To demonstrate feasibility of computational analysis detection of seizure clusters (SCs) in free text notes (FTNs) in an electronic seizure diary data set.

FTNs in electronic diaries reflect concerns and priorities of patients and provide supplemental information to the diary forms. These notes have been largely unused due the difficulty of data reduction, requiring an automated method. This project evaluated feasibility of using computational analysis to identify FTNs relevant to SCs.

Data was extracted over 2010-2019 from EpiDiary, a free epilepsy diary with 44,799 unique users, generating 1,096,168 entries. Expert reviewers rated the first 10 chronological notes from a sample of 100 patients for SCs (defined as ≥ 2 seizures/24 h). Later, the same FTNs were filtered by computational analysis for keywords related to SCs.

13,987 (32.1%) individual users had at least one FTN. The average number of notes per user was 30.7, median 9, totaling 245,792 FTNs.  Reviewers inspected a sample of 100 diaries with a total of 3,973 notes. Inter-rater agreement of the three expert reviewers was 0.94 for a definite cluster, according to Fleiss’ kappa. A computational analysis algorithm processed the same notes and the ratings were added to the inter-rater analysis as a 4^th rater, resulting in a Kappa score of 0.85, equivalent of “near perfect agreement.” Categorization, including the computer, agreed in 784 notes (88.59%) and disagreed in 101 (11.41%). FTNs identified SCs without a corresponding patient-selected check-mark in the diary cluster box in 64 instances (7.23%). Several notes had medical or personal information that might have been important to their medical care.

Patient-generated FTNs are mostly untapped resources. This study demonstrates feasibility of utilizing computational analysis to identify FTNs with useful information, such as SCs. This method of analysis shows the potential to gain new insights from our patients using FTNs.