基於神經發展治療及深度學習之自動化步態訓練機研發

Abstract

本論文針對中風患者的復健，整合先前開發的自動化神經發展治療(Neuro Developmental Treatment;NDT)步態訓練系統，進一步開發一套可以有效提升受測者骨盆側向對稱性的訓練方法，並以實驗驗證其效果。另外，我們嘗試應用深度學習技術於中風步態分類上，以協助步態的篩檢，並有效地提升步態評估之效率。 NDT步態訓練整合各種支撐、誘發、引導、與抑制等手法，對於中風病患是相當有效的復健方式，然而對於治療師而言，實施這種治療方式相當耗時且費力；對於病患而言，則因訓練本身的複雜度及練習時間受限，因此臨床上實際NDT訓練的質與量皆相當不夠。故本研究在過去開發一套自動化NDT步態訓練機，降低治療師的負擔並同時增加中風患者的訓練時間，以提升NDT訓練的效果。 本論文首先取得治療師與中風受測者進行訓練時各種運動行為的量化資訊，將臨床know-how轉譯建立治療師專家系統，並根據研究結果發現，自動化NDT步態訓練機訓練對增進縱向擺動期對稱性的效果顯著，但是在骨盆側向對稱性訓練的效果則不如治療師訓練，因此本論文進一步探討，治療師的訓練方法與骨盆側向對稱性進步之間的關係，並提出一套增進骨盆側向偏移對稱性之控制流程，整合至自動化NDT步態訓練系統中。為了能讓系統提供良好的系統響應，本論文使用系統識別的方式來建立數學模型，並針對此系統模型設計控制器。最後本論文招募受測者，以擬中風的方式進行NDT步態訓練，以驗證所提出的骨盆側向對稱性訓練方法對於受測者的訓練效果。 隨著深度學習演算法的成熟與電腦硬體的快速發展，深度學習可以處理許多複雜之非線性問題，故本論文應用深度學習於中風步態分類及篩檢，以提高步態評估之效率。本論文首先收集治療師評估中風異常步態的臨床方法，再分析治療師於步態評估過程中的判斷方法與病患之動作資料，以建立專家資料庫。並利用此資料庫，開發出一套中風步態分類模型，來進行中風步態的偵測與分類，結果顯示模型的平均正確率達到94%。未來，可望結合深度學習步態分析與自動化NDT訓練機，以發展客製化的步態訓練。

This thesis develops a training method for the rehabilitation of stroke patients that can effectively improve the symmetry of lateral pelvic displacement (LPD). And this method is integrated into the automated neural development treatment(NDT) gait trainer developed by the previous research. At the end, we apply deep learning techniques to the classification of stroke gaits, which can help screening gaits and improving the efficiency of gait assessment. NDT is an effective rehabilitation method for stroke patients, because it can let patients have the feeling on walking with minimal intervention. However, it is very labor-intensive and time-consuming for therapists. As a result, patients usually receive insufficient training due to the shortage of therapist assistance. In the previous research, an automated NDT gait trainer was developed to reduce therapists’ workloads and to increase the training time of stroke patients, thus improving the effectiveness of NDT training. Based on the experimental results, the designed gait trainer was effective in improving the symmetry of swing phase. However, the symmetry of LPD was not improved much compared to therapist guiding. The first contribution of this thesis is to develop a training method improve the symmetry of LPD. First, we record the therapist’s actions and the test subject’s motions during the training. Then we use these data to translate the clinical know-how to build an expert dataset and propose an effective method to improve the symmetry of lateral pelvic displacement. Second, in order to improve the system response, we use the identification techniques to derive the transfer function and design robust controllers to cope with disturbances and noises during gait training. Finally, we recruit the subjects to perform NDT gait training to verify the effectiveness of the proposed method. The second contribution of this thesis is to apply deep learning techniques to the classification of stroke gaits. With the maturity of deep learning algorithms and the rapid development of computer hardware, deep learning can solve many complex nonlinear problems. This thesis applies deep learning to assist gait screening and to improve the efficiency of gait classification. First, we collect the clinical methods used by the therapist to assess the abnormal gaits of stroke patients, then analyze the gait data to establish an expert dataset. Using this dataset, we develop a gait classifier to detect and classify stroke gaits. Based on the results, the average accuracy of the proposed model was greater than 94%. In the future, the gait classifier is expected to integrated within the NDT trainer to develop customized rehabilitation method.