TL;DR
This paper introduces an interpretable fully connected neural network (FCNN) for classifying Autism Spectrum Disorder (ASD) from resting-state functional connectivity (rsFC) data. Beyond achieving state-of-the-art accuracy, the FCNN's piecewise linear nature allows for precise interpretation, revealing specific rsFC connections that drive the classification for individual subjects and the ASD group.
What Problem Does This Paper Address?
Autism Spectrum Disorder (ASD) is a complex developmental disorder with diverse presentations, making diagnosis challenging. Functional Magnetic Resonance Imaging (fMRI) offers a non-invasive way to study brain function, and machine learning has been applied to fMRI data for ASD detection. However, deep learning models, while powerful, often act as "black boxes," making it difficult to understand why they make certain predictions. This lack of interpretability hinders trust and limits the ability to extract meaningful biological insights. This paper aims to bridge this gap by developing an interpretable deep learning model for ASD classification using rsFC data, allowing for both accurate prediction and clear explanation of the underlying brain connectivity patterns.
Methods
The core of this work is a novel Interpretable Fully Connected Neural Network (FCNN). This FCNN is built upon a Piecewise Linear Neural Network (PLNN) architecture, specifically using the LeakyReLU activation function. LeakyReLU, being a piecewise linear function, allows the entire network to be mathematically represented as a set of local linear classifiers. This property is crucial for interpretability.
The FCNN consists of:
- An input layer receiving rsFC features.
- Two hidden layers, each comprising a fully connected layer, Batch Normalization (BN), and LeakyReLU activation.
- An output layer using a sigmoid function for probability prediction.
A key innovation is the interpreting method developed alongside the FCNN. This method allows for:
- Instance-level interpretation: For any given subject, the model can generate a precise linear formula explaining its prediction, highlighting the specific rsFC connections (features) that contributed most to that individual's classification.
- Group-level interpretation: By analyzing the weights across multiple subjects in the ASD group, the method identifies a set of decision features (rsFC connections) that are consistently important for classifying ASD.
Datasets
The study utilizes the Autism Brain Imaging Data Exchange (ABIDE) I database. This large, multi-site dataset contains resting-state fMRI data from 871 subjects, including 403 individuals with ASD and 468 healthy controls (HC). The data was preprocessed using the Configurable Pipeline for the Analysis of Connectomes (CPAC) and parcellated into 116 regions of interest (ROIs) based on the AAL atlas.
Key Insights and Ideas
- Piecewise linearity is key to interpretability: By using LeakyReLU, the authors create a deep neural network that can be understood through linear models, making the decision-making process transparent.
- Beyond accuracy: Understanding why: The paper demonstrates that achieving high classification accuracy is only half the battle. The ability to explain which brain connections are driving the classification is equally, if not more, important for scientific discovery and clinical trust.
- Individualized explanations: The model can provide a specific linear formula for each subject, offering insights into their unique connectivity patterns that contribute to their ASD classification.
- Identifying group-level biomarkers: The interpretation method can pinpoint a set of rsFC connections that are consistently important for distinguishing the ASD group from controls, potentially revealing underlying neurobiological differences.
- Batch Normalization improves performance: The study found that incorporating Batch Normalization layers in the FCNN not only improved performance but also enhanced model stability.
Key Results
- State-of-the-art classification accuracy: The proposed FCNN achieved the highest classification accuracy of 69.81% on the ABIDE I dataset, outperforming traditional methods like SVM and Random Forest, as well as other deep learning approaches.
- Strong performance metrics: The FCNN also demonstrated superior sensitivity (63.05%), specificity (75.63%), F1 score (65.82%), and AUC (0.7262) compared to benchmark models.
- Identified decision features: The interpretation method successfully identified specific rsFC connections that are crucial for ASD classification. For instance, when K=20, 15 decision features involving 23 brain regions were identified, with two of these connections showing statistically significant differences between ASD and HC groups (p<0.05). This suggests the model can find subtle but important patterns.
Limitations and Future Work
- Dataset size and generalizability: The authors acknowledge that the dataset, while large, is limited to 871 participants. Future work should involve testing the approach on larger, more diverse datasets to ensure generalizability.
- Model complexity: The current FCNN is a relatively compact network. Investigating the effectiveness of the interpretation approach with deeper and more complex neural network architectures is a promising direction.
- Broader applications: While focused on ASD, the proposed interpretable learning framework has the potential to be generalized to other brain science and medical applications that utilize deep neural networks, offering a valuable tool for biomarker discovery and diagnosis.
Summary updated