Physical Function Key to Predicting Heart Failure Survival, Machine Learning Study Finds

Latest research from Juntendo University demonstrates that incorporating physical function metrics into machine learning models significantly improves the accuracy of predicting one-year mortality risk in elderly patients undergoing treatment for heart failure (HF). The findings challenge the reliance on traditional cardiac-specific variables and highlight the importance of comprehensive geriatric assessment in managing this complex condition.

The Limitations of Current Risk Models

Existing heart failure mortality risk models, such as AHEAD (Atrial fibrillation, Hemoglobin, Elderly, Abnormal renal parameters, Diabetes mellitus) and BIOSTAT compact, have proven clinically valuable. However, these models were primarily developed using data from European and North American populations and often underestimate risk in older East Asian patients. This disparity underscores the necessitate for more nuanced and population-specific prognostic tools.

A New Approach: Machine Learning and Physical Function

Researchers led by Professor Tetsuya Takahashi, Assistant Professor Kanji Yamada and Associate Professor Nobuyuki Kagiyama at Juntendo University utilized an eXtreme Gradient Boosting (XGBoost) algorithm – a sophisticated machine learning technique – to identify key predictors of mortality in heart failure patients. Their work, published in The Lancet Regional Health—Western Pacific, focused on data from nearly 10,000 elderly patients treated for HF at 96 institutions across Japan.

Key Findings: The Importance of Non-Cardiac Factors

The study revealed that physical function metrics are critically important determinants of survival, rivaling the importance of traditional cardiovascular risk factors. The team developed two XGBoost models: one using all available variables and a second, “Top-20 XGBoost” model, focusing on the 20 most impactful variables. Notably, seven of these top 20 variables were related to physical function and other non-cardiac factors.

“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,” explained Dr. Yamada. “Unlike subjective assessments, performance-based assessments, such as the Barthel Index (BI) and Short Physical Performance Battery (SPPB), offer greater reproducibility and capture functional limitations more directly.”

Implications for Clinical Practice

Both XGBoost models demonstrated similar accuracy in predicting one-year mortality risk. The Top-20 XGBoost model, however, more effectively classified patients according to their risk of death compared to the AHEAD and BIOSTAT compact models. This suggests that a more targeted approach to patient care, informed by physical function assessments, could improve outcomes.

The researchers are developing a tool based on the Top-20 XGBoost model that will allow physicians to input patient data and receive an accurate estimation of their mortality risk. This could facilitate more efficient allocation of medical resources and enable tailored post-discharge care for those at highest risk.

Future Directions

While the findings are promising, the research team emphasizes the need for further testing and refinement of the model, both within Japan and in other countries. The study underscores the essential value of integrating comprehensive geriatric and functional assessments into the routine management and risk stratification of older patients with heart failure. This research highlights the potential of physical rehabilitation as a crucial component of long-term heart failure management and the importance of maintaining physical function before and after hospitalization.

Source: Juntendo University