AI-Powered Health Monitoring (CPU-only)

Streamlit + FastAPI app for vitals monitoring and fall detection. Designed to run on laptops with CPU only (no GPU required).

---

Features

• Vitals Stream (CSV/API): real-time charts, rule-based risk, alert table.
• Video Test (Upload): upload a short clip → classify fall/normal using a Random-Forest on sliding-window features.
• Event-aware decision rule: FALL if mean(prob) ≥ mean_thresh OR p(peak_pct) ≥ peak_thresh.
• Tunable thresholds from the UI (no restart) or via environment variables.
• Dockerized (UI + API) or pure Python virtualenv.
• Education-ready: simple baselines, reproducible, CPU-friendly.

---

Architecture

• Streamlit UI (streamlit_app/app.py)
  • Vitals Stream tab: KPI tiles, rolling charts, alerts table.
  • Video Test (Upload) tab: file uploader, threshold controls, results panel.
• FastAPI Backend (src/api.py)
  • /predict/video — classify a clip; returns label, score, stats, and params used.
  • /stream/* — simple vitals stream helpers for the dashboard.
• Inference (src/inference.py)
  • Loads models from models/ if present; otherwise uses safe heuristics.
  • Implements the mean + pXX event-aware rule.

Tip: See src/train_baselines.py for CPU-friendly training pipelines and src/preprocess_video.py for frame extraction.

---

Repository layout (typical)

.
├── data/
│   ├── raw/
│   │   ├── toy/                  # toy_fall_sim.mp4 + extracted frames
│   │   └── urfd/                 # (optional) download datasets here
│   └── vitals/                   # vitals CSVs (simulated or real)
├── docs/
│   └── img/                      # screenshots for README
├── models/                       # saved .joblib files
├── notebooks/                    # EDA + training walkthroughs
├── src/
│   ├── api.py                    # FastAPI server
│   ├── inference.py              # model loading + prediction logic
│   ├── preprocess_video.py       # frame extraction (CPU-friendly)
│   ├── simulate_vitals.py        # HR/Temp simulator
│   ├── train_baselines.py        # vitals + video baselines
│   └── fetch_urfd.py             # dataset fetch helper (URLs + licensing notes)
├── streamlit_app/
│   └── app.py                    # dashboard
├── Dockerfile.api
├── Dockerfile.app
├── docker-compose.yml
├── Makefile
├── README.md
├── DATASETS.md
├── MODEL_CARD.md
└── requirements.txt

---

Quickstart (venv)

python -m venv .venv
# macOS/Linux
source .venv/bin/activate
# Windows PowerShell
# .\.venv\Scripts\Activate.ps1

pip install --upgrade pip
pip install -r requirements.txt

# Generate vitals & frames (CPU-only)
python src/simulate_vitals.py --out data/vitals/vitals_train.csv --minutes 120 --seed 42
python src/preprocess_video.py --video data/raw/toy/toy_fall_sim.mp4 --out data/raw/toy/frames

# Train simple baselines (saves models/*.joblib)
python src/train_baselines.py --vitals data/vitals/vitals_train.csv --frames_dir data/raw/toy/frames --model_out models

# Run services
uvicorn src.api:app --reload --port 8000
streamlit run streamlit_app/app.py

Open the UI → API Mode (http://localhost:8000) → Video Test (Upload) → upload a 10–30s clip.

---

Thresholds (UI or env)

You can tune thresholds from the Video Test (Upload) tab — no restart required.
Alternatively, set environment variables before starting the API:

export VIDEO_MEAN_THRESH=0.45
export VIDEO_PEAK_THRESH=0.60
export VIDEO_PEAK_PCT=95

The API echoes the parameters actually used in the JSON response under "params" so students can record settings.

---

Docker (optional)

docker compose up --build
# UI:  http://localhost:8501
# API: http://localhost:8000

---

API examples

Classify a clip with custom thresholds:

curl -s -X POST "http://localhost:8000/predict/video"   -F file=@data/raw/toy/toy_fall_sim.mp4   -F mean_thresh=0.45 -F peak_thresh=0.60 -F peak_pct=95 | jq

Health check:

curl http://localhost:8000/health

---

Datasets & ethics

See DATASETS.md for recommended sources (URFD, SisFall, UP-Fall, PhysioNet) and licensing notes.
Always document consent, anonymize where possible, and include a short data sheet with sources, preprocessing, splits, biases, and limitations.

---

Teaching alignment

This repo maps to a standard 16-week capstone:

• Milestone 1: Proposal + system diagram.
• Milestone 2: Data/EDA + vitals simulator + sample video frames.
• Milestone 3: Baseline models + metrics.
• Milestone 4: Integrated MVP (Streamlit + FastAPI).
• Milestone 5: Final demo (video upload) + documentation.

---

Screenshots

Place PNGs in docs/img/ and reference here, e.g.:

![Vitals stream](docs/img/vitals_stream.png)
![Video test](docs/img/video_test.png)

---

License

MIT © 2025 Sitaram Ayyagari. See LICENSE for details.