Monitoring and treating heart failure (HF) is a challenging condition at any age. Several models, such as Atrial fibrillation, Hemoglobin, Elderly, Abnormal renal parameters, Diabetes mellitus (AHEAD), and BIOlogy Study to TAilored Treatment in Chronic Heart Failure (BIOSTAT) compact, have been developed to predict the likelihood of a patient's survival based on clinical factors such as arrhythmia, anemia, age, diabetes, and ejection fraction. However, previous studies have shown that these tools, which were developed for European and North American populations, consistently underestimate the risk among older East Asian patients. Could incorporating other factors improve predictions of patient survival?
A team of researchers from Juntendo University has developed a better model to predict long-term survival after HF. This research project was led by Professor Tetsuya Takahashi and Assistant Professor Kanji Yamada from the Faculty of Health Science, and Associate Professor Nobuyuki Kagiyama from the Graduate School of Medicine. The team used machine learning algorithms to find the most important metrics for gauging the odds of survival. Their findings were published on February 3, 2026, in Volume 67 of the journal The Lancet Regional Health – Western Pacific .
Describing the deficiencies of existing models of HF severity, Dr. Yamada says, "These models rely primarily on cardiac-specific and biomedical variables, often underestimating the impact of non-cardiac factors such as physical function, frailty, and nutritional status, which are critical determinants of prognosis in older adults and, unlike fixed factors such as age, may represent modifiable targets through rehabilitation and supportive care."
The research team turned to the nationwide J-Proof HF registry that tracks elderly patients treated for HF at 96 institutions across Japan. Using data from 9,700 patients treated between December 2020 and March 2022 and discharged from the hospital, the team trained an eXtreme Gradient Boosting (Full XGBoost) algorithm to predict the risk of mortality within one year of treatment.
The team also developed a second model (Top-20 XGBoost) using the 20 most important variables from the first model. 7 of the 20 variables were related to physical function and other non-cardiac factors. "The prominence of the BI [Barthel Index] and SPPB [Short Physical Performance Battery] in our analysis is clinically coherent," said Dr. Yamada, adding, "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."
Both XGBoost models were similarly accurate in predicting the risk of death within one year. In addition, the Top-20 XGBoost model more effectively classified patients according to their risk of death compared to the AHEAD and BIOSTAT compact. As the model was developed using data from a nationwide Japanese cohort, it may provide a more context-specific tool for risk assessment in older patients with HF in Japan.
Instead of using a "one-size-fits-all" approach to treating elderly patients with HF, doctors and other healthcare professionals can use Top-20 XGBoost to accurately identify patients who could benefit from closer monitoring or more tailored post-discharge care. This would also be a more efficient use of medical resources. The prominence of physical function metrics in this model highlights the importance of physical rehabilitation as part of long-term heart failure management, as well as the potential value of maintaining physical function before and after hospitalization.
"Our findings reveal that physical function at discharge is a critically important determinant of survival, rivaling the importance of traditional cardiovascular risk factors. This study underscores the essential value of integrating comprehensive geriatric and functional assessments into the routine management and risk stratification of older patients with HF," remarked Dr. Yamada.
The team is cautiously optimistic, noting that the model will need to be refined with more testing, both in Japan and other countries. Nevertheless, they have begun developing a tool based on Top-20 XGBoost, where physicians and other healthcare professionals can enter information about a patient and get an accurate estimation of their risk of mortality.
