A1244
Title: XGBoost-based prediction of self-reported diseases and QTs from multimodal image reports and PRS in Taiwan biobank
Authors: Yi-Ju Lee - Academia Sinica (Taiwan) [presenting]
Sung-Lin Tsai - Academia Sinica (Taiwan)
Chia-Wei Chen - Academia Sinica (Taiwan)
Shih-Kai Chu - National Cheng Kung University (Taiwan)
Chih-Ting Yang - Vanderbilt University (United States)
Jenn-Hwai Yang - Academia Sinica (Taiwan)
Yu-Jen Liang - Academia Sinica (Taiwan)
Yin-Chun Lin - Academia Sinica (Taiwan)
Han-Ni Hsieh - Academia Sinica (Taiwan)
Sheng-Yuan Hu - Academia Sinica (Taiwan)
Tso-Jung Yen - Academia Sinica (Taiwan)
Chun-houh Chen - Academia Sinica (Taiwan)
Hsin-Chou Yang - Academia Sinica (Taiwan)
Abstract: Taiwan Biobank (TWB) multimodal medical imaging data offers rich predictive value, yet structured image reports across diseases and quantitative traits (QTs) remain underexplored. A systematic evaluation of five imaging modalities integrated with polygenic risk scores (PRS) and demographic data via XGBoost was conducted among 22,067 TWB first-wave follow-up participants (mean age 56.08, range 32.52 to 79.60). Five modalities -abdominal ultrasound (ABD), bone densitometry (BMD), electrocardiography (ECG), carotid ultrasound (CAU), and thyroid ultrasound (THY)- yielded 155 structured features, integrated with 3,636 PRS and demographic variables. XGBoost models predicted 34 self-reported diseases and 73 QTs using 64/16/20\% splits. Five diseases met criteria (AUC > 0.7; over 5\% improvement over demographics): hypertension, stroke, diabetes, cataract, and osteoporosis. CAU features drove cardiovascular prediction; BMD dominated metabolic and skeletal outcomes. Multimodal integration improved AUC by 8 to 15\%. Image data outperformed PRS alone in 63/73 QTs; top-predicted QTs included TSH, HbA1c, and creatinine. A force-directed network positioned diabetes near glycemic QTs, revealing system-level pathological associations. Multimodal image reports enable robust disease and QT prediction, supporting smart health applications and advancing precision medicine for the Han Chinese population.
