Identification of Low-Mortality/High-Decline Risk Phenotypes in Heart Failure via Machine Learning

Article Impact Level: HIGH
Data Quality: STRONG
Summary of  Journal of Cardiac Failure https://doi.org/10.1016/j.cardfail.2026.08.005
Dr. Kanji Yamada et al.

Points

Juntendo University researchers evaluated six thousand three hundred eighty two older hospitalized heart failure patients from the prospective J Proof HF Registry using eXtreme Gradient Boosting machine learning.
One year post discharge evaluations revealed that two thousand seventy six patients experienced functional decline and nine hundred twenty eight died while three thousand three hundred seventy eight maintained independence.
Quantitative predictive modeling demonstrated that discharge physical performance metrics including gait speed SPPB and Barthel Index exhibited greater predictive importance than left ventricular ejection fraction.
Phenotypic risk clustering identified a low mortality high decline subgroup of seventeen hundred thirty two patients exhibiting a fifty seven percent functional decline rate despite eighty nine percent survival.
Authors concluded that machine learning combining physical function metrics enables targeted post discharge rehabilitation and patient centered planning to preserve functional independence and quality of life.

Summary

This study evaluated an eXtreme Gradient Boosting (XGBoost) machine learning model designed to predict functional maintenance, functional decline, and mortality one year after hospital discharge in older patients with heart failure (HF). Published in the Journal of Cardiac Failure by researchers at Juntendo University using the prospective J-Proof HF Registry, the investigation addressed limitations in traditional prognostic models that focus exclusively on mortality or rehospitalization. Given that loss of functional independence severely reduces quality of life, increases caregiver burden, and raises care needs, the research sought to identify patients likely to survive while experiencing severe functional loss.

The multicenter registry cohort comprised 6,382 patients aged 65 years or older who were hospitalized for HF, received physical therapy, and were functionally independent prior to admission. The XGBoost model evaluated 77 candidate predictors, incorporating discharge physical performance measures (maximum gait speed, Short Physical Performance Battery [SPPB], handgrip strength, and Barthel Index [BI]) alongside preadmission frailty via the Kihon Checklist (KCL). At one year, 3,378 patients maintained functional independence, 2,076 experienced functional decline, and 928 died. Nested leave-one-site-out validation confirmed that a parsimonious Top-10 XGBoost model significantly outperformed traditional modified and total KCL scoring systems, with maximum gait speed, preadmission KCL score, discharge BI, and SPPB demonstrating greater predictive power than left ventricular ejection fraction.

The authors conclude that machine learning integrating physical performance metrics accurately identifies post-discharge functional trajectories in older HF patients. The model stratified four distinct risk phenotypes, notably isolating a low-mortality/high-decline cluster (n=1,732) exhibiting a 57% rate of functional decline despite an 89% one-year survival probability. By identifying patients who survive but risk losing functional autonomy, this clinical decision framework and its prototype web simulator enable targeted outpatient rehabilitation and patient-centered discharge planning to preserve long-term independence.

Link to the article: https://onlinejcf.com/article/S1071-9164(26)00496-3/fulltext

References

Yamada, K., Kagiyama, N., Morisawa, T., Saitoh, M., Iwata, K., Kato, M., Sakurada, K., Kono, Y., Iida, Y., Taya, M., Funami, Y., Kamiya, K., & Takahashi, T. (2026). Machine learning predicts functional decline and risk phenotypes in older patients with heart failure. Journal of Cardiac Failure, S1071916426004963. https://doi.org/10.1016/j.cardfail.2026.08.005