⚡ electrical engineering

Detection of Autonomic Dysreflexia in Individuals With Spinal Cord Injury Using Multimodal Wearable Sensors

This study presents a non-invasive, explainable machine learning framework using multimodal wearable sensors that effectively detects Autonomic Dysreflexia in individuals with spinal cord injury, achieving high accuracy through a stacked ensemble model with heart rate and ECG features as the most informative predictors.

Original authors: Bertram Fuchs, Mehdi Ejtehadi, Ana Cisnal, Jürgen Pannek, Anke Scheel-Sailer, Robert Riener, Inge Eriks-Hoogland, Diego Paez-Granados

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bertram Fuchs, Mehdi Ejtehadi, Ana Cisnal, Jürgen Pannek, Anke Scheel-Sailer, Robert Riener, Inge Eriks-Hoogland, Diego Paez-Granados

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

🚨 The Problem: The Silent Alarm

Imagine your body is a high-tech house with a sophisticated security system. For most people, if something goes wrong (like a fire), the smoke alarm goes off, and you feel it immediately.

But for people with a Spinal Cord Injury (SCI), the "wiring" between the brain and the body is cut. Specifically, a dangerous condition called Autonomic Dysreflexia (AD) can happen. This is like a fire starting in the basement (the bladder or bowel) that causes the pressure in the house to skyrocket instantly.

What happens? The blood pressure shoots up dangerously high, which can lead to a stroke or even death.
The Catch: Because the spinal cord is damaged, the person often doesn't feel the pain or the warning signs. It's a silent alarm.
The Current Fix: Right now, doctors have to stick a cuff on the arm every few minutes to check the pressure, or wait for the patient to guess they feel "off." This is invasive, uncomfortable, and doesn't work well when the person is at home.

🛠️ The Solution: A "Smart Watch" Detective Team

The researchers from ETH Zurich and Swiss Paraplegic Research wanted to build a non-invasive, automatic detective that wears on the body. Instead of one detective, they built a team of sensors (like a squad of specialized spies) working together to spot the danger before it gets too late.

They strapped three different types of devices onto 27 patients:

A Smart Wristband: Like a Fitbit, but super-powered. It measures heart rate, skin temperature, and blood flow (PPG).
A Chest Patch: A small sticker that listens to the heart's electrical rhythm (ECG) like a stethoscope that never sleeps.
A Temperature Patch: A sensor under the arm that measures the body's core heat.

🧠 The Brain: How the Computer Learned

The team didn't just look at the data; they taught a computer to be a super-detective.

The Training Camp: They induced mild AD episodes in a hospital setting (by filling the bladder, which is a standard medical test) while recording everything. They knew exactly when the "fire" started because they had a medical blood pressure cuff as the "truth."
The Feature Hunt: The computer looked at millions of data points. It asked: "Does the heart beat change shape? Does the skin get colder? Does the breathing speed up?"
- The Analogy: Imagine trying to guess if a storm is coming. You could look at the clouds (skin temp), the wind (breathing), or the birds' behavior (heart rhythm). The computer realized that how the heart beats (the rhythm and variability) was the most reliable "bird behavior" to predict the storm.
The Teamwork (Ensemble Learning): Instead of trusting just one sensor, they used a voting system.
- Sensor A (Heart) says: "I think danger is coming!"
- Sensor B (Skin) says: "I'm not sure, but I feel a bit weird."
- Sensor C (Chest) says: "Definitely danger!"
- The Decision: The computer combines these votes. If enough sensors agree, it sounds the alarm. This makes the system very hard to fool, even if one sensor gets covered by a blanket or falls off.

🏆 The Results: Who Won the Race?

The study found some surprising winners in the "detective squad":

The MVP (Most Valuable Player): Heart Rate (HR) and ECG (the electrical heart signal). These were the best at spotting the danger, with a success rate of about 93%. It turns out, the heart's rhythm changes in a very specific way right before the blood pressure spikes.
The Runner-Up: The Smart Wristband (using light to measure blood flow) was also very good.
The Underachievers: Temperature and Breathing were less reliable. They were like the detectives who sometimes get distracted; they helped a little, but weren't the main reason the alarm went off.
The "Magic" Trick: The computer learned that even if one sensor failed (e.g., the wristband fell off), the other sensors could still do the job. The system is resilient, like a sports team that can win even if one player gets injured.

🚀 Why This Matters

This isn't just a lab experiment; it's a step toward freedom.

Before: A person with SCI might have to sit with a blood pressure cuff on all day, or rely on a caregiver to constantly check them.
After: They could wear a small, comfortable patch and a wristband. If the "Detective Team" senses a problem, it could send an alert to a phone or a nurse before the blood pressure gets dangerous.

In a nutshell: The researchers built a digital safety net. By listening to the heart and skin with smart sensors and using a clever computer brain, they can now spot a life-threatening blood pressure spike early, giving people with spinal cord injuries a better chance to stay safe and live independently.

1. Problem Statement

Autonomic Dysreflexia (AD) is a life-threatening condition affecting up to 90% of individuals with spinal cord injuries (SCI) at or above the T6 level. It is characterized by sudden, severe spikes in systolic blood pressure (SBP) triggered by noxious stimuli below the injury level (e.g., bladder distension).

Current Limitations: Existing detection methods rely on invasive monitoring or subjective patient/caregiver reporting. These approaches lack scalability, are often delayed, and fail to provide continuous, objective monitoring in daily life.
Research Gap: Previous studies on automated AD detection have suffered from small sample sizes, reliance on self-reported ground truth, lack of cross-validation, or the use of invasive/non-human data. There is a critical need for a non-invasive, objective, and generalizable detection system using wearable sensors.

2. Methodology

2.1. Study Population and Data Collection

Participants: 27 individuals with chronic, motor- and sensory-complete SCI (AIS A) at or above T6.
Protocol: Data were collected during scheduled urodynamic studies (UDS), where the bladder was filled with saline to induce AD episodes.
Ground Truth: Objective AD labels were derived from synchronized, medical-grade cuff-based blood pressure (BP) measurements. AD was defined as a sustained SBP increase of $\ge$ 20 mmHg above baseline.
Wearable Sensors: Three commercial devices were used to capture multimodal data:
1. Multimodal Wristband (CardioWatch): PPG, Skin Temperature, Respiratory Rate (RR), Bioimpedance (BioZ).
2. ECG Patch (VivaLNK): ECG, Heart Rate (HR), RR.
3. Temp Patch (CORE): Core and Skin Temperature.
Final Dataset: After screening for data completeness and signal quality, 17 participants were retained for analysis, yielding 3,696 samples (210 AD events, 3,486 normal).

2.2. Signal Processing and Feature Extraction

Preprocessing: Signals were synchronized, denoised, and quality-controlled.
Sliding Windows: Multiple window configurations were tested (5s to 60s). The optimal configuration was determined to be 60s window with a 10s step.
Feature Types: A comprehensive set of features was extracted, including:
- Time-domain: Mean, variance, RMSSD, pNN50.
- Frequency-domain: LF/HF ratios, FFT harmonics, spectral entropy.
- Morphological: Wavelet coefficients, Local Binary Patterns (LBP), fiducial points (e.g., QRS duration, PPG systolic peak).
- Non-linear: Poincaré plot descriptors (SD1, SD2).

2.3. Machine Learning Framework

Feature Selection: BorutaSHAP was employed to robustly select features. This method compares SHAP values of real features against "shadow" (permuted) features to eliminate noise and ensure statistical significance.
Ensemble Architecture:
- Weak Learners: Multiple Random Forest models were trained, each on a specific modality (e.g., ECG, HR) or device (e.g., ECG-patch, Wristband).
- Aggregation Strategies:
  1. k-Threshold Voting: Predicts AD if $k$ weak learners agree.
  2. Stacked Ensemble: Outputs from weak learners serve as input to a meta-learner (tested over 40 classifiers).
Validation: Leave-One-Subject-Out (LOSO) cross-validation was used to ensure generalizability across individuals. The dataset was stratified by participant to prevent data leakage.

3. Key Contributions

Objective Ground Truth: First study to validate AD detection using objective, continuous BP measurements rather than subjective self-reports.
Multimodal Ensemble Framework: Development of a robust stacked ensemble classifier capable of handling sensor dropout and modality failure, making it suitable for real-world deployment.
Explainable AI (XAI): Utilization of SHAP (SHapley Additive exPlanations) and BorutaSHAP to identify which physiological features and sensor modalities are most predictive, providing clinical interpretability.
Comprehensive Dataset: A larger, rigorously validated dataset compared to prior works, including detailed analysis of sensor reliability (ECG vs. PPG HR).

4. Results

4.1. Feature Importance

Dominant Modalities: Heart Rate (HR) and ECG-derived features were the most informative.
- Key features included HR variability metrics (SDNN, RMSSD), rhythm morphology, and cumulative mean HR.
- Morphological ECG features (e.g., Uniform Local Binary Patterns of QRS) were highly predictive.
Secondary Modalities: PPG and BioZ contributed significantly.
Low Impact: Respiratory Rate (RR) and Temperature (Temp) features showed lower predictive power and were often excluded in top-performing subsets, likely due to missing data and lower specificity.

4.2. Performance Metrics

Best Model: The Nearest Centroid Stacked Ensemble achieved the highest performance:
- Macro F1-Score: $0.77 \pm 0.03$
- Significance: Significantly outperformed the baseline Dummy classifier ($0.48$).
Modality Performance (AUC):
- HR: 0.93 (Outstanding)
- ECG: 0.88 (Excellent)
- PPG: 0.86 (Excellent)
- Temp: 0.75 (Moderate)
- BioZ: 0.69
- RR: 0.59
Device Performance:
- ECG-patch: AUC 0.94
- Multimodal Wristband: AUC 0.90
- Temp-patch: AUC 0.78
Robustness: The model remained robust even when excluding less critical modalities (e.g., a subset excluding Temp achieved an F1 of 0.78), demonstrating resilience to sensor failure.

5. Significance and Conclusion

Clinical Impact: This study demonstrates the feasibility of non-invasive, real-time AD monitoring using off-the-shelf wearable technology. Early detection can prevent catastrophic complications like stroke or seizures.
Methodological Rigor: By using stratified cross-validation and objective BP labels, the study addresses the reproducibility and generalizability issues plaguing previous research.
Future Directions: The authors highlight the need for larger datasets and testing in free-living (ambulatory) environments to further mitigate motion artifacts. However, the current results represent a critical proof-of-concept for personalized, continuous cardiovascular surveillance in the SCI population.

In summary, the paper establishes that HR and ECG signals are the primary drivers for AD detection, and a stacked ensemble approach using multimodal wearables can achieve high accuracy, offering a viable path toward proactive clinical management of Autonomic Dysreflexia.