iTalkSign - Hearing-Loss Risk Regression

Mission & problem: iTalkSign builds offline, low-end-device AI for real-time communication for deaf and hard-of-hearing individuals. This project predicts a person's pure-tone-average hearing threshold (PTA4, dB HL) from demographic and noise-exposure survey answers, so the app can flag likely hearing-loss severity without a full audiometric exam.

Dataset: NHANES 2015–2016 (CDC/NCHS open data), merging demographics, an audiometry questionnaire (noise exposure, tinnitus, hearing-aid use), and measured hearing thresholds:

• Demographics — docs | direct download
• Audiometry Questionnaire — docs | direct download
• Audiometry Exam — docs | direct download

Live links

• API (Swagger UI): https://linear-regression-model-nrpk.onrender.com/docs
• YouTube demo video (≤7 min): https://youtu.be/rLYb9EgqlDI

Repository layout

linear_regression_model/
├── summative/
│   ├── linear_regression/
│   │   ├── multivariate.ipynb      # EDA, feature engineering, model comparison, best-model save
│   │   ├── build_notebook.py       # generates multivariate.ipynb programmatically
│   │   ├── data/                   # raw NHANES XPT files (DEMO_I, AUQ_I, AUX_I)
│   │   ├── best_model.joblib       # best-performing regressor (RandomForestRegressor)
│   │   ├── scaler.joblib, feature_columns.joblib
│   │   └── *.png                   # saved plots
│   ├── API/
│   │   ├── main.py                 # FastAPI app: /predict, /retrain, CORS, Pydantic validation
│   │   ├── prediction.py           # loads model, builds feature vector, predicts
│   │   ├── requirements.txt
│   │   └── Procfile
│   └── FlutterApp/                 # single-page Flutter app (hearing_predictor)
├── pyproject.toml / uv.lock
└── README.md

Running the notebook / retraining locally

cd linear_regression_model
uv sync
uv run jupyter nbconvert --to notebook --execute --inplace summative/linear_regression/multivariate.ipynb

Running the API locally

cd linear_regression_model/summative/API
uv run --project ../.. uvicorn main:app --reload
# Swagger UI at http://127.0.0.1:8000/docs

Running the Flutter app

cd linear_regression_model/summative/FlutterApp
flutter pub get

kApiBaseUrl in lib/main.dart is already set to the deployed Render API above. To point it at a local API instead, change it to http://10.0.2.2:8000 for the Android emulator or http://127.0.0.1:8000 for iOS simulator/desktop. Then:

flutter run            # pick your connected device/simulator

The app is a single page: enter the 13 input values (dropdowns for categorical fields, text fields for age and income-to-poverty ratio), tap Predict, and the predicted hearing threshold (dB HL) and severity band are shown, or a validation/error message if a value is out of range or missing.

CORS configuration rationale

The API restricts allow_origins to explicit local/dev origins (no allow_origins=["*"]), since one of the two POST endpoints (/retrain) mutates the deployed model and should not be reachable from arbitrary third-party web origins. Native Flutter mobile builds don't send a browser Origin header so they're unaffected; the origin list only matters for a Flutter web build or browser-based Swagger testing. allow_methods is limited to GET/POST (all this API implements) and allow_headers to Content-Type; allow_credentials is left off since the API uses no cookies/auth headers that need cross-origin exposure.