運用倒傳遞類神經網路與鑑別分析分類精神科日間病房病患日常生活功能

Using Back-propagation Artifi cial Neural Network (BPN) and Discriminant Analysis (DA) to Classify the Functioning Level of Psychiatric Daycare Ward Patients’ Activities of Daily Living

研究目的：臨床上，精神病患的日常生活功能測驗較費時，且許多病患經多次施測已熟悉測驗，可能使得日常生活功能測驗的結果失準。本研究期透過傳統統計及人工智慧方法，建構電腦輔助系統，以預測精神病患的日常生活功能。研究方法：以台北市某醫學中心日間病房54筆資料，利用「鑑別分析(DiscriminantAnalysis)」與「倒傳遞類神經網路(Back-PropagationNeuralNetwork)」，依病患「性別、學歷、診斷、年齡」等變數，與「褚氏日常生活評量表」的等級，建構其「日常生活功能」之預測模式，並比較兩者預測之準確性。研究結果：本研究透過「鑑別分析」發現男性、診斷非精神分裂症、年齡較輕者，較有可能有較高的日常生活功能。而「倒傳遞類神經網路模式」之平均整體鑑別率遠高於「鑑別分析」，且準確率高達92.55%。研究結論：本研究發現「倒傳遞類神經網路」比傳統的「鑑別分析」更能有效地預測精神病患的日常生活功能，未來可多將「倒傳遞類神經網路」應用於臨床研究中。建議未來研究可多增加研究樣本、研究變數，以更精準預測病患日常生活功能。

Objectives: Evaluating daily functioning of people with mental illnesses with questionaires is time-consuming. Additionally, patients’ familiarity with the context of function tests compromises test accuracy. Applying a computer-assisted support assessment system in this study, we intended to predict the daily functioning of the mentally ill objectively and conveniently. Methods: We collected 54 patients attending a psychiatric daycare ward at a medical center in the Taipei city. A fi ve-fold cross-validation scheme was applied to minimize possible bias and to provide reliable estimates. We used discriminant analysis (DA) and a back-propagation neural network (BPN), to predict patients’ daily functioning, according to gender, educational background, diagnosis, and age, on Chu’s daily function scale. Results: Both models achieved high average overall accuracy of more than 70%. The BPN model had a high overall classifi cation accuracy of 92.55%, 16.55% better than that of the DA model. Additionally, the discriminant function showed that young males not diagnosed with schizophrenia had better daily function. Conclusion: This study was found that the BPN as a computer-assisted assessment support system predicted daily functioning more effectively than DA. To predict daily functioning relatively more precisely, we suggest that future research need to expand the sample population and to use additional variables, such as patients’ personality, family support, and living status.