利用深度學習與數據增強技術提升脈搏分類準確度

Abstract

把脈是中醫師判斷患者重要的依據，藉由手部按壓感受脈搏，可以診斷患者的生理狀況。但隨著科技的進步，開始有許多研究利用電腦資訊技術來輔助脈搏判讀。過去已經有研究將量測的脈搏數據進行健康者與不健康者的分類，也有探討特定脈搏與特定疾病的關聯性。這些研究大部分探討脈搏與病理的對應關係，並且搭配大量不同演算法做脈搏特徵的擷取。然而這種擷取特徵的方式會有許多應用上的限制，因為必須先了解脈搏哪些特徵是具有這類病理的資訊，才接續探討演算法的選用，最後提取這類特定的特徵進行分類。 本研究將利用深度學習(deep learning)建構脈搏分類器，進行脈搏多分類的研究。藉由深度學習自主擷取特徵的優勢，改善過往耗費大量時間尋找、選用特定演算法來擷取脈搏特徵。本研究建立七個深度學習的模型並加以比較，在實驗中發現卷積神經模型(convolutional neural network, CNN)搭配卷積的長短期記憶模型(convolutional long short-term memory, CNN-LSTM)分類效果最好，準確率可以達到94 %。最後利用數據增強的技術，增加訓練模型的脈搏數量，產生更大的數據集。最終藉由數據增強之技術提供分類效果最好的模型更多穩健的數據，並強化模型分類。

Pulse taking is an important skill for Chinese medicine practitioners to diagnose a patient. By feeling the pulse fluctuation, the patient's symptoms can be diagnosed. Owing to advances in technology, using computers to assist pulse interpretation has gained popularity. In the past, there had been studies to classify the measured pulse data between healthy and unhealthy people and explored the correlation between pulse patterns and diseases. Most of these studies discussed the correspondence between pulse and pathology and used many different algorithms to extract pulse features. However, extracting pulse features have many application limitations, because you must first understand which features containing with information needs to be understood first. Then algorithms for extracting specific features for classification can be selected. In this study, deep learning will be applied to construct a pulse classifier to conduct pulse multi-classification research. Deep learning extracting features that are not designed by human engineers, time to find and select specific algorithms to extract pulse features could be avoid. First, seven deep learning models were established and compared. In the experiment, CNN (convolution neural network) with CNN-LSTM (convolutional long short-term memory) classification had the best accuracy of 94%. Then, data augmentation technology was applied to increase pulse datasets. In the study, data augmentation technology provided more robust data for the model with the best classification of 95%.