Machine Learning Predicts Treatment Response in Bipolar & Major Depressive Disorders

Mustafa S. Salman, Eric Verner, H. Jeremy Bockholt, Zening Fu, Maria Misiura, Bradley T. Baker, Elizabeth Osuch, Jing Sui, Vince D. CalhounbioRxiv 2022 · DOI

TL;DR

This paper presents a novel "DSM-free" machine learning approach using resting-state fMRI data to predict treatment response in patients with bipolar disorder (BD) and major depressive disorder (MDD). By leveraging a robust independent component analysis (ICA) framework and kernel support vector machines (SVMs), the researchers identified replicable biomarkers and achieved promising accuracy in predicting medication class response, offering a potential alternative to traditional diagnostic methods.

What Problem Does This Paper Address?

Diagnosing and treating mood disorders like bipolar disorder (BD) and major depressive disorder (MDD) is notoriously challenging. A key issue is that patients with BD often present with more depressive than manic symptoms, leading to potential misdiagnosis as MDD. This misdiagnosis can result in ineffective or even harmful medication choices, as treatments for MDD (like antidepressants) can sometimes worsen BD. The current diagnostic system, the DSM, is based on symptom clusters rather than underlying biological mechanisms, further complicating accurate diagnosis and treatment. The authors aim to develop a biologically-grounded, "DSM-free" approach to predict treatment response, moving beyond symptom-based diagnoses to improve patient care.

Methods

The core of this study lies in a two-pronged approach:

  • Neuromark ICA for Feature Extraction: The researchers utilized the "neuromark" framework, which employs spatially-constrained independent component analysis (scICA). This method takes resting-state fMRI data and extracts meaningful features in the form of spatial maps (SMs) and time courses (TCs) representing brain network activity. A key aspect is the use of pre-defined "neuromark templates" derived from large datasets, which act as guides to identify reproducible brain networks. This allows for a more standardized and robust extraction of features across different datasets.
  • Kernel SVM for Classification: These extracted features (SMs, functional network connectivity (FNC), or a combination) are then fed into a kernel support vector machine (SVM) classifier. SVMs are powerful tools for finding decision boundaries between different classes. The use of a "kernel function" allows the SVM to map the data into a higher-dimensional space, making it easier to find a linear separation, which is crucial for complex biological data. The study also explores a "multiple kernel SVM" to integrate features from different modalities (SMs and FNC).

The novelty lies in applying this robust ICA framework across multiple datasets and using it to train an SVM for predicting medication response, aiming for a "DSM-free" classification.

Datasets

The study utilized data from three distinct datasets:

  • Western Data: This dataset was used for training the initial models and for replication studies. It included patients with known medication-class responses (mood stabilizers/MS or antidepressants/AD).
  • LA5C Data: This dataset was used for validation. It included individuals with bipolar disorder (BD) and controls.
  • EMBARC Data: This dataset was also used for validation and included patients with major depressive disorder (MDD) and controls. It also provided data on patient improvement scores.

The authors note that the LA5C and EMBARC datasets suffered from imbalanced sample sizes (more patients than controls), which they addressed by selecting an equal number of subjects for classification.

Key Insights and Ideas

  • "DSM-Free" Prediction is Possible: The most significant insight is that it's possible to predict treatment response using neuroimaging data without relying on traditional DSM diagnoses. This opens doors for more objective and biologically-informed clinical decision-making.
  • Replicable Biomarkers: The neuromark ICA approach, combined with SVM, successfully identified "salient biological markers" (specific brain network patterns) that are consistently associated with treatment response across different datasets. This suggests these patterns could serve as reliable biomarkers.
  • FNC Outperforms SMs: Interestingly, functional network connectivity (FNC) features generally outperformed spatial maps (SMs) in classification accuracy across most experiments, suggesting that the dynamic interactions between brain regions are particularly informative for predicting treatment response.
  • Generalizability Across Datasets: The model demonstrated good generalizability, performing well on independent datasets (LA5C and EMBARC) even without site or scanner harmonization techniques, highlighting the robustness of the neuromark approach.
  • Predicting Improvement Scores: Beyond just medication class, the model showed promise in predicting the degree of symptom improvement in MDD patients, offering a more nuanced understanding of treatment outcomes.

Key Results

  • High Accuracy in Replication: The neuromark framework replicated previous findings with high accuracy (95.45% in initial tests).
  • Validation Accuracy: When validated on independent datasets (LA5C and EMBARC), the model achieved testing accuracies of up to 89% without site harmonization.
  • Specific Classification Results:
    • Classification of MS vs. AD responders in Western data: 84.33% accuracy (N=83).
  • Classification of nonresponder subjects from MS: 93.27% accuracy.
  • Classification of nonresponder subjects from AD: 92.38% accuracy.
  • 3-way classification (MS, AD responders, nonresponders): 80.24% accuracy.
  • BD vs. Controls in LA5C: 85.88% accuracy.
  • MDD vs. Controls in EMBARC: 87.48% accuracy.
  • MDD Improvement Score Prediction: Up to 70% accuracy in predicting improvement scores for MDD patients.

Limitations and Future Work

  • Imbalanced Datasets: While addressed by subsampling, the inherent imbalance in some datasets could still influence results.
  • Age Differences: The study acknowledges age differences across datasets as a potential factor affecting performance.
  • Treatment Response Information: Treatment response data was primarily available from the Western dataset, while other sites provided DSM diagnoses. This difference in target labels could impact cross-dataset comparisons.
  • Future Work: The authors suggest several avenues for future research:
  • Multi-site Harmonization: Implementing techniques like ComBat to account for site effects and improve cross-site consistency.
  • Weighted/Multiple Kernel Learning: Exploring more advanced kernel functions to better integrate features from different modalities.
  • Orthonormal Features for TCs: Investigating methods to extract orthonormal features from time courses (TCs) to better align with the Riemannian distance measure.
  • Multi-modal Analysis: Incorporating other neuroimaging modalities like structural MRI (sMRI) and diffusion tensor imaging (DTI) for a more comprehensive analysis.

Summary updated

Abstract

Abstract Diagnosis of bipolar disorder (BD) patients with complex symptoms presents a challenge to clinicians. Patients tend to spend more time in a depressive state than a manic state. In such complex cases, the current Diagnostic and Statistical Manual (DSM), which is not based on pathophysiology, can lead to misdiagnosis as major depressive disorder (MDD) and an imperfect or even harmful medication response. A biologically-based classification algorithm is needed to improve the accuracy of diagnosis. Osuch et al. (2018) presented a kernel support vector machine (SVM) algorithm to predict the medication-class of response from new patient samples whose diagnoses were unclear. Here we also utilize the kernel support vector machine (SVM) algorithm but with a few novel contributions. We applied the robust, fully automated neuromark independent component analysis (ICA) framework to extract comparable features in a multi-dataset setting and learn a kernel function for support vector machine (SVM) on multiple feature subspaces. The neuromark framework successfully replicates the prior result with 95.45% accuracy (sensitivity 90.24%, specificity 92.3%). To further evaluate the generalizability of our approach, we incorporated two additional datasets comprising bipolar disorder (BD) and major depressive disorder (MDD) patients. We validated the trained algorithm on these datasets, resulting in a testing accuracy of up to 89% (sensitivity 0.88, specificity 0.89) without using site or scanner harmonization techniques. We also translated the model to predict improvement scores of major depressive disorder (MDD) with up to 70% accuracy. This approach reveals some salient biological markers of medication-class of response within mood disorders. Highlights We demonstrate a DSM-free approach for predicting treatment response from resting-state functional magnetic resonance imaging (fMRI) data. We identify several replicable biomarkers using the approach. Our work has potential for clinical application by replacing trial-and-error in treating complex psychiatric disorders.
BibTeX
@misc{salman2022machine,
  author = {Mustafa S. Salman and Eric Verner and H. Jeremy Bockholt and Zening Fu and Maria Misiura and Bradley T. Baker and Elizabeth Osuch and Jing Sui and Vince D. Calhoun},
  title = {{Machine Learning Predicts Treatment Response in Bipolar \& Major Depressive Disorders}},
  year = {2022},
  doi = {10.1101/2022.05.13.491903},
  howpublished = {bioRxiv},
}
Figure