.png)
02
A Predictive Model for Senior Institutionalization Risk
Domain: Public Health | Geriatrics | Machine Learning
Timeline: Spring 2025
Tools: R, GLMM, XGBoost, Cluster-based Undersampling, NHATS Data
Abstract
This study investigates the predictive relationship between declining social participation and transitions to institutional care among older adults with mental health conditions, leveraging longitudinal data from the National Health and Aging Trends Study (NHATS). By analyzing nine annual survey waves from 9,844 participants, we employed advanced statistical methods to address severe class imbalance and model temporal dynamics. A two-stage framework combining Generalized Linear Mixed Models (GLMM) and machine learning techniques revealed that faster declines in social participation significantly increased institutionalization risk (OR=2.3, *p*<0.001), with XGBoost achieving strong predictive performance (AUC-PR=0.81). The findings highlight the clinical relevance of social engagement monitoring and provide methodological insights for handling imbalanced longitudinal data in geriatric research.
Introduction
Social participation is a cornerstone of healthy aging, yet its decline often precedes critical transitions to institutional care among older adults. This project examines whether longitudinal reductions in social engagement predict institutionalization in individuals with mental health conditions—a population disproportionately affected by isolation and care needs. Using NHATS data spanning 2011–2020, we addressed three challenges: quantifying individual trajectories of social participation decline, disentangling its predictive value from health and cognitive factors, and developing actionable models despite extreme class imbalance (4.9% institutionalization rate). The analysis advances methodologies for rare-event prediction in longitudinal datasets while offering evidence to guide preventive healthcare strategies for vulnerable aging populations.
Data & Methods
The study utilized data from 12,427 NHATS participants aged 65+, narrowed to 9,844 individuals with mental health conditions (PHQ-4≥3), and at least two survey rounds. Institutionalization was operationalized as transitions from community residences to skilled nursing facilities or long-term hospitalization. Social participation metrics included event attendance, family visits, and community activities, supplemented by health, cognitive, and socioeconomic variables.
Data preprocessing involved harmonizing nine survey rounds, addressing missing values through multiple imputations, and calculating composite scores for mental health (PHQ-4) and cognitive function. Temporal alignment was achieved by recentering observation windows to each participant’s first survey round. To mitigate class imbalance, we implemented cluster- based undersampling, grouping non-transitioned participants via k-means clustering (*k*=20) and proportionally sampling representatives to create balanced training data (764 non- transitioned vs. 382 transitioned). Analytically, the project employed a two-stage framework. First, GLMMs with random intercepts and slopes modeled individual trajectories of social participation decline. Second, logistic regression and machine learning models (XGBoost, Random Forest) predicted institutionalization using these trajectories alongside health and demographic covariates. Model performance was evaluated through 80/20 train-test splits, precision-recall curves (PR-AUC), and sensitivity-specificity analyses.
Results & Discussion
Participants who transitioned to institutional care exhibited significantly steeper declines in social participation, with 69% reporting limitations in their final pre-transition survey compared to 39% among stable community dwellers (*p*<0.001). GLMM estimates revealed that each unit increase in negative social participation slope (faster decline) raised institutionalization odds by 130% (OR=2.3, 95% CI:1.1–4.8), independent of health and cognitive status. This relationship intensified in later survey rounds, suggesting a critical window for intervention.
Predictive modeling underscored the value of machine learning in handling complex interactions. XGBoost achieved the highest performance (AUC-PR=0.81), identifying key risk factors including interactions between healthcare access and cognitive function (SHAP value=0.015) and advanced age (>90 years, OR=7.8). Paradoxically, financial assistance recipients showed elevated social participation limitations, potentially reflecting unmet needs or reverse causality. Despite these insights, precision remained modest (17–26%), emphasizing the challenges of rare-event prediction. Methodologically, cluster-based undersampling improved model sensitivity without sacrificing specificity. However, residual class imbalance and annual survey gaps limited temporal resolution, suggesting opportunities to integrate real-time wearable data in future studies.
Conclusion
This study establishes declining social participation as a critical predictor of institutionalization in older adults with mental health conditions while demonstrating innovative approaches for analyzing imbalanced longitudinal data. The two-stage framework—combining interpretable trajectory modeling with robust machine learning—provides a template for studying rare
6
outcomes in aging populations. Clinically, findings advocate for integrating social participation assessments into geriatric care protocols to identify at-risk individuals earlier.
Future research should validate these insights through prospective studies incorporating continuous activity monitoring and explore targeted interventions to sustain social engagement. By pairing predictive models with clinical judgment, healthcare systems can better allocate resources to delay or prevent institutional transitions, ultimately enhancing the quality of life for aging populations.