Detecting functional performance deterioration in activities of daily living using Conditional Preference Networks and Sequential Pattern Mining

Abstract

Aging in place depends on maintaining the ability to perform Activities of Daily Living (ADLs), yet clinicians can miss early functional decline because assessments are infrequent, subjective, and rarely capture performance in natural settings. This thesis examines whether a skeleton-based sensing pipeline can detect subtle deterioration within an individual by comparing ADL execution under baseline and aging-simulated conditions. Two Orbbec Femto depth cameras recorded ADL trials and produced 3D skeletons with 32 joints. Temporal encoding converted joint trajectories into activity sequences for two evaluators: (1) Sequential Pattern Mining in the SCOUT codebase, which outputs confidence scores that quantify similarity to baseline, and (2) a Conditional Preference Network model of baseline movement preferences, which uses ordering queries to compute an ADL-fit score. Across no-, low-, medium-, and high-gear conditions and multiple ADLs, both evaluators reflected expected degradation with increasing gear while revealing participant- and activity-specific variability. However, frame loss and limited per-participant training data reduced sensitivity.