Computer Science Department , PhD Dissertation Defense, Atifa Sarwar "Machine Learning Models for Passive Pre-symptomatic Detection of Covid-19 from Smart Wearable Data" "

Atifa Sarwar

PhD Candidate

WPI– Computer Science

Monday, July 15, 2024

Time: 10:00AM – 11:00AM

Location: Fuller Labs 141

Zoom: https://wpi.zoom.us/j/2635545596

Committee Members:

Dissertation Advisor: Prof. Emmanuel Agu, WPI– Computer Science

Prof. Erin Solovey, WPI – Computer Science

Prof. Yanhua Li, WPI– Computer Science

External Committee Member: Prof. Bashima Islam, WPI – Electrical and Computer Engineering

Abstract:  

Covid-19, an infectious Influenza-Like Illness (ILI) caused by Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) recently led to a global pandemic. Covid-19 vaccinations have significantly reduced adverse outcomes, including hospitalizations and deaths. However, due to the mutating nature of the virus, Covid-19 remains a global concern. Passive, low-burden methods for early Covid-19 detection can inform proactive interventions and improve outcomes. This doctoral dissertation investigates Machine Learning models for detecting Covid-19 from abnormal deviations in longitudinal physiological signs (such as heart rate and steps) collected passively from consumer-grade smart wearables, without requiring patient medical history or human-reported symptoms. Three categories of ML models are proposed: 

I)     Leveraging physiological Signs and Features Predictive of Covid-19: We explored traditional ML algorithms to predict Covid-19 infection from heart rate, physical activity (steps), and sleep pattern features. Additionally, we proposed CovidRhythm, a deep Gated Recurrent Unit (GRU)-based Multi-Head Self-Attention (MHSA) model that identifies Covid-19 infection from disruptions in biobehavioral rhythmic features, before biological symptoms manifest.

II)     Addressing Inter-individual Differences in Vital Sign Manifestations with minimal data: Given significant individual differences in vital sign manifestations that hinder the generalizability of AI models, we proposed MetaCovid. This deep adaptation framework leverages meta-learning to address inter-subject differences with minimal data, enabling the detection of Covid-19 before symptom onset.

III)    Earliest Possible Covid-19 Infection Detection: Delayed diagnoses were a major contributor to the Covid-19 pandemic, as most testing was reactive, occurring only after symptom onset. Early detection of Covid-19 could substantially reduce infection rates by 35.7% and decrease mortality by 46.2%. We explored the earliest possible Covid-19 infection detection through EarlyDetect, a Reinforcement Learning-based Early Time Series Classification method, and ECovGNN, a Graph Neural Network-based method that boosts model performance by leveraging intra- and inter-subject similarities within vital signs for early Covid-19 detection. 

This dissertation achieves breakthroughs in pre-symptomatic Covid-19 screening by demonstrating effective detection three days before biological symptom onset using 72-hour non-overlapping windows of physical activity, circadian rhythm, and physiological features. We believe that findings of this dissertation will pave the way for timely disease detection, clinical management, and improved public health response to future infectious disease outbreaks.