TLDR: This paper argues that current approaches to identifying brain disease biomarkers are flawed due to the brain’s degeneracy and reliance on cohort comparisons and single data types. Using a thought experiment, it demonstrates why more data and powerful AI alone are insufficient. The authors propose a shift towards using multimodal and longitudinal brain data from individuals to guide subgrouping, enabling the definition of multidimensional biomarkers, and advocate for interpretable, model-driven AI analysis.
For years, the quest to find reliable biomarkers for brain diseases has been a challenging endeavor. Despite significant advancements in brain data acquisition and sophisticated AI algorithms, features derived directly from the brain are surprisingly underutilized in clinical diagnosis and prognosis. A recent paper by Pascal Helson and Arvind Kumar argues that the field’s continued reliance on comparing patient groups to healthy controls, alongside the inherent complexity of brain features, is a fundamental flaw hindering progress. This perspective suggests that simply gathering more data or employing more powerful algorithms won’t be enough to unlock the secrets of brain disease biomarkers.
The Problem of Degeneracy and Flawed Comparisons
The core of the issue, as highlighted by Helson and Kumar, lies in the “degeneracy” of brain features. This means that the same behavioral outcome can be achieved through many different patterns of brain activity, and conversely, the same brain activity pattern can lead to different behaviors depending on the context and individual. To illustrate this, the authors present a thought experiment using two identical AI models (Recurrent Neural Networks, RNNs) trained to perform the same task. While both models learned the task equally well, their internal “brain activity” (synaptic weights) diverged due to the stochastic nature of training. Crucially, when the “brain activity” from one healthy RNN was paired with the “readout” (output interpretation) of the other, the performance degraded significantly, making the first RNN appear “dysfunctional.”
This experiment translates directly to the human brain: a healthy brain activity pattern in one person might appear dysfunctional if applied to another, due to individual differences in brain structure and function. This degeneracy undermines the common practice of “cohort comparison,” where researchers try to identify biomarkers by finding statistical differences between a group of patients and a group of healthy controls. This approach implicitly assumes that all healthy brains are statistically identical, which the degeneracy principle refutes. While cohort comparisons have yielded some insights in rare cases, like specific patterns in epilepsy or Parkinson’s disease, they have largely failed to produce novel biomarkers useful in clinical settings for most brain disorders.
The Peril of Partial Information: Why Multimodal Data is Essential
Another critical limitation is the tendency to rely on single types of data. The authors use a simple analogy: imagine trying to determine if a hotel room light is on or off by looking at only one of its two-way switches. It’s impossible because the light’s state depends on both switches together. Similarly, the brain is a multi-scale system, with function and dysfunction determined by interactions across genetic, molecular, cellular, network, and behavioral levels. Measuring only one or a few of these “switches” is unlikely to provide a complete or accurate picture of a brain disease.
This is why genetic, chemical, or morphological features alone often fall short in classifying patients from healthy individuals. Just as a doctor considers multiple symptoms and clinical indicators for a differential diagnosis, understanding brain disorders requires integrating diverse data types—brain activity, neurotransmitters, neuromodulators, and imaging—to capture the full context of neural function and dysfunction.
Also Read:
- Unlocking Brain Connectivity Patterns for Data-Efficient Psychiatric Diagnosis
- Neuromorphic Intelligence: A Dynamical Systems Perspective on Brain-Inspired AI
A Way Forward: Individualized, Multimodal, and Interpretable Approaches
Helson and Kumar propose a paradigm shift. Instead of a “third-person perspective” that compares cohorts, they advocate for a “first-person perspective” focusing on individual-specific, longitudinal measurements of multimodal brain data. This means tracking an individual’s brain over time, across various levels of biological organization, to identify significant deviations from their own baseline.
Central to this new approach is the “subgrouping hypothesis.” The idea is that despite individual uniqueness, functional brains operate within certain “solution islands” in a high-dimensional data space. Diseases might cause individuals to shift within or between these islands. By collecting rich, multimodal, and longitudinal data, researchers can group subjects (not just patients) based on common biological features. This allows for more precise identification of disease-related changes within these subgroups, moving away from a “one-size-fits-all” search for biomarkers.
Finally, the paper emphasizes the role of “interpretable AI” and “model-driven data analysis.” AI should not merely classify but also reveal what features it uses for classification. Guiding feature extraction with biologically plausible models can help uncover underlying dysfunctions, such as imbalances in excitation-inhibition or critical slowing. This combination of model-assisted, AI-driven strategies, integrating diverse datasets from genetics to digital health monitoring, represents the most promising path for advancing brain health. For more detailed insights, you can read the full research paper here.
