請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74365
標題: | 預測早期思覺失調症患者之治療效果:利用大腦神經纖維束及皮質體積之機器學習模型 Machine learning prediction of treatment response in early schizophrenia based on cortical volume and white matter tract integrity |
作者: | Wei-Chia Lu 呂維家 |
指導教授: | 曾文毅(Wen-Yih Tseng) |
關鍵字: | 思覺失調症,治療成效,擴散頻譜造影,大腦白質神經纖維束,全腦白質神經纖維自動化分析,機器學習, Schizophrenia,treatment outcomes,diffusion spectrum image,tract-based automatic analysis,white matter tract,machine learning, |
出版年 : | 2019 |
學位: | 碩士 |
摘要: | 介紹: 對於思覺失調症患者來說,達到症狀的緩解是非常重要的,然而每位患者對於抗精神藥物的治療反應不盡相同,有些患者經過藥物的治療後卻無法達到症狀緩解的情形。而根據之前的文獻指出不同的治療效果會反映在白質及灰質微結構上(Ashburner & Friston, 2005; Bora et al., 2011; Reis Marques et al., 2014),在這次的研究中,我們找到在早期思覺失調症中對於用藥緩解和非緩解兩組間有顯著差異的幾個部位,包含利用規範性模型計算出的z值來得到的三條神經纖維束以及利用LPBA40的模板計算出的皮質體積,我們將得到的三條神經纖維束以及皮質體積並藉由機器學習的模型來達到個別化預測早期思覺失調症藥物治療後的症狀反應情形。
材料及方法: 本次實驗納入65位的思覺失調症患者(病程皆小於三年內,29位男性,36位女性,年齡:26.9±6.2年),和50位健康的受試者,後分為52位於訓練組(年齡:26.4±6.1年,26位男性,26位女性)以及13位於測試組(年齡:28.7±6.6年,男性3位,女性10位)。所有的受試者皆在台灣大學醫院的3T西門子Tim Trio磁振造影系統以32通道線圈來掃描獲得T1加權影像以及擴散頻譜影像。影像品質確保:僅有影像的訊號雜訊比大於20以及T1-綜合非等項性指標之相關係數大於0.5的個案會被納入此研究分析:我們利用散頻譜影像與全腦白質纖維神經束分析方式(Chen et al., 2015)來獲得大腦主要76條白質神經束的完整度,並利用779位其年齡涵蓋5至80歲的正常族群資料,依據性別分組,並計算出患者的標準偏差值,經由統計分析後得到三條神經束在緩解與非緩解組中有相當程度的差異,此三條神經纖維為右側弓狀束、右側下枕額束及右側鉤束。此外由CAT12中的LPBA40模板計算出皮質體積的資訊,並得到在緩解與非緩解組中尾狀核有顯著性的差異。患者被分類為緩解及非緩解是藉由活性與負性症狀量表、思覺失調症緩解工作小組(Nancy C. Andreasen et al., 2005)以及台灣大學醫院的精神科醫師來診斷分類,最後藉由MATLAB中的classification learner來得到最有潛力的機器學習模型以預測治療的反應。 結果: 65位思覺失調症患者皆完成至少6個月的追蹤評定,根據思覺失調症緩解工作小組以及台灣大學醫院的精神科醫師診斷,52位訓練組中有37位為症狀緩解,15位為非緩解; 13位測試組中有9位為症狀緩解,4位為非緩解。而在統計分析後得到在緩解與非緩解中有顯著性差異的三條神經束及腦區,利用機器學習的方式來得到最有潛力分辨緩解與非緩解的模型,在訓練組中的正確分辨率為: Decision Tree: 0.923, RUSBoosted Trees: 0.827, Fine KNN: 0.808,套用到訓練組中得到個案的預測準確率為Decision Tree: 0.846, RUSBoosted Trees: 0.615, Fine KNN: 0.615. 結論: 我們的目標是找到對於早期思覺失調症治療反應的生物標記並且將組間比較邁向個體化預測,而在這次實驗中我們得到藉由決策樹訓練出來的機器學習模型在於個案的預測準確率有84.6%,是非常有潛力來達到個體化預測的模型。對於思覺失調症而言,一個長期的研究計畫是非常重要且有價值的,我們可以在此實驗中得到神經束以及皮質體積有其貢獻在預測治療反應。並冀望結合更多臨床的評估及其他神經影像資訊,此模型將可作為醫師擬定個案治療計畫之參考。 Introduction: It is important to achieve symptomatic remission for the treatment of schizophrenia. However, some patients don't achieve remission state even after antipsychotics treatment. Different treatment outcomes may be reflected in the white matter microstructure (Ashburner & Friston, 2005; Bora et al., 2011; Reis Marques et al., 2014). In this study, we identified three white mater tracts calculate by adjusted Z-scores of diffusion indices and cortical volume by LPBA40 atlas that showed significant difference between remitted and non-remitted patients in early schizophrenia. We combined those three bundles and cortical volume with machine learning to promote prediction accuracy for treatment outcomes in early schizophrenia. Materials and Methods: sixty-five patients with schizophrenia (course of disease < 3 years, age: 26.864 ± 6.224, 29 males and 36 females) and 50 healthy control were recruited. Separate into training group (age: 26.404 ± 6.026, 26 males and 26 females) and testing group (age: 28.705 ± 6.65, 3 males and 10 females). Imaging: All participants received T1-weighted imaging and DSI on a 3T Siemens Tim Trio MRI System with 32-channel phased array head coil in National Taiwan University Hospital. Image Quality Assurance: Only images with signal-to-noise ratio higher than 20 and T1-GFA correlation coefficient higher than 0.5 were included for subsequent analysis. Analysis: We used TBAA to obtain profiles of diffusion indices (Chen et al., 2015), and calculated GFA, AD, RD, MD adjusted Z-scores made by 779 healthy subjects, corresponding to three tracts of interest. These three tracts included the right arcuate fasciculus, the right inferior frontal occipital fasciculus, and the right uncinate fasciculus. Furthermore, calculate cortical volume using LPBA40 atlas by CAT12. Right caudate have significant difference. Patients were grouped into good or poor treatment outcomes according to the PANSS scores, RSWG criteria (Nancy C. Andreasen et al., 2005) and clinical diagnosis by National Taiwan University Hospital. We used the classification learner from MATLAB to determine the most effective model, which had great potential for predicting treatment outcomes. Result: 65 patients had completed the 6-month follow-up assessment. According to the RSWG criteria and clinical diagnosis, 37 patients were assigned to the remission group and 15 patients were assigned to the non-remission group inside the training group and 9 and 4 patients were assigned to remission and non-remission group inside the testing group. According to the statistics, three white matter tracts and the cortical volume of right caudate show significant difference between remitted and non-remitted patients in early schizophrenia. Based on these three tract bundles and cortical volume, we used machine learning to calculate those feature together to determine the most potential model to predict treatment outcomes. The accuracy of the prediction model in training group: Decision Tree: 0.923, RUSBoosted Trees: 0.827, Fine KNN: 0.808. With the testing data, the accuracy of each model as follow: Decision Tree: 0.846, RUSBoosted Trees: 0.615, Fine KNN: 0.615. Conclusion: Our goal is to find biomarkers of treatment response in early schizophrenia. We approach the goal from group comparison to individualized prediction. Model training by decision tree is potentially useful in treatment response prediction for each individual patient. A longitudinal study on early state schizophrenia is required to validate the clinical value of these tracts and cortical volume in predicting treatment response. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/74365 |
DOI: | 10.6342/NTU201902802 |
全文授權: | 有償授權 |
顯示於系所單位: | 醫療器材與醫學影像研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-108-1.pdf 目前未授權公開取用 | 2.5 MB | Adobe PDF |
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。