藉由大腦白質微結構特性建構腦年齡預測模型及其潛在臨床應用

Abstract

基於醫學影像的腦年齡預測提供了一種嶄新的方式來評估相對於健康族群的個人化腦年齡資訊。腦年齡預測模型估計的”腦年齡”具有做為臨床上神經退化性疾病的生物標記之潛在價值。過去有關腦年齡預測的研究，往往使用腦部較為巨觀的變化，如灰質、白質的體積改變等。然而，一些研究表示，白質神經束的微結構變化，像是神經束完整性，對於老化的進程更為敏銳。因此，本研究旨在開發基於白質神經束微結構的腦年齡預測模型。利用擴散頻譜造影獲得腦部白質神經纖維束的擴散資訊，此擴散資訊可以反映出神經束的微結構特徵。再者，腦年齡預測模型所預測出的年齡若與實際年齡有所差異，可能反映出潛在神經相關性疾病的存在。過去研究報告指出，輕度認知障礙和思覺失調症對於腦部皆會造成加速老化的現象。因此，本研究除了建構腦年齡預測模型外，並將預測模型應用於輕度認知障礙患者與思覺失調症患者上，觀察模型所預測的年齡是否能反映出加速老化的現象，作為腦年齡預測模型的潛在臨床應用開發。 本研究中採用了四組獨立樣本: 192名健康成年人(年齡範圍:31至92歲)，作為訓練模型組、30名健康成年人(年齡範圍:31至80歲)，作為測試模型組、35名輕度認知障礙患者(年齡範圍:61至83歲)與44名思覺失調症患者(年齡範圍:32至62歲)，作為臨床測試組。透過MAP-MRI組件將擴散頻譜造影的影像資料重建為七種擴散指標，包含概化部分不等向性(GFA)、軸向擴散係數(AD)、徑向擴散係數(RD)、平均擴散係數(MD)、非高斯係數(NG)、正交非高斯係數(NGO)以及平行非高斯係數(NGP)。其後，利用全腦白質自動化神經束分析，獲得大腦主要76條白質神經數的七種擴散指標所構成的三維腦聯結圖(3D-connectogram)。為了從三維腦聯結圖提取和年齡相關的特徵，作為建模時候的變數，我們在三維腦聯結圖上的每一個資料點建構與年齡相關的迴歸模型，將與年齡有關的顯著片段提取出來，而後利用主成分分析減少特徵的維度，再將降維後的變數代入高斯過程迴歸模型(GPR model)，以擴散指標為預測因子，將年齡作為反應變數，進行監督式機械學習。因建構模型是在每一種擴散指標上進行，其後我們將每一種擴散指標所得到的預測年齡進行加權平均，得到最後的預測年齡。此外我們建構模型時使用了六重交叉驗證，以檢驗訓練組的代表性與模型估計參數的表現。評量模型預測能力的一致性與準確性我們主要利用皮爾森相關係數(r)與平均絕對誤差(MAE)作為衡量的依據。而在臨床應用的試驗中，通過從預測年齡減去實際年齡來計算的預測年齡差異(PAD)，評估不同臨床疾病組相較於健康對照組是否出現年齡高估的現象。PAD越高，表示年齡高估越多，可以反映加速老化的現象。PAD在臨床疾病組與對照組中我們使用共變數分析(ANCOVA)檢驗族群差異，並控制性別因子。 模型評估的實驗結果中，訓練組(r = 0.86, MAE = 5.6年)與健康測試組(r = 0.92, MAE = 4.3年)的表現皆十分準確。在臨床應用模型試驗中，輕度認知障礙患者(r = 0.69, MAE = 6.8年)與思覺失調症患者(r = 0.61, MAE = 9.4年)在PAD的反映上，與健康對照組(平均PAD = -1.92年)相比，輕度認知障礙患者(平均PAD = 2.96年)與思覺失調症患者(平均PAD = 8.65年)皆顯著性高估於對照組。 臨床疾病組的腦年齡高估，可能反映出加速老化的現象。 我們的研究結果顯示透過GPR建模的方式在訓練組與測試組中，獲得了相當高的準確性。而在臨床疾病組中，平均預測的腦年齡有高估的趨勢，反映出加速老化的情形，和過去文獻的結果一致。總結: 基於白質微結構特徵的腦年齡預測模型在健康族群中能有高準確度的預測，可以個體化進行腦年齡評估。此外，該模型具有助於發掘在個體上加速老化效應的潛在價值。

Imaging-based brain-age prediction provides a promising approach to assess an individual’s brain age relative to healthy populations. The estimated “brain-age” potentially offers clinically relevant biomarkers of neurodegenerative diseases which often manifest accelerated aging process. Studies on brain-age prediction depend on morphological changes of cerebral macrostructure, like volumes in gray or white matter. However, some studies suggest that alterations of white matter microstructure, like tract integrity, are more sensitive to aging effects. Therefore, we aimed to develop a brain-age prediction model based on white matter microstructure, using diffusion spectrum imaging to acquire characteristics of white matter. In addition, the disparity between chronological age and the corresponding predicted brain age might signal the presence of neurodegenerative disease. Studies reported that mild cognitive impairment (MCI) and schizophrenia were both engaged in accelerated aging of the brain. Therefore, to explore the potential clinical applications of the prediction model, we applied the prediction model to MCI and schizophrenia patients. We aimed to investigate whether the overestimated brain age could be observed in these two patient groups, reflecting the effect of accelerated aging. Four independent samples were recruited in the study: 192 healthy controls (age: 31–92 years) as the training set, 30 healthy controls (age: 31–80 years) as the model testing set, 35 MCI patients (age: 61–83 years) and 44 patients with schizophrenia (age: 32–62 years) as the clinical testing sets. MAP-MRI framework was used to reconstruct diffusion spectrum imaging (DSI) datasets into 7 diffusion indices, namely generalized fractional anisotropy (GFA), axial diffusivity (AD), radial diffusivity (RD), mean diffusivity (MD), non-Gaussianity (NG), NG orthogonal (NGO), and NG parallel (NGP). Whole-brain tract-based automatic analysis was implemented to obtain 3D-connectograms of the 7 diffusion indices. To extract age-related features, general linear models were estimated at each step of the connectograms using linear and quadratic age as the independent variables. Continuous steps with significant aging effect were selected as segments. The segments underwent principal component analysis to reduce the dimensions. Gaussian process regression (GPR) models were employed to fit each diffusion index using the age as the response variable and the principal segments as the predictors. An integrative model was defined to integrate the GPR models for each diffusion index into a unified model. Six-fold cross-validation on the training set was conducted to validate the robustness of the model. Model performance was assessed by Pearson’s correlation coefficient and mean absolute error (MAE) between the predicted age and chronological age. In the model test for clinical applications, predicted age difference (PAD) was calculated by subtracting chronological age from predicted age. The higher the PAD, the more overestimation of the brain age is. The PAD scores were used to test group differences among the three study groups using analysis of covariance (ANCOVA), controlling sex. In the model assessment, Pearson’s correlation coefficients and MAE in the training and testing sets were r = 0.86, MAE = 5.6 years, and r = 0.92, MAE = 4.3 years, respectively. In the model test for clinical applications, Pearson’s correlation coefficients and MAE in the MCI and schizophrenia groups were, r = 0.69, MAE = 6.8 years and r = 0.61, MAE = 9.4 year, respectively. Compared to the healthy controls (-1.92 years), the MCI and schizophrenia groups had significantly increased PAD by 2.96 and 8.65 years, respectively Our results showed that the GPR modeling approach achieved equally high accuracy in the training group and the testing group of healthy controls. In the MCI and schizophrenia groups, the average predicted brain age was overestimated with respect to their chronological age. The results are consistent with previous studies that MCI and schizophrenia may accelerate the aging process. In summary, a model of brain age prediction based on white matter microstructural properties was developed with high accuracy in the healthy population, allowing brain age assessment on individual basis. Moreover, this model might be helpful in detecting individuals with accelerated aging effects.