Exercise training is a cornerstone of cardiac rehabilitation (CR) for patients with coronary artery disease (CAD), and current guidelines rate it as a top-tier recommendation for reducing the risk of future cardiovascular events. Yet a substantial share of patients, often estimated at one in five or more, show little or no measurable improvement in their exercise capacity despite completing a full rehabilitation program. Identifying these "non-responders" before training begins could allow clinicians to adjust treatment plans early, rather than discovering the lack of benefit only after weeks of therapy have passed.

A new study from researchers at the University of Witten/Herdecke and DRV Clinic Königsfeld in Germany, in collaboration with colleagues from FORTH in Greece, used machine learning to address this challenge. The researchers analyzed data from 353 patients with CAD who had experienced a heart attack and/or undergone procedures such as angioplasty or bypass surgery, all of whom completed 3 to 4 weeks of inpatient cardiac rehabilitation.

Using only information collected at the start of rehabilitation including cardiopulmonary exercise testing and pulse wave analysis (a non-invasive measure of arterial stiffness and vascular function) the team trained ten different machine learning algorithms to predict which patients would and would not show a clinically meaningful improvement in peak oxygen uptake (V̇O₂peak), a key marker of cardiovascular fitness and long-term survival. Among the algorithms tested, a Random Forest model performed best, correctly classifying responders and non-responders with 77% accuracy. The study led by Professor Boris Schmitz and Professor Frank Mooren from University of Witten/Herdecke, Germany, was made available online on May 05, 2026, in the Journal of Sport and Health Science.

Notably, at the start of rehabilitation, responders and non-responders looked very similar on standard clinical measures such as age, sex, body mass index, baseline fitness, and medical history which did not reliably distinguish the two groups. Using explainable AI techniques (SHAP analysis), the researchers found that the most influential predictors instead came from how efficiently patients breathed during exercise testing and how stiff their arteries were. Patients who used more ventilation per unit of oxygen consumed, had less breathing reserve, or showed greater arterial stiffness (higher pulse wave velocity) were less likely to improve their fitness through training. Two classes of blood pressure medication, angiotensin II receptor blockers and calcium channel blockers, also influenced predictions, while the type or severity of underlying heart disease had little bearing on the outcome.

"Our model shows that we can identify likely non-responders using data that is already routinely collected at the start of cardiac rehabilitation," said corresponding author Prof. Schmitz of the University of Witten/Herdecke. "This opens the door to tailoring exercise programs to the individual, rather than applying the same training plan to everyone, so that patients unlikely to benefit from standard training can be identified early and offered adjusted or more intensive interventions."

The authors note that the findings point toward a more personalized model of cardiac rehabilitation, in which baseline vascular and respiratory characteristics and not just traditional risk factors like age or disease severity, help guide how exercise programs are designed for each patient. They caution that the model was developed in a specific clinical population and will need further validation in older patients, those with additional health conditions, and rehabilitation programs structured differently from the one studied. The team's next step is a randomized controlled trial testing whether predicted non-responders benefit from individually adjusted aerobic interval training, with the goal of reducing the number of patients who complete rehabilitation without meaningful fitness gains.

Source:

Journal reference: