以演化式機器學習建構可解釋性脂肪肝風險預測模型

An interpretable steatotic liver disease risk prediction model based on evolutionary machine learning

脂肪肝為近年快速增加之代謝性肝病，具有高盛行率與疾病進展風險，早期辨識與風險預測對臨床決策具重要意義。本研究使用花蓮慈濟醫院匿名化血液檢驗資料，共426筆完整樣本（脂肪肝234筆；非脂肪肝192筆），以腹部超音波診斷為黃金標準，用以發展脂肪肝風險預測模型。
模型方法包含支援向量機（SVM）、決策樹（J48、PART）與演化式基因規劃（Genetic Programming, GP），並採用10-fold cross-validation評估效能。結果顯示，配合統計檢定進行特徵顯著性篩選後，多數模型效能略有提升，其中以GP表現最佳（ACC=75.35%、AUROC=0.7958）。此外，J48模型提供可視化決策規則，有助臨床理解決策邏輯；GP則以算式形式呈現分類結構，揭示eGFR與HDL-C等指標對預測的影響，並具可追溯與解譯優勢。結果也發現部分變數方向與臨床文獻不同，顯示評估模型的發展仍有改善空間。
整體結果支持可解釋AI在脂肪肝風險辨識之潛力，未來可藉由系統化解譯模組、擴充運算子設計與導入臨床知識，以發展兼具準確性與臨床可信度之決策支援模型。

Steatotic liver disease (SLD) is a rapidly increasing metabolic liver disorder with high prevalence and progressive risk. Early identification and risk prediction are important for clinical decision-making. This study used anonymized blood test data from Hualien Tzu Chi Hospital, consisting of 426 complete samples (234 steatotic liver vs. 192 non-steatotic liver), with abdominal ultrasound diagnosis as the reference standard for developing a risk prediction model.
Machine learning methods included Support Vector Machine (SVM), Decision Tree (J48 and PART), and evolutionary Genetic Programming (GP). Model performance was evaluated using 10-fold cross-validation. Results showed that feature selection based on statistical significance slightly improved model performance across most methods, with GP achieving the best results (ACC=75.35%, AUROC=0.7958). In addition, the J48 model produced clear decision rules that support clinical interpretability, while GP generated explicit mathematical expressions that highlight key predictors such as eGFR and HDL-C, providing traceable and explainable decision logic. Some feature directions differed from clinical literature, indicating the need for further model refinement.
Overall, findings support the potential of explainable AI for steatotic liver disease risk identification. Future work may focus on enhanced interpretation modules, extended function operators, and integration of clinical knowledge to build a decision support model with improved accuracy and trustworthiness in clinical practice.