TL;DR
This paper introduces a deep learning approach to automatically classify children with spastic diplegia into four distinct clinical forms based on their gait patterns. By leveraging Multilayer Perceptrons (MLPs) and Recurrent Neural Networks (RNNs) on optoelectronic gait data, the study demonstrates that these AI models can achieve classification accuracy comparable to human experts for most forms, offering a promising tool for objective clinical assessment.
What Problem Does This Paper Address?
Cerebral Palsy (CP) is a group of movement disorders affecting children, and spastic diplegia is a specific subtype characterized by impaired coordination, particularly in the lower limbs. Clinically, diplegic patients are categorized into four forms based on their gait patterns, which aids in diagnosis, treatment planning, and communication among clinicians. However, this classification relies on expert observation, which can be subjective and time-consuming. The paper aims to bridge this gap by developing an automated, objective system using deep learning to classify these four forms of diplegia based on quantitative gait analysis. This would provide rehabilitation centers with a reliable tool for quick diagnosis confirmation and decision-making.
Methods
The researchers employed two deep learning architectures:
- Multilayer Perceptron (MLP): A standard feedforward neural network with multiple layers. The paper details a specific architecture that achieved the best performance, including a first hidden layer with 256 units and subsequent layers doubling the number of perceptrons. A dropout layer was incorporated to mitigate overfitting.
- Recurrent Neural Network (RNN): Specifically, a Long Short-Term Memory (LSTM) network was used, which is well-suited for analyzing sequential data like gait patterns. The proposed RNN architecture utilizes a single LSTM layer, chosen to balance performance and prevent overfitting.
Both networks were trained on processed gait data, which involved transforming 3D marker coordinates into 27 angular parameters per frame. Further preprocessing included extracting frequency domain information using Fast Fourier Transform (FFT) and organizing data into sequences.
Datasets
The study utilized a dataset comprising gait data from 174 patients collected in an Italian Motion Analysis Lab (LAMBDA). This dataset consists of 1121 trials (walks). The data was acquired using an optoelectronic system with 19 markers and subsampled to 50 frames/sec. The distribution of patients across the four diplegia forms in the dataset was uneven. The data was partitioned into training and testing sets, with data augmentation (doubling occurrences for form 1) applied to the training set to address the imbalance.
Key Insights and Ideas
- Deep Learning for Clinical Gait Classification: The paper demonstrates the feasibility of using deep learning, specifically MLPs and LSTMs, to automate a clinically established classification system for diplegia. This moves beyond traditional feature engineering by allowing the models to learn relevant patterns directly from the data.
- LSTM's Suitability for Sequential Gait Data: The RNN (LSTM) approach showed superior performance compared to the MLP, highlighting the advantage of models designed to capture temporal dependencies inherent in walking sequences.
- Form 3 Classification Challenge: The study identified that classifying "form 3" diplegia was particularly challenging for both models. This aligns with clinical observations, suggesting that perceptual impairments, which are not directly captured by motion data, play a significant role in this form.
- Potential for Objective Clinical Support: The results suggest that these AI models could serve as objective tools to support clinicians in diagnosing and subtyping diplegia, potentially reducing subjectivity and improving efficiency.
Key Results
- RNN Performance: The LSTM network achieved an overall accuracy of approximately 0.869 on test trials, outperforming the MLP network's accuracy of 0.804.
- MLP Performance: The best-performing MLP architecture achieved an accuracy of about 0.587 on test trials.
- Comparison to Experts: The classification results were comparable to those provided by clinical experts in 3 out of 4 forms.
- Form 3 Difficulty: Both models struggled with form 3, with low accuracy scores (e.g., 0.111 for MLP T1 and 0.333 for LSTM T1).
- SVM Baseline: Both the MLP and RNN methods outperformed a Support Vector Machine (SVM) baseline.
- Overfitting Mitigation: The use of dropout in the MLP and the choice of a single LSTM layer in the RNN helped to mitigate overfitting.
Limitations and Future Work
- Form 3 Classification: The paper acknowledges the difficulty in classifying form 3, attributing it to the potential perceptual impairments that are not captured by the motion data. Future work could explore incorporating other modalities or features to address this.
- Dataset Size and Distribution: While the dataset is substantial (174 patients, 1121 trials), the uneven distribution of patients across forms, particularly the scarcity of form 1 cases, presented a challenge. Data augmentation was used, but larger, more balanced datasets could further improve performance.
- Data Confidentiality: The authors note that the dataset could not be made public due to confidentiality, limiting direct comparison with other studies. They commit to releasing the anonymized dataset in the future.
- Overlap Between Forms: The authors mention that diplegia forms can partially overlap, making classification inherently complex. The models' ability to provide probability scores for each form could assist in identifying intermediate or non-nominal cases.
- Future Directions: The paper suggests that these deep learning techniques could become valuable tools in motion analysis labs (MALs) for objective classification and could be extended to other CP subtypes or movement disorders.
Summary updated