Oura Ring vs. Fitbit Charge 4

Project Type: Applied Data Science / Machine Learning
Focus Areas: Wearables analytics, feature importance, model interpretability
Tools & Methods: Python, Linear Regression, Permutation Feature Importance, SHAP

Project Overview

Consumer wearables promise actionable insights into sleep quality and daily readiness, yet their scoring algorithms remain proprietary and opaque. This project reverse-engineers and compares the sleep and readiness scoring systems of two leading devices—the Oura Ring and Fitbit Charge 4—to understand how similar inputs can produce meaningfully different outcomes.

Using multi-year, real-world data from both devices worn simultaneously by the same user, I analyzed score alignment, modeled Oura’s sleep and readiness algorithms, and evaluated which physiological signals truly drive each score. The work culminated in interpretable machine learning models that approximate Oura’s scoring logic and quantify feature importance.

Key Questions Addressed

• Do Oura Ring and Fitbit Charge 4 produce comparable sleep scores for the same user on the same night?

• Which physiological features most influence Oura’s Sleep Score and Readiness Score?

• Does removing the Oura Ring during training materially impact readiness outcomes?

• How well do quantified readiness scores align with human intuition?

Methodology Highlights

• Data Sources:

  • Multi-year daily sleep and readiness data from Oura Ring (Gen 2 & 3)

  • Concurrent Fitbit Charge 4 sleep data

  • Validation dataset from a second Oura user

  • Short subjective readiness survey for intuition comparison

• Modeling Approach:

  • Linear regression models to approximate proprietary scoring algorithms

  • Feature importance assessed via coefficients, permutation importance, and SHAP values

  • Model validation against unseen user data

• Evaluation Metrics:

  • R², RMSE, MAE, Max Error (rounded to match device score granularity)

Key Findings

• Sleep scores trend similarly but diverge frequently.
  While 74% of daily scores fall within ±7 points, the overall correlation is only moderate (~0.57). The largest discrepancies stem from how devices interpret REM vs. deep sleep.

• Total sleep duration dominates sleep scoring.
  For Oura, total sleep time is the most influential factor, followed by sleep latency and deep sleep. Restfulness appears important by coefficient but less impactful in practice when evaluated with SHAP.

• HRV balance is critical for readiness.
  Heart Rate Variability (HRV) balance and sleep balance emerge as the strongest drivers of readiness, outweighing activity metrics.

• Not wearing the ring during workouts is unlikely to penalize readiness.
  Because readiness relies more heavily on recovery signals (HRV, resting heart rate, sleep balance), missing activity data has limited impact.

• Devices cannot replace intuition.
  Subjective readiness matched device scores only about half the time, reinforcing that wearables should inform—not dictate—daily decisions.

Why This Project Matters

This project demonstrates how machine learning and explainability techniques can be used to:

• Reverse-engineer black-box consumer algorithms

• Translate physiological signals into actionable insights

• Evaluate product claims with real-world data

• Bridge quantitative metrics with human judgment

It also reflects a broader theme in my work: turning raw data into interpretable, decision-ready insights, particularly in health, fitness, and wearable technology contexts.

Full Paper & Presentation

• Capstone_Presentation_PDF
• Final_Project_Report