Personal Physiology Digital Twin Project

ISEF Category: Biomedical Engineering

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: Full Year

The Hook

Your watch already collects a lot of clues about your body. The trick is turning those clues into a forecast you can test, not just a pretty graph. A digital twin tries to do that by acting like a mini version of your physiology. You can build one that predicts tomorrow’s energy and recovery from today’s data.

What Is It?

A digital twin is a model that stands in for a real system. In this project, the system is a person’s daily physiology. You feed in health data from Apple Health or Fitbit, then your model tries to predict next-day energy or recovery scores. Think of it like a weather forecast for the body, where the inputs are sleep, heart rate, heart rate variability, and maybe glucose data if you have it.

The math often uses ODEs, or ordinary differential equations. That sounds scary, but the idea is simple. You write rules for how one state changes over time, then let the model update itself step by step. For example, poor sleep might lower next-day recovery, while better heart rate variability might point the other way. Your job is not to make a perfect clone of the body. Your job is to test whether a simple model can predict something useful better than a guess.

Why This Is a Good Topic

This is a strong science fair topic because you can test it with real data, clear metrics, and honest validation. You do not need a wet lab, but you do need careful modeling, clean data handling, and a plan for measuring prediction quality. The topic connects to wearable health tech, personalized medicine, and decision support for recovery and performance. You can learn data cleaning, model building, forecasting, and how to judge whether a model really works.

Research Questions

• How does including sleep duration change next-day energy prediction accuracy?
• What is the effect of adding heart rate variability to a model that already uses sleep data?
• Does a simple ODE model predict recovery scores better than a baseline moving average?
• To what extent does one person’s model transfer to a family member with different sleep and activity patterns?
• Which input set, sleep only, sleep plus resting heart rate, or sleep plus HRV, gives the lowest prediction error?
• How does model accuracy change when you train on one month versus several months of wearable data?
• To what extent do Apple Health and Fitbit export formats produce different predictions for the same person?

Basic Materials

• Laptop or desktop computer with internet access.
• Apple Health export data or Fitbit export data.
• Spreadsheet software such as Google Sheets or Excel.
• Python with Jupyter Notebook.
• Free visualization tool such as Plotly or Matplotlib.
• Data dictionary for the wearable export fields.
• Consent form for each family or friend participant.
• Secure folder for storing private health data.

Advanced Materials

• Laptop or desktop computer with internet access.
• Apple Health export data or Fitbit export data.
• Python with Jupyter Notebook.
• NumPy, Pandas, SciPy, and Statsmodels.
• Plotly or Matplotlib for visual checks.
• GitHub private repository or local version control.
• Secure de-identified database structure.
• Optional continuous glucose monitor export data, if available and ethically approved.
• Statistical testing tools for cross-validation and error analysis.

Software & Tools

• Python: Cleans wearable exports, fits the model, and calculates prediction error.
• Jupyter Notebook: Lets you test code, document decisions, and show your workflow.
• Pandas: Organizes time-stamped health data into usable tables.
• SciPy: Solves ordinary differential equations and supports model fitting.
• Plotly: Makes interactive plots that help you compare predictions with real outcomes.

Experiment Steps

1. Define the one outcome you will predict, such as next-day energy or recovery, and decide how you will score prediction accuracy.
2. Choose the smallest set of inputs that still makes biological sense, such as sleep, resting heart rate, and heart rate variability.
3. Build a baseline model first so you know whether the digital twin adds value beyond a simple average or trend line.
4. Design your ODE structure so each variable has a clear role, then check whether the equations match the story you want to test.
5. Plan a validation scheme that uses held-out days or held-out people, not the same data used to fit the model.
6. Set up error analysis so you can compare performance across different people, different devices, and different data windows.

Common Pitfalls

• Mixing daily averages and overnight values in the same model, which makes the inputs describe different time windows.
• Training and testing on the same days, which makes the prediction score look better than it really is.
• Using raw wearable exports without cleaning missing timestamps, which creates fake spikes or gaps in the model.
• Treating energy or recovery scores as exact measurements, which ignores that many wearable scores are indirect estimates.
• Building one model for all family members without checking person-to-person differences, which hides how much the model depends on the individual.

What Makes This Competitive

A stronger project goes beyond a simple prediction chart. You can compare several model structures, test them on held-out people, and report error with proper statistics. You can also ask whether one variable matters more than the others, or whether a simple model beats a more complex one. If you explain your assumptions clearly and show where the model fails, your project will feel much more like real biomedical engineering.

Project Variations

• Use sleep, activity, and resting heart rate only, then test whether the model still predicts next-day recovery well.
• Swap Apple Health for Fitbit and compare whether export format or sensor style changes prediction quality.
• Add glucose data from a continuous glucose monitor and test whether metabolic signals improve energy forecasting.

Learn More

• Apple Health Export documentation: Search Apple’s support pages for how to export health data and understand the XML structure.
• Fitbit Web API docs: Search Fitbit developer documentation for field definitions and data export formats.
• NIH PubMed: Search for review articles on wearable-derived heart rate variability, sleep, and recovery prediction.
• NIH RePORTER: Search funded projects on digital twins, wearable biomarkers, and personalized health modeling.
• MIT OpenCourseWare, system dynamics and differential equations materials: Use course notes to review ODE modeling basics.
• IEEE Journal of Biomedical and Health Informatics: Search recent papers on wearable sensing, forecasting, and personalized health analytics.