Real-time bacterial risk monitoring for hospital sink traps

Built at MPCHacks 2026 — a low-cost IoT + ML system that continuously monitors hospital drain water to flag pathogen outbreak conditions before patients are exposed.

Hospital water infrastructure — sinks, drains, utility rooms — is a primary reservoir for opportunistic pathogens and Antibiotic Resistance Genes (ARGs) in clinical settings like NICUs. A 2026 study by Bourdin et al. found S. maltophilia in 67% of NICU drain samples.

The only way to confirm what's growing — qPCR and metagenomic sequencing — is slow, manual, and far too expensive to run continuously. Outbreaks get detected after patient exposure, leaving clinical staff entirely reactive.

In the US alone, antibiotic-resistant infections cause 2.8 million infections and 35,000 deaths yearly, costing over $4.6 billion annually for just the six most aggressive pathogens (CDC).

BioWatch is a physical sensor node that sits in a sink's P-trap and continuously monitors turbidity and temperature — the environmental conditions that precede pathogen proliferation. When conditions cross risk thresholds, alerts escalate through a tiered system (Passed → Warn → Warn Lvl 2 → PANIC) so staff know exactly which sinks need attention and how urgently.

ESP32-S3 Node                Raspberry Pi Backend              Dashboard
┌──────────────┐           ┌──────────────────────┐         ┌──────────────┐
│ DS18B20 temp │──┐        │  FastAPI server      │         │ React + Vite │
│ Turbidity    │──┤ WiFi   │  ┌─────────────────┐ │  WS     │ Live gauges  │
│ sensor       │──┼──────► │  │ ML Risk Engine  │ │ ◄─────► │ Sparklines   │
│              │  │  WS    │  │ (Decision Tree) │ │         │ Flag log     │
└──────────────┘  │        │  └─────────────────┘ │         │ Tier alerts  │
                  │        └──────────────────────┘         └──────────────┘
                  │                    │
                  │                    │
                  │        ┌───────────▼──────────┐
                  └──────► │ ILI9341 240×320      │
                           │ on-device display    │
                           └──────────────────────┘

Data flow: ESP32 (temp + turbidity) → WebSocket → Backend (risk scoring + SQLite) → WebSocket push → Dashboard (display only) + Pi display

BioWatch/
├── deployment/               # ESP32-S3 Arduino firmware
│   └── deployment.ino        #   Sensor reads → WiFi → WebSocket JSON
├── server.py                 # FastAPI backend (port 8001) — ML inference service
├── calculate_real_time_risk.py  # AMRRiskEngine — deterministic risk scorer
├── data_gen.py               # Pi Zero ILI9341 display client
├── simulation/
│   └── esp32_sensor_data.py  # Simulated sensor data generator (CSV + JSON)
├── ML/                       # Training pipeline (offline)
│   ├── retrieval.py          #   USGS data retrieval
│   ├── water_quality_portal.py  # WQP data fetch
│   ├── build_features.py     #   24h/48h feature engineering
│   ├── compiling_training_data.py  # Match env data ↔ biological data
│   └── visualize_tree.py     #   Decision tree visualization
├── src/
│   ├── App.jsx               # Main dashboard component
│   ├── main.jsx              # React entry point
│   ├── config.js             # Thresholds, tier definitions, constants
│   ├── seed.js               # Demo sensor seed data
│   ├── mockBackend.js        # Client-side mock for offline dev
│   ├── styles.js             # Dashboard CSS-in-JS
│   └── components/
│       ├── StatTile.jsx      # Temp/turbidity card with sparkline
│       ├── Sparkline.jsx     # Recharts mini line graph
│       ├── FlagMeter.jsx     # Escalation progress bar
│       ├── TierBadge.jsx     # Risk tier indicator
│       └── WaterPipe.jsx     # Animated pipe visualization
├── package.json
├── vite.config.js
└── requirements.txt

Total per sink node: ~$25

• USGS NWIS — 103,526 instantaneous water quality records (Station 01646500, Potomac River near Washington, DC)
• Zenodo — Shenandoah Valley Water Transect Data (Graber Neufeld, 2025) — biological validation targets

1. Data Collection — Retrieve raw 15-min interval flow, temperature, and turbidity data from USGS via retrieval.py and water_quality_portal.py
2. Feature Engineering — Convert raw readings into 24h/48h trend deltas using build_features.py (how do conditions change over time?)
3. Training Set — Match 103 days of environmental data against biological targets using compiling_training_data.py
4. Feature Selection — mRMR (Maximum Relevance, Minimum Redundancy) to identify the 4 most informative, non-redundant indicators
5. Classification — Decision tree for interpretable risk thresholds that run directly on edge hardware
6. Deployment — calculate_real_time_risk.py produces a deterministic 0–100 risk score from live sensor input

Calibrated against regional environmental medians (T_median = 16.5°C, Turb_median = 8.2 FNU):

We initially explored LASSO for feature selection, but it consistently eliminated turbidity — our single most important signal — because the relationships between environmental parameters and biological outcomes are nonlinear. A decision tree handles this naturally and, critically, produces interpretable thresholds — exactly what you need when the model triggers a real alert on real hardware.

The dashboard uses a flag-based escalation system. A flag fires when turbidity ≥ 4.5 NTU AND temperature ≥ 26°C simultaneously, with a 30-second debounce. Flags accumulate in a sliding 6-hour window:

• Node.js ≥ 18
• Python ≥ 3.10
• Arduino IDE (for ESP32 firmware)

npm install
npm run dev        # Vite dev server on :5173

The dashboard works offline with the built-in mock backend for demo purposes.

pip install -r requirements.txt
python server.py   # FastAPI ML service on :8001

1. Install Arduino libraries: WebSockets, ArduinoJson, OneWire, DallasTemperature
2. Update WiFi credentials and Pi IP in deployment/deployment.ino
3. Flash to ESP32-S3

# Enable SPI: sudo raspi-config → Interface Options → SPI → enable → reboot
pip3 install adafruit-circuitpython-rgb-display adafruit-blinka pillow websockets
python data_gen.py

We ran a dilution series experiment to characterize the turbidity sensor's response curve, fitting it with R², establishing the limit of detection, and measuring repeatability across trials. This calibration maps raw ADC values to meaningful NTU readings using the relationship:

NTU = max(0, (3.0V - V_measured) / 3.0V × 100)

• U.S. Geological Survey. National Water Information System (NWIS) Instantaneous Values Data. Station 01646500, Potomac River near Washington, DC. Retrieved May 30, 2026.
• Graber Neufeld, D. (2025). Shenandoah Valley Water Transect Data [Data set]. Zenodo. https://doi.org/10.5281/zenodo.15865614

• CDC — Antibiotic Resistance Threats and Partner Estimates
• Bourdin, T. et al. (2026). Abundance of opportunistic pathogens in NICU drain microbiomes. mSystems. https://journals.asm.org/doi/10.1128/msystems.01546-25

Hardware: ESP32-S3 · DS18B20 · SEN0189 Turbidity Sensor · Raspberry Pi Zero W · ILI9341 TFT

Backend: Python · FastAPI · scikit-learn · joblib · SQLite · WebSockets

Frontend: React 18 · Vite · Recharts · Lucide React

ML: mRMR feature selection · Decision tree classification · Spearman rank correlation (validation)

Built at MPCHacks 2026