Abstract
The use of Artificial Intelligence and machine learning in basic research and clinical neuroscience is increasing. AI methods enable the interpretation of large multimodal datasets that can provide unbiased insights into the fundamental principles of brain function, potentially paving the way for earlier and more accurate detection of brain disorders and better informed intervention protocols. Despite AI's ability to create accurate predictions and classifications, in most cases it lacks the ability to provide a mechanistic understanding of how inputs and outputs relate to each other. Explainable Artificial Intelligence (XAI) is a new set of techniques that attempts to provide such an understanding, here we report on some of these practical approaches. We discuss the potential value of XAI to the field of neurostimulation for both basic scientific inquiry and therapeutic purposes, as well as, outstanding questions and obstacles to the success of the XAI approach.
Introduction: Moving Beyond the Black Box
While artificial intelligence (AI) and machine learning (ML) have become remarkably proficient at classifying complex brain states and predicting clinical outcomes, they frequently operate as "black boxes." In the field of neurostimulation—where devices are designed to modulate brain activity in real-time—this lack of transparency represents a significant barrier to clinical adoption and scientific discovery.
Current clinical standards for conditions like epilepsy often rely on relatively simple "open-loop" systems. However, treating more complex psychiatric and neurological disorders, such as treatment-resistant depression or obsessive-compulsive disorder (OCD), requires "closed-loop" systems. These systems must sense, interpret, and modulate neural circuits at the "speed of thought." High-performing models like deep neural networks are notoriously difficult to interpret, making it challenging for researchers to extract new scientific theories or for clinicians to trust the system’s autonomous decisions. Integrating explainability into the AI pipeline is essential to ensure these interventions are both effective and scientifically grounded.
An XAI-Enabled Closed-Loop Framework
The future of neurostimulation lies in a framework that bridges the gap between high-performing predictive models and mechanistic biological understanding. This paper proposes an XAI-enabled closed-loop neurostimulation process divided into four critical phases:
- System-level recording: Capturing high-resolution, multimodal brain signals including spikes, EEG, and fMRI.
- Multimodal fusion: Integrating neural data with behavioral and cognitive assessments to create a holistic view of the patient's state.
- XAI Algorithm: Utilizing unbiased biomarker discovery to provide mechanistic explanations. This stage is crucial for improving model performance and effectively rejecting stimulation artifacts.
- Validation: Applying complex spatio-temporal stimulation patterns to validate the model's predictions and directly affect behavior.
Categorizing XAI Techniques
To implement this framework, XAI techniques are categorized into three distinct stages of the modeling pipeline:
- Pre-modeling: This stage focuses on characterizing input data through exploratory analysis and feature engineering. Techniques like t-SNE or UMAP are used to identify latent structures within complex datasets before a model is even built.
- Modeling: This involves designing inherently interpretable architectures. Examples include decision trees, rule sets, or "hybrid models" that explicitly combine biophysical theories with data-driven machine learning.
- Post-modeling: These are post-hoc tools applied to existing models to explain specific outputs. Common methods include LIME (Locally Interpretable Model-Agnostic Explanations), saliency maps, and sensitivity analysis.
Key Insights for Neurostimulation
The Accuracy-Explainability Trade-off
A historical tension exists where the most accurate models (e.g., deep learning) are often the least explainable, while the most explainable models (e.g., linear regression) often lack predictive power in complex biological systems. XAI aims to break this trade-off by creating "hybrid" models that maintain high performance while providing interpretable insights.
Causality vs. Correlation
A critical distinction in neuroscience is between neural activity that is merely correlated with a behavior and activity that causes it. XAI is essential for identifying actionable variables—biomarkers that, when targeted by stimulation, actually change the patient's clinical or behavioral state.
Human-in-the-Loop
The authors advocate for an iterative "Human-in-the-Loop" process. In this model, human experts evaluate AI-generated explanations, inject domain-specific knowledge back into the system, and refine stimulation protocols based on both data and clinical intuition.
Challenges and Future Directions
Despite the promise of XAI, several hurdles remain. Current human recording technology often lacks the spatial and temporal resolution (single-neuron, millisecond level) required for optimal XAI models. Furthermore, there is a significant need for a cultural shift toward open data sharing and reporting "valuable failures" to refine these complex systems.
Regulatory hurdles also loom large. Continuously evolving, AI-driven medical devices pose unique challenges for FDA approval and legal liability. Addressing these issues will require close collaboration between machine learning engineers, neuroscientists, and clinicians to move XAI from a theoretical framework to a clinical reality.
Summary updated