Convulsive Movement Detection using Low-Resolution Thermopile Sensor Array

Ouday Hanosh, Rashid Ansari, Naoum P. Issa, A. Enis ÇetinWorkshops @ CVPR 2020 · OpenReview

Abstract

Sudden Unexplained Death in Epilepsy (SUDEP) is a fatal threat to patients who suffer from convulsive seizures. The causes of the SUDEP are still ambiguous, and the patients who suffer from epileptic seizures may face death in well sleeping, likely after an unwitnessed convulsive seizure. An important step towards SUDEP prevention is reliable seizure detection during sleep that is inexpensive and unobtrusive. In this work, we developed a non-contact, non-intrusive, privacy-preserving system that can detect convulsive movements experienced by human subjects. Detection is accomplished by a combination of uncooled low-cost, low-power, low-resolution 8 x 8 IR array sensor, and a deep learning algorithm implemented with a Convolutional Neural Network (CNN). The thermopile sensor array is placed 1m from subjects who are reclining in bed. The CNN training set consists of thermal video streams from 40 healthy subjects mimicking convulsive movements or lying in bed without making convulsive movements. After training, the CNN was tested on thermal video streams not included in the training set and had a 99.2% accuracy in classifying convulsive movements and non-convulsive episodes, with no false negatives to distinguish between the occurrence and non-occurrence of convulsive movements. The performance results show that the thermopile sensor array has the potential to detect convulsive seizures while maintaining patient privacy and not requiring direct patient contact.

Introduction

Sudden Unexplained Death in Epilepsy (SUDEP) remains a critical challenge in neurology, often occurring during sleep following an unwitnessed convulsive seizure. To mitigate this risk, reliable nighttime monitoring is essential. However, existing solutions face significant barriers to adoption: EEG is uncomfortable for long-term home use, wearables can cause skin irritation or be forgotten, and high-resolution video cameras raise substantial privacy concerns.

This research presents a "privacy-by-design" system that is contact-free, inexpensive, and privacy-preserving. By utilizing an ultra-low-resolution ($8 \times 8$) thermal sensor array and a deep learning classifier, the system can distinguish life-threatening convulsions from normal nocturnal movements without identifying the subject.

Methodology and System Design

The proposed system architecture integrates low-cost hardware with a specialized deep learning pipeline.

Hardware Configuration

The core of the system is a Panasonic Grid-EYE thermopile sensor. Unlike standard optical cameras, this sensor captures a $8 \times 8$ grid of temperature values. At this resolution, a human subject appears only as a generic "heat blob," making it impossible to identify facial features or personal details, thus ensuring inherent privacy. The sensor is positioned approximately 1 meter from the subject, as shown in Figure 1.

The experimental setup featuring a subject at a 1m distance from the 8x8 Grid-EYE thermopile sensor, which transmits data to an Arduino Uno and a processing unit for classification.

Signal Processing and CNN Architecture

A key innovation in this work is the use of temporal difference imaging. Rather than processing raw thermal frames, the system calculates the absolute difference between consecutive frames ($|I[n] - I[n-1]|$). This preprocessing step filters out static heat sources—such as the bed or a stationary body—and emphasizes the dynamic motion signatures characteristic of seizures.

These motion signatures are fed into a 2D Convolutional Neural Network (CNN) consisting of two convolutional layers for spatial feature extraction, max-pooling layers for down-sampling, and fully connected layers for final binary classification.

Experimental Dataset

To evaluate the system, the researchers compiled a dataset from 40 healthy volunteers (28 male, 12 female). Subjects performed three distinct activities while reclining: stationary sleep, normal incidental movements (like turning over), and mimicked convulsive seizures based on clinical instructional videos.

Thermal image sequences illustrating the three classes of data: (a) sleeping, (b) normal movement, and (c) convulsive movement.

A total of 9,520 preprocessed images were generated, and the model was validated using 10-fold cross-validation to ensure robustness.

Results and Key Insights

The system demonstrated exceptional performance, particularly when utilizing temporal difference images.

  • Accuracy Boost: Switching from raw thermal images to difference images improved classification accuracy from 83.3% to 99.2%. This confirms that the dynamics of heat change are significantly more informative than static heat distribution for movement detection.
  • Safety Priority: In medical monitoring, false negatives (missing a seizure) are critical. The proposed CNN achieved a 0% false negative rate on the test set, successfully identifying every simulated convulsive event.
  • Low False Alarm Rate: Only 1.6% of normal movements were misclassified as convulsions, minimizing unnecessary caregiver disruptions.
  • Computational Efficiency: The low-resolution input ($8 \times 8$) is computationally lightweight, suggesting that the system could be deployed on low-power edge devices, such as smart nightstand lamps.

Conclusion and Future Work

This study validates the potential of low-resolution thermopile arrays for non-intrusive seizure monitoring. While the results are promising, several limitations remain. The current model was trained on mimicked seizures; future work must involve clinical trials with epilepsy patients to capture the full variability of real-world seizures. Additionally, further testing is needed to assess system performance under varying environmental conditions, such as the use of heavy blankets or the presence of multiple occupants in the bed.

Despite these challenges, the system offers a significant step toward an unobtrusive, privacy-preserving solution for SUDEP prevention.

Summary updated

Abstract

Sudden Unexplained Death in Epilepsy (SUDEP) is a fatal threat to patients who suffer from convulsive seizures. The causes of the SUDEP are still ambiguous, and the patients who suffer from epileptic seizures may face death in well sleeping, likely after an unwitnessed convulsive seizure. An important step towards SUDEP prevention is reliable seizure detection during sleep that is inexpensive and unobtrusive. In this work, we developed a non-contact, non-intrusive, privacy-preserving system that can detect convulsive movements experienced by human subjects. Detection is accomplished by a combination of uncooled low-cost, low-power, low-resolution 8 x 8 IR array sensor, and a deep learning algorithm implemented with a Convolutional Neural Network (CNN). The thermopile sensor array is placed 1m from subjects who are reclining in bed. The CNN training set consists of thermal video streams from 40 healthy subjects mimicking convulsive movements or lying in bed without making convulsive movements. After training, the CNN was tested on thermal video streams not included in the training set and had a 99.2% accuracy in classifying convulsive movements and non-convulsive episodes, with no false negatives to distinguish between the occurrence and non-occurrence of convulsive movements. The performance results show that the thermopile sensor array has the potential to detect convulsive seizures while maintaining patient privacy and not requiring direct patient contact.
BibTeX
@inproceedings{hanosh2020convulsive,
  author = {Ouday Hanosh and Rashid Ansari and Naoum P. Issa and A. Enis Çetin},
  title = {{Convulsive Movement Detection using Low-Resolution Thermopile Sensor Array}},
  booktitle = {Workshops @ CVPR},
  year = {2020},
  url = {https://openreview.net/forum?id=4A6oyCGPUC},
}
Figure