AN INTERPRETABLE AI-DRIVEN FRAMEWORK FOR MONITORING BEHAVIOR CHANGES IN CARE ENVIRONMENTS

Abstract

The growing prevalence of functional and cognitive impairments among older adults presents significant societal challenges, particularly because these conditions often remain undetected until they progress into more serious health concerns. Traditional clinical assessments, which rely primarily on self-reported data, can be hindered by recall bias and subjectivity, limiting their utility for early detection. To address these gaps, this thesis proposes an interpretable, AI-driven framework that integrates ambient sensor data with machine learning (ML) and large language models (LLMs) to support the identification of behavioral changes in smart home environments. Rather than replacing self-report, this approach aims to complement it, with the ultimate goal of enabling timely clinical intervention and promoting aging-in-place with dignity and autonomy. Despite recent advancements, current approaches to behavior anomaly detection face critical limitations, including underutilization of temporal dependencies, narrow focus on intra-activity anomalies, reliance on labeled data, poor model generalizability, and lack of interpretability. This research addresses these challenges by proposing a novel, multi-component framework that integrates: (1) inverse reinforcement learning (IRL) models for scalable, label-efficient behavior change detection; (2) Transformer-based architectures with transfer learning to improve generalizability and mitigate cold-start issues; (3) a synthetic data generation model (BehavGAN) to augment training data diversity; and (4) an LLM-based interpretability layer to translate activities of daily living (ADL) logs and anomaly detections into human-readable, clinician-friendly summaries. Grounded in Fogg’s Behavior Model, the proposed system captures both point and collective anomalies by modeling inter-activity and temporal patterns of ADLs. Experiments on public smart home datasets (CASAS-Twor, CASAS-Aruba, and Kastaren) demonstrate high performance across modules: over 90% recall in behavior change detection with an 11% false positive rate, effective cross-user generalization, and successful near real-time monitoring capabilities. LLM integration further bridges the gap between quantitative sensor data and qualitative clinical insight, while human-in-the-loop (HITL) mechanisms and risk mitigation strategies address challenges related to bias, hallucinations, and ethical oversight. This thesis contributes a scalable, explainable, and ethically aware solution to preventive geriatric care, demonstrating how AI and generative technologies can be responsibly deployed in eldercare to enhance quality of life, reduce healthcare burden, and empower clinicians with actionable, real-time insights.