[Paper Review] Deep learning from wristband sensor data: towards wearable, non-invasive seizure forecasting

Christian Meisel, Rima El Atrache+4 more|arXiv (Cornell University)|Jun 3, 2019

EEG and Brain-Computer Interfaces24 references8 citations

TL;DR

This study proposes a non-invasive seizure forecasting method using deep learning on multi-modal wristband sensor data (EDA, BVP, temperature, actigraphy) from 50 epilepsy patients. Using 1D and 2D convolutional neural networks on raw sensor inputs, it achieves better-than-chance prediction of pre-ictal states in 60% of patients via leave-one-out and 43% via pseudo-prospective validation, demonstrating feasibility for wearable, non-invasive seizure risk assessment.

ABSTRACT

Seizure forecasting may provide patients with timely warnings to adapt their daily activities and help clinicians deliver more objective, personalized treatments. While recent work has convincingly demonstrated that seizure risk assessment is possible, these early approaches relied largely on complex, often invasive setups including intracranial electrocorticography, implanted devices and multi-channel EEG, which limits translation of these methods to broad clinical application. To facilitate broader adaptation of seizure forecasting in clinical practice, non-invasive, easily applicable techniques that reliably assess seizure risk, in combination with clinical information, are crucial. Wristbands that continuously record physiological parameters, including electrodermal activity, body temperature, blood volume pressure and actigraphy, may afford monitoring of autonomous nervous system function and movement relevant for such a task, hence minimizing potential complications associated with invasive monitoring, and avoiding stigma associated with bulky external monitoring devices on the head. Here, we use deep learning to analyze long-term, multi-modal wristband sensor data from 50 patients with epilepsy (total duration $>$1400 hours) to assess its capability to distinguish preictal from interictal states. Prediction performance is assessed using area under the receiver operating charateristic (AUC) and improvement over chance (IoC) based on F1 scores. Using one- and two-dimensional convolutional neural networks, we identified better-than-chance predictability in out-of-sample test data in 60\% of the patients in leave-one-out and 43\% of patients in pseudo-prospective approaches. These results provide a step towards developing easier to apply, non-invasive methods for seizure risk assessments in patients with epilepsy.

Motivation & Objective

• To evaluate whether non-invasive wristband sensor data can enable reliable seizure risk forecasting using deep learning.
• To assess the performance of 1D and 2D convolutional neural networks in distinguishing pre-ictal from interictal states using raw, multi-modal physiological signals.
• To determine if the method generalizes to out-of-sample, prospective-like testing conditions, simulating real-world deployment.
• To explore the potential of wearable, continuous monitoring for improving patient autonomy and clinical treatment personalization in epilepsy.

Proposed method

• Collected multi-modal wristband data (EDA, BVP, temperature, 3D actigraphy) from 50 epilepsy patients during in-hospital monitoring, with a total duration exceeding 1400 hours.
• Defined pre-ictal states as 30-second segments 61 to 1 minute before seizures and interictal states as segments ≥2 hours from any seizure.
• Used 1D and 2D convolutional neural networks (CNNs) with raw sensor data as input, avoiding handcrafted feature engineering.
• Applied leave-one-out cross-validation and pseudo-prospective validation to assess generalization and prospective performance.
• Evaluated model performance using area under the ROC curve (AUC) and improvement over chance (IoC) based on F1 scores.
• Selected the longest wristband recording per patient when multiple devices were used, ensuring data consistency.

Experimental results

Research questions

• RQ1 Can deep learning models trained on raw, multi-modal wristband sensor data distinguish pre-ictal from interictal states in epilepsy patients?
• RQ2 Does the proposed method achieve better-than-chance prediction performance in both leave-one-out and pseudo-prospective validation settings?
• RQ3 How does the performance of 1D CNNs on raw data compare to 2D CNNs on power spectral density representations of the same data?
• RQ4 To what extent can this non-invasive, wearable approach support real-world, prospective seizure forecasting?

Key findings

• The proposed method achieved better-than-chance predictability in 60% of patients using a leave-one-out cross-validation approach.
• In a more realistic pseudo-prospective validation setting, the method demonstrated better-than-chance performance in 43% of patients.
• 1D convolutional neural networks using raw sensor inputs outperformed 2D CNNs using power spectral density representations in the tested setup.
• The study confirms that multi-modal wristband sensor data combined with deep learning can detect pre-ictal states with statistically significant accuracy, even in out-of-sample, prospective-like conditions.
• The results suggest that non-invasive, wearable monitoring systems could enable scalable, patient-friendly seizure forecasting in clinical practice.