Room-Aware Transformer for Activities of Daily Living Recognition in Adults with Mild Neurocognitive Disorder Using Ambient Sensors: Development and Evaluation Study

ABSTRACT Background: Unobtrusive recognition of activities of daily living (ADLs) from ambient sensors may support longitudinal assessment of everyday functioning in adults with mild neurocognitive disorder (mild NCD). However, real-world activity recognition is complicated by heterogeneous home layouts, sensor configurations, routines, and class distributions. Objective: This study evaluated whether a room-aware Transformer that preserves location-specific sensor structure improves participant-level ADL recognition relative to a global Transformer and a stacked unidirectional long short-term memory (LSTM) network. Methods: Following pilot feasibility testing with two adult volunteers, we analyzed five homes of adults with mild NCD, comprising 92,765 minute-level observations across 67 annotated home-days. Motion, temperature, illuminance, humidity, and door-contact data were aligned to participant activity logs. The original labels were harmonized into eight categories. Models were developed separately for each home using 10-, 30-, and 60-minute windows. Complete calendar days were assigned to mutually exclusive training, validation, and chronological test sets, and no sequence crossed a day or partition boundary. The same partitions and evaluation procedures were used for all three architectures. Performance was assessed on held-out test days using accuracy, per-class F1 score, and micro- and macro-averaged one-versus-rest area under the receiver operating characteristic curve (AUC). Results: The room-aware Transformer achieved the highest or tied accuracy in 9 of 15 home-window comparisons. Its accuracy ranged from 0.27 to 0.57, micro-AUC from 0.57 to 0.82, and macro-AUC from 0.54 to 0.78. Across the 15 settings, mean accuracy was similar for the room-aware and global Transformers and higher than for the LSTM. Mean micro-AUC and macro-AUC were marginally higher for the global Transformer than for the room-aware Transformer. The room-aware model showed its clearest gains in selected contexts, including Homes C and E at 10 minutes and Home A at 60 minutes, but no architecture or temporal window was uniformly optimal. Sleep/Rest was the most consistently recognized activity, whereas low-support activities frequently had low F1 scores despite moderate or high AUC values. Conclusions: Room-specific feature processing provided context-dependent gains in participant-level ADL classification, particularly for accuracy in several home-window settings. The marked variation across homes, activities, and temporal windows supports personalized model development, class-sensitive evaluation, and validation in larger and more diverse residential samples before clinical use. Keywords: activities of daily living; mild neurocognitive disorder; ambient sensors; deep learning; artificial intelligence