¹Gifu University Graduate School of Medicine, Department of Psychiatry, Gifu, Japan, ²Teikyo University, Department of Psychiatry, Tokyo, Japan, ³National Center for Global Health and Medicine (NCGM), Genome Medical Science Project (Toyama), Tokyo, Japan, ⁴Warakukai Medical Corporation, Panic Disorder Research Center, Tokyo, Japan, ⁵The University of Tokyo, Center for Research on Counseling and Support Services, Tokyo, Japan, ⁶The University of Tokyo, Department of Physical and Health Education, Graduate School of Education, Tokyo, Japan, ⁷Mie University, Center for Physical and Mental Health, Mie, Japan, ⁸Gifu University Graduate School of Natural Science and Technology, Department of Intelligence Science and Engineering, Gifu, Japan

Background: Panic disorder (PD) is a modestly heritable condition. Currently, diagnosis is based only on clinical symptoms; identifying objective biomarkers and a more reliable diagnostic procedure is desirable.

Aims & Objectives: We investigated whether people with PD can be reliably diagnosed utilizing combinations of multiple polygenic scores (PGSs) for psychiatric disorders and their intermediate phenotypes, compared to single PGS approaches, by applying specific machine learning techniques.

Method: PGSs for 48 psychiatric disorders and intermediate phenotypes based on large-scale genome-wide association studies (n=7,556–1,131,881) were calculated for people with PD (n=718) and healthy controls (HCs, n=1,717). Discrimination between people with PD and HCs was based on the 48 PGSs using five methods for classification: logistic regression (LR), neural networks (NN), quadratic discriminant analysis (QDA), random forests (RF), and a support vector machine (SVM). Differences in discrimination accuracy (AUC) due to an increased number of PGS combinations and differences in the accuracy across five classifiers were investigated.

Results: All five classifiers performed relatively well for distinguishing people with PD from HCs by increasing the number of PGSs. Of the 48 PGSs, the PGS for anxiety UK Biobank (UKBB) was the most useful for discrimination by the classifiers. In combinations of two or three PGSs, the PGS for anxiety UKBB was included as one of PGSs in all classifiers. When all 48 PGSs were used in combination, the greatest AUCs significantly differed among the five classifiers. SVM and LR had higher accuracy than QDA and RF. For each classifier, the greatest AUC was 0.600±0.030 for LR (PGS combinations N=14), 0.591±0.039 for NN (N=9), 0.603±0.033 for QDA (N=10), 0.572±0.039 for RF (N=25), and 0.617±0.041 for SVM (N=11). The greatest AUCs at the best PGS combination significantly differed among the five classifiers. RF had the lowest accuracy among classifiers. SVM had higher accuracy than NN.

Discussion & Conclusions: These findings suggest that increasing the number of PGS combinations up to approximately 10 effectively improved the discrimination accuracy and that SVM exhibited greater accuracy among classifiers. However, the discrimination accuracy for PD, when based solely on PGS combinations, was found to be modest.