| 摘要: | 神經科學在人工智慧的發展上扮演相當關鍵角色，也一直是構建人工智慧技術的主要來源。一般而言，有兩種形成路徑：第一，模擬相當人類的智力，第二，建立模擬大腦結構的神經網絡。以往在醫學影像辨識及分類的效果仍差強人意，所以在一些難以判讀的異常人體細胞時，往往讓醫師因無法確實掌握正確病因的關鍵確診證據而產生遺憾。故若能正確將人工智慧在檢驗醫學領域的應用加以推廣，讓操作軟體具有訓練及學習能力的演算功能，對於圖像資料庫的正確判讀是有幫助的。
因為人工智慧數位化神經理學檢查判讀在以往相關研究較為缺乏，有鑑於此議題的重要性，而手指碰鼻 (Finger to nose) 測試為在臨床評估上肢功能常見的測試方式之一，故本研究擬此測試為研究對象，以探討人工智慧數位化神經理學檢查判讀效果，例如左右手食指移動之預測力，同時探討機器學習之檢測正確率與Loss率。
在這項研究中，本研究採用準實驗研究設計 (Quasi-experimental Research Design) 方法裡的病例對照研究 (Case‐control study) 進行。研究實際總收案數為44人(18人為手部抖動，26人為手部不抖動)，共拍攝88部影片，而影片則以左右手區分是否抖動，共有39部為抖動，49部為不抖動。實驗過程則採用錄影資料，後續將建立類神經網路、羅吉斯迴歸模型，以人工智慧建立機器學習模型產生分數辨識正常或異常。
主要分析成果，則說明如後：1.差異分析部分，無論是以Person或是以Hand為分析單位，年齡部分均呈現顯著差異，性別部分均呈現不顯著差異。2.羅吉斯迴歸分析結果，(1)檢測左手食指移動的預測能力約為88.6%；檢測右手食指移動的預測能力約為81.8% - 84.1%，檢測左手有明顯相關的特徵為：食指的垂直(Y軸)移動平均值，呈負向影響。(2)檢測食指移動的預測能力約為79.5% - 84.1%，檢測有明顯相關的特徵為食指的垂直(Y軸)移動平均值，呈負向影響。3.機器學習分析結果，在30輪次的迭代學習後其正確率已相當接近100%，而經過50輪次的迭代學習則達到100%，在Loss率部分，20輪次的迭代學習後仍有3次的Loss，經過30次亦為零Loss。
從以上結果可知，本研究所提之人工智慧數位化神經理學檢查判讀機制，在評估手指碰鼻的上肢功能測試中能達到不錯的效果，未來發展上可嘗試與其它各類手部顫抖徵狀，如意向性手抖、運動性手抖、姿勢性手抖等，並配合相關檢驗機制，以提高診斷之正確率。
Neuroscience plays a key role in the development of artificial intelligence, and has always been the main source of artificial intelligence technology. Generally speaking, there are two ways of formation: first, to simulate human intelligence; second, to establish a neural network to simulate brain structure. In the past, the effect of medical image recognition and classification is still unsatisfactory, so when some abnormal human cells are difficult to read, doctors often regret that they can not grasp the key evidence of correct etiology. Therefore, if the application of artificial intelligence in the field of laboratory medicine can be promoted correctly, and the operation software can have the calculation function of training and learning ability, it will be helpful for the correct interpretation of image database.
Due to the lack of previous studies on the interpretation of artificial intelligence digital neurology examination, in view of the importance of this topic, and finger to nose test is one of the common test methods in clinical evaluation of upper limb function, this study intends to take this test as the research object to explore the effectiveness of artificial intelligence digital neurology examination interpretation, such as the predictive power of the index finger movement of the left and right hand, and discuss the detection accuracy and loss rate of machine learning at the same time.
In this study, the study was conducted using case-control studies in the Quasi-experimental Research Design methodology. The actual total number of cases received in the study was 44 (18 people were shaking hands and 26 people were not shaking hands). A total of 88 films were shot, and whether the films were shaking or not was distinguished by left and right hands. A total of 39 films were shaking and 49 films were not shaking. The experimental process will use video materials, followed by the establishment of neural networks, Rogis regression model, artificial intelligence to establish machine learning models to produce scores to identify normal or abnormal. The main analysis results are described as follows: 1. In the difference analysis part, whether taking person or hand as the analysis unit, the age part shows significant differences, and the gender part shows no significant differences. 2. 2. Logistic regression analysis results: (1) the predictive ability of detecting left index finger movement is about 88.6%; The predictive ability of detecting the movement of the right index finger is about 81.8% - 84.1%. The obvious related characteristics of detecting the left hand are: the average value of the vertical (Y-axis) movement of the index finger is negatively affected (2) The predictive ability of detecting the movement of the index finger is about 79.5% -84.1%. The detection is obviously related to the average value of the vertical (Y-axis) movement of the index finger, which has a negative impact. 3. According to the analysis results of machine learning, after 30 rounds of iterative learning, the accuracy rate is quite close to 100%, while after 50 rounds of iterative learning, it reaches 100%. In the loss rate part, there are still 3 times of loss after 20 rounds of iterative learning, and after 30 rounds, it is also zero loss.
From the above results, it can be seen that the artificial intelligence digital neurological examination and interpretation mechanism proposed in this study can achieve good results in the evaluation of the upper limb function test of finger touching the nose. It can be tried with other various hand tremor symptoms in future development. , Such as intentional hand tremor, sports hand tremor, postural hand tremor, etc., and cooperate with related inspection mechanisms to improve the accuracy of diagnosis. |
