Classification of kinematic data from a robotic upper limb rehabilitation orthosis

Abstract

In today’s world, the role of computer science has become so essential that it can be considered mandatory in various aspects of life. Over time, it has seamlessly integrated into nearly every domain, from engineering to biology and even medicine. there are many examples in this matter such as using algorithms to enhance the clarity of radiographic images, improve the quality of echo-cardiograms, and assist doctors in diagnosing patients. These advancements have allowed for more precise diagnoses, quicker decision-making, and overall better patient care. However, what happens when we combine the power of both engineering and computer science? The possibilities for innovation and problem-solving expand tremendously, offering new pathways for creating cutting-edge solutions or improving existing technologies, particularly in fields like rehabilitation. Rehabilitation, while not a novel concept in medicine, has historically been underdeveloped due to limitations in materials and a lack of comprehensive knowledge. Until recently, advancements in rehabilitation had been modest at best. However, with the rapid progress in medical technology, rehabilitation methods have evolved significantly. New techniques, such as brain-computer interfaces (BCIs), stem cell therapy, and functional electrical stimulation, have emerged as promising approaches to patient recovery. Despite these innovations, significant challenges still remain in modern rehabilitation practices. One of the primary issues is the stagnation or plateauing of progress during therapy sessions, often compounded by a heavy dependence on human therapists. Another notable problem is the difficulty in recognizing improvement until much later in the rehabilitation process, making it hard to track incremental progress. These ongoing challenges underscore the need for more advanced and reliable solutions. One promising approach is the fusion of computer science and engineering to develop more efficient and effective rehabilitation tools. This can be achieved through the creation of robotic rehabilitation systems, where engineering plays a role in the design and development of the robotic device, and computer science, particularly through machine learning, is employed to analyze data collected during rehabilitation sessions. Machine learning algorithms can extract and process vast amounts of kinematic and physiological data from the robotic device, using this data to monitor patient progress, predict outcomes, and adapt therapy protocols in realtime. This integration of computer science and engineering offers a more personalized, datadriven approach to rehabilitation, potentially minimizing reliance on therapists, providing more immediate feedback on patient progress, and leading to better long-term recovery outcomes. By merging these fields, we can build a future where rehabilitation is not only more accessible but also far more efficient, offering real-time improvements and adjustments based on patient data. This represents a significant leap forward from traditional methods, moving toward a more autonomous and intelligent form of therapeutic care, ultimately leading to better recovery experiences for patients. The robot we are discussing is called the Armeo robot and is created by the company HACOMA. This robot is used during rehabilitation sessions for patients with upper limb injuries or abnormalities. It helps patients exercise by engaging in mini-games on a computer that encourage movement of the upper limb. Sensors attached to the patients’ limbs record data on the angles of movement, which is then used in machine learning models to predict outcomes. As part of our research, we aimed to determine which of the well-known algorithms (KNN, SVM, Linear regression) is best suited for this goal. We found that linear regression could not be used for classification, so we used logistic regression instead. During the research, several models were created and divided into two groups. The first group was regression models, creating two models for each of the three algorithms. The models Abstract calculated the range of motion for the entire upper limb using the range of motion of each angle, and predicted the healthiness of each angle as a percentage. In this task, SVM performed better than the other two algorithms according to accuracy tests. The second group consisted of classification models, with three models created for each algorithm. The first model had four classes (good, bad, excellent, severely bad) and classified the patient’s upper limb based on the range of motion obtained from the regression task. Of the three algorithms, KNN performed better according to accuracy test results. We further tested the trajectories and durations to classify the limb into two classes (healthy and unhealthy), with KNN performing well in this task too. It’s worth noting that we employed code enhancements such as incremental learning and dynamic thresholds to achieve our goals. The differences found were attributed to the complexity of the data, as we were working with atypical data that still requires refinement for accurate results. Overall, these experiments provided insights into how the classification algorithms would perform if clinically integrated.