Automated Rehabilitation Exercise Assessment by Genetic Algorithm-optimized CNN
Background
Neurodegenerative disorders like Parkinson's Disease (PD) and stroke significantly impact motor function,
affecting millions of people worldwide. With over 6 million individuals living with PD and more than 80 million
people having experienced a stroke, the need for effective rehabilitation is critical. Physical therapy plays a
vital role in recovery, helping patients regain movement through exercises like trunk rotation, squatting, and
knee extensions. However, traditional rehabilitation requires continuous clinician supervision, making it
resource-intensive and often impractical, especially for long-term therapy. The growing shift towards home-based
rehabilitation, accelerated by the COVID-19 pandemic, has highlighted the need for automated assessment frameworks
to ensure accurate exercise evaluation while reducing healthcare costs. By integrating artificial intelligence
with motion analysis, an automated system can provide objective assessments, minimize clinician workload, and make
rehabilitation more accessible and scalable. This study presents an innovative framework that leverages 1D Local
Binary Pattern (LBP) for feature extraction and a Genetic Algorithm (GA)-optimized Convolutional Neural Network
(CNN) to predict exercise quality, offering a cost-effective and efficient solution for remote rehabilitation.
Workflow
The workflow begins with processing the publicly available KIMORE dataset, which is particularly valuable for
developing and evaluating AI-based rehabilitation assessment models. What sets this dataset apart is its inclusion
of both healthy individuals and patients suffering from stroke, back pain, and Parkinson’s disease. This diversity
ensures that the model is trained and tested on a realistic distribution of subjects, making it more robust and
generalizable to real-world clinical applications. The dataset includes a total of 78 subjects, with 44 healthy
individuals and 34 patients performing five different rehabilitation exercises, providing a well-balanced
foundation for training and evaluating the model in real-world rehabilitation scenarios.
Following this, feature extraction is performed using the 1D Local Binary Pattern (LBP) method. The extracted
features are then fed into a Convolutional Neural Network (CNN) that has been optimized using a Genetic Algorithm
(GA). The CNN predicts an exercise quality score, allowing clinicians to track a patient’s rehabilitation progress
over time. Finally, performance evaluation is conducted using metrics such as Mean Absolute Deviation (MAD) to
ensure model reliability. The automated nature of this process minimizes manual assessment efforts while
maintaining high accuracy in exercise evaluation.
Feature Extraction
Feature extraction is a crucial step in processing raw movement data into meaningful information for exercise
assessment. We employ 1D Local Binary Pattern (LBP), a method known for its ability to capture local variations in
time-series data. This technique involves selecting a sliding window over the motion signal and comparing each
data point with its neighboring values. Based on this comparison, a binary code is generated for each segment of
the motion sequence, effectively encoding movement patterns.
As illustrated in the extracted figure, the 1D LBP method efficiently transforms motion signals into binary
patterns that highlight significant changes in joint movement. This process ensures that fine-grained motion
variations are preserved, making it highly effective for rehabilitation exercises, where even minor changes in
movement patterns can indicate progress or potential issues. By converting raw motion signals into structured
features, we enable the CNN to better learn movement trends and improve assessment accuracy.
Model Optimization
The model's ability to accurately assess rehabilitation exercises relies heavily on optimizing the CNN
architecture. Instead of relying on manually selected hyperparameters, we employ a Genetic Algorithm (GA) to
automatically tune the CNN’s structure and parameters. GA, inspired by the principles of natural selection,
iteratively evolves the model by selecting the best-performing configurations and refining them over multiple
generations.
The CNN is structured with three key layers: convolutional layers, pooling layers, and dense layers. The Genetic
Algorithm searches for the best combination of filters, kernel sizes, pooling strategies, and dropout rates to
minimize error. As shown in the CNN architecture diagram, the optimization process results in a model with
fine-tuned filter sizes, kernel dimensions, and stride values, improving its ability to capture motion
intricacies.
The fitness evolution graph demonstrates how Mean Absolute Deviation (MAD) decreases over generations, proving
that the optimization process leads to a better-performing model. The final CNN architecture, optimized by GA,
achieves lower MAD values on training and testing datasets, ensuring it generalizes well to unseen rehabilitation
exercises.
Results
The model’s performance is evaluated based on its ability to accurately predict exercise quality scores. The
training and validation loss graphs (previous figure) illustrate how the optimized CNN achieves a steady reduction
in error,
indicating that it effectively learns meaningful movement patterns. The scatter plot
comparing predicted and true scores further highlights the model's accuracy, showing a close correlation between
ground truth and predictions.
Benefits and Future Scope
This automated exercise assessment framework offers several advantages. By replacing manual evaluation with
AI-driven automation, it significantly reduces the workload of clinicians, allowing them to focus on patient care
rather than tedious assessments. In the future, integrating real-time feedback mechanisms will enable adaptive
exercise recommendations, allowing patients to adjust their therapy sessions dynamically. The implementation of
multi-objective optimization techniques could further refine CNN hyperparameters, improving both accuracy and
computational efficiency.
