations. Reinforcement learning policies tended to recommend coronary artery bypass grafting more frequently than physicians, especially under the least conservative settings.
Transfer learning was then applied to adapt reinforcement learning models to new target populations.
Fine tuning a base model with only 10% of target group data produced outcomes that approached those achieved by models trained using the full target dataset.
According to the researchers, this demonstrates the potential value of transfer learning in addressing distribution shifts that commonly limit the deployment of clinical artificial intelligence tools.
The researchers concluded that reinforcement learning and transfer learning may help optimise the treatment of coronary artery disease, although further validation is required before clinical implementation.
Reference
Ghasemi P et al. Investigating clinical practice variations in coronary artery disease treatment using reinforcement and transfer learning. npj Cardiovasc Health. 2026;DOI:10.1038/s44325-026-00147-0
Featured image: Art_Photo on Adobe Stock
Each article is made available under the terms of the Creative Commons Attribution-Non Commercial 4.0 License.
