Machine learning can identify older heart failure patients at risk of losing independence

For older adults with heart failure (HF), leaving the hospital alive does not always mean returning to an independent life. Loss of independence can reduce quality of life, increase caregiver burden, and raise care needs. Yet most prognostic approaches focus on death or rehospitalization, making it difficult to identify patients likely to survive while losing independence.

Addressing this challenge, a research team led by Assistant Professor Kanji Yamada from the Department of Physical Therapy, Faculty of Health Science, Juntendo University, Japan, including Associate Professor Nobuyuki Kagiyama and Specially-Appointed Professor Tetsuya Takahashi, both from the same university, analyzed 6,382 patients aged 65 years or older who were hospitalized for HF, received physical therapy, and were independent in activities before admission. Using the prospective J-Proof HF Registry, they developed an eXtreme Gradient Boosting (XGBoost) model to predict functional maintenance, functional decline, or death 1 year after discharge. The study was published online in the Journal of Cardiac Failure on August 20, 2026.

The team used 77 candidate predictors. The model incorporated assessments performed by physical therapists at discharge, including maximum gait speed, the Short Physical Performance Battery (SPPB), handgrip strength, and the Barthel Index (BI), as well as preadmission frailty measured by the Kihon Checklist (KCL). A parsimonious Top-10 model was evaluated using nested leave-one-site-out validation and compared with KCL-based scores.

At 1 year, 2,076 patients experienced functional decline and 928 died, while 3,378 maintained functional independence. The Top-10 XGBoost model significantly outperformed the modified KCL score and KCL total score. Gait speed, preadmission KCL score, discharge BI, and SPPB were among the influential predictors, with greater predictive importance than cardiac measures such as ejection fraction.

Being able to live independently is just as important to many patients as how long they live. Our findings show that physical function at discharge can reveal a risk that mortality-focused models may overlook, identifying patients who are likely to survive but lose independence."

Dr. Kanji Yamada, Assistant Professor, Department of Physical Therapy, Faculty of Health Science, Juntendo University

The model identified four risk phenotypes by combining predicted risks of functional decline and mortality. Most notably, a low-mortality/high-decline group of 1,732 patients had a 57% rate of functional decline despite an 89% probability of survival at 1 year. Identifying this group could help medical professional prioritize post-discharge support, including more intensive outpatient or home-based rehabilitation.

"Instead of asking only whether a patient will survive, we need to ask whether they will be able to maintain the life they value after discharge," Dr. Yamada says. "A practical risk estimate could help clinicians and patients discuss rehabilitation needs earlier and make better use of limited resources."

To support clinical use, the researchers developed a prototype web application that estimates individual risk at discharge and simulates how changes in physical function, such as faster gait speed, may affect predicted risk. The application is a prototype, and implementation studies are needed to determine whether model-guided decisions improve patient outcomes.

Overall, the study demonstrates that machine learning can help make functional decline visible alongside mortality risk in older adults with HF. By combining clinical information with direct measures of physical performance, the approach could support targeted rehabilitation and patient-centered discharge planning, with the goal of preserving independence and quality of life.