An Explainable Machine Learning Framework to Inform Integrative Psychosocial Correlates of Sleep Functioning in Adults with Cancer and their Caregivers

Jerry Bonnell, Nikhita Guhan, Thomas C. Tsai, Robert Moulder, Mitsunori Ogihara, Youngmee Kim

Confirmation: I have read and agree with the IEEE BHI 2025 conference submission's policy on behalf of myself and my co-authors.

Keywords: Explainable AI, machine learning, sleep health, oncology dyads, psychosocial predictors, preprocessing, multi-level, multicollinearity

Abstract: Sleep disturbance in cancer patients and caregivers is a substantial challenge in survivorship care. As a dyadic process influenced by daytime experiences, understanding interdependent sleep health by identifying critical dyadic stress regulatory factors and psychosocial predictors is crucial for informing effective interventions. This supervised machine learning (ML) study utilized multi-modal, multi-level data from patients with colorectal cancer and spousal caregivers (n = 149 dyads; 298 persons). The dataset integrated psychosocial characteristics, dyadic cardiovascular and psychological responses to laboratory-induced stress, and 20 self-report and actigraph-derived sleep markers. Preprocessing consisted of three optional techniques: principal component analysis (P) for high dimensionality, correlated feature selection (C) for multicollinearity, and data augmentation (A) for small sample size. Four regression algorithms (Linear Regression, Ridge, Random Forest, and Support Vector Regression) were trained independently for each sleep outcome, evaluating optimal performance across different preprocessing combinations. SHAP analysis was subsequently utilized on best-fitted models to identify key predictors. Linear Regression best predicted caregivers' actigraph-derived interdaily stability ($R^2$=31.2%, via P+C+A), while Ridge best predicted patients' self-reported sleep efficiency ($R^2$ = 18.0%, via P+A), improving non-preprocessed baselines by 30.4% and 11.7%, respectively. SHAP identified dyadic stress regulatory indices as key predictors with complex dynamics and subsequent commonality analysis revealed phase-specific suppressor effects among such indices. This study identified optimal preprocessing combinations that significantly improved sleep prediction in oncology dyads. Complex interactions observed among key dyadic stress regulation indices map out pathways of stress response that shape sleep health, advancing understanding of its interdependent nature and informing targeted intervention strategies.

Track: 7. General Track

Registration Id: GGNTGVNJ5MY

Submission Number: 57