Wearable Data Models for Adolescent Burnout

ISEF Category: Behavioral and Social Sciences

Subcategory: Other  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: 1 to 2 Months

The Hook

Burnout does not always arrive as a crash. Tiny shifts in sleep, heart-rate variability, and mood often show up first. If you track both wearable data and one daily question, you can test whether the pair predicts low-energy days better than either signal alone. That gives you a clean human data project with a clear story.

What Is It?

This project asks whether wearable data can warn you before a student feels burned out. Fitbit exports can give you sleep duration, bedtime timing, resting heart rate, and heart-rate variability, or HRV, which tracks how much beat-to-beat timing changes.

A daily mood question adds the human side. Think of the wearable as a smoke alarm and the mood survey as the person saying how bad the fire feels. Your job is to see whether the two signals together predict the next day's low-mood or high-strain days better than either one alone.

Why This Is a Good Topic

This topic works well for science fair because you can turn a fuzzy feeling into data. You can test clear variables, compare simple models, and see whether personal baselines beat group averages. It also connects to sleep, stress, and student well-being, so the results matter beyond the notebook. A student can learn data cleaning, prediction, and ethical handling of human data without needing a wet lab.

Research Questions

• How does adding sleep duration change next-day burnout prediction compared with mood survey only?
• What is the effect of including HRV on model accuracy for low-mood days?
• Does a personalized baseline predict burnout better than one model trained on all participants?
• To what extent do weekday and weekend signals differ in burnout risk?
• Which single feature, sleep duration, bedtime consistency, resting heart rate, or HRV, explains the most variation?
• How does a simple interpretable model compare with a more flexible classifier on this cohort?

Basic Materials

• Fitbit account with export access to sleep, heart-rate, and HRV data
• Daily one-question survey form in Google Forms or Microsoft Forms
• Spreadsheet software for cleaning timestamps and merging records
• Laptop with internet access for analysis and charts
• Consent and assent forms approved by your school or mentor

Advanced Materials

• Raw Fitbit CSV exports with timestamped sleep, heart-rate, and HRV fields
• De-identified participant ID key stored separately from survey data
• Python environment with pandas, scikit-learn, and matplotlib
• RStudio with tidyverse and caret for model comparison
• SHAP or permutation-importance tool for explaining predictions

Software & Tools

• Google Forms: Collects one daily mood response with a timestamped record.
• Google Sheets: Merges wearable exports and survey rows in a simple table.
• Python: Cleans data, builds models, and makes trend plots.
• scikit-learn: Trains baseline classifiers and tests prediction quality.

Experiment Steps

1. Define the burnout outcome you will predict, such as a low-mood day or a strain threshold.
2. Choose the wearable signals you will compare, then decide which ones count as same-day versus next-day features.
3. Plan a data table that matches each survey response to the correct wearable record.
4. Build a baseline model with the survey answer alone, then test whether wearable features improve it.
5. Add an explanation method so you can see which signals drive each prediction.
6. Set aside a holdout split or later time window so you can judge how well the model generalizes.

Common Pitfalls

• Mixing up Fitbit sleep records with calendar days, which shifts features to the wrong survey response.
• Letting missing survey answers pile up on stressful days, which makes the sample look calmer than it is.
• Comparing raw HRV values across students without a personal baseline, which hides normal individual differences.
• Training and testing on the same days, which makes the model look better than it performs on new data.
• Using too many wearable features for a short study, which lets the model memorize noise instead of burnout patterns.

What Makes This Competitive

A stronger version of this project does more than predict a label. It compares a simple, explainable model with a more flexible one, then shows where each breaks. It also tests whether personal baselines beat pooled baselines, which is a real question in student well-being data. Strong handling of missing days, time splits, and feature explanation can move the work well past a basic class demo.

Project Variations

• Use sleep timing and bedtime consistency instead of HRV if your wearable export is limited.
• Swap burnout for next-day stress, fatigue, or low focus, then compare which outcome is easiest to predict.
• Compare a logistic regression model with a decision tree to see how accuracy and clarity trade off.

Learn More

• PubMed: Search review articles on adolescent sleep, stress, HRV, and mood prediction.
• NIH National Library of Medicine: Find plain-language summaries on sleep, stress, and wearable health data.
• OpenIntro Statistics: Free textbook with clear chapters on regression, sampling, and model checking, available on the OpenIntro site.
• MIT OpenCourseWare: Free statistics and machine learning lectures for learning prediction basics.
• scikit-learn documentation: Read model and evaluation guides for logistic regression, decision trees, and metrics.
• CDC Youth Risk Behavior Survey: Use national data on teen well-being and sleep as background context.