Low-Cost Triage Device With TinyML

ISEF Category: Translational Medical Science

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

A simple device can miss a sick patient if it only checks one sign at a time. That is why clinics often watch oxygen, temperature, and breathing together. You can build a small triage system that copies that idea. The hard part is making it accurate enough to trust.

What Is It?

This project asks you to build a low-cost triage device that measures a few basic health signals, then combines them into one risk score. Think of it like a smoke alarm with more than one sensor. One sensor watches oxygen level, another watches temperature, and another estimates breathing rate. A tinyML model, which is a small machine learning model that runs on a low-power device, can turn those readings into a warning score.

The goal is not to diagnose disease. The goal is to spot when someone may need faster attention. That matters in rural clinics, where staff may have limited tools and many patients to see. Your job is to test whether a cheap, portable system can track clinical risk well enough to be useful.

Why This Is a Good Topic

This is a strong science fair topic because you can ask clear, measurable questions about sensor accuracy, model performance, and tradeoffs between cost and reliability. It connects directly to real-world care in places where medical equipment is limited. You can learn data collection, signal processing, model validation, and statistics, which gives your project real depth. The project also has room for original work, since you can compare design choices, model types, or sensor combinations.

Research Questions

• How does adding respiratory rate to pulse oximetry and temperature change early-warning-score accuracy??
• What is the effect of using a tinyML model instead of a simple rule-based score on triage classification performance??
• Does sensor fusion reduce false alarms compared with using each sensor alone??
• To what extent does motion or poor finger placement change the device's measurement error??
• Which model inputs give the best balance of accuracy, speed, and power use on a low-cost microcontroller??
• How does calibration with a reference dataset change agreement with MIMIC-derived ground truth??

Basic Materials

• Microcontroller with Bluetooth or Wi-Fi support, such as an ESP32 board.
• Low-cost pulse oximeter sensor module.
• Digital temperature sensor with known accuracy.
• Respiratory-rate measurement method, such as a pressure sensor, IMU, or camera-free chest motion sensor.
• Breadboard and jumper wires.
• USB cable and laptop.
• Rechargeable battery pack.
• Digital kitchen scale or basic bench scale for cost tracking.
• Phone camera or digital camera for documenting builds.
• Notebook or spreadsheet for data logging.

Advanced Materials

• Microcontroller development board with enough memory for tinyML inference.
• Medical-grade reference pulse oximeter for comparison.
• Calibrated thermometer or clinical temperature reference.
• Reference respiratory monitoring system, such as a spirometer or respiratory belt, if available.
• 3D-printed or laser-cut enclosure.
• Signal conditioning components for analog sensor cleanup.
• Access to de-identified clinical reference datasets, such as MIMIC.
• ECG-compatible reference tools if the model includes broader early warning features.
• Bench power supply.
• Oscilloscope or logic analyzer for debugging sensor timing.

Software & Tools

• Python: Cleans data, fits models, and compares sensor readings against reference labels.
• Jupyter Notebook: Lets you explore datasets, plot trends, and test scoring methods step by step.
• TensorFlow Lite for Microcontrollers: Runs tinyML models on a small device with limited memory.
• ImageJ: Measures visual changes if you test display readability or color-based indicators.
• R: Helps you run statistics, agreement tests, and confusion matrix analysis.

Experiment Steps

1. Define the exact triage outcome you want to predict, such as a low, medium, or high risk category.
2. Choose the sensor set you will compare, then decide which variables are measured directly and which are estimated.
3. Build a reference dataset plan so you can compare your device output against labeled ground truth.
4. Design a calibration approach that turns raw sensor signals into comparable numbers across users and conditions.
5. Select one model family to test first, then plan how you will compare it with a rule-based baseline.
6. Set evaluation metrics before collecting data, including error, sensitivity, specificity, and agreement.

Common Pitfalls

• Trusting raw pulse-oximeter readings without checking for motion artifacts, which can make oxygen estimates look better than they are.
• Using a respiratory-rate method that works only when the subject stays still, which can fail in real clinic settings.
• Training the model on one narrow dataset, which makes it overfit and weak on new users.
• Mixing clinical labels and device labels without a clear timestamp match, which ruins ground truth alignment.
• Ignoring battery, memory, and compute limits, which can make the tinyML model impossible to run on the device.

What Makes This Competitive

A class-level version might just build a working sensor and report readings. A stronger project tests whether the device stays accurate across different users, lighting, motion, and skin tones, then compares a few model designs with proper statistics. You can also add a real engineering angle by measuring power use, latency, and cost, not just accuracy. That mix of clinical relevance, device design, and careful validation is what makes the work stand out.

Project Variations

• Compare the device against a phone-based triage app to see which approach handles motion and poor lighting better.
• Test whether adding heart-rate variability improves early-warning scoring beyond oxygen, temperature, and respiratory rate.
• Build separate models for adults and children, then compare whether one shared model or two age-specific models performs better.

Learn More

• NIH PubMed: Search for review articles on pulse oximetry accuracy, early warning scores, and clinical decision support.
• MIMIC Database: Find de-identified intensive care data through the PhysioNet portal and search its documentation.
• PhysioNet: Explore open biomedical signal datasets and tutorials for respiratory and vital-sign analysis.
• TensorFlow Lite for Microcontrollers: Read the official guides for deploying tinyML models on small devices.
• MIT OpenCourseWare: Search for free courses on biomedical signal processing, machine learning, and embedded systems.
• NOAA Education Resources: Use basic sensor and data analysis materials if you need a refresher on measurement and uncertainty.