具有自動化特徵選取方法之機器/深度學習型慢性與代謝疾病早期預測及其病因防治系統

Abstract

本研究旨在根據所提出的基於特徵向量的自動特徵選擇方法(Eigenvector- based Feature Selections, EFSs)，來實現滿足專家觀點的輸入特徵子集的快速優化，以提高例如脂肪肝 (Fatty Liver Disease, FLD) 與高血壓 (Hypertension)等慢性或代謝疾病基於機器或深度學習進行早期預測模型的效能，並據此為單一或多種具有共病性或併發症因子的慢性或代謝疾病，就治療或生活方式改變提供簡單可行之建議。 在實驗中，針對脂肪肝及高血壓的預測與防治為例，我們分別探索來自台灣北部的健檢中心的大規模且高維度數據集(包括從1999到2017年約50萬筆記錄)。進行資料前處理及清理後，以所提出的三種植基於特徵向量的自動特徵選擇方法(即EFS-TW、EFS-TRW及EFS-RW)、聯合交集 (Intersection of Union, IoU) 和覆蓋率 (coverage) 最高值來確定包含單一疾病病因的最佳特徵子集，並使用各種與長短期記憶 (LSTM) 相關的分類器進行此種單一疾病預測及評估模型與系統性能。實驗結果顯示，針對脂肪肝及高血壓的預測與防治，EFS-TW及EFS-RW分別可選出對應最佳的特徵子集且耗費最短的總計算時間，且相比之下，最高的IoU、覆蓋率和計算時間都優於講求精確卻耗時的序列前向特徵選擇 (Sequence Forward Feature Selection, SFS)。 此外，從預防醫學的角度來看，早期預測並從眾多致病因素中選擇關鍵因素來制定簡單易行的預防計劃是有必要的，因此，本研究進一步提出了模式匹配和概率樹（Pattern Matching and Probability Tree, PMPT）方法，用於根據此潛在的高血壓患者和數據集內與他具有相同致病特徵與模式的罹病者的統計值，來預測他未來基於排名前n個特徵的罹病概率與擬訂醫療或生活型態的防治計畫，使其易於遵循、預防高血壓等慢性和代謝疾病的發生，且有助於醫院的家庭醫學科或健檢中心達成疾病早期預防、精準醫療、決策輔助及健康管理等目的。

This study aims to achieve rapid and automatic optimization of input feature subsets that satisfy expert’ domain knowledge based on the proposed Eigenvector-based Feature Selections (EFSs) to improve the performance of early prediction models based on machine or deep learning for chronic or metabolic diseases, such as fatty liver disease (FLD) and hypertension, and accordingly provide simple and feasible suggestions for the treatment or lifestyle changes of single or multiple types of chronic or metabolic diseases with the comorbidity or the complication factors. In the experiment, for the prediction and prevention of FLD and hypertension, we separately explored a large-scale and high-dimensional dataset including around 500,000 records from 1999 to 2017 from a health examination center in Taipei. After data preprocessing and cleaning, we obtained the optimal feature subset by using the proposed EFSs (That is, EFS-TW, EFS-TRW, and EFS-RW) and the best intersection of union (IoU) and coverage, and then we used Long Short-Term Memory (LSTM)-related classifiers for FLD and hypertension prediction and model performance evaluation. Experimental results show that the optimal feature subset selected by EFS-TW for FLD prediction and EFS-RW for hypertension prediction have the highest IoU and coverage, and the corresponding feature selections consume the far shorter total computation time than that of Sequence Forward Feature Selection (SFS). From the perspective of preventive medicine, it is necessary to develop a simple and easy-to-follow prevention plan by early prediction and selection of the key risk factors. Accordingly, we further propose a pattern matching and probability tree (PMPT) method, based on the statistics of diagnosed and non-diagnosed hypertension patients with the same pathogenic features, trends and patterns in the data set as the potential patient, to predict the probability based on each of the top n features in the future and formulate the prevention plan in medical arrangement or life style change. The proposed method makes the prevention and treatment plan easy to follow, prevents the occurrence of chronic and metabolic diseases, and helps the family medicine department of the hospital or health examination center to achieve the purposes of early disease prevention, precision medicine, decision-making support and health management.