基於 FDG-PET 與機器學習量化重度憂鬱症患者之臨床嚴重性及關鍵腦區特徵剖析

Abstract

重度憂鬱症是嚴重且易反覆發作的精神疾病，對患者的生活品質會造成重大影響，若不及早正確治療將可能發展成慢性病。憂鬱症的治療是新興關注議題，世界衛生組織 (WHO) 更預估於 2030 年，憂鬱症將會成為影響全人類最嚴重的疾病之一。而部分重度憂鬱症患者使用多種常規抗憂鬱症藥物的治療效果不佳，需要採用如 rTMS(重覆經顱磁刺激) 等方式或個別化進一步治療，為頑固型憂鬱症(Treatment-resistant Depression)。近年來的大量研究更指出，除了頑固性、嚴重性不一外，重度憂鬱症在不同腦區代謝之異常亦有其異質性。因此，如何在臨床評估時提供更為精準的頑固性、嚴重性預測及腦區特異對於患者的治療進程至關重要。在此研究中，將共 219 位重度憂鬱症之患者 FDG-PET 腦部正子造影作為資料集，分別針對:(1) 頑固性及 (2) 嚴重性運用七個機器學習模型，並透過 optuna調整超參數並進行集成學習投票，預測、量化出患者結果；其中在頑固性模型部分，我們更進一步分析整體訓練集之腦區特徵 SHAP Value，運用在測試集腦區顯著特徵預測。在獨立測試集上我們的模型達到之預測準確度結果分別為: 頑固性83.3%，嚴重性 87.0%，並分析出對於頑固性模型來說重要的腦區特徵，包括前額葉皮質、梭狀回、赫氏迴、顳極等，並和過往相關研究比較進行討論。

Major depressive disorder (MDD) is a severe and recurrent mental illness that significantly impacts patients’ quality of life. If not treated correctly and promptly, it can develop into a chronic condition. The treatment of depression has gained significant attention, with the World Health Organization (WHO) projecting that by 2030, it will rank among the most serious health conditions impacting humanity. Some patients with major depressive disorder do not respond well to standard antidepressant medications and require personalized treatments or advanced methods such as repetitive transcranial magnetic stimulation (rTMS). These cases are termed treatment-resistant depression (TRD). Recent studies have highlighted the heterogeneity in metabolic abnormalities across different brain regions in MDD, including variability in resistance and severity. Thus, providing precise predictions for resistance, severity, and specific brain regions is crucial for improving treatment outcomes. This study involved FDG-PET brain imaging of 219 MDD patients, employing seven machine-learning models to predict and quantify (1) resistance and (2) severity. Hyperparameters were tuned using Optuna, and ensemble learning voting was used for predictions. For the resistance model, SHAP values of brain region features were analyzed on the training set and applied to significant features in the test set. The achieved prediction accuracies on the independent test set were 83.3% for resistance and 87.0% for severity. Important brain region features for the resistance model included the prefrontal cortex, fusiform gyrus, Heschl’s gyrus, temporal pole, etc., with comparisons made to previous related studies.