請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/84475
標題: | 使用電子病歷資料預測抗憂鬱藥的治療困難度 Predicting the Level of Treatment Difficulty with Antidepressants Using Electronic Health Records |
作者: | Ju-Yi Tung 董如意 |
指導教授: | 郭柏秀(Po-Hsiu Kuo) |
共同指導教授: | 蕭朱杏(Chuhsing Kate Hsiao) |
關鍵字: | 電子病歷,治療困難度,治療反應,抗憂鬱藥,抗憂鬱藥治療耐性評估量表, Electronic health record,Treatment difficulty,Treatment response,Antidepressant,Treatment Resistance to Antidepressants Evaluation scal, |
出版年 : | 2022 |
學位: | 碩士 |
摘要: | 重鬱症是最常見和負擔最重的疾病之一。抗憂鬱藥是最常見的一線治療策略。然而,接受抗憂鬱藥治療的患者往往不能完全緩解,並且經常復發。如果醫生能夠更早地了解患者的治療結果,則可以分配更好的治療方案或整合資源來幫助患者。電子病歷數據提供了豐富的患者處方和健康信息。將機器學習應用於 電子病歷資料已成為預測治療反應的一種有前途的手段。 本研究旨在使用電子病歷資料預測(基於抗憂鬱藥治療抗性評估量表 TRADES)治療重鬱症的抗憂鬱藥治療難度。我們試圖探索不同時間區間中的模型表現的並找到重要變項。我們還探討了不同變項組合對模型性能的影響。最後,我們嘗試將模型使用在未評估 TRADES的案例。 本研究共收入179位在2010年10月至2021年5月在台大醫院就診的重鬱症國際疾病分類診斷代碼患者。符合條件的病人為20至75歲。本研究基於TRADES,該量表評估使用抗憂鬱藥治療的重鬱症病人的困難治療程度。TRADES的評分基於從完整和詳細的醫療病歷紀錄得出的終生治療概況。使用TRADES評估所需的最短持續時間是重鬱症首次發作後2年。我們將TRADES評分超過6分定義為難以治療,並使用電子病歷資料構建預測模型,而不是查看病歷。我們執行了200次並記錄了每個group LASSO模型在每個預測時間區間(6個月、1年、1.5年、2年、全時段)中選擇的變項。較高頻率被group LASSO留下的變項會被選中進入Random Forest和Logistic Regression預測治療難度。使用獨立測試集中的 AUC、準確性、敏感性和特異性評估模型性能。我們試圖將模型表現最好的全追蹤時間的完整EHR模型運用在64 名新重鬱症患者來預測其治療困難度。在次要分析中,我們將基因資料納入預測模型以提高模型性能。 在本研究AUC在全追蹤時間達到0.77 (0.76~0.78),在6個月的prediction window中達到0.68 (0.66~0.69)。雖然模型性能隨著prediction window的延長而緩慢提升,但沒有明顯的進步。模型3(僅EHR代碼)在不同prediction window中通常不是最佳性能,但6個月prediction window AUC與模型4(完整電子病歷模型)相似。在2年prediction window中,最佳AUC是模型2(手刻變相)。就如預期,模型4在不同的prediction window中具有更好的模型性能且添加的遺傳數據提高模型性能。Bupropion, Tinten, Alprazolam,便秘相關診斷和藥物治療、多階段心理測驗(電子病歷代碼)、藥物當量(augmentation)、使用超過4週的augmentation種類(A3_2)、使用超過 2 種抗抑鬱藥的數量種類(A3_1)和精神科急性病房住院次數是最常見的重要變項。我們試圖外推模型(model 4),預測新病人的難治率約為43.3%。 本研究的研究限制為缺乏對預測治療難度可能很重要的其他因素,如家庭支持、遵囑性、非結構化電子病歷(臨床記錄、圖像數據、護理記錄)。與其他電子病歷研究相比,本研究樣本量較小,受限於需要評估過TRADES的個案,但我們後來將模型運用到少量之前未評估 TRADES 的個案,並試圖證明預測結果是否可靠。 儘管未來臨床適用性可能受限於模型表表現,但電子病歷研究仍然展現其潛力,值得未來投注更多心力在這個方向 Major depression disorder (MDD) is one of the most common and burdened disease. Antidepressants are the most common first-line treatment strategy. However, patients treated with antidepressants often did not achieve full remission and had recurrence and relapse frequently. If physicians can know earlier about treatment outcomes of patients, a better treatment regimen or integrating resources could be allocate to help patients. Electronic health record (EHR) data provide abundant prescription and health information of patients. Applying machine learning to EHR data has emerged as a promising means to predict treatment response. This study aimed to used EHR to predict antidepressant treatment difficulty which based on TRADES (Treatment Resistance to Antidepressants Evaluation Scale) for MDD. We sought to evaluated the model performance in different time window and identified important factors. We also explored the effect of different variable combinations on the model performance. Finally, we tried to apply the trained model in new patients who had not evaluated TRADES. 179 patients with International Classification of Disease diagnostic code for MDD in Nation Taiwan University Hospital in October, 2010 to May, 2021 were enrolled. Eligible cases were 20 to 75 years old. Present study based on TRADES which evaluate the level of difficulty in treating MDD with antidepressants. The scoring of TRADES grounded on the lifetime treatment profile which was derived from the complete and detailed medical chart review. The minimum duration that was required for an assessment with the TRADES was 2 years after the first onset of MDD. We defined TRADES score more than 6 as difficulty-to-treat and used EHR to construct prediction models instead of review medical charts. We performed 200 times and recorded variables that each group LASSO model selects in each prediction time windows (6-month, 1-year, 1.5-year, 2-year, full time period). Top frequency variables were selected to Random Forest and Logistic Regression models to predict treatment difficulty. The model performance was evaluated using AUC, accuracy, sensitivity, specificity, positive predictive value, and negative predictive value in an independent testing set. We tried to apply full EHR model of full time period, the trained model with the best model performance, 46 new MDD patients. In secondary analysis, we incorporated genetic data to prediction models in order to improve model performances. AUC achieved 0.77 (0.76~0.78) in full time period and 0.68 (0.66~0.69) in 6-month prediction window. Although model performance increased slowly as the prediction time window extends but there is no obvious progress. Model 3 (EHR code only) usually not the best performance in different time window, but 6-month prediction window AUC is similar to model 4 (full EHR model). In 2-year prediction window best AUC is model 2. Unsurprising, model 4 have the better model performance in different time windows and added genetic data can improve model performance. Bupropion, Tinten, Alprazolam, constipation related diagnosis and medication, multiphasic psychological test (EHR code), total define daily dose of augmentations, the number of augmentations which use more than 4 weeks (A3_2), the number of antidepressants which use more than 2 weeks (A3_1), and the number of psychiatry acute ward visits are the most common important variables. We tried to apply full EHR model and the difficulty-to-treat rate was about 43.3%. This study lack of other factors that may be important to predict treatment difficulty, such as family support, compliance, unstructured EHR (clinic notes, image data, nursing record). Compared with other EHR studies, this study had a small sample size and was limited by the need to be evaluated TRADES, but we later applied the model to a small number of cases that did not evaluate TRADES before and tried to prove whether the predictions were accurate. Future clinical use is limited due to model performance. It is worth investing more effort in the future. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/84475 |
DOI: | 10.6342/NTU202204209 |
全文授權: | 同意授權(限校園內公開) |
電子全文公開日期: | 2027-09-29 |
顯示於系所單位: | 流行病學與預防醫學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
U0001-2809202218045800.pdf 目前未授權公開取用 | 1.31 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。