Characterizing Autonomic Dysfunction during Resuscitation in Sepsis using Multiscale Entropy

This retrospective study demonstrates that Multiscale Entropy (MSE) features derived from heart rate variability during the first 24 hours of ICU admission predict 7-day mortality and 28-day organ dysfunction in sepsis patients with significantly higher accuracy than traditional severity of illness scores, particularly in those requiring vasopressors.

The Gist

The Big Picture: Listening to the Body's "Jazz"

Imagine your body's nervous system is like a jazz band. When you are healthy, the band is playing complex, improvisational jazz. The drummer (your heart) isn't just keeping a steady beat; they are reacting to the bassist, the saxophonist, and the room's energy. This "complexity" and "flexibility" are signs of a healthy, resilient system.

Now, imagine a patient gets sepsis (a severe, life-threatening infection). It's like a sudden storm hits the jazz club. The band starts to panic. The music becomes rigid, repetitive, and boring. The drummer just hits the same beat over and over. In medical terms, this loss of "complexity" is called autonomic dysfunction.

This paper asks a simple question: Can we listen to this "music" (heart rate) to predict if the patient will survive, even better than the doctors' current tools?

The Problem with Current Tools

Right now, when doctors assess how sick a sepsis patient is, they use "scorecards" (like the SOFA or APACHE II scores). Think of these scorecards like a weather report that only tells you the temperature and wind speed right now. They are static snapshots. They tell you it's raining, but they don't tell you if the storm is about to get worse or if the clouds are about to break.

The researchers wanted to know if looking at the pattern of the heart rate over time could give a better forecast.

The New Tool: "Multiscale Entropy" (MSE)

The researchers used a fancy math tool called Multiscale Entropy (MSE).

• The Analogy: Imagine looking at a forest.
  • A standard heart rate test looks at the trees one by one (how far apart they are).
  • MSE looks at the whole forest from a drone. It checks the patterns of the trees from the ground level, up to the canopy, and all the way to the horizon. It asks: "Is this forest chaotic and alive, or is it a rigid, dead grid?"

They analyzed the heartbeats of sepsis patients during their first 24 hours in the ICU. They didn't just look at one second; they looked at patterns over seconds, minutes, and hours.

What They Found

1. The "Dead Beat" Signal: Patients who were sicker and died within a week had heart rhythms that were much more "rigid" and "boring" (low entropy). Their jazz band had stopped improvising. Patients who survived kept a more complex, flexible rhythm.
2. Better Than the Scorecard: The MSE "listening" tool was much better at predicting death than the traditional "scorecard" tools.
   • The Scorecard was like guessing the weather based on a single temperature reading (64% accuracy).
   • The MSE Tool was like analyzing the pressure systems, humidity, and wind patterns (84% accuracy).
3. The Drug Effect: The study also looked at how medicine changed the "music."
   • When patients needed stronger drugs (vasopressors) to keep their blood pressure up, their heart rhythms became even more rigid.
   • It was like the storm got worse, and the band had to play even more mechanically. The more drugs they needed, the less "jazz" their hearts had left.

Why This Matters

This research suggests that we can use a simple ECG monitor (which is already in every ICU) to "listen" to the body's complexity.

• If the music is complex: The body is still fighting and adapting. The patient is likely to survive.
• If the music becomes rigid: The body is losing its ability to adapt. The patient is at high risk of dying or developing organ failure (kidney, lung, or brain issues).

The Takeaway

Think of this study as a new way to listen to the heartbeat's "personality." Instead of just counting how fast the heart beats, this method checks how flexible and alive the heart rhythm is.

The researchers found that this "flexibility" is a powerful crystal ball. It can tell doctors early on which patients are in deep trouble and need immediate help, potentially saving lives by catching the deterioration before it's too late. It turns the heart rate from a simple number into a rich story about the patient's resilience.

Technical Summary

1. Problem Statement

Sepsis is characterized by a dysregulated host response leading to autonomic dysfunction, which is a critical predictor of mortality and organ failure. While Heart Rate Variability (HRV) is a known marker of autonomic health, traditional HRV metrics (time-domain, frequency-domain, and single-scale non-linear measures) often fail to capture the complex, dynamic evolution of neuroendocrine-immune interactions over time.

• Limitations of Current Methods: Prior studies have largely focused on short-term HRV metrics, which may not fully represent the physiological complexity required to predict deterioration in septic shock. Furthermore, the influence of concurrent treatments (e.g., vasopressors, corticosteroids) on HRV makes it difficult to distinguish between disease severity and treatment effects.
• Research Gap: There is a need for a robust, scale-sensitive biomarker that can quantify autonomic dysfunction across multiple temporal scales, independent of or in conjunction with standard severity scores, to predict short-term mortality and long-term organ dysfunction in septic patients.

2. Methodology

Study Design and Cohort:

• Type: Retrospective cohort study.
• Population: Adult patients (≥18 years) admitted to the ICU at Emory University Hospital (2016–2019) meeting Sepsis-3 criteria.
• Cohorts:
  • Fluids-only: Received ≥30 mL/kg fluid resuscitation.
  • Fluids-plus-Vasopressor: Received fluids + norepinephrine.
• Matching: Propensity score matching (1:1) was performed based on age, sex, APACHE II, and SOFA scores to create a balanced cohort of 958 patients (479 pairs).

Data Processing:

• Source: High-resolution ECG waveforms (240 Hz) extracted from GE bedside monitors.
• Feature Extraction:
  • Multiscale Entropy (MSE): Computed across 20 temporal scales using the PhysioNet Cardiovascular Signal Toolbox.
  • Aggregation: Scales were grouped into seven features: Ultra Short (1–2), Very Short (3–4), Short (5–6), Intermediate (7–9), Long (10–12), Very Long (13–15), and Ultra Long (16–20).
  • Baseline: MSE features were calculated over a 1-hour baseline window (starting 24 hours post-admission or nearest available).
• Comparison Models:
  1. MSE-only: Entropy features alone.
  2. MSE-DV: Entropy + Demographics/Vitals (age, sex, race, MAP, SBP, DBP, Temp, RR).
  3. HRV Model: Traditional time, frequency, and non-linear HRV features.
  4. SOAP Model: Standard severity scores (SOFA + APACHE II).

Machine Learning Approach:

• Algorithms: Logistic Regression, XGBoost, Random Forest, Gradient Boosting.
• Validation: 80/20 train/test split with 5-fold cross-validation and stratified bootstrapping (1,000 iterations).
• Metrics: Primary metric was AUROC; secondary metrics included Sensitivity, Specificity, PPV, NPV, and Brier score for calibration. SHAP (SHapley Additive exPlanations) values were used for feature interpretability.

Outcomes:

• Primary: 7-day mortality.
• Secondary: 28-day persistent organ dysfunction (renal, neurological, respiratory).

3. Key Contributions

• Novel Biomarker Application: Demonstrated that Multiscale Entropy (MSE), specifically capturing complexity across long temporal scales, is a superior prognostic marker for sepsis compared to traditional HRV and severity scores.
• Treatment-Aware Analysis: Explicitly analyzed how MSE profiles change with vasopressor escalation (norepinephrine vs. epinephrine/vasopressin/phenylephrine) and corticosteroid use, revealing a progressive loss of physiological complexity with treatment intensity.
• Model Parsimony: Showed that a model using only seven aggregated MSE features could achieve high discrimination (AUROC 0.77–0.84), outperforming complex models relying on full HRV suites or clinical severity scores.
• Interpretability: Utilized SHAP analysis to identify that both ultra-short-term (rapid fluctuations) and long-term (slow regulatory) complexity losses are critical drivers of mortality risk.

4. Results

Predictive Performance (7-Day Mortality):

• Fluids-plus-Vasopressor Cohort:
  • MSE-DV Model: Achieved the highest AUROC of 0.84 (95% CI: 0.73–0.92).
  • MSE-only Model: AUROC of 0.77 (95% CI: 0.63–0.88), significantly outperforming the SOAP score (AUROC 0.64).
  • HRV Model: AUROC of 0.75, slightly lower than MSE.
  • Specificity: All models were tuned for 80% specificity; MSE-DV maintained a high Negative Predictive Value (NPV) of 0.93, indicating strong utility for ruling out mortality.
• Pooled Cohort (Fluids + Vasopressor):
  • MSE-DV (AUROC 0.74) and MSE-only (AUROC 0.72) continued to outperform SOAP (AUROC 0.60) and HRV (AUROC 0.66).

Organ Dysfunction (28-Day):

• In the fluids-plus-vasopressor cohort, the MSE-DV model predicted:
  • Neurological Dysfunction: AUROC 0.79.
  • Renal Failure: AUROC 0.75.
  • Respiratory Failure: AUROC 0.71.

Physiological Phenotyping:

• Mortality Stratification: Non-survivors (expired within 24 hours) exhibited significantly lower MSE across all scales compared to survivors, with large effect sizes in mid-range and long-term scales.
• Treatment Escalation: Patients requiring second-line (epinephrine, vasopressin) and third-line vasopressors, or those receiving corticosteroids, showed progressively lower MSE values. The combination of Norepinephrine + Phenylephrine + Hydrocortisone resulted in the most profound reduction in complexity, particularly at mid-to-long scales (6–15).
• Feature Importance: In the high-risk vasopressor cohort, very-short-term and ultra-short-term entropy were the top predictors. In the pooled cohort, demographic/vital signs gained prominence, but long-term entropy remained a key predictor.

5. Significance and Conclusion

• Clinical Utility: MSE provides a dynamic, non-invasive measure of physiological resilience that outperforms static severity scores (SOFA/APACHE II) in predicting early deterioration and organ failure in septic shock.
• Mechanistic Insight: The findings suggest that sepsis-induced autonomic dysfunction manifests as a loss of complexity across multiple time scales. The progressive decline in MSE with vasopressor escalation reflects a shift toward rigid, less adaptive physiology and "autonomic exhaustion."
• Future Directions: The study supports the integration of entropy-based analytics into ICU decision-support systems for early risk stratification. Future work should focus on multicenter validation, longitudinal tracking of MSE trends, and accounting for confounders like mechanical ventilation and sedation depth.

Conclusion: Multiscale Entropy derived from the first 24 hours of ICU admission is a robust, independent predictor of mortality and organ dysfunction in sepsis, offering superior discrimination over traditional clinical scores and HRV metrics, particularly in patients requiring vasopressor support.