TL;DR
This paper introduces a novel method to identify subtypes within schizophrenia using dynamic functional network connectivity (dFNC). They demonstrate that a deep learning classifier's performance and the features it relies on vary significantly across these subtypes, paving the way for more reliable and personalized clinical decision support systems.
What Problem Does This Paper Address?
Clinical decision support systems (CDSS) powered by neuroimaging data, like deep learning classifiers, aim to aid in diagnosing and understanding neurological and psychiatric disorders. However, a major hurdle to their widespread adoption is their reliability and robustness across diverse patient populations. A key challenge is that these disorders are heterogeneous, meaning they manifest differently in different individuals. Simply training a model on the average presentation of a disorder might lead to poor performance in specific patient subgroups. This paper addresses this by proposing a method to identify distinct subtypes within schizophrenia based on brain activity patterns and then evaluating how a diagnostic classifier performs and what it learns for each subtype. This is crucial because understanding subtype-specific performance is essential for ensuring that CDSS are not only accurate but also equitable and trustworthy for all patients.
Methods
The researchers employed a two-stage clustering approach to identify schizophrenia subtypes. First, they clustered dynamic functional network connectivity (dFNC) data from individuals with schizophrenia to identify distinct "states" of brain network activity. dFNC captures how brain network correlations change over time, offering a richer picture than static connectivity. They identified 4 such states. Second, they calculated the "occupancy rate" (OCR) for each participant – essentially, the percentage of time they spent in each of these dFNC states. These OCRs were then used in a second round of k-means clustering to define data-driven subtypes of schizophrenia.
For classification, they developed an explainable 1-dimensional convolutional neural network (1D-CNN). This model was trained to distinguish between individuals with schizophrenia (SZ) and healthy controls (HCs). To understand why the model made its predictions, they utilized Layer-wise Relevance Propagation (LRP), a technique that attributes the model's output to specific input features (in this case, dFNC patterns). This allowed them to examine which brain network interactions were most important for classifying individuals within each identified subtype.
Datasets
The study utilized the Functional Imaging Biomedical Informatics Research Network (FBIRN) rs-fMRI dataset. This dataset comprises resting-state fMRI recordings from 151 individuals with schizophrenia (SZs) and 160 healthy controls (HCs), collected across multiple sites. This is a publicly available dataset commonly used in neuroimaging research.
Key Insights and Ideas
- Heterogeneity is Key: The study strongly reinforces the idea that neuropsychiatric disorders are not monolithic. Identifying distinct subtypes based on dynamic brain activity is a critical step towards understanding this heterogeneity.
- Classifier Performance Varies: A "one-size-fits-all" deep learning model for schizophrenia diagnosis is insufficient. The classifier's accuracy and reliability can differ significantly depending on which subtype of schizophrenia a patient belongs to.
- Explainability Reveals Subtype-Specific Mechanisms: By using LRP, the researchers found that the deep learning model relies on different dFNC features to classify individuals from different subtypes. This suggests that the underlying brain mechanisms contributing to the classifier's decision-making are also subtype-specific.
- Novel Subtyping Approach: The combination of dFNC state identification followed by OCR-based clustering for subtyping is presented as a novel contribution, moving beyond previous studies that primarily focused on identifying dFNC states or clustering static connectivity.
Key Results
- Four Schizophrenia Subtypes Identified: The OCR-based clustering successfully identified 4 distinct subtypes of schizophrenia, with varying participant numbers (28, 56, 27, and 40). These subtypes were characterized by different patterns of time spent in the identified dFNC states.
- Classifier Performance Above Chance: The 1D-CNN classifier achieved an overall accuracy of 75.00% ± 7.26%, with sensitivity of 74.38% ± 11.24% and specificity of 75.63% ± 14.64%.
- Subtype-Specific Performance Differences: The model performed better for subtypes 2 and 4, which were larger and spent more time in less connected states. Subtypes 1 and 3 showed more variable performance, with subtype 1 exhibiting a significantly lower median performance.
- Subtype-Specific Feature Importance: LRP analysis revealed distinct patterns of feature importance across subtypes. For example, subtype 1 relied heavily on SCN interactions, while subtypes 2 and 4 showed reliance on CBN connectivity, and subtype 3 on VSN/SCN and VSN/SMN interactions.
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Spatial vs. Temporal Explanations: The current LRP analysis focused on spatial differences in relevance. Future work could explore temporal differences in relevance between subtypes.
- Clustering Explainability: Applying clustering explainability approaches could further quantify the relative importance of network pairs for each subtype.
- Clinical and Neuropsychological Correlates: A crucial next step is to investigate how these identified dFNC subtypes relate to neuropsychological assessments and clinical symptoms.
- Model-Specific Subtypes: The current study identified data-specific subtypes of schizophrenia. Future research could explore identifying model-specific subtypes, which might offer a different perspective on classification.
Summary updated