| 摘要: | 情感性疾患常伴隨嚴重認知障礙，對患者的康復與功能性結果產生顯著影響。儘管神經退化性疾病的認知障礙預測研究已相當豐富，情感性疾患領域卻仍缺乏整合人口學變項與多模態特徵的預測模型。本研究提出多模態分析框架，結合人口學資料、語音特徵及文本特徵來預測威斯康辛卡片分類測驗（Wisconsin Card Sorting Test, WCST）指標，並透過對比手工特徵與學習型特徵，探究哪些特徵類型最適合預測特定認知領域。研究分析了來自台灣北部某醫院的133名情感性疾患病人與63名健康對照組的人口學資料、主述語音檔案及WCST量表分數。統計結果顯示，精神科住院史與所有WCST指標皆有顯著相關，而教育程度、年齡、婚姻狀態、酒精使用史、高血壓及高血脂則與特定認知領域相關。模型預測結果發現，不同模態與特徵類型各有優勢：結合人口學變項後，語音為基礎的單模態模型在總錯誤數預測上表現最佳；文本為基礎的單模態模型則在預測固著性反應與固著性錯誤方面表現優異；而多模態整合模型在預測完成類別數量上成效最佳。研究成果為臨床應用提供了新方向，未來可望透過軟體整合，協助臨床醫師進行快速認知評估，並發展便於一般民眾使用的自我認知監測工具。;Mood disorders frequently accompany severe cognitive impairment (CI) that impacts recovery and functional outcomes. Despite extensive research on CI prediction in neurodegenerative disorders, a notable methodological gap exists for mood disorders, with current literature lacking robust predictive models that integrate demographic variables with multimodal features. This study proposed a multimodal framework integrating demographic, audio, and text features to predict Wisconsin Card Sorting Test (WCST) measures, while conducting comprehensive examination of both handcrafted and learned features to identify which feature types best predict specific cognitive domains. We analyzed a dataset comprising demographic data, speech recordings, and WCST scale scores from 133 mood disorder patients and 63 healthy controls recruited from a hospital in northern Taiwan. Statistical analysis revealed psychiatric hospitalization as significantly associated with all WCST metrics, while education, age, marital status, alcohol consumption history, hypertension, and hyperlipidemia correlated with specific cognitive domains. Predictive modeling revealed modality and feature-specific advantages: when combined with demographic variables, audio-based unimodal models excelled for total errors, text-based unimodal models excelled for perseverative responses and errors, while multimodal models optimally predicted categories completed. These findings offer promise for future clinical applications through software integration, potentially enabling rapid cognitive assessment by clinicians and accessible self-screening tools for monitoring cognitive changes in mood disorders. |
