三幅場景插畫：老人對照屏幕跟讀發聲訓練、面部與聲譜圖形的分析畫面、以及手持設備的測試場景。它們與團隊“聲音作爲研究對象”的方向主題相關；現有證據未將其對應到具體研究項目。
圖背後的故事
帕金森病由多種風險因素引起，臨牀表現多樣，不同因素與疾病特徵之間的關係及其對患病的影響仍不明確。研究採用貝葉斯網絡建立預測模型，使用 35 名患者與 26 名對照的數據訓練；通過樹增強網絡完成結構學習，使用 Netica 軟件進行參數學習。研究分別針對單音節、雙音節、多音節及未分段音節建立四個模型，曲線下面積分別爲 0.95、0.83、0.80 和 0.84。在單音節測試中，時長是最強預測特徵，其後驗概率爲 92.70%；雙音節測試中最低基頻的預測表現最好，多音節與未分段音節中則分別爲末端基頻及最高基頻。橫斷面比較顯示，單音節測試中各變量的預測表現總體優於其他組別，單音節模型的預測表現最高。作者認爲，這種網絡方法可能成爲帕金森病臨牀預測的有用工具。
論文摘要（中文譯文 · 原文）
中文譯文 帕金森病（PD）是一種由多種危險因素引起、臨牀表現多樣的神經退行性疾病。然而，不同因素的作用、與PD相關的不同特徵之間的關係，以及這些因素導致PD發病的程度仍不明確。我們採用貝葉斯網絡構建預測模型，以35名患者與26名對照的數據訓練預測系統；網絡的結構學習與參數學習分別通過樹增強樸素貝葉斯網絡（TAN）與 Netica 軟件完成。就音節層面構建了四個貝葉斯網絡：單音節、雙音節、多音節與未分音節。單音節、雙音節、多音節與未分音節模型的曲線下面積（AUC）分別爲0.95、0.83、0.80與0.84。單音節測試中PD的最佳預測因子是時長（duration），其後驗概率爲92.70%；雙音節測試中最低基頻（minimum f0，61.60%）預測效果最佳，多音節與未分音節測試中預測最佳的變量分別爲末端基頻（end f0，59.40%）與最大基頻（maximum f0，58.40%）。橫向比較顯示，單音節測試中各變量的預測效果普遍高於其他測試組，單音節模型對PD的預測性能最高；在聲學參數中，時長是單音節測試中預測PD患病最強的特徵。我們認爲這一網絡方法將成爲帕金森病臨牀預測的有用工具。
Parkinson's disease (PD) is a neurodegenerative disease with various clinical manifestations caused by multiple risk factors. However, the effect of different factors and relationships between different features related to PD and the extent of those factors leading to the incidence of PD remains unclear. we employed Bayesian network to construct a prediction model. The prediction system was trained on the data of 35 patients and 26 controls. The structure learning and parameter learning of Bayesian Network was completed through the tree-augmented network (TAN) and Netica software, respectively. We employed four Bayesian Networks in terms of the syllable, including monosyllables, disyllables, multisyllables and unsegmented syllables. The area under the curve (AUC) of monosyllabic, disyllabic, multisyllabic, and unsegmented-syllable models were 0.95, 0.83, 0.80 and 0.84, respectively. In the monosyllabic tests, the best predictor of PD was duration, the posterior probability of which was 92.70%. Meanwhile, minimum f0 (61.60%) predicted best in the disyllabic tests and the variables that predicted best in multisyllables and unsegmented syllables were end f0 (59.40%) and maximum f0 (58.40%). In the cross-sectional comparison, the prediction effect of each variable in the monosyllabic tests was generally higher than that of other test groups. The monosyllabic models had the highest predicted performance of PD. Among acoustic parameters, duration was the strongest feature in predicting the prevalence of PD in monosyllabic tests. We believe that this network methodology will be a useful tool for the clinical prediction of Parkinson's disease.