TL;DR
This paper introduces a novel Hierarchical Attention-based Temporal Convolutional Network (HA-TCN) for diagnosing Myotonic Dystrophy from handgrip time series data. The HA-TCN not only achieves competitive diagnostic accuracy but also provides interpretable insights by highlighting specific segments of the time series, particularly the muscle relaxation phase, which is crucial for clinical diagnosis.
What Problem Does This Paper Address?
The diagnosis of Myotonic Dystrophy, a progressive neuromuscular disorder, often relies on analyzing handgrip time series data. A key symptom is myotonia, characterized by delayed muscle relaxation after contraction. Traditional methods often use handcrafted features derived from this data, but these can be sensitive to noise and variability in data collection, leading to inaccurate estimates and requiring expert verification. While deep learning models have shown promise in medical time series analysis, they often suffer from a lack of interpretability, making it difficult for clinicians to understand why a particular diagnosis is made. This paper aims to bridge this gap by developing a deep learning model that is both accurate and interpretable for Myotonic Dystrophy diagnosis, specifically by focusing on the clinically relevant muscle relaxation phase.
Methods
The core of this work is the proposed Hierarchical Attention-based Temporal Convolutional Network (HA-TCN). This architecture builds upon Temporal Convolutional Networks (TCNs) by incorporating a two-level attention mechanism:
- Within-Layer Attention: This mechanism operates on the convolutional activations within each layer of the TCN. It learns to assign weights to different time steps within that layer's output, effectively summarizing the most relevant information at that specific temporal resolution.
- Across-Layer Attention: This higher-level attention mechanism then combines the summarized information from all the within-layer attention outputs. It learns to weigh the importance of information from different TCN layers, creating a final representation for classification.
This hierarchical attention allows the model to not only capture temporal dependencies but also to pinpoint which parts of the time series are most informative for the diagnosis. The HA-TCN uses causal and dilated convolutions, similar to standard TCNs, to efficiently process sequential data and build a large receptive field. A key aspect is that this attention mechanism is designed to be more interpretable than those in some other deep learning models, allowing for the identification of specific time segments.
Datasets
The study utilized a dataset comprising 467 individual handgrip time series samples from 37 patients diagnosed with Myotonic Dystrophy and 270 samples from 18 healthy subjects. Each sample was collected using a Quantitative Myotonia Assessment (QMA) device, involving a squeezing and release of a handgrip. The data was preprocessed by truncating or padding to 750 time steps and normalizing amplitudes to [0,1]. Ten-fold cross-validation at the subject level was employed for evaluation.
Key Insights and Ideas
- Interpretability through Attention: The HA-TCN's hierarchical attention mechanism is crucial for interpretability. Unlike some other attention models where relevant segments are hard to pinpoint, the HA-TCN's design allows for the identification of specific time steps and layers that contribute most to the diagnosis.
- Focus on the Relaxation Phase: The model consistently highlights the muscle relaxation phase of the handgrip time series as the most relevant segment for diagnosing Myotonic Dystrophy. This aligns perfectly with how clinicians diagnose the condition, lending credibility to the model's interpretability.
- Efficiency and Performance: The HA-TCN demonstrates competitive accuracy while being more computationally efficient than standard TCNs, especially with shallower architectures. This suggests a good trade-off between performance and resource utilization.
- Robustness to Noise: The HA-TCN shows increased robustness to noise compared to attention-based LSTM models, likely due to its TCN foundation and hierarchical attention structure.
Key Results
- Accuracy: The HA-TCN achieved an average classification accuracy of 93.82% ± 2.30% and an F1 score of 0.95 ± 0.01. This performance is comparable to other deep learning models like LSTMs and Bi-LSTMs with attention, and all deep learning models outperformed the traditional SVM with handcrafted features.
- Interpretability: Visualizations (Figure 5) show that the HA-TCN consistently identifies the decreasing portion of the handgrip strength during the relaxation phase as the most relevant segment for diagnosis, whereas other attention-based models sometimes highlighted irrelevant parts of the signal.
- Computational Efficiency: The HA-TCN was found to be more computationally efficient for training than standard TCNs, particularly when using fewer hidden layers.
Limitations and Future Work
- Dataset Size: While the dataset is specific to Myotonic Dystrophy, the authors acknowledge that testing the HA-TCN on multiple datasets is a crucial next step to validate its generalizability.
- Comparison with Non-Attention Methods: The paper suggests comparing the HA-TCN's explainability with other non-attention-based interpretable approaches in the future.
- Generalizability: The current study focuses on a specific medical condition (Myotonic Dystrophy) and a specific type of time series data (handgrip). Further research could explore its application to other medical time series and conditions.
Summary updated