A prior-knowledge-guided feature selection method and its application to biomarker identification of schizophrenia

Ying Xing, Godfrey D. Pearlson, Peter Kochunov, Vince D. Calhoun, Yuhui DuNetwork 2025 · DOI

TL;DR

This paper introduces PriFS, a novel feature selection method that leverages prior knowledge about brain connectivity to identify more specific and accurate biomarkers for schizophrenia from fMRI data. By integrating diverse brain-derived information, PriFS significantly improves classification accuracy compared to existing methods, offering a more nuanced understanding of the disorder.

What Problem Does This Paper Address?

Identifying reliable biomarkers for psychiatric disorders like schizophrenia using neuroimaging data is challenging due to the complexity and high dimensionality of the data. Existing feature selection methods often overlook the inherent biological knowledge embedded within brain data, leading to biomarkers with limited specificity and interpretability. This gap hinders our ability to accurately diagnose and understand the underlying neural mechanisms of these disorders. The authors aim to bridge this gap by developing a method that can flexibly incorporate various forms of prior knowledge to guide the feature selection process, leading to more targeted and biologically meaningful biomarkers.

Methods

The core of this paper is the Prior-Knowledge-Guided Feature Selection (PriFS) method. PriFS builds upon a constrained $L_{2,1}$-norm sparse regularization framework. The novelty lies in the introduction of a constraint matrix (B), which is constructed based on prior knowledge. This matrix allows the method to prioritize features that are relevant to the target condition (schizophrenia) by differentially penalizing them.

In addition to the prior-knowledge-guided regularization, PriFS incorporates two other key regularization terms:

  • Graph-based regularization: This term preserves the local structure of the data, ensuring that similar samples in the original feature space remain similar in the transformed space.
  • Redundancy-removal regularization: This term penalizes highly similar features, promoting independence among the selected biomarkers.

These three components are integrated into a single objective function, which is then optimized using an efficient iterative algorithm. The authors also detail how four different types of constraint matrices (BPOS, BNEG, BDIF, BAVG) can be derived from functional network connectivity (FNC) patterns, representing different types of prior knowledge.

Datasets

The study utilizes four multisite resting-state fMRI (rs-fMRI) datasets:

  • COBRE: 68 SZ patients and 89 healthy controls (HCs).
  • MPRC: 150 SZ patients and 238 HCs.
  • FBIRN: 137 SZ patients and 144 HCs.
  • B-SNIP: 182 SZ patients and 237 HCs.

In total, the study includes data from 537 schizophrenia patients and 708 age-matched healthy controls. The authors note that these datasets were obtained from previously published studies and that all human data were de-identified.

Key Insights and Ideas

  • Prior knowledge is crucial for specific biomarker identification: The study strongly suggests that simply applying general feature selection methods to neuroimaging data is insufficient. Incorporating domain-specific knowledge about brain connectivity significantly enhances the specificity and accuracy of identified biomarkers.
  • Different types of prior knowledge reveal different aspects of the disorder: The four types of prior knowledge (positive connectivity, negative connectivity, differences, and average connectivity) led to the selection of distinct sets of features, highlighting that different aspects of brain network dysfunction are captured by different types of prior information.
  • Schizophrenia exhibits complex connectivity alterations: The findings point to both reduced positive connectivity (e.g., within visual and subcortical domains) and altered negative connectivity patterns in individuals with schizophrenia, suggesting a multifaceted dysregulation of brain networks.
  • PriFS offers a flexible framework: The method's ability to integrate diverse prior knowledge makes it adaptable for exploring various hypotheses about brain function and dysfunction in psychiatric disorders.

Key Results

  • Improved Classification Accuracy: PriFS consistently outperformed nine state-of-the-art feature selection methods, achieving an average improvement in classification accuracy of 3.89% to 11.24% across the four datasets.
  • Identification of Specific Abnormalities: The method identified specific schizophrenia-related brain abnormalities, including reduced interactions within the visual domain and between subcortical and visual domains.
  • Reduced Feature Set: PriFS achieved superior classification performance with fewer selected features compared to many other methods.
  • Cross-Dataset Generalizability: The method demonstrated robust performance across different datasets, indicating good generalizability.
  • Ablation Studies: The results from ablation experiments confirmed that each of the three regularization terms (constrained $L_{2,1}$-norm, graph-based, and redundancy-removal) contributes significantly to the overall performance of PriFS.

Limitations and Future Work

The authors acknowledge the following limitations:

  • Manual Parameter Tuning: The optimal performance of PriFS requires manual adjustment of three regularization parameters ($\lambda_1, \lambda_2, \lambda_3$).
  • Limited Scope of Validation: While theoretically generalizable, the current validation focused solely on FNC features in schizophrenia.
  • Potential Confounding Factors: Observed FNC abnormalities might reflect a combination of disease-specific effects and treatment-related effects, as medication confounds were not explicitly regressed out.

Future work suggested by the authors includes:

  • Developing automated parameter optimization techniques.
  • Validating the framework across multimodal data, diverse psychiatric disorders, and comorbid populations.
  • Investigating disease-specific neural signatures and disentangling them from treatment effects.

Summary updated

Abstract

Despite considerable efforts to uncover the neural basis of psychiatric disorders using neuroimaging, few methods utilize intrinsic brain-derived knowledge, leading to limited specificity and discriminability in biomarker identification. To leverage the inherent characteristics within the brain, we propose a prior-knowledge-guided feature selection method to flexibly unveil discriminative and target-oriented biomarkers of psychiatric disorders. Specifically, we construct a constrained sparse regularization allowing for the flexible integration of diverse prior knowledge to identify sparse neuroimaging features linked to specific psychopathology. Additionally, we simultaneously integrate graph-based regularization and redundancy-removal regularization to further ensure the discriminability and independence among the selected features. Different priors hold varying significance in identifying specific biomarkers. Four functional magnetic resonance imaging (fMRI) datasets from 708 healthy controls and 537 schizophrenia patients are used to evaluate our method integrated with various prior knowledge, revealing specific schizophrenia-related brain abnormalities. Compared with nine advanced feature selection methods, our method improves mean classification accuracy by 3.89% to 11.24%, particularly revealing reduced interactions within the visual domain and between subcortical and visual domains in schizophrenia patients. The proposed method offers flexible and precise biomarker identification tailored to specific targets, advancing the understanding and diagnosis of psychiatric conditions.
BibTeX
@inproceedings{xing2025prior,
  author = {Ying Xing and Godfrey D. Pearlson and Peter Kochunov and Vince D. Calhoun and Yuhui Du},
  title = {{A prior-knowledge-guided feature selection method and its application to biomarker identification of schizophrenia}},
  booktitle = {Network},
  year = {2025},
  doi = {10.1162/netn.a.37},
}
Figure