Rationale:
Timely detection of seizures is crucial for intervention and may alleviate complications and injury from seizures (Shellhaas, 2019). Wearable sensors have shown promising results for automated seizure detection, but this method is limited by patient tolerance, is not practical for use in newborns, and devices currently require recharging. Additional non-contact video and audio-based technologies have become available, but a detailed literature review of automated video and audio-based seizure detection publications is lacking.
Methods:
An initial PubMed search was aimed to identify relevant keywords. We combined keywords using Boolean logic with appropriate controlled vocabulary mined from known relevant articles. This strategy was translated into six other databases. Searches were run from 2006 to June 27th, 2022. All results were deduplicated and imported into Covidence (Melbourne, Australia). Titles and abstracts were screened by two independent reviewers for assessment against the inclusion criteria: primary research on automated audio and/or video detection of any seizure type, patients with a diagnosis of epilepsy, and articles written in English. The full text of selected citations was then assessed again against the inclusion criteria by the same two independent reviewers. Figure 1 details the study design from database searches to final article selection.
Results:
Initial search terms identified 4,487 deduplicated articles. Thirty-eight studies met inclusion criteria and varied in design, algorithm methods, and performance metrics. Most studies focused on myoclonic, generalized tonic-clonic (GTC), and clonic seizures, but several studies also analyzed absence seizures as well as other minor seizures.
Sensitivity ranged from 100% to 78%, with optical flow analysis showing the highest sensitivity. Other detection methods included convolutional neural networks (CNN), long short-term memory (LSTM), and leave-one-out cross validation, among many other machine learning methods. Audio-based methods achieved a sensitivity of 98%, suggesting their potential complementarity to video-based approaches. Specificity of all included articles ranged from 97% to 70%. Limited studies reported accuracy, but when reported, optical flow and audio-based algorithms demonstrated the highest accuracy. Table 1 includes a complete list of results.
Conclusions:
This scoping review demonstrates feasibility of video seizure detection, and improvement in detection sensitivity and sensitivity by integrating audio-based detection into video detection algorithms. Limitations in sample sizes and standardized evaluation protocols highlight the need for further research. Further testing and algorithm improvements will contribute to the advancement of non-invasive seizure detection and improve the management of epilepsy.
References
Shellhaas RA. Seizure classification, etiology, and management. Handb Clin Neurol. 2019;162:347-361. doi: 10.1016/B978-0-444-64029-1.00017-5. PMID: 31324320.
Covidence systematic review software, Veritas Health Innovation, Melbourne, Australia. www.covidence.org
Funding: N/A