Individualized reaching training framework using model-based evaluation for robot-assisted stroke rehabilitation

국문 초록 (Abstract)

상지 재활 로봇은 뇌졸중 환자에게 반복적인 도달 운동 훈련을 제공할 수 있으나 초기 로봇 재활 시스템은 단순한 도달 환경과 획일화 된 스케줄만을 제공하기 때문에 환자 개인의 특성을 반영하지 못하였다. 본 학위논문은 모델 기반 평가법을 활용하여 환자별 맞춤형 도달 훈련 스케줄을 자동 생성하는 개인화 도달 운동 훈련 프레임워크를 제안한다. 먼저, 시간적·공간적 편향 없이 작업공간 전반의 도달성능을 평가할 수 있는 정량적 평가 기법을 개발하였다. 해당 기법은 1. Fitts 법칙 (속도–정확도), 2. Almanji 모델 (뇌성마비 대상 마우스 포인팅 동작), 3. 본 연구의 제안 모델 등 세 가지 후보 모델을 비교 하고자. 모델 검증을 위해 건강인 12 명, 뇌졸중 환자 7 명으로부터 로봇 기반 운동학 데이터를 수집·분석하였다. 건측의 데이터를 이용해 정상 도달 모델을 추정하고, 이를 기준으로 환측 팔을 평가하도록 설계하였다. 제안 모델은 건강인 12 명 전원 및 뇌졸중 환자 19 명의 덜 손상된 팔 중 16 명에서 R2 > 0.7 이상의 설명력을 보였으나, 환측 팔의 비정상 궤적은 올바르게 식별하지 못해 평가 지표로서의 정상 움직임에 대한 특이성 (specificity) 을 입증하였다. 시각화된 평가 결과는 환측 팔의 고유 운동 특성을 직관적으로 제시하여 맞춤형 훈련 과제 선정에 활용 가능함을 확인하였다. 이후 운동 학습(motor-learning) 원리를 반영한 스케줄링 알고리즘을 제안하고. 알고리즘은 적응성 (adaptiveness)과 무작위성 (randomness) 을 동시에 반영해 회복 촉진과 기술 유지 (skill retention) 를 도모한다. 실제 환자 데이터를 기반으로 한 컴퓨터 시뮬레이션 결과, 제안 스케줄은 기존 방법 대비 훈련 적응도 및 변동성에서 유의미한 우수성을 보였다 (p < 0.001). 또한 기존 정상 도달 운동 모델 고도화를 위해 건강인 및 뇌졸중 환자의 광범위한 기구학적 변수를 분석하였다. 교차검증과 AIC 등을 활용해 비선형 회귀모델을 탐색하여, 정상 도달 시간 예측에 가장 영향력 있는 변수 집합을 도출하였다. 개선된 모델은 기존 접근법과 비교해 통계적 설명력과 특이성이 향상되었음을 검증하였다.

다국어 초록 (Multilingual Abstract)

Upper-limb rehabilitation robots provide repetitive reaching movement training to poststroke patients. However, the early version of robot-aided robot training provides a simple reaching environment and fixed scheduling training which is not individualized to patients. This dissertation introduces a novel individualized reaching training framework that utilizes a model-based evaluation method to schedule individualized reaching training to enhance stroke survivors’ rehabilitation. We first developed an objective evaluation method based on a normal reaching model, assessing reaching performance across one’s workspace without temporal and spatial bias. This method uses three candidate models: Fitts’law for the speed-accuracy relationship, the Almanji model for mouse-pointing tasks in cerebral palsy, and our proposed model. Kinematic data were obtained from healthy (n = 12) and post-stroke (n = 7) subjects using a robot to validate the model and evaluation method. The models derived from the less-affected arm’s performance were used to predict the patients’normal reaching performance, setting the standard for evaluating the affected arm.
Our proposed normal reaching model successfully identified the reaching performance of all healthy subjects (n = 12) and the less-affected arm of post-stroke patients (n = 19; with 16 showing an R2 > 0.7), but did not identify erroneous reaching in the affected arm. Furthermore, our evaluation method visually and intuitively demonstrated the unique motor characteristics of the affected arms, offering insights for individualized training. Once we developed the model-based evaluation method, we proposed a scheduling method acoording to motor learning principles. By incorporating adaptiveness and randomness, the framework aligns with motor learning principles to optimize recovery and skill retention.
Supported by computational modeling based on actual patient data (n = 6), simulations confirmed the framework’s ability to adapt and introduce variability in training schedules compared to conventional methods (p < 0.001) in theoretical level. We will present the scheduling method in the practical level which attempted to overcome the evaluation validity issues due to unreachable movements. Furthermore, we aim to enhance the normal reaching model by establishing a rigorous theoretical and empirical foundation for a data-driven approach. By analyzing a broad set of kinematic features from both healthy and stroke populations, we identified the most informative variables for predicting normal movement time. An exhaustive search of non-linear regression models was conducted, with model performance evaluated using cross-validation and statistical metrics such as R2 and AIC. The resulting models were compared against existing approaches to assess their explanatory power and specificity.

목차 (Table of Contents)

• Chapter 1. INTRODUCTION 1
• 1.1 Background 1
• 1.1.1 Robot-aided therapy 3
• 1.1.2 Toward effective robot-aided therapy 4
• 1.1.3 Challenges 6
• 1.2 Literature Survey 8
• 1.2.1 Overview of Individualized Training Frameworks 8
• 1.2.2 Individualized Mapping Performance 10
• 1.2.3 Robot training with vector field 11
• 1.2.4 Motor Improvement Tracking Method 12
• 1.2.5 Balancing the difficulty mechanism 12
• 1.3 Objectives of Research 15
• 1.4 Problem Statements 15
• 1.4.1 Model-based Evaluation 15
• 1.4.2 Training Scheduling Method 17
• Chapter 2. Model-based Evaluation 20
• 2.1 Introduction 20
• 2.2 Individually scaled evaluation method 21
• 2.2.1 Model formulation for normal reaching movement 22
• 2.2.2 Evaluation method based on normal reaching model 24
• 2.2.3 Visualization of evaluation results 27
• 2.3 Experiments 27
• 2.3.1 Experimental design 27
• 2.3.2 Participants 31
• 2.3.3 Protocols 32
• 2.3.4 Data analysis 33
• 2.4 Results 36
• 2.4.1 Normal Reaching Model Performance 36
• 2.4.2 Evaluation Visualization 46
• 2.4.3 Discussion 49
• 2.4.4 Conclusion 53
• Chapter 3. Individualized Training Framework 55
• 3.1 Introduction 55
• 3.1.1 Proposed Framework 56
• 3.2 Theoretical: Simulation of Scheduling Method 56
• 3.2.1 Simulation Architecture 56
• 3.2.2 Reaching Training Environment 57
• 3.2.3 Recovery Model 57
• 3.2.4 Training Scheduling 60
• 3.2.5 Data Collection and Virtual Patients 61
• 3.2.6 Data Analysis 62
• 3.2.7 Result 63
• 3.2.8 Discussion 65
• 3.3 Practical: Scheduling method accounting for evaluation validity 69
• 3.3.1 Introduction 69
• 3.3.2 Method 70
• 3.3.3 Preliminary Test and Discussion 76
• Chapter 4. Model Modification 80
• 4.1 Introduction 80
• 4.2 Model Exploration 82
• 4.2.1 Feature Exploration 82
• 4.2.2 Data Preparation 83
• 4.2.3 Exhaustive Search of Non-linear Regression Model 83
• 4.2.4 Data Selection Method 91
• 4.2.5 Data Analysis 93
• 4.3 Result 95
• 4.3.1 Kinematic Features Explaining Healthy Reaching Movement Time 95
• 4.3.2 Candidate Models in Healthy Reaching Data 97
• 4.3.3 Candidate Models in Stroke Reaching Data 98
• 4.3.4 Individualized Training Simulation 100
• 4.4 Discussion 102
• Chapter 5. Conclusion 107
• 5.1 Summary 107
• 5.2 Limitations 109