HLA-B27 陽性患者中，併發僵直性脊椎炎之患者特性：基於機器學習的研究

HLA-B27已知為與僵直性脊椎炎密切相關的基因。僵直性脊椎炎（Ankylosing Spondylitis，簡稱AS）患者中約七至九成為此基因的陽性者；然而，此基因的陽性者卻僅一到兩成患有AS，這樣的現象讓人好奇HLA-B27基因與AS中間的致病機轉為何。本研究旨在利用TriNetX全球大型病歷資料庫的資料，並以監督式機器學習演算法對HLA-B27的陽性者進行「是否罹患僵直性脊椎炎」的預測，藉以觀察哪些特徵（預測變數）對於預測HLA-B27陽性者的僵直性脊椎炎具有較高的重要性，以及建立具有潛力的HLA-B27陽性者僵直性脊椎炎預測模型。本研究使用TriNetX資料庫，從中挑選出4525位HLA-B27基因陽性者樣本，以：在資料中是否顯示「至少被診斷過三次任意部位的僵直性脊椎炎」為標準，定義出本研究的確診者1803位與非確診者2722位，作為監督式學習的標籤（待預測問題的答案）。接著以邏輯式回歸、隨機森林、梯度提升決策樹、深度學習等四種機器學習演算法建立模型，對樣本進行「是否罹患僵直性脊椎炎」的預測。由隨機森林演算法建立的模型為本研究中表現最佳的模型，且多數實驗室檢測結果特徵對模型的預測重要性較診斷碼特徵高。本研究隨機森林模型的預測準確率約71%、精確率約60%、敏感度約82%、F1 score約70%、ROC/AUC約80%、特異度約64%。邏輯式回歸與梯度提升決策樹模型在本研究中缺乏隨機性，而深度學習模型的穩定性較差。在原始特徵資料集加入C反應蛋白作為特徵後提升了隨機森林模型的預測能力。以隨機森林模型為基礎，較好的敏感度和準確率對於HLA-B27陽性者的僵直性脊椎炎初步預測可起到一定的作用。TriNetX平台資料庫的齊全程度、結果定義的準確性以及選樣偏差對於模型準確度等各項指標可能造成的影響，將來進行進一步研究時應更加深度探討，以發展出作為HLA-B27患者預測AS的實用工具，並擴大使用族群，增添研究的外推性。

HLA-B27 has been identified in numerous studies as a gene highly associated with Ankylosing Spondylitis (AS), with approximately 70% to 90% of AS patients carrying this gene. However, only 10% to 20% of HLA-B27 positive individuals develop Ankylosing Spondylitis. This research aims to utilize TriNetX, a global electronic medical record database, and supervised machine learning techniques to predict the occurrence of AS in HLA-B27 positive patients. By analyzing the predictions, we can identify the most significant features contributing to the model's accuracy and potentially develop a reliable AS prediction model for HLA-B27 positive patients.This research utilized the TriNetX database, selecting 4,525 HLA-B27 positive samples. Based on the criterion of being diagnosed with AS at least three times, 1,803 individuals were classified as AS positive, and 2,722 were classified as AS negative. Subsequently, I employed four algorithms—Logistic Regression, Random Forest, Gradient Boosting Decision Tree（GBDT）, and Deep Learning—to build prediction models and applied these models to our sample data.The model built with Random Forest demonstrated the best performance in this research, with laboratory result features contributing significantly to the model's accuracy.The Random Forest model achieved an accuracy of approximately 71%, a precision of around 60%, a sensitivity of about 82%, an F1 score of approximately 70%, an ROC/AUC of around 80%, and a specificity of about 64%. The Logistic Regression and Gradient Boosting Decision Tree models showed no significant improvement, while the deep learning model exhibited the least stability.The predictive ability of the original feature dataset improved after adding CRP as a feature.The results of this research establish a foundation using the Random Forest model for predicting AS in HLA-B27 positive individuals. However, there are limitations to consider: the data integrity from TriNetX, the accuracy of the criteria used to define the samples, and selection bias can all impact model performance. These issues will require deeper investigation in future research.The Random Forest model serves as a preliminary screening tool for predicting AS in HLA-B27 positive populations, demonstrating good sensitivity and accuracy. By addressing the current limitations, such as the selection of study subjects, data filtering, and supervised machine learning label definitions, there is potential to enhance the model's performance in the future. This would enable its practical use in predicting AS among HLA-B27 positive patients, extend its applicability to a broader population, and enhance the generalizability of the research.