Project plan · Senior Design · Fall 2026
COVID-19 Detection
A web app that screens a cough recording and returns a real-time risk assessment — with longitudinal health tracking for patients and medical professionals.
Team
Drew Quashie · Amanda Ogbonna · Loleyi Oluwatomisin · Richard Alonso Garcia
Advisor
Dr. Zahra Nematzadeh — znematzadeh@fit.edu
Client
Medical Professionals · Potential COVID Patients
Objective
Catch COVID-19 early using the devices people already carry.
Our web app predicts infection by analyzing cough sounds captured in a browser — no specialized hardware required. It returns real-time risk assessments and longitudinal health tracking to support both patients and medical professionals. The system is explicitly a screening tool, not a diagnosis.
[01] Problem
Early detection today depends on clinics and lab tests. A phone or laptop microphone reaches much farther — the goal is to turn a cough recording into a defensible risk assessment, despite noisy audio and skewed class balance.
[02] Related Work
Training runs on the CoughVid dataset (~2,800 recordings with status and metadata). We follow the spectrogram + CNN convention, adding SpecAugment, pitch/time shifting, and additive noise so the model survives real, noisy audio from any browser.
[03] Plan
Phase 1 · Data Preprocessing
- · Extract MFCCs & spectrogram images from CoughVid (~2,800 recordings)
- · Augment: SpecAugment, pitch/time shifting, additive noise
- · 70/15/15 train/val/test split with k-fold cross-validation
Phase 2 · Model Development
- · Build, train, fine-tune CNNs on spectrogram input
- · Transfer learning with ResNet
- · Classify COVID-19 status from audio
Phase 3 · Evaluation
- · Metrics: Accuracy, Precision, Recall, F1, ROC-AUC
- · Measure on an independent test set
Phase 4 · Web Integration
- · Full-stack app: real-time capture via Web Audio API
- · Feature extraction + FastAPI backend inference
- · Risk feedback (Low/Medium/High) with next-step guidance
[04] Milestones · Task Matrix
M1 · Sep 28 — M2 · Oct 26 — M3 · Nov 23
| Task | Owner | Status |
|---|---|---|
| Requirements & project plan documentation | Amanda Ogbonna | Completed |
| CoughVid preprocessing & MFCC/spectrogram pipeline | Drew Quashie | In Progress |
| Audio augmentation (noise, pitch & time shifts) | Drew Quashie | In Progress |
| Baseline PyTorch CNN architecture | Loleyi Oluwatomisin · Amanda Ogbonna | Pending |
| React audio recording interface | Richard Alonso Garcia | Pending |
[05] Challenges
Audio noise variability
High ambient noise and mic discrepancies across browsers demand robust preprocessing, filtering, and augmentation.
Real-time latency
Optimize the pipeline from browser recording and feature extraction to server-side PyTorch for near-instant feedback.
Class imbalance & generalization
Cough-audio distribution biases could hurt accuracy across demographics and symptom severities — address in the data pipeline.
Kickoff meeting
Tue, Aug 25 · 11:15 am — Olin Engineering Complex, room 353.