運用機器學習方法進行心血管疾病的直接與間接預測

Abstract

本研究著重於建構心血管疾病直接與間接因子預測模型之模型與流程之建構。心血管疾病在世界是造成傷殘與死亡的重大因素，若可以及早預測罹患風險，即可提供患者相關預先治療措施以降低傷亡率。本研究分成兩大部分，第一部分提出心血管疾病長短期風險預測模型，第二部分提出心血管疾病風險因子- 夜間動態血壓值之預測模型，並建構特徵工程與特徵選擇之處理流程。 在心血管疾病的研究中，展示了心血管疾病預測模型的良好性能，尤其是在長期預測方面。其中XGBoost模型表現出最佳預測性能。在夜間動態血壓的研究方面，三個不同地區與種族的資料集一致地建議早上6點和晚上9點為家庭血壓測量的最佳時間，這是有別於過往居家血壓量測的時間組合，可以在未來提供臨床相關建議。研究顯示使用特徵工程策略2搭配順序性特徵選擇方法組合所產生的人工特徵有助於提升預測效能，而在模型方面，MLP神經網絡相較於傳統機器學習與迴歸模型表現最佳。此外，本研究發現損失函數的調整有助於預測表現的提升。相較於過往研究，本研究廣泛的納入地域性、種族性的比較，並發現在地域性的比較上，兩資料集皆強調了數值特徵與人工特徵的重要性，並證明其有助於提升預測表現；在種族性比較上，研究與TCHC資料集的比較凸顯了性別在預測夜間動態血壓方面的重要性。本研究在夜間動態血壓的預測主題上無論是模型、特徵工程、特徵選擇、損失函數調整皆有有別於過往研究的嘗試與突破，並廣泛的進行比較與闡述，可以對未來夜間動態血壓領域的研究提供研究幫助。

This study focuses on constructing models and workflows for the prediction of both direct and indirect factors associated with cardiovascular disease (CVD). CVD is a major cause of disability and mortality worldwide. Early prediction of the risk of developing CVD can provide patients with relevant preventive measures to reduce morbidity and mortality. The study is divided into two main parts. The first part proposes a model for the prediction of long-term and short-term CVD risk. The second part presents a predictive model for cardiovascular risk factors, specifically nighttime ambulatory blood pressure values. Feature engineering and feature selection processes are implemented in the construction of these models. The CVD research demonstrates the favorable performance of the cardiovascular disease prediction models, particularly in long-term predictions. Among the models assessed, XGBoost emerges as the optimal predictive model. Concerning the investigation of nocturnal ambulatory blood pressure, three distinct datasets representing different regions and ethnicities consistently suggest that 6 AM and 9 PM are the optimal times for home blood pressure measurements, this finding contrasts with previous combinations of home blood pressure measurement times and could offer valuable clinical recommendations in the future. The study reveals that the utilization of artificial features generated through Feature Engineering Strategy 2, combined with sequential feature selection, contributes to enhanced predictive performance. Among the model architectures explored, the MLP neural network outperforms traditional machine learning and regression models. Furthermore, adjustments to the loss function are shown to improve prediction performance. In comparison to prior research, this study extensively incorporates regional and ethnic comparisons and demonstrates the importance of numerical and artificial features in both datasets, contributing to enhanced predictive performance. In terms of ethnic comparisons, the study highlights the significance of gender in predicting nocturnal ambulatory blood pressure, particularly concerning the TCHC dataset. The research offers novel approaches and breakthroughs in model construction, feature engineering, and feature selection, providing valuable insights for future studies on nighttime ambulatory blood pressure prediction.