Computer Science Department , PhD Dissertation Defense Kavin Chandrasekaran "  Deep Learning Approaches for Passive, Multi-Scale Activity Recognition for Healthcare"

DATA SCIENCE

PhD Dissertation  Defense

Student Name: Kavin Chandrasekaran

Date: Wednesday, July 23rd, 2025

Time: 1pm to 3pm Eastern Time

Fuller Beckett Conference Room

Zoom Link: https://wpi.zoom.us/j/99250582988

Committee:

• Advisor: Dr. Emmanuel Agu, Professor, Computer Science, WPI
• Co-Advisor: Dr. Elke Rundensteiner, Professor, Data Science, WPI
• Committee Member: Dr. Nima Kordzadeh, Assistant Professor, Business School, WPI
• Committee Member: Dr. Victor Robles, Senior Member of Technical Staff | AI, AMD

Abstract:

Passive Human activity recognition (HAR) from Inertial Measurement Unit (IMU) sensor data is important in many critical healthcare domains including patient health, rehabilitation, and mental health monitoring. The three main types of ambulatory patient activities that need to be monitored are: simple activities (standing, walking), activity transitions (sit to stand), and complex activities (going to the bathroom, cooking). While prior work has extensively researched simple activity recognition from IMU and smartphone sensor data, complex activities and transitions, which this dissertation focuses on, are under-researched.

Transitions: Monitoring the occurrence and duration of transitions between physical activities such as sitting to standing can provide valuable clues for assessing patient mobility, fall-risk, and health. Transitions are challenging to detect because they are brief (a few seconds), yielding limited data for analyses.

Complex activities: The types, duration, and frequency of complex activities performed daily are valuable indicators of patients’ physical and mental health. Complex activity recognition (CAR) faces two major challenges: i) high intra-class and inter-subject variability in activity performance styles and ii) scarcity of labeled datasets. Variability occurs because the constituent simple activities can occur concurrently, be interleaved with other activities, or in different sequences each time the same complex activity is performed. Such variability poses challenges to pattern matching approaches such as machine learning. In addressing these challenges, this dissertation makes the following contributions:

1. A deep transition recognition model: Incorporating a Gated Recurrent Unit (GRU) with a self-attention mechanism for robust recognition of short-duration transitions.
2. A new comprehensive CAR dataset (MUSIC-CAR): With labels for complex activities as well as fine-grained labels of the constituent simple activities, was collected and publicly released.
3. NLP-inspired CAR models: That draw analogies between activity sequences and natural language:
   1. CARTMAN (Complex Activity Recognition using Topic Models for feAture geNeration): Leverages topic models to generate powerful features that are classified by a custom convolutional-recurrent network.
   2. P-HART (Personalized-Health Activity Recognition Transformer): A state-of-the-art, personalized Transformer architecture that learns individual complex activity performance nuances to address intra-class and inter-subject variability

The proposed models are part of a comprehensive framework for understanding nuanced human activities, moving beyond simple activity recognition to personalized, proactive healthcare monitoring.