Early Identification of Hospital Visit Risk in Heart Failure Using Wearable-Derived Data

This study demonstrates that passive monitoring of wearable-derived data, specifically reduced step counts and elevated resting heart rates, can effectively identify heart failure patients at increased risk of hospitalization up to three days in advance.

The Gist

Imagine your heart is like a car engine that's starting to sputter. Usually, you only find out the engine is about to stall when you pull over to the side of the road (a hospital visit) or when the smoke starts billowing out (severe symptoms). By then, it's often too late to prevent the breakdown.

This research paper asks a simple but powerful question: Can we spot the "check engine" light before the car breaks down, using a smartwatch?

Here is the story of how they tried to do that, explained in plain English.

The Experiment: A 6-Month Watch Party

The researchers gathered 249 people with heart failure (a condition where the heart struggles to pump blood effectively). They gave everyone two things:

1. A Fitbit (to track steps and heart rate).
2. A smart scale (to track weight).

They split the group into three teams to see what worked best:

• Team A: Just got the gadgets.
• Team B: Got gadgets + a phone app to fill out daily surveys about how they felt.
• Team C: Got gadgets + app + cash rewards for sticking to the plan.

They watched them for 180 days (6 months) to see who ended up in the hospital.

The Big Discovery: The "Silent Siren"

Out of the 249 people, 51 ended up visiting the hospital. The researchers looked at the data from the weeks leading up to those visits and found some very clear patterns. It's like looking at the weather forecast before a storm hits.

1. The "Step Count" Drop (The Best Predictor)
Think of your daily steps as the fuel gauge on your car.

• People who ended up in the hospital started taking fewer steps about three weeks before their visit.
• It wasn't just a bad day; it was a steady decline. If you were usually a "high-mileage" walker and suddenly started walking less, it was a huge red flag.
• The Metaphor: Imagine a bird that usually flies high and far. If you notice it suddenly staying close to the ground and hopping slowly, you know something is wrong long before it falls out of the sky.

2. The "Heart Rate" Spike (The Last Warning)
Think of your resting heart rate as the engine temperature.

• About three days before a hospital visit, the participants' resting heart rates started to creep up.
• This was a short-term warning sign, like the engine getting hot right before it overheats.

3. The "Symptom" Report (The Driver's Complaint)

• People who went to the hospital also reported feeling worse (more shortness of breath, fatigue, etc.) in the weeks leading up to the event.
• However, the researchers found that the Fitbit data (steps and heart rate) was actually better at predicting the trouble than the people's own reports. Sometimes, people don't realize how sick they are until it's too late, but the watch knows.

The Surprising Twist: You Don't Need the App

The researchers thought the phone app and the daily surveys would be the key to saving lives. They were wrong.

• The Fitbit data alone (passive monitoring) was just as good, if not better, at predicting who would get sick.
• The Metaphor: You don't need to ask the driver "How do you feel?" every five minutes. You just need to watch the speedometer and the fuel gauge. The watch was doing the talking for them.

Why This Matters

Currently, heart failure patients usually only see their doctor once a month or when they feel terrible. By then, the "engine" is already failing.

This study suggests that if doctors could look at a patient's Fitbit data, they could spot the "check engine" light weeks in advance.

• Early Intervention: Instead of waiting for a hospital visit, a doctor could call the patient, adjust their medication, or tell them to rest before the crisis happens.
• Saves Money and Lives: Preventing a hospital visit saves thousands of dollars and keeps people out of the hospital bed.

The Bottom Line

Wearable devices are like 24/7 bodyguards for your heart. They can hear the subtle whispers of your body changing (walking less, heart beating faster) long before you hear the scream. By paying attention to these digital clues, we might be able to stop heart failure emergencies before they even start.

Technical Summary

1. Problem Statement

Heart failure (HF) is a leading cause of mortality and a significant economic burden, with hospitalizations accounting for the majority of the ~$31 billion annual cost in the U.S. Early identification of patients at risk of health exacerbation is critical for timely clinical intervention. While previous studies have utilized wearable data, many focused on specific cardiac devices or younger cohorts. There is a need to validate whether consumer-grade wearable data (specifically Fitbit) combined with patient-reported outcomes can effectively predict hospital visits (emergency department visits or admissions) in a population with an age distribution typical of HF incidence (50–80 years).

2. Methodology

Study Design and Population

• Cohort: 249 English-speaking adults (aged 50–80) with a confirmed HF diagnosis, recruited between July 2021 and September 2024.
• Duration: Prospective 180-day monitoring period.
• Randomization: Participants were randomized into three arms:
  • Group D (n=72): Received devices only (Fitbit + wireless scale).
  • Group D+A (n=89): Received devices + a mobile app with daily surveys.
  • Group D+A+F (n=88): Received devices + app + financial incentives for adherence.
• Data Collection:
  • Passive: Fitbit data (steps, resting heart rate) retrieved via Fitabase API.
  • Active: Daily surveys (for Groups D+A and D+A+F) assessing symptom severity, diet adherence, and medication usage on a scale of 0–2.
  • Outcomes: Hospital visits (ED or admission) were self-reported or verified via Electronic Health Record (EHR) review.

Data Processing and Analysis

• Preprocessing: Continuous Fitbit data (e.g., steps) was smoothed using a 2-day lagging rolling mean and Z-score normalized per patient (initialized from the first three non-missing values).
• Statistical Comparisons:
  • Mann-Whitney U test: Compared 3-week windows prior to a visit (Visit group) against random 3-week windows (No Visit group).
  • Wilcoxon signed-rank test: Compared the pre-visit window against random windows within the Visit group to check for persistent trends.
• Risk Stratification: Participants were split into "High Risk" and "Low Risk" based on median baseline metrics (first 2 weeks).
  • Survival Analysis: Kaplan-Meier curves with Log-rank tests to compare time-to-event (hospital visit) between risk groups.
• Predictive Modeling:
  • Algorithm: Logistic regression with balanced class weights.
  • Features: Daily steps, resting heart rate, and symptom severity from the first two weeks (baseline).
  • Validation: Leave-one-out cross-validation (LOOCV). Performance evaluated via AUROC, AUPRC, and Balanced Accuracy. Significance tested against random performance via permutation tests ($n=1000$) and DeLong's test.

3. Key Results

Demographics and Outcomes

• Visits: 51 unique participants (20%) experienced a total of 80 hospital visits during the 180-day period.
• Group Differences: Group D (devices only) had a significantly higher visit rate (25%) compared to Groups D+A (17%) and D+A+F (16%), though the app did not confound the physiological trends identified.

Physiological and Behavioral Trends

• Step Count:
  • Participants with visits took significantly fewer steps than those without ($p=.002$).
  • A decreasing trend in steps was observable starting ~2 weeks prior to an event.
  • Baseline Prediction: Participants with lower average steps in the first two weeks had a significantly higher probability of a hospital visit over time ($p=.003$).
• Resting Heart Rate (RHR):
  • RHR increased significantly in the three days immediately preceding a visit compared to the preceding 18 days ($p=.022$) and the No Visit cohort ($p=.022$).
  • However, baseline RHR alone did not significantly stratify risk ($p > .78$).
• Symptom Severity:
  • The Visit group reported higher symptom severity overall ($p=.044$).
  • Symptoms were significantly higher in the 3 weeks prior to a visit compared to the No Visit group ($p=.005$).
  • Baseline symptom severity showed a trend toward higher risk but did not reach statistical significance ($p=.073$).

Predictive Performance (Logistic Regression)

• Best Single Feature: Steps alone yielded the best performance (AUROC = 0.671, Balanced Accuracy = 0.641).
• Feature Combinations:
  • Adding Heart Rate or Symptoms to Steps did not significantly improve AUROC (e.g., Steps + HR + Symptoms AUROC = 0.642).
  • However, adding these features modestly improved the Area Under the Precision-Recall Curve (AUPRC), increasing from 0.273 (Steps only) to 0.322 (All features).
• Conclusion on Modeling: Passive activity data (steps) is the strongest predictor; adding active survey data provides marginal gains in precision but does not drastically alter overall classification accuracy.

4. Key Contributions

1. Validation of Consumer Wearables: Demonstrated that consumer-grade Fitbit data, specifically daily step counts, can effectively identify HF patients at risk of hospitalization in a real-world setting.
2. Temporal Dynamics: Identified specific temporal patterns: a long-term decline in activity (baseline risk) and a short-term spike in resting heart rate (3 days pre-event) are distinct markers of exacerbation.
3. Passive vs. Active Monitoring: Showed that passive physiological monitoring (steps) is a robust predictor, suggesting that continuous daily surveys may not be strictly necessary for initial risk stratification, potentially reducing patient burden.
4. Age-Appropriate Cohort: Addressed the gap in literature by studying a cohort (50–80 years) that aligns with the typical incidence age of HF, rather than focusing solely on younger populations.

5. Significance and Future Directions

• Clinical Impact: The study supports the use of wearable-derived metrics as a triage tool. Baseline activity levels can identify high-risk patients for increased monitoring, while acute changes in RHR can trigger immediate clinical intervention.
• Economic Potential: By enabling early intervention, this approach has the potential to reduce hospitalization rates, thereby addressing the substantial economic burden of HF.
• Limitations & Future Work:
  • Noise: Wearable data is inherently noisy; larger sample sizes are needed to distinguish low-intensity signals.
  • Intervention: Future studies should utilize these risk stratifications in pragmatic trials to test if targeted interventions actually reduce hospitalization rates.
  • Engagement: Providing apps to all participants in future studies could improve engagement and reduce confounding variables related to adherence.

In summary, the paper provides strong evidence that daily step counts serve as a powerful, passive biomarker for heart failure exacerbation risk, while resting heart rate serves as a critical short-term warning signal, offering a viable pathway for continuous, remote patient monitoring.