non;Bipolar Disorder Prediction with Transfer Learning on Acoustic and Linguistic Embeddings

NCUIR > School of Management at National Central University > Graduate Institute of Information Management > Electronic Thesis & Dissertation > Item 987654321/95540

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/95540

Title:	non;Bipolar Disorder Prediction with Transfer Learning on Acoustic and Linguistic Embeddings
Authors:	黃永璿;Huang, Yun-shuan
Contributors:	資訊管理學系
Keywords:	躁鬱症;音訊特徵;文字特徵;深度學習特徵;多模態學習;Bipolar Disorder;Acoustic Feature;Linguistic Feature;Learned Feature;Multi-modal Learning
Date:	2024-07-22
Issue Date:	2024-10-09 17:00:01 (UTC+8)
Publisher:	國立中央大學
Abstract:	躁鬱症是一種精神疾病，對個人的生活有深遠影響。準確且早期的診斷至關重要，然而誤診為抑鬱症可能導致錯誤的治療。開發協助臨床醫生進行精確診斷的工具可以減少誤診的機會。機器學習提供了這樣的解決方案。近年來，機器學習在音訊領域的研究日益增多，許多研究探索了使用音訊資料來預測精神疾病的可能性。然而，要收集到足夠的音訊資料來建立分類模型的成本非常高，這也成為使用音訊作為模型學習資料來源的相關研究中的一大限制。遷移學習提供了一個可行的解決方案來應對小型資料集的問題。本研究針對音訊和文本特徵在診斷躁鬱症的有效性進行了全面研究，比較了傳統的特徵工程技術與預訓練模型所提取的學習特徵。本研究的結果顯示學習特徵顯著優於傳統的特徵工程術。此外，本研究還探討了結合音訊和文本資料多模態方法的成效。雖然這些多模態方法未能超越單獨使用音訊特徵的表現，但它們在單獨使用文本特徵方面提供了顯著的提升。這些發現為未來進行相關研究提供了有價值的方法論參考。;Bipolar disorder is a mental disorder that can seriously affect individuals. Early and accurate diagnosis is crucial; however, misdiagnosis of bipolar disorder as depression can lead to incorrect treatment. Therefore, it is important to develop tools that support clinicians in making accurate diagnoses. Machine learning approaches can provide such solutions. Recently, audio has become an important domain for research, with increasing studies exploring the use of audio data to predict mental disorders. However, collecting a sufficient amount of audio data to build a classifier model is costly and impractical, presenting a challenge in utilizing audio data. To address the issue of limited datasets, transfer learning offers a viable solution. In this paper, we conduct a comprehensive study on the effectiveness of audio and textual features, comparing conventional hand-crafted features with learned features. Our results show that learned features outperform conventional hand-crafted features. Additionally, we explore multimodal approaches that combine audio and textual data, finding that while multimodal techniques do not surpass the performance of audio features alone, they do provide improvements over textual features alone. These findings highlight the potential of learned features and multimodal approaches in supporting the accurate diagnosis of bipolar disorder, suggesting a promising direction for future research.
Appears in Collections:	[Graduate Institute of Information Management] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	69	View/Open

社群 sharing

Loading...