- 期刊
使用統計分析篩選有效語音資料與特徵值以發展HIV感染者之情緒辨識模型
Development of an Emotional Recognition Model for People Living with HIV through Statistical Analysis to Filter Valid Speech Data and Voice Features
摘要
HIV感染者比一般人更容易出現憂鬱或情緒低落之狀況，也容易發生輕生等事件。人們的情緒起伏會反應在說話時特定的聲音特徵上，因此若研究情緒與聲音之間的關係，便能建構人工智慧模型來辨別處於情緒低落的狀態，而個案管理師就可給予有效的關懷，避免不幸發生。為此目的，吳允彤（2022）使用於衛生福利部桃園醫院感染科合作期間蒐集之資料，發展HIV與物質成癮患者情緒辨識模型，研究中發現1.部分受測者回報的情緒程度不夠精準以及2.大部份受測者回報之情緒過度集中於特定程度，出現分布不均的現象。上述兩個資料相關限制導致在使用統計分析篩選重要特徵值時效果不佳，進而影響後續建模成效。為了克服上述兩點資料限制，第一階段研究著重於個別HIV受測者之聲音分析，針對個別受測者之資料個別進行單因子變異數分析（ANOVA）與相關性分析（Correlation），篩選在所有受測者中顯著性頻率較高的特徵值，並用於決策樹與隨機森林兩種模型建構。然而，此方法並沒有顯著提升模型對於心情值的辨識成效，因此在第二階段研究中改以複迴歸分析建構個別受測者的心情預測模型。吳允彤（2022）所發展的模型若要能辨別80%情緒低落的資料，則會產生將近75%的誤警，而本研究第二階段所使用建模的模型，若一樣是要成功辨別80%情緒低落的資料，則能夠降低誤警機率至40%。這代表原本個管師及護理師需撥打大量的電話才能確保有關心到情緒低落的患者，若改用本研究複迴歸分析模型，則能撥打較少的電話即可關懷到情緒低落的患者，不但減少個案管理師的負擔，也能夠降低院方的人力成本。
並列摘要
Patients living with Human Immunodeficiency Virus (HIV) are more often depressed or have risk of suicide than the general population. To prevent unfortunate events in HIV patients, Wu (2022) attempted to study the relationship between mood and voice properties to build a model for determining patient's mood. However, due to the limitations of collected data in Taoyuan Hospital, Ministry of Health and Welfare, the general model could not discriminate bad mood of the HIV patients effectively. While Wu (2022) used the overall data from thirty-six patients to build a general model, this study aimed to apply another approach to increase the model's accuracy. This study used four patients’ data that had relatively more data and were evenly distributed in all mood levels. Using ANOVA and correlation two kinds of statistical analysis to find out important voice features. Instead of building a general model, individual models were built for the four patients using multiple regression analysis modeling to reduce individual differences. The new approach makes models obtain an accuracy of 80% in detecting bad moods with a false alarm rate of 40%. Compared to the general model built by Wu (2022) that resulted in a false alarm rate of 75% when keeping the same accuracy of detecting bad moods, this study show the new approach performs better and can save nurses and clinic case managers' workload and the hospital's costs.