Walking Speed at Hospital Discharge Rivals Heart Function in Predicting Post-HF Survival

A Predictive Gap in a Global Disease

Heart failure affects roughly 64 million people worldwide, and predicting which patients will die within a year of hospitalization has direct consequences for care. Physicians need that information to decide who warrants intensive monitoring after discharge, who benefits most from aggressive rehabilitation, and where limited clinical resources should concentrate. Several validated scoring tools exist for that purpose -- AHEAD, BIOSTAT compact, and others -- built on factors like age, kidney function, ejection fraction, and arrhythmia history.

These tools work reasonably well in the populations for which they were developed. The problem is that they were developed almost exclusively from European and North American patient data. Applied to older East Asian patients, they systematically underestimate mortality risk. Why that gap exists -- and whether a model built specifically for this population could outperform them -- is what a team from Juntendo University in Tokyo set out to investigate.

Building the Model from 9,700 Patients

The researchers drew on the J-Proof HF registry, a nationwide Japanese database tracking elderly patients hospitalized for heart failure at 96 institutions. They worked with data from 9,700 patients treated between December 2020 and March 2022 who were subsequently discharged. The outcome they sought to predict: death within one year of discharge.

The modeling approach was eXtreme Gradient Boosting (XGBoost), a machine learning algorithm that excels at ranking predictors by their contribution to outcome prediction while handling the messy, heterogeneous data that real-world patient registries contain. The team first trained a full model using all available variables, then built a leaner second model using only the 20 most important variables that model identified.

Of those 20 top variables, 7 were related to physical function or other non-cardiac factors -- a finding that surprised the researchers in its scale. Prominent among them were two standardized assessment tools: the Barthel Index (BI), which scores a patient's ability to perform daily activities like dressing, toileting, and walking, and the Short Physical Performance Battery (SPPB), which tests balance, walking speed over four meters, and chair-rise ability.

What Physical Function Predicts

"The prominence of the BI and SPPB in our analysis is clinically coherent," said Assistant Professor Kanji Yamada, the study's lead author. "Unlike subjective activities of daily living assessments included in some scores, performance-based assessments, such as the BI and SPPB, offer greater reproducibility and capture functional limitations more directly."

A patient who leaves the hospital unable to walk independently or who scores poorly on chair-rise tests is at substantially higher mortality risk than a patient with similar cardiac parameters but better physical function -- and the data from 9,700 patients make that effect large enough to quantify reliably. "Our findings reveal that physical function at discharge is a critically important determinant of survival, rivaling the importance of traditional cardiovascular risk factors," Yamada said.

The 20-variable model performed as well as the full model in predicting one-year mortality. When both were compared against AHEAD and BIOSTAT compact in their ability to stratify patients by risk tier, the XGBoost models proved more discriminating.

From Research to Clinical Tool

The practical ambition of the project goes beyond academic publication. The team has begun developing a web-based tool that would allow clinicians to enter a patient's scores and receive an estimated mortality risk in real time. If validated, such a tool could help hospitals identify which discharged heart failure patients need the closest follow-up and targeted rehabilitation rather than applying uniform monitoring to everyone.

Several caveats apply. The model was built and tested on Japanese patients; its performance in other East Asian populations or non-Asian groups is unknown and would need independent validation before adoption. Machine learning models also tend to reflect the data they were trained on, and the registry captures only patients at participating centers. External validation, both within Japan and internationally, is needed before clinical deployment. The study was published on February 3, 2026, in The Lancet Regional Health - Western Pacific (Volume 67).