Juntendo University ML Model Improves Heart Failure Survival Prediction by 20% Incorporating Physical Function

Breakthrough in Heart Failure Prognosis: Juntendo's AI Model Integrates Physical Function

Researchers at Juntendo University have developed a groundbreaking machine learning model that significantly enhances the prediction of one-year survival rates in elderly patients with heart failure (HF). By incorporating objective measures of physical function at hospital discharge, the model achieves superior accuracy compared to traditional risk scores, offering a 20% improvement in patient reclassification for mortality risk. This advancement addresses a critical gap in HF management, where conventional models often overlook non-cardiac factors like frailty and muscle strength, which are prevalent in Japan's super-aging population.

Heart failure, a condition where the heart cannot pump blood effectively, affects over 1.28 million people in Japan, with mortality rates reaching 16-30% within one year post-hospitalization for those over 65. The new model, powered by XGBoost—a robust gradient boosting algorithm—leverages data from nearly 10,000 patients, highlighting how simple tests like grip strength and gait speed can rival complex cardiac biomarkers in prognostic power.

The Growing Burden of Heart Failure in Japan's Elderly

Japan's rapidly aging society faces an escalating HF crisis. By 2026, nearly 30% of the population is over 65, and HF prevalence has surged due to improved survival from acute cardiac events but persistent risk factors like hypertension and diabetes. Hospitalization for acute decompensated HF is common, with readmission rates high and one-year mortality around 20% for super-elderly patients (85+).

Frailty exacerbates outcomes; up to 70% of hospitalized HF patients exhibit sarcopenia or reduced physical function, leading to non-cardiovascular deaths like infections or falls. Traditional prognostic tools like the AHEAD or BIOSTAT-CHF scores, designed for Western cohorts, underperform in Asian elderly, with AUCs as low as 0.60. Juntendo's research underscores the need for models tailored to geriatric syndromes.

• HF hospitalization rates: 25.9/1000 in men, 5.3/1000 in women (age-standardized).
• 1-year post-discharge mortality: 16.5% in J-Proof cohort.
• Frailty prevalence: 50-70% in elderly HF patients.

For professionals in higher education jobs in health sciences, this highlights opportunities in interdisciplinary research at institutions like Juntendo.

Introducing the J-Proof HF Registry: A Gold Standard Dataset

The foundation of this innovation is the J-Proof HF Registry, a nationwide prospective cohort led by the Japanese Society of Cardiovascular Physical Therapy. Enrolling 10,620 patients aged 65+ hospitalized for HF across 96 institutions from December 2020 to March 2022, it captures comprehensive data on rehabilitation-prescribed cases. After exclusions, 9,700 patients were analyzed (median age 83, 51% male), with 16.5% one-year mortality.

This real-world dataset includes demographics, labs, echocardiograms, and crucially, functional assessments performed during rehab—Barthel Index (BI), Short Physical Performance Battery (SPPB), gait speed, and handgrip strength. Unlike retrospective databases, its prospective design and multicenter scope minimize bias, making it ideal for AI training.

Juntendo's Faculty of Health Science played a pivotal role, with experts like Prof. Tetsuya Takahashi driving data curation and model validation.

How the ML Model Was Built: From Data to XGBoost Mastery

The team employed eXtreme Gradient Boosting (XGBoost), an ensemble method excelling in tabular data with mixed types. Starting with 77 predictors, they used SHAP (SHapley Additive exPlanations) values to rank importance, deriving a parsimonious top-20 model. Validation via leave-one-site-out (LOSO) prevented overfitting across sites.

1. Data Preprocessing: No imputation for functional measures; missing gait speed treated as severe impairment.
2. Training: Full cohort split; hyperparameter tuning via Bayesian optimization.
3. Feature Selection: Top features: BI at discharge (SHAP rank 1), SPPB, gait speed, handgrip strength, BNP, age, eGFR.
4. Evaluation: AUC, AUPRC, NRI, DCA, calibration plots.

Both full and top-20 models hit AUC 0.76 (LOSO), vs. 0.60 for benchmarks—a leap forward. A web app prototype allows clinicians to input data for instant risk scores.

Spotlight on Physical Function Metrics: Simple Tests, Profound Insights

Physical function emerged as king. Here's a breakdown:

• Barthel Index (BI): Assesses activities of daily living (ADL) like feeding, bathing (0-100 score). Discharge BI was top predictor; higher scores lower risk.
• Short Physical Performance Battery (SPPB): Composite of balance, gait speed, chair stands (0-12). Low SPPB signals frailty.
• Gait Speed: Max walking speed (m/s) over 4m; <0.8 m/s poor prognosis.
• Handgrip Strength: Max kg force; sarcopenia marker, modifiable via training.

These outperform labs like BNP in SHAP analysis, as they capture holistic vulnerability. Pre-hospital BI/Kihon Checklist also factored in.

In Japan, where rehab is standard, these metrics are feasible, enabling step-by-step assessment: measure at admission/discharge, track deltas for rehab efficacy.

Photo by Sean Chen on Unsplash

Superior Performance: 20% Reclassification Boost and Risk Strata

The model's edge shines in metrics:

NRI quantifies reclassification: 21-24% more patients correctly shifted to high/low risk categories—the '20% better' hallmark. Tertile risks: low 2.5%, intermediate 11.5%, high 35.5% mortality. DCA confirms net benefit for interventions at 10-40% thresholds.

Juntendo University's Legacy in Cardiac Rehab and Frailty Research

Juntendo, a Tokyo-based leader in health sciences, has pioneered HF rehab for 20+ years. Teams under Prof. Takahashi and Dr. Yamada focus on geriatric HF, with prior work on CR impact in frail elderly, GLIS sarcopenia model (outperforms AWGS), and Clinical Frailty Scale (CFS) for quick prognosis.

PubMed lists dozens of Juntendo HF papers, emphasizing multidisciplinary approaches. For aspiring researchers, research jobs here blend AI, PT, and cardiology.

Clinical Implications: Tailored Rehab and Precision Care

This model shifts paradigms: high-risk patients get intensified rehab, home monitoring, nutrition. Benefits include:

• Resource optimization in strained systems.
• Modifiable targets: grip/gait training cuts mortality 20-30%.
• Holistic geriatric care integrating PT early.

Proposed pathway: screen at discharge, risk-stratify, refer high-risk to outpatient CR. In Japan, universal insurance supports scalability.

AI's Rising Role in Japanese Cardiology

Japan leads Asia in AI-cardiology, with models for CHD risk, AF-HF prediction. Juntendo's tool joins UTokyo's HF monitoring AI, Shimadzu-funded startups. Future: federated learning across registries, wearable integration for real-time gait tracking.

Explore academic CV tips for AI-health roles.

Challenges, Future Directions, and Global Potential

Challenges: missing data handling, external validation beyond Japan. Outlook: prospective trials, app deployment. Globally, adaptable for Asia's aging boom; code available via authors.

Stakeholders—clinicians, policymakers—gain actionable insights. Juntendo exemplifies university innovation driving patient outcomes.

Photo by Bridget Adolfo on Unsplash

Conclusion: Empowering Better HF Outcomes Through Innovation

Juntendo's ML model marks a milestone, proving physical function's primacy in HF survival. By blending AI with rehab, it promises reduced mortality, better QOL. Researchers and educators, check Rate My Professor, explore higher ed jobs, university jobs, career advice. For Japan opportunities, visit AcademicJobs Japan.