Integrating dynamic nomogram and machine learning for personalized disability prediction in elderly cardiometabolic multimorbidity: routine blood markers and mental health

Technical Summary: Integrating Dynamic Nomogram and Machine Learning for Personalized Disability Prediction in Elderly Cardiometabolic Multimorbidity

Problem Statement
Disability in elderly patients with cardiometabolic multimorbidity (CMM)—defined as the co-occurrence of two or more cardiometabolic diseases such as hypertension, heart disease, stroke, diabetes, and dyslipidemia—poses a significant public health burden. While previous studies have identified general risk factors (e.g., age, BMI, depression), existing clinical prediction models often lack individualized precision and fail to address specific complexities: the predictive value of routine blood markers, the comparative dominance of mental health versus physical function, and whether the risk of disability increases linearly with the number of comorbidities or reaches a plateau. Current tools, such as standard ADL/IADL scales, provide general assessments but lack dynamic, personalized risk stratification.

Methodology
The study utilized data from the China Health and Retirement Longitudinal Study (CHARLS), analyzing 1,424 patients aged ≥60 with CMM for model development (training and internal validation sets) and 3,335 patients for external validation. Disability was defined as impairment in Activities of Daily Living (ADL) or Instrumental Activities of Daily Living (IADL), or self-reported disability.

The research employed a multi-faceted analytical approach:

1. Variable Selection: From 46 candidate predictors (including demographics, lifestyle, physical exams, clinical history, and 20+ laboratory markers like HbA1c, TyG, CVAI, and renal function), variables were screened using univariate logistic regression, followed by Least Absolute Shrinkage and Selection Operator (LASSO) regression to reduce dimensionality.
2. Nomogram Development: A dynamic nomogram was constructed using multivariate logistic regression on the final set of independent predictors. This model was implemented as an online tool (Shiny application) for real-time risk calculation.
3. Advanced Statistical Analysis:
   • Mediation Analysis: Explored pathways where physical function (grip strength, walking speed) and falls mediated the effects of depression, cognition, and stroke on disability.
   • Nonlinear Analysis: Restricted Cubic Spline (RCS) and threshold effect analysis were used to identify inflection points in relationships (e.g., cognitive function).
   • Dose-Response Assessment: Linear trend tests and Tukey-adjusted pairwise comparisons evaluated the relationship between the number of CMM conditions and disability risk.
4. Machine Learning (ML): Five algorithms (Logistic Regression, Random Forest, SVM, LightGBM, XGBoost) were trained using the Boruta algorithm for feature selection. The Random Forest (RF) model was selected as optimal based on performance metrics. SHapley Additive exPlanations (SHAP) were applied to interpret feature importance and directionality in the ML model.

Key Results

• Predictive Model: The final dynamic nomogram incorporated six independent predictors: depression symptoms, cognitive function, stroke, number of CMM, age, and falls. The model demonstrated robust discrimination with Area Under the Curve (AUC) values of 0.746 (training), 0.740 (internal validation), and 0.754 (external validation). Calibration curves and Decision Curve Analysis (DCA) confirmed excellent agreement between predicted and actual probabilities and significant clinical net benefit.
• Routine Blood Markers: In traditional logistic regression, routine blood markers were not statistically significant. However, SHAP analysis of the Random Forest model identified HbA1c, creatinine, uric acid, hematocrit, fasting glucose, TyG, and CVAI as important risk markers, suggesting ML captures nonlinear interactions missed by conventional models.
• Mental vs. Physical Health: Depression symptoms emerged as a dominant predictor, outperforming physical function metrics like grip strength and walking speed in both traditional and ML models. Mediation analysis revealed that left-hand grip strength mediated 12.3% of the effect of stroke on disability, while walking speed mediated the effects of cognitive function and stroke.
• CMM Count Plateau: The relationship between the number of CMM conditions and disability risk was not purely linear. Risk significantly increased from 2 to 3 and 4 conditions but appeared to plateau after CMM ≥ 3. Sensitivity analysis merging CMM=4 and 5 confirmed that while CMM≥4 carries higher risk than CMM=2, the difference between CMM=3 and CMM≥4 was not significant.
• Machine Learning Insights: The Random Forest model achieved the highest AUC (0.782) in the internal validation set. Notably, in the ML model, kidney disease ranked higher in predictive value than stroke or the number of CMM, contrasting with the traditional regression where stroke was the strongest predictor (OR=2.55).
• Nonlinear Relationships: Cognitive function showed an inverted U-shaped relationship with disability, with an inflection point at 12.043.

Significance and Claims
The authors claim this study provides the first practical dynamic nomogram specifically for predicting disability in elderly patients with CMM. The study's significance lies in its integrated approach, which:

1. Validates Routine Markers: Demonstrates that routine blood markers (e.g., HbA1c, creatinine) possess predictive value for disability when analyzed through machine learning, even if they are non-significant in linear regression.
2. Prioritizes Mental Health: Establishes that depression symptoms are a more critical predictor of disability risk than physical function metrics (grip strength, walking speed) in this population.
3. Refines Risk Stratification: Identifies a plateau effect in the CMM-disability relationship, suggesting that clinical risk stratification may not need to distinguish linearly beyond three comorbidities.
4. Clinical Utility: Offers a clinically accessible online tool for personalized risk stratification, potentially guiding early interventions focusing on mental health assessment and routine metabolic screening.

The paper concludes that while stroke remains a primary traditional predictor, machine learning reveals the critical role of kidney disease and metabolic dysregulation. The authors acknowledge limitations, including the retrospective nature of the data, potential selection bias regarding blood markers, and the need for external validation in diverse populations, but maintain that the developed nomogram is a valuable step toward early risk identification and personalized management strategies.