Wearable-derived physiological features for trans-diagnostic disease comparison and classification in the All of Us longitudinal real-world dataset

This study demonstrates that incorporating wearable-derived physiological features, such as heart rate, activity, and sleep patterns, significantly improves the classification and risk stratification of multiple chronic physical and mental health conditions in a large longitudinal real-world dataset, with particularly strong gains observed for major depressive disorder and anxiety.

The Gist

Imagine your body is like a high-tech car that constantly sends out signals about how it's running—how fast the engine is spinning (heart rate), how much you're driving (activity), and how well the battery recharges while parked (sleep). For years, doctors have mostly looked at these signals to check for just one specific problem, like a flat tire (a heart issue) or a clogged fuel line (diabetes). But this study asked a bigger question: Can we look at all these signals at once to understand the whole car, even when it has multiple, different kinds of problems?

Here is the story of what the researchers found, broken down simply:

The Big Experiment: A "Trans-Diagnostic" Detective

Think of the researchers as detectives who didn't just look for one type of crime. Instead, they gathered a massive group of 9,301 people from the "All of Us" dataset (a huge digital library of real-world health data). They looked at people with a wide mix of conditions:

• Physical: Heart trouble, diabetes, and sleep apnea (trouble breathing while sleeping).
• Mental: Depression, anxiety, bipolar disorder, and ADHD.

They wanted to see if the "digital footprints" left by Fitbits could tell the difference between these conditions and healthy people, and if those footprints looked similar or different across the board.

The Magic Ingredient: Adding the "Wearable" Lens

The team built two types of computer "brains" (models) to guess who had which disease:

1. The Old Way: They fed the computer basic info like age, gender, and lifestyle habits. This is like trying to guess a car's problem just by looking at the owner's driver's license and the color of the car.
2. The New Way: They added the Fitbit data—heart rate, sleep patterns, and movement. This is like plugging a high-tech diagnostic computer directly into the car's engine to read the live data.

The Result? The "New Way" was almost always a better detective. By adding the wearable data, the computer got much better at spotting who was sick and who was healthy.

The Surprise Winners: Mental Health

While the wearable data helped with physical diseases, the biggest "aha!" moment happened with mental health.

• For Depression and Anxiety, adding the Fitbit data was like turning on a flashlight in a dark room. The computer's ability to correctly identify these conditions jumped significantly.
• It's as if the data revealed that when someone is struggling with anxiety or depression, their body starts whispering secrets through their sleep cycles and heart rate that we couldn't hear before.

Note: The one exception was ADHD. The study didn't find a huge boost there, but the researchers suspect it's just because they didn't have enough ADHD patients in the group to make the pattern clear yet.

What the Data Actually "Said"

The researchers didn't just get a "yes/no" answer; they looked at why the computer made its guesses. They found some fascinating clues:

• Sleep is a Storyteller: It wasn't just about how much people slept, but the quality and consistency. For example, the amount of time spent in "Deep Sleep" and "REM sleep" (the dreamy part of sleep), and how much those times changed from day to day, were strong indicators of risk for certain conditions.
• Variation Matters: If your sleep or heart rate is all over the place (very inconsistent), it might be a red flag for your health, much like a car engine that idles erratically.

The Bottom Line

This study is a proof-of-concept that our smartwatches and fitness trackers are more than just pedometers for counting steps. They are continuous health monitors that can act as a "second opinion" for doctors.

Imagine a future where, alongside your annual check-up, your doctor also checks your wearable data to see if your body is showing early signs of stress, sleep issues, or mood changes before you even feel sick. This research suggests that by listening to these continuous signals, we can catch health issues earlier and understand the complex links between our physical and mental well-being.

Technical Summary

1. Problem Statement

Current research utilizing wearable-derived physiological features (e.g., heart rate, activity, sleep) for disease risk assessment predominantly focuses on single-condition studies. This siloed approach limits the understanding of cross-disease patterns and shared physiological signatures across different chronic conditions. Furthermore, there is a lack of comprehensive evaluation regarding the utility of these features for trans-diagnostic disease classification in large-scale, real-world longitudinal datasets. The study aims to bridge this gap by investigating whether wearable data can capture both shared and condition-specific physiological signatures across a spectrum of physical and mental health disorders and evaluating their efficacy in improving disease classification models.

2. Methodology

The study employed a rigorous data-driven approach using the All of Us Controlled Tier Dataset (Version 8), a large-scale real-world health database.

• Cohort Selection:
  • Total Sample: 9,301 patients.
  • Inclusion Criteria: Participants must have at least 21 days of consecutive FitBit data.
  • Disease Subcohorts: Seven chronic conditions were selected based on prevalence and relevance, spanning physical and mental health:
    • Cardiovascular Disease (CVD)
    • Diabetes
    • Obstructive Sleep Apnea (OSA)
    • Major Depressive Disorder (MDD)
    • Anxiety
    • Bipolar Disorder
    • Attention-Deficit/Hyperactivity Disorder (ADHD)
• Feature Engineering:
  • Baseline Features: Demographic and lifestyle variables.
  • Wearable-Derived Features: Physiological signals including heart rate, physical activity metrics, and sleep architecture (specifically duration and day-to-day variation of deep and REM sleep).
• Modeling Strategy:
  • Approach: A trans-diagnostic framework comparing each disease subcohort against a control cohort.
  • Algorithms: Two distinct machine learning models were fitted for each condition:
    1. Logistic Regression (Linear baseline).
    2. XGBoost (Gradient boosting for non-linear relationships).
  • Experimental Design: Models were compared in two configurations:
    1. Baseline features only.
    2. Baseline features + Wearable-derived features.
• Evaluation Metrics: Primary performance metric was ROC-AUC (Receiver Operating Characteristic - Area Under the Curve). Statistical significance was assessed via p-values.

3. Key Contributions

• Trans-Diagnostic Framework: This is one of the first comprehensive studies to evaluate wearable-derived features across a broad spectrum of physical and mental health conditions simultaneously, moving beyond single-disease silos.
• Real-World Validation: The study leverages a massive, diverse longitudinal dataset (All of Us) rather than controlled clinical trial data, enhancing the generalizability of the findings to real-world settings.
• Feature Interpretability: The study goes beyond black-box classification by analyzing feature contributions, identifying specific physiological markers (e.g., sleep duration and variability) associated with disease risk.
• Clinical Utility Demonstration: It provides evidence that wearable data serves as a practical, complementary tool to traditional clinical assessments for risk stratification.

4. Key Results

• General Performance Improvement: Incorporating wearable-derived features significantly improved classification performance (ROC-AUC) for all disease subcohorts in both Logistic Regression and XGBoost models, with one notable exception.
• Exception (ADHD): For the ADHD subcohort, performance improvement was observed primarily in the Logistic Regression model's ROC-AUC. The authors attribute the limited gains in XGBoost and the overall smaller effect size to the smaller sample size available for the ADHD cohort.
• Top Performers: The most substantial performance gains were observed in mental health conditions:
  • Major Depressive Disorder (MDD):
    • Logistic Regression: +0.077 ROC-AUC ($p < 0.001$)
    • XGBoost: +0.071 ROC-AUC ($p < 0.001$)
  • Anxiety:
    • Logistic Regression: +0.077 ROC-AUC ($p < 0.001$)
    • XGBoost: +0.108 ROC-AUC ($p < 0.001$)
• Physiological Insights: Feature analysis revealed that specific sleep metrics, particularly the duration and day-to-day variation of deep sleep and REM sleep, were strongly associated with increased disease risk in certain conditions.

5. Significance

This study validates the potential of wearable technology to act as a continuous, non-invasive monitoring tool for trans-diagnostic risk stratification.

• Clinical Integration: The findings support the integration of real-time, automated wearable assessments into clinical workflows to complement standard evaluations, potentially allowing for earlier detection and monitoring of chronic conditions, particularly in mental health.
• Future Research: The study establishes a foundation for exploring more fine-grained wearable data, suggesting that continuous physiological monitoring can reveal complex, shared biological signatures across diverse disease states that are not apparent through traditional cross-sectional studies.
• Scalability: By demonstrating success in a large, real-world dataset, the research paves the way for scalable, digitized health interventions that leverage consumer-grade devices for population health management.