Early Detection of COVID-19 Using Wearable Device Data

Overview

In this project, I explored whether physiological data from consumer wearable devices such as heart rate, heart rate variability, sleep, activity, etc… can be used to predict COVID-19 infection before symptom onset.

  • Implemented a sliding-window algorithm on publicly available wearable datasets
  • Achieved early detection performance of around 80% up to ~5 days before symptom onset (proof-of-concept level)
  • Built an end-to-end pipeline from data ingestion to anomaly detection
  • Turned the work into an award-winning startup concept, winning the Big Idea Challenge at the University of North Dakota and reaching the finals of the Middleton School of Entrepreneurship business plan competition

Problem Background

During the COVID-19 pandemic, most screening relied on:

  • Spot checks (temperature, pulse oximetry)
  • Lab tests (PCR, antigen)
  • Self-reported symptoms

These are often:

  • Late (after symptoms appear)
  • Infrequent (once a day, or only when people feel sick)
  • Sometimes inconvenient or expensive

At the same time, many people already wear devices that continuously capture signals closely tied to early illness:

  • Resting heart rate (RHR)
  • Heart rate variability (HRV)
  • Sleep disruption
  • Activity patterns

Previous studies have shown that changes in these metrics can precede SARS-CoV-2 symptom onset by several days. My goal was to turn these ideas into a concrete, reproducible modeling pipeline, and then explore what a real product around this might look like.

Technical Approach

Data & Signals

I worked with publicly available wearable datasets where:

  • Participants wore smartwatches or similar devices for multiple months
  • The data included RHR, HRV, steps, and sleep metrics
  • COVID-19 infection dates and symptom onset days were known or approximated

This allowed me to:

  • Align time series around infection (“day 0” = symptom onset / diagnosis)
  • Study patterns in the days leading up to infection
  • Prototype detection algorithms without needing to run my own large prospective study

Modeling

  1. Preprocesses the time-series data
    • Normalize signals (e.g., resting heart rate deviations from personal baseline)
    • Handle missing data and noisy segments
  2. Slices the data into overlapping windows
    • Windows of fixed length (e.g., several days)
    • Slide forward one day at a time
    • Each window is labeled relative to infection status (pre-infection, infection, or healthy period)
  3. Extracts features per window
    • Baseline-adjusted RHR
    • Short-term and long-term HRV statistics
    • Sleep duration and interruptions
    • Activity changes (e.g., steps)
  4. Detects anomalies
    • Compares each window to that person’s usual patterns
    • Flags periods where signals deviate significantly
    • Aggregates anomalies to mark potential “early warning” zones
  5. Evaluates performance
    • Measures how often warning periods overlap with the days leading up to infection
    • Reports sensitivity/recall for detecting infection in advance
    • Compares performance across different feature sets (RHR only vs. multi-signal)

The modeling approach is directly inspired by prior work (e.g., RHR-Diff and HROS-AD style methods) but focuses on multi-signal integration and sliding-window framing as a clean way to think about early detection.

Developing the Venture Around the Technology

Although this started as a research exercise, it quickly grew into a startup-oriented concept.

Big Idea Challenge – 1st Prize

I used this project to participate in the Big Idea Challenge at the University of North Dakota and won first place.

In that context, I:

  • Defined the product as a digital health platform that:
    • Integrates with existing wearables
    • Computes early-risk scores for COVID-like infections
    • Provides dashboards for users and potentially clinicians
  • Highlighted key differentiators:
    • Focus on pre-symptomatic detection, not just monitoring
    • Ability to integrate new signals (temperature, SpO₂) as devices evolve
    • Potential extension beyond COVID-19 to broader infectious diseases
  • Presented:
    • A business model (subscription / B2B2C SaaS)
    • A rough product roadmap (from proof-of-concept to validated medical device)
    • Early thoughts about regulatory and clinical validation

Some of the slides I used in that pitch included:

  • An “80% prediction 5 days early” result summary
  • A market overview of the wearable and remote patient monitoring space
  • A competition and unique value proposition (UVP) slide showing how this sits next to Apple, Google, and other digital health companies
  • A roadmap slide outlining steps toward integrating with third-party devices, developing dedicated hardware, and pursuing FDA clearance

Middleton School of Entrepreneurship – Finalist

Later, I expanded the idea into a 30-page business plan and became a finalist in the UND Middleton School of Entrepreneurship Business Plan Competition.

In that process, I:

  • Conducted market analysis (TAM/SAM/SOM, key players, trends in remote patient monitoring and digital biomarkers)
  • Defined a go-to-market strategy (clinical partners, pilot projects, early adopters)
  • Built financial projections
  • Outlined a regulatory and validation path:
    • Clinical trials
    • Partnerships with hospitals / clinics
    • Steps toward FDA clearance and broader deployment

I am not publishing the business plan text here, but it is available upon request in the right context (e.g., for serious collaboration or review).

UND Biomedical Research Conference Poster (2021)

Where I’d like to take this next

Although this specific project grew out of the first wave of COVID-19, the ideas behind it are still very relevant:

  • Using wearables for digital biomarkers
  • Developing early warning systems for infectious disease
  • Bridging technical modeling, clinical validation, and productization

If you work on:

  • Digital health / remote monitoring
  • Illness prediction from biosignals
  • Digital biomarkers and anomaly detection
  • Or you are building products in this space

and you think our interests align, feel free to reach out, my contact info is at the bottom of this page.