The Pragmatic Classification of Upper Extremity Motion in Neurological Patients: A Primer

Useless as is. Assessments NOT SOLUTIONS. After you have assessed the disability what does the survivor need to do to get 100% recovered?

The Pragmatic Classification of Upper Extremity Motion in Neurological Patients: A Primer

Avinash Parnandi¹, Jasim Uddin², Dawn M. Nilsen³ and Heidi M. Schambra^1,4*

• ¹Department of Neurology, New York University School of Medicine, New York, NY, United States
• ²Department of Neurology, Columbia University Medical Center, New York, NY, United States
• ³Department of Rehabilitation and Regenerative Medicine, Columbia University Medical Center, New York, NY, United States
• ⁴Department of Rehabilitation Medicine, New York University School of Medicine, New York, NY, United States

Recent advances in wearable sensor technology and machine learning (ML) have allowed for the seamless and objective study of human motion in clinical applications, including Parkinson's disease, and stroke. Using ML to identify salient patterns in sensor data has the potential for widespread application in neurological disorders, so understanding how to develop this approach for one's area of inquiry is vital. We previously proposed an approach that combined wearable inertial measurement units (IMUs) and ML to classify motions made by stroke patients. However, our approach had computational and practical limitations. We address these limitations here in the form of a primer, presenting how to optimize a sensor-ML approach for clinical implementation. First, we demonstrate how to identify the ML algorithm that maximizes classification performance and pragmatic implementation. Second, we demonstrate how to identify the motion capture approach that maximizes classification performance but reduces cost. We used previously collected motion data from chronic stroke patients wearing off-the-shelf IMUs during a rehabilitation-like activity. To identify the optimal ML algorithm, we compared the classification performance, computational complexity, and tuning requirements of four off-the-shelf algorithms. To identify the optimal motion capture approach, we compared the classification performance of various sensor configurations (number and location on the body) and sensor type (IMUs vs. accelerometers). Of the algorithms tested, linear discriminant analysis had the highest classification performance, low computational complexity, and modest tuning requirements. Of the sensor configurations tested, seven sensors on the paretic arm and trunk led to the highest classification performance, and IMUs outperformed accelerometers. Overall, we present a refined sensor-ML approach that maximizes both classification performance and pragmatic implementation. In addition, with this primer, we showcase important considerations for appraising off-the-shelf algorithms and sensors for quantitative motion assessment.

Introduction

Wearable sensors, such as inertial measurement units (IMUs) and accelerometers, provide an opportunity for the objective, and seamless capture of human motion. Machine learning (ML) enables computers to learn without being explicitly programmed, and provides an opportunity to rapidly identify patterns in data. ML is potentially a powerful tool for clinical application because of its ability to automatically recognize categories of interest. These categories could be used for diagnostic purposes (e.g., severity of disease, disease identification) or therapeutic purposes (e.g., dose quantitation during stroke rehabilitation).

Given recent technological and computational advances, combining wearable sensor data with ML algorithms has the potential for rapid, automated, and accurate classification of motion. Researchers have begun using this combined sensor-ML approach in a number of applications. These include human activity recognition (1–3), gesture analysis (4), assessment of bradykinesia in Parkinson's disease (5, 6), motor function assessment in multiple sclerosis (7), and differentiating between functional and non-functional arm usage in stroke patients (8, 9). While many of these studies showcase the application of sensors and ML in clinical populations, no previous work has detailed the various hardware and software considerations for using the sensor-ML approach. Furthermore, no guide currently exists to advise investigators in building and troubleshooting this approach, which sits at the intersection of human movement science, data science, and neurology. With the potential for the sensor-ML approach to have widespread applicability to neurological disorders, understanding how to develop this approach for one's own area of inquiry is paramount.

One possible application of the combined sensor-ML approach is the monitoring of rehabilitation dose in stroke patients. Quantifying the dose of rehabilitation entails classifying units of measurement, which are subsequently tallied. In our previous proof-of-principle study, we used IMUs worn by stroke subjects performing a structured tabletop activity to capture motion data. Our units of measurement were functional primitives, elemental motions that cannot be further decomposed by a human observer. We applied an ML algorithm (hidden Markov model with logistic regression) to the IMU motion data to recognize primitives embedded in this activity, achieving an overall classification performance of 79% (10). While promising, this sensor-ML approach had variable classification performance among the primitives (62–87% accuracy). It also did not address research implementation challenges such as the computational complexity and computational costs of the ML approach, or clinical implementation challenges such as the expense (11) and electromagnetic intolerance of the IMUs.

In the present study, we address these limitations in the form of a primer, outlining deliberations that researchers developing their own sensor-ML approach would need to consider. We describe our rationale and steps for identifying (1) an algorithm that is highly accurate but computationally tractable, and (2) the type and array of sensors that minimize cost but maximize accuracy. We use functional primitives as the motion type to be classified, and describe our approach for both capturing and identifying these motions. We also use off-the-shelf algorithms and sensors, providing an accessible framework for investigators seeking to address new scientific and clinical questions with the sensor-ML approach.