IPM Take
Cardiac rehabilitation is evidence-based.
That does not mean every patient responds to the same exercise programme in the same way.
In this study, 225 of 353 analysed patients with coronary artery disease, 63.7%, did not improve peak oxygen uptake beyond the study’s threshold for a measurable training response during a short inpatient rehabilitation programme. Yet standard characteristics such as age, BMI and underlying diagnosis did little to distinguish them beforehand.
Machine learning did better.
The provocative implication is not that AI should decide who receives cardiac rehabilitation. Exercise-based rehabilitation remains strongly recommended because its benefits extend well beyond VO₂peak, including cardiovascular risk reduction and improved outcomes.
The real opportunity is more precise: stop assuming that referral to rehabilitation is the end of personalisation.
If a patient is unlikely to improve with the standard training prescription, the clinically useful question is what should be changed before several weeks are lost.
Executive Summary
Researchers from Germany and Greece developed machine-learning models using baseline data from patients with coronary artery disease undergoing three to four weeks of inpatient cardiac rehabilitation after myocardial infarction, PCI and/or coronary artery bypass grafting.
Of 393 participants initially included, 353 remained after data cleaning. Only 128 were classified as exercise responders, while 225, or 63.7%, were considered non-responders because their improvement in peak oxygen uptake did not exceed the study’s predefined threshold based on measurement variability. Training participation was similarly high in both groups, suggesting that attendance alone did not explain the difference.
Ten supervised machine-learning approaches were tested. A Random Forest model performed best, achieving 77% balanced accuracy in model development and an area under the ROC curve of 0.84. The study also reported successful validation in independent coronary artery disease cohorts from other centres.
The most influential predictors included breathing and ventilation characteristics, oxygen uptake combined with pulse wave velocity, cardiac output and vascular measures. Primary diagnosis and disease severity contributed comparatively little.
The next test is clinical utility. The researchers plan a randomized trial to determine whether patients predicted not to respond to standard rehabilitation benefit from individually adjusted aerobic interval training.
Why it matters
- HTA bodies: An accurate prediction model is not automatically a valuable health technology. Future assessment would need evidence that using the algorithm changes exercise prescriptions, improves functional capacity or outcomes, and does so cost-effectively.
- Payers: Cardiac rehabilitation is already an evidence-based secondary-prevention intervention. Predicting poor response could help target additional rehabilitation resources toward patients most likely to need modified or intensified programmes rather than waiting until discharge to discover limited fitness improvement.
- Industry / innovation partners: CPET, vascular measurements and explainable machine learning create an opportunity for clinical decision support in precision rehabilitation. The commercial challenge will be demonstrating prospective benefit, interoperability and transportability across different rehabilitation systems.
Cardiac rehabilitation is personalised medicine only up to a point.
Patients receive assessment, exercise prescription, risk-factor management and education. Yet exercise programmes still frequently begin from broadly standardised protocols, with response assessed after rehabilitation is already underway.
A new study suggests machine learning could move that decision earlier.
Researchers studied patients with coronary artery disease who entered three to four weeks of inpatient cardiac rehabilitation after myocardial infarction, angioplasty, bypass surgery or combinations of these events and procedures.
Cardiopulmonary exercise testing was performed before and after rehabilitation.
The outcome was peak oxygen uptake, or VO₂peak, a major measure of cardiorespiratory fitness with prognostic importance in coronary disease. Patients were considered responders only when their improvement exceeded the study’s estimate of normal measurement variability, approximately 0.17 L/min.
After data cleaning, 353 patients were analysed.
Only 128 met that response criterion.
The remaining 225, or 63.7%, were classified as non-responders.
That number needs careful interpretation.
It does not mean nearly two-thirds received no benefit from cardiac rehabilitation. The study examined one specific outcome over a relatively short programme. Cardiac rehabilitation also targets symptoms, risk factors, medication adherence, psychosocial health, physical activity and subsequent cardiovascular events.
Current ESC guidance recommends multidisciplinary exercise-based programmes for patients with chronic coronary syndromes, and the 2025 ACC/AHA acute coronary syndrome guideline gives cardiac rehabilitation referral a Class I recommendation.
The question is therefore not whether rehabilitation works.
It is why the fitness response varies so much.
The usual clinical characteristics were not the best clues
Responders and non-responders looked surprisingly similar before rehabilitation.
Age, sex, BMI, baseline fitness and much of the conventional disease history were poor discriminators.
Instead, the Random Forest algorithm placed greater weight on physiological information already available from cardiopulmonary exercise testing and pulse wave analysis.
Important signals included ventilation relative to oxygen consumption, breathing reserve and frequency, pulse wave velocity, cardiac output and augmentation time. Greater arterial stiffness and less favourable ventilatory characteristics tended to point toward poorer training response.
That is where the study becomes interesting for precision cardiology.
A diagnosis such as coronary artery disease tells clinicians why someone belongs in rehabilitation.
It may say much less about how that individual should train.
Two patients with apparently similar coronary disease can have different vascular stiffness, ventilatory limitations and cardiovascular responses to exertion. A standard programme may therefore impose the same intervention on physiologically different patients.
Machine learning offers a way to find patterns that conventional categorisation misses.
But 77% balanced accuracy is not certainty.
Misclassification matters. A patient predicted to be a “non-responder” could still benefit from standard exercise, while another predicted to respond may not. Any clinical model would therefore need to guide adaptation, not ration access.
That distinction should shape implementation.
The future of AI in cardiac rehabilitation should not be an algorithm deciding who deserves exercise.
It should be an algorithm helping clinicians decide what kind of exercise each patient needs.
The researchers’ planned randomized trial will be the more consequential experiment: whether changing training on the basis of predicted response actually converts non-responders into responders.
If it does, cardiac rehabilitation could move from measuring variability after the fact to designing around it from the beginning.
That would be a far more meaningful use of AI than simply predicting who is likely to struggle.
