基於 FDG-PET 和線性與非線性降維方法的機器學習之頑固型憂鬱症分類模型

Abstract

重度憂鬱症（MDD）對全球造成了嚴重的公共衛生負擔，其中難治型憂鬱症 （TRD）患者在經過多次治療後仍未見改善，其症狀持續且惡化，對患者生活品質 產生多方面負面影響。本研究旨在利用正子斷層掃描（FDG-PET）腦部影像與機 器學習方法，建立一個穩定且高準確度的 TRD 預測模型，同時分析關鍵腦區的功 能異常。 本研究從 315 名 MDD 患者中， 根據 Maudsley Staging Method-Treatment （MSM-T） 評分， 將其分為 TRD 組 （208 名） 與非 TRD 組 （107 名）， 並納入 104 名健康對照組（HC）進行 Z-map 分析。我們透過 Z-map 分析，將每位患者的腦部代謝活動與 HC 樣本進行標準化比較，藉此量化特定腦區的異常程度。為了應對神經影像資料常見的「維度災難」與模型過度擬合問題，我們採用了主成分分析（PCA）、Isomap 和局部線性嵌入（LLE）等降維技術。此外，本研究也比 較了包含小腦的 AAL-116 腦圖譜與 AAL-90 腦圖譜的預測效能，以探討不同腦區 劃分對模型性能的影響。 結果顯示，本研究建立的機器學習模型在 TRD 預測任務上達到 81% 的準確率。值得注意的是，儘管 AAL-116 腦區特徵包含了小腦，其表現並未優於 AAL-90，這點為憂鬱症的腦區功能研究提供了新的見解。本研究提出的方法能有 效提升模型的泛化能力與臨床實用價值，為 TRD 的精準診斷與治療提供了更穩固的科學基礎，同時也為神經影像學與機器學習的跨領域研究開闢了新的方向。

Major Depressive Disorder (MDD) poses a significant global public health burden. Among these patients, those with treatment-resistant depression (TRD) experience per- sistent and worsening symptoms despite multiple treatment attempts, leading to various negative impacts on their quality of life. This study aims to use FDG-PET brain imaging and machine learning methods to build a stable and highly accurate TRD prediction model while also analyzing functional abnormalities in key brain regions. Based on the Maudsley Staging Method-Treatment (MSM-T) score, we divided 315 MDD patients into a TRD group (208 individuals) and a non-TRD group (107 individu- als), and included 104 healthy controls (HC) for Z-map analysis. Through Z-map analysis, we standardized and compared each patient’s brain metabolic activity with HC samples to quantify the degree of abnormality in specific brain regions. To address the ”curse of dimensionality” and overfitting issues common in neuroimaging data, we employed di- mensionality reduction techniques such as Principal Component Analysis (PCA), Isomap, and Locally Linear Embedding (LLE). Additionally, this study compared the predictive performance of the AAL-116 brain atlas (which includes the cerebellum) with the AAL- 90 brain atlas to investigate the impact of different brain region parcellations on model performance. The results showed that the machine learning model developed in this study achieved an 81% accuracy rate for the TRD prediction task. Notably, although the AAL-116 brain atlas included cerebellar features, its performance was not superior to that of the AAL-90 atlas, which provides new insights for brain functional research in depression. The method proposed in this study effectively enhances the model’s generalization ability and clinical utility, offering a more solid scientific foundation for the precise diagnosis and treatment of TRD. It also opens up new directions for interdisciplinary research in neuroimaging and machine learning.