Multimodal Sparse Classifier for Adolescent Brain Age Prediction

Peyman Hosseinzadeh Kassani, Alexej Gossmann, Yu-Ping Wang

TL;DR

This paper introduces a novel sparse machine learning model, RES-ELM, to predict adolescent brain age using multimodal fMRI data. By cleverly leveraging the numerical "residual errors" generated during the pseudo-inverse computation of an Extreme Learning Machine (ELM), the model effectively prunes redundant neural network nodes. This approach overcomes the severe overfitting typical of high-dimensional neuroimaging data, achieving state-of-the-art classification accuracy while using a fraction of the computational nodes.

What Problem Does This Paper Address?

Understanding how the brain matures from childhood to adulthood is crucial for studying neurodevelopment and psychiatric disorders. Researchers often use functional MRI (fMRI) to predict a subject's "brain age" based on how different regions of the brain communicate (functional connectivity).

However, this creates a massive data problem: fMRI data is extremely high-dimensional. A single scan can produce tens of thousands of connectivity features, but a typical study might only have a few hundred subjects. This high variable-to-instance ratio ($p \gg N$) causes standard machine learning models to severely overfit—they memorize the training data but fail to generalize. The authors specifically tackle the overfitting problem in Extreme Learning Machines (ELMs), a type of fast, single-hidden-layer neural network, by figuring out how to identify and drop useless features across multiple fMRI modalities.

Methods

The authors build on the Extreme Learning Machine (ELM), a neural network where input weights are assigned randomly, and output weights are calculated in a single step using a least-squares estimation (specifically, the Moore-Penrose pseudo-inverse).

To prevent the ELM from overfitting, the authors propose RES-ELM (Residual Error based Sparse ELM). Their novel contribution lies in how they prune the network:

  1. The "Bug" as a Feature: When computing the pseudo-inverse of a massive matrix using Singular Value Decomposition (SVD), tiny singular values are usually truncated to zero to prevent numerical instability. This truncation introduces a small mathematical discrepancy, which the authors call the residual error.
  2. Neuron Ranking: The authors realized that specific hidden neurons in the ELM are responsible for the bulk of this residual error. They designed an algorithm to calculate the relative error contributed by each neuron.
  3. Pruning: RES-ELM ranks the hidden neurons by their residual error and aggressively prunes the ones causing the highest errors. This leaves a sparse, highly optimized network that only retains the most informative features.

Prior to the ELM, the fMRI data is preprocessed using Group Independent Component Analysis (ICA) to reduce spatial dimensions and extract functional connectivity matrices.

Datasets

The study uses data from the Philadelphia Neurodevelopmental Cohort (PNC).

  • Subjects: ~900 adolescents aged 8 to 21. To frame this as a binary classification problem ("low age" vs. "high age"), the authors converted ages to z-scores and discarded the middle of the distribution, leaving 568 subjects.
  • Modalities: Three distinct fMRI scans were used for each subject:
  • Resting-state fMRI (rs-fMRI)
  • Working memory n-back task (nb-fMRI)
  • Emotion identification task (em-fMRI)

Key Insights and Ideas

  • Repurposing Numerical Error: The most elegant insight of the paper is taking a standard linear algebra headache—truncation errors in SVD—and turning it into a powerful heuristic for neural network pruning.
  • Curing ELM's Fatal Flaw: In a standard ELM, as the number of hidden neurons increases to match the number of training instances, the network perfectly memorizes the data but its test accuracy plummets. RES-ELM's residual-error pruning completely stabilizes this; the network maintains high accuracy even as the base pool of hidden neurons grows.
  • Active Brains Reveal More Age: When analyzing which features survived the pruning process, the working memory task (nb-fMRI) proved to be the most predictive individual modality for brain age, slightly edging out resting-state data.

Key Results

  • Top Accuracy: RES-ELM achieved the highest overall accuracy across almost all tests. The best combination was the bi-modal pairing of working memory and resting-state fMRI (nb-fMRI + rs-fMRI), hitting 85.45% accuracy.
  • Network Efficiency: RES-ELM achieved these top-tier results while using significantly fewer hidden neurons than its competitors. For example, it often used only ~100 to 200 neurons out of a possible 1500 candidate nodes.
  • Beating the Baselines: RES-ELM consistently outperformed standard ELM (which suffered heavily from overfitting, hovering around 80%), Randomly Pruned ELM (RP-ELM), and was highly competitive with the computationally expensive Sparse Bayesian ELM (SB-ELM).

Limitations and Future Work

  • The Limits of Multimodality: Interestingly, the authors found that combining all three fMRI modalities actually resulted in a slight drop in accuracy compared to the best two-modality combination. Adding a third modality increased the feature-to-instance ratio so much that even the sparse classifier began to struggle with overfitting.
  • Binary Simplification: To simplify the machine learning task, the authors discarded subjects in the middle of the age distribution to create a binary "young vs. old" classification. In clinical reality, brain age is a continuous regression problem.
  • Future Work: While not explicitly detailed in a dedicated section, the results imply a need for even more robust multimodal integration techniques that can handle an arbitrary number of data streams without succumbing to the curse of dimensionality.

Summary updated

Abstract

The study of healthy brain development helps to better understand both brain transformation and connectivity patterns, which happen during childhood to adulthood. This study presents a sparse machine learning solution across whole-brain functional connectivity measures of three datasets, derived from resting state functional magnetic resonance imaging (rs-fMRI) and two task fMRI data including a working memory n-back task (nb-fMRI) and an emotion identification task (em-fMRI). The fMRI data are collected from the Philadelphia Neurodevelopmental Cohort (PNC) for the prediction of brain age in adolescents. Due to extremely large variable-to-instance ratio of PNC data, a high-dimensional matrix with several irrelevant and highly correlated features is generated, and hence a sparse learning approach is necessary to extract effective features from fMRI data. We propose a sparse learner based on the residual errors along the estimation of an inverse problem for extreme learning machine (ELM). Our proposed method is able to overcome the overlearning problem by pruning several redundant features and their corresponding output weights. The proposed multimodal sparse ELM classifier based on residual errors is highly competitive in terms of classification accuracy compared to its counterparts such as conventional ELM, and sparse Bayesian learning ELM.
BibTeX
@article{kassani2020multimodal,
  author = {Peyman Hosseinzadeh Kassani and Alexej Gossmann and Yu-Ping Wang},
  title = {{Multimodal Sparse Classifier for Adolescent Brain Age Prediction}},
  journal = {IEEE J. Biomed. Health Informatics},
  year = {2020},
  url = {https://openreview.net/forum?id=J73fJ5n_5X},
}
Figure