EXPLORING THE FEASIBILITY OF INSOMNIA SUBTYPING USING MACHINE LEARNING CLUSTERING BASED ON DIGITAL PHENOTYPE DATA

초록

Background: Insomnia is diagnosed through patients' voluntary reports. This feature leads to ambiguity in the diagnosis and subtyping of insomnia.(Ferini-Strambi et al., 2019) Digital phenotyping means to regard behavior patterns that appear on digital devices or online as a kind of phenotype and is attracting attention as a means of future medical care, but there are not many cases applied to insomnia.(Kim et al., 2023, Lee et al., 2023) Aims & Objectives: Therefore, our study performs data-driven subtyping of insomnia to confirm the relationship between clinical characteristics and digital phenotypes, and this will enable a new approach to the disease. Method: Participant recruitment is done by Korea University Anam Hospital. In this sample clustering, 73 subjects were included. All subjects are adults between 19 and 70 with an ISI insomnia severity index score of 15 or higher. Digital phenotypes related to the quality of sleep, life patterns, and environment of the subject are recorded on the application (for 4 weeks). Using 16 variables (working pattern, life pattern, number of steps/exercise time/movement distance, average heart rate, average total sleep time), K-means clustering of sci-kit learn was done to find the best grouping model. Then, a between-group analysis was conducted using data from the initial evaluation questionnaire of subjects. ANOVA and T-test or Chi-square test were conducted depending on the shape of the variable. Results: We have applied the K-means clustering algorithm, and the optimal cluster number (k) was determined using the elbow curve method. According to the curve, the elbow point is at K=3. Therefore, 3 subgroups were divided (cluster0; n=32, cluster1; n=31, cluster2; n=10). Among 16 axes, steps, exercise time, and movement distance for each 4 weeks illustrated the best data point distributions to form clusters. Average heart rate, average total sleep time, working pattern, and life pattern had lower correlation coefficients with other variables, and cluster boundaries were unclear. After 10-fold cross-validation, the accuracy score was 80%, and the SHAP value plot was visualized. According to the plot, movement distance, exercise time, and steps were important features. Then demographic and clinical characteristics were compared of each cluster using ANOVA and Chi-square test. Among 28 features (age, lifestyle, BMI, and case report form of 25 self-rating questionnaires including sleep-related scale, moodrelated scale, and health-related scale), only the DBAS score showed a significant difference between groups (p-value=0.04). A higher DBAS score indicates more dysfunctional beliefs and attitudes about sleep. The box plot shows the DBAS scale difference between clusters. A student’ s t-test was performed to show C0=C1, C1=C2, C0