Abstract Objective: Inadequate daytime illumination and excessive nighttime light exposure in home environments are key risk factors for sleep disturbances and circadian rhythm disruption in older adults. This pre-post pilot study aimed to evaluate the feasibility and acceptability of a machine learning-based intelligent lighting intervention among older adults and obtain preliminary evidence of its efficacy on sleep quality. This study also evaluated the external validity of a predictive model of sleep efficiency trained in a laboratory setting. Methods: A single-group, pre-post design was employed. In phase 1, a multimodal laboratory data set (n = 129), comprising demographic data, sleep-related questionnaire scores, and detailed light exposure characteristics, was used to develop machine learning models—including Classification and Regression Trees (CART), Random Forest, and XGBoost—to predict sleep efficiency and melatonin suppression. Model performance was evaluated using a train–test split and 5-fold cross-validation; XGBoost was selected for subsequent application due to its superior generalization performance for sleep efficiency. In phase 2, 6 healthy older adults aged 60–75 years (n = 5 completed the study) were recruited for a 6-week in-home study consisting of a 2-week baseline, a 2-week intervention, and a 2-week autonomous adjustment period. The intervention implemented a dynamic lighting strategy through a smart lighting system: high illuminance and high correlated color temperature (CCT) during daytime, and low illuminance and low CCT at night. Sleep monitoring devices and questionnaires were used to assess the primary outcome—sleep efficiency—and secondary outcomes, including total sleep time and wake after sleep onset. Repeated -measures ANOVA was used to analyze within-subject changes across 3 time points: baseline (T0), mid-intervention (T1), and end-of-intervention (T2). Interviews were conducted during the adjustment period to understand user experiences and preferences. Results: The intervention significantly optimized the home lighting environment, achieving daytime eye-level illuminance of ∼1000 lx and reducing nighttime levels to below 50 lx. Compared with baseline, daytime illuminance during the intervention increased from typical levels of 0–350 lx to peak exposures near 1000 lx, while evening and pre-sleep illuminance decreased from up to ~300–0.05), likely due to limited statistical power (n = 5). When the XGBoost model, developed from laboratory data, was applied to an independent in-home data set, predicted and measured sleep efficiency exhibited a strong, statistically significant correlation (r = 0.987, P = 0.001, n = 5), supporting the model’s external validity. During the autonomous adjustment phase, participants systematically reduced daytime illuminance and increased nighttime levels relative to the standardized protocol, frequently reporting the prescribed settings as “too bright during the day” and “too dim at night.” Nevertheless, their preferred settings maintained a healthier day–night illuminance contrast than baseline, indicating both heightened awareness of circadian lighting principles and an ongoing tension between physiological optimality and subjective comfort. Conclusions: This pilot study demonstrates that a personalized, machine learning-guided intelligent lighting intervention is technically feasible and subjectively acceptable in aging-in-place home environments. Although statistically significant improvements in sleep outcomes were not detected, the observed directional trends and robust laboratory-to-home generalization of the predictive model provide encouraging preliminary evidence. Future research should prioritize larger, controlled trials and the development of adaptive algorithms that co-optimize physiological efficacy and promote long-term adherence.
Li et al. (Tue,) studied this question.