[1K3-GS-10] AI application: Medicine / Healthcare

[1K3-GS-10-03] Comparison of Feature Extraction Methods for Prognosis Prediction in Hereditary Arrhythmia Disease

〇Takeshi Mitani^1,2, Yuji Okamoto¹, Kohjitani Hirohiko^1,3, Takanori Aizawa³, Takeru Makiyama^1,3, Seiko Ohno⁴, Minoru Horie⁵, Yasushi Okuno¹ (1. Graduate School of Medicine, Kyoto University, Kyoto, 2. Department of Education and Training, Kurashiki Central Hospital, Kurashiki, 3. Department of Cardiovascular Medicine, Kyoto University, Kyoto, 4. Department of Bioscience and Genetics, National Cerebral and Cardiovascular Center, Suita, 5. NCD Epidemiology Research Center, Shiga University of Medical Science, Otsu)

Keywords:Hereditary Disease, Foundation Model, Explainable AI

This study presented multiple prognosis prediction methods for Long QT Syndrome type 2 (LQT2), a hereditary arrhythmic disease. Specifically, we compared the effectiveness of a traditional method using Multiple Sequence Alignment (MSA) with that of a Foundation model (ProtBert) pre-trained on a large dataset without MSA. The results indicated that the method using ProtBert with reconstruction showed the highest prognostic accuracy, suggesting that it is effective in predicting LQT2 prognosis. It is also applicable to the analysis of genetic variants, and this method may be particularly useful for prognosis prediction in situations where annotation costs are high and labeled data sets are scarce.