Rationale:
Home videos are increasingly presented in clinical practice, and secure digital tools for clinical video sharing are urgently needed. In infantile epileptic spasm syndrome (IESS), for example, the use of smartphone videos has been shown to accelerate time to diagnosis. Timely and accurate recognition of seizures like epileptic spasms (ES) is critical since delayed treatment may lead to severe cognitive and developmental impairments. In patients with known epilepsy home long-term videos may provide more objective seizure counts and classification since patient seizure self-reports are often inaccurate. However, video seizure evaluation still requires analysis by a specialist which limits its broader applicability. Furthermore, there is a lack of low-cost, scalable video-diagnostic solutions applicable both to early epilepsy diagnosis and long-term monitoring. This study evaluated the feasibility of a secure video-sharing digital platform supported by artificial intelligence (AI)-based analysis of videos for automated seizure detection from home videos.
Methods:
We first conducted a prospective pilot study that included 60 children (median age 5.5 years) with suspected seizures and their caregivers at a tertiary center. Participants were invited to submit videos of suspected events through a secure smartphone application. Videos were reviewed by an epileptologist and in parallel participants underwent standard diagnostic evaluation. Outcome measures included video quality, concordance of video assessment with diagnostic workup, and user evaluation questionnaires. Second, we used a large video dataset of different seizure types (ES, tonic-clonic seizures (TCS), hyperkinetic seizures (HKS), psychogenic non-epileptic seizures (PNES)) and controls to evaluate the performance of AI-based seizure detection.
Results:
Of recruited participants, 16 (26.7%) used the app to submit 81 videos (median 5.1 videos per patient), with 94% (76/81 videos) having sufficient quality for meaningful clinical review. Among these participants, 11 (68.8%) had confirmed epilepsy as their final diagnosis, while 5 (31.2%) did not. Expert review classified 48.1% of videos as likely epileptic events, with generalized tonic-clonic seizures being most common (24.7%). Concordance between submitted videos and final clinical diagnosis was demonstrated in 68.8% of cases, with higher concordance significantly related to the number of videos submitted (mean 6.4 vs. 2.2 videos per patient, p=0.04). User evaluation showed favorable ratings for app usability (mean score 5.4/7). AI-based seizure detection indicated high performance for detection of ES and TCS followed by lower performances for HKS and PNES.
Conclusions:
This study demonstrates the feasibility of secure digital video transfer in epilepsy diagnostics, emphasizing the importance of multiple video submissions for accurate diagnosis. AI-based video seizure detection may reduce the workload in a shared human-AI decision support setting. Future implementation should address barriers to wider adoption and investigate clinical impact.
Funding: N/A