Machine Learning-Based Survival Prediction in Advanced Cancer Using Actigraphy-Derived Rest-Activity, Sleep, and Routine Clinical Data

Predicting survival remains a critical challenge in oncology and palliative care, with few reliable tools available. In this study, we explored whether combining wrist actigraphy, patient-reported sleep diaries, and standard clinical parameters could improve prognostic accuracy in advanced cancer. Fifty outpatients with an anticipated survival of under one year participated, wearing an Actiwatch® device for eight days and recording sleep patterns. Data from 66 variables were analyzed using univariate and Lasso-regularized multivariate regression to identify key predictors of survival. Of the 49 patients who completed the study, 34 died within a year. Although 42 participants showed disrupted rest-activity rhythms (dichotomy index I < O ≤ 97.5%), this measure alone did not predict survival in univariate analysis. The optimized Lasso model successfully separated patients into short- and long-survival groups (log-rank p < 0.0001). Longer survival was linked to factors including wake-up time, sleep efficiency, subjective sleep quality, clinician-estimated prognosis, global health score, and hemoglobin levels. Conversely, shorter survival was associated with sleep disturbances, elevated neutrophils, serum urea, creatinine, and C-reactive protein. These findings indicate that integrating machine learning with circadian activity data, patient-reported sleep, and routine clinical measurements may offer a promising approach for individualized prognostic assessment in advanced cancer.