Event
Ph.D. Research Proposal: Hossein Khayami
Friday, September 4, 2026
10:00 a.m.
AVM 2328
Souad Nejjar
301 405 8135
snejjar@umd.edu
ANNOUNCEMENT: Ph.D. Research Proposal Exam
Name: Hossein Khayami
Committee:
Professor Hernisa Kacorri (Chair)
Professor Behtash Babadi
Professor Ang Li
Date/time: Friday, September 4, 2026 at 10 AM
Location: AVM 2328
Title: Personalizing Activity Tracking: Enabling Older Adults to Fine-Tune Their Wearables
Abstract: Commercial wearable activity trackers are increasingly used for health and behavioral monitoring, yet their adoption among older adults remains comparatively low. One factor contributing to this gap is the lower activity recognition accuracy of current trackers for older adults. This issue is partly rooted in common practices for developing and evaluating machine learning models for human activity recognition (HAR). These models are often trained and benchmarked on datasets collected from younger adults performing researcher-scripted activities, which may not adequately represent the variability and activities encountered by older adults in their daily lives.
This proposal investigates the challenges of improving activity tracking for older adults through a series of studies addressing data collection, performance disparities, personalization, and recognition of meaningful activities. First, we investigate the challenges of collecting and labeling activity data from older adults in free-living environments, where researchers cannot continuously observe and annotate activities. We find that triangulating users’ self-reports with predictions from a thigh-worn sensor can improve label quality and provide a practical basis for developing personalized models. However, obtaining high-quality labels remains labor-intensive for both users and researchers, and stricter labeling criteria further reduce the amount of usable data. We therefore examine how the data collection and labeling process can be made more practical and user-centered.
Next, we investigate whether advances in machine learning for HAR are closing the performance gap between younger adults, who are predominantly represented in existing benchmarks, and older adults in real-world settings. Across multiple modern HAR architectures, we show that improvements on younger-adult benchmarks do not translate proportionally to older adults, leaving a persistent and sometimes widening performance gap. These findings motivate the investigation of personalization as a means of adapting HAR models to individual older adults. We compare multiple personalization strategies, including a multi-stage fine-tuning approach, to understand how effectively models can be adapted using limited user-provided labeled data.
Finally, improving low-level posture-based activity recognition alone may not fully address what older adults want from activity trackers. Building on our co-design findings, which suggest that higher-level activities could be a valuable addition to existing tracking capabilities, we propose methods for inferring semantically meaningful activities beyond postures and step counts from streams of raw wearable sensor data. Together, these studies follow a progression from asking whether activity can be measured reliably, to how recognition can be personalized with realistic amounts of user effort, and ultimately to how activity tracking can become more relevant to users’ everyday lives. This progression aims to provide a better foundation for developing wearable activity tracking systems that are accurate, practical, personalized, and meaningful for older adults.
