Machine learning for triage of strokes with large vessel occlusion using photoplethysmography biomarkers
- Joachim A. Behar
- 2 days ago
- 4 min read
Updated: 13 hours ago
By Márton Áron Goda, Helen Badge, Jasmeen Khan, Yosef Solewicz, Moran Davoodi, Rumbidzai Teramayi, Dennis Cordato, Longting Lin, Lauren Christie, Christopher Blair, Gagan Sharma, Mark Parsons and Joachim A Behar
This study represents a joint collaboration between Pázmány Péter Catholic University’s Faculty of Information Technology and Bionics (Hungary), the Faculty of Biomedical Engineering at the Technion–Israel Institute of Technology (Israel), and the Ingham Institute for Applied Medical Research / Sydney Brain Center at UNSW, Liverpool Hospital (Australia).
Published in IOP Physiological Measurement:
Introduction: The Race Against Time in LVO Stroke
Large vessel occlusion (LVO) stroke is one of the most time-critical medical emergencies. Caused by a blockage in a major cerebral artery, LVO strokes are associated with severe neurological damage, long-term disability, and high mortality. Time is essential. The faster clinicians can diagnose an LVO and deliver endovascular thrombectomy, the better the chances of recovery.
However, diagnosing LVO before hospital arrival is challenging. Existing clinical scales such as variants of the NIH Stroke Scale require patient cooperation, take time to administer, and do not always effectively distinguish between LVO, non-LVO strokes, and stroke mimics. In patients with dementia, severe neurological deficits, or communication impairments, these scales may not be usable.
As healthcare systems face increasing demand, there is a critical need for fast, reliable, and scalable early triage tools that can function even in the noisy and time-pressured prehospital environment.
Our latest study asks a simple question: Can a 30-second photoplethysmography (PPG) signal help identify LVO stroke?
Why PPG? A Simple Signal With Hidden Potential
PPG is a widely used optical biosignal that measures changes in blood volume in the microvasculature. It is noninvasive, inexpensive, and available in almost every ambulance through standard fingertip sensors.
LVO strokes disrupt cerebral blood flow. This can trigger systemic hemodynamic changes and sympathetic activation. These effects may subtly alter the shape and variability of peripheral pulse waveforms captured by PPG.
This leads to a promising idea:PPG waveforms might contain physiological signatures that help identify LVO stroke.
Study Overview
Our research team recorded fingertip PPG signals from 88 patients presenting with stroke symptoms at Liverpool Hospital in Sydney, Australia. The cohort included:
25 patients with confirmed LVO stroke
36 patients with non-LVO stroke
27 patients with stroke mimics
Each participant had a 10-minute PPG recording. We then segmented the recordings into 30-second windows, a duration suitable for real-world emergency settings.
From each window, we extracted three types of features:
PPG morphological features (MOR): 101 waveform descriptors
Beat-rate variability features (BRV): 17 measures of beat-to-beat dynamics
Patient metadata (META): age and sex
We trained logistic regression models using different combinations of these features. To ensure robustness, we repeated the train-test split 100 times with strict patient-level separation to prevent information leakage.
Key Findings
The best performing model, which combined MOR, BRV, and META features, achieved:
AUROC: 0.77 (interquartile range 0.71 to 0.82)
Sensitivity: 74 percent
Specificity: 66 percent
Precision: 62 percent
For context, the widely used Hunter-8 clinical scale in this dataset had:
Very high specificity (approximately 0.96)
Lower sensitivity (approximately 0.57)
Missing scores for 13 percent of patients due to communication limitations.
This comparison shows that machine learning applied to PPG can provide a more balanced sensitivity and specificity, especially in early triage situations where missing an LVO can have severe consequences.
Feature importance analysis further showed that PPG waveform morphology plays the most critical role, suggesting that LVO strokes affect pulse-wave shape in measurable ways.
Implications for Emergency Medicine
This proof-of-concept study shows that a short PPG recording contains informative physiological patterns that can support LVO triage. The potential impact is significant.
1. Progress Toward Stroke-Capable Ambulances
Ambulances equipped with a simple PPG sensor and onboard algorithms could generate rapid LVO risk assessments, allowing paramedics to choose the most appropriate hospital destination.
2. Faster and More Accessible Assessment
PPG can be collected quickly and does not rely on patient cooperation, making it useful for patients who are unresponsive or have impaired communication.
3. Scalable and Cost Effective
PPG sensors are inexpensive and already widely available, making this approach suitable for both high-resource and low-resource health systems.
Limitations and Future Directions
Although the results are promising, this study has several limitations that future work should address:
The dataset is small. Larger and more diverse cohorts are essential.
Real-world ambulance recordings may include more noise and motion artifacts.
More advanced models such as deep learning may improve accuracy when sufficient data become available.
Combining PPG with additional signals such as ECG or audio may yield stronger triage tools.
Our long-term goal is to support the development of next-generation smart ambulances that can provide early, automated clinical insights.
Conclusion
This study demonstrates that a 30-second PPG signal can support machine learning models capable of identifying large vessel occlusion stroke with promising accuracy. With further research, development, and validation, this approach could improve early stroke recognition, reduce treatment delays, and ultimately save lives.
Digital biomarkers and artificial intelligence are beginning to reshape emergency medicine. This work highlights how a simple and widely available signal such as PPG may contribute to faster and more accurate stroke care.
Comments