Tdlr

Building interpretability methods directly into the GNN architectures used to predict treatment outcomes in stem-cell-based epilepsy therapy—so that black-box clinical AI models can be trusted, audited, and acted on rather than just believed.

Project summary

AI models are increasingly proposed for high-stakes clinical decisions—predicting which patients will respond to a treatment, whether a therapeutic intervention is working, and when to escalate care. Graph Neural Networks (GNNs) in particular are emerging as the natural architecture for modeling biological systems that are fundamentally relational: neural circuits, protein interactions, and disease progression networks.

But these models are black boxes. A clinician or regulator cannot act on a prediction they cannot interrogate. This is not a hypothetical concern—it is the single most-cited barrier standing between promising clinical AI methods and real deployment. The consequences of getting this wrong are not abstract: misplaced trust in an uninterpretable model's prediction can directly affect a treatment decision for a real patient. This is a small but concrete instance of a much larger challenge that AI safety research broadly concerns itself with—ensuring that AI systems making consequential decisions can be understood, audited, and trusted before they're allowed to act on that trust.

What are this project's goals? How will you achieve them?

Drug-resistant epilepsy (DRE) affects roughly 30% of the global epilepsy population of 50 million people. A promising emerging treatment—transplanting GABAergic inhibitory interneurons to repair the excitatory/inhibitory imbalance that drives seizures—has reached early first-in-human trials, but a central unresolved problem blocks its path to reliable clinical use: how do we determine, from noisy multimodal data (EEG, imaging, molecular markers), whether transplanted cells have actually integrated and formed the correct functional connections?

I've previously developed a formal mathematical framework proposing Graph Neural Networks as the computational architecture for this integration-decoding problem—modeling neural circuits as dynamic, heterogeneous graphs and defining a composite "Integration Success Score" from electrophysiological, structural, and molecular biomarkers. That framework is theoretical: it specifies what the model should predict, but not how to make its predictions interpretable enough for clinical trust.

This project is the next step: implementing and testing explainable-AI methods natively inside that architecture, rather than bolted on after the fact.

What This Project Would Fund

1. Implementation of the proposed heterogeneous, dynamic GNN architecture using existing open epilepsy datasets (OpenNeuro IDEAS, EBRAINS) as a starting substrate, since dedicated stem-cell integration datasets remain scarce.

2. Native interpretability methods—attention-weight visualization, gradient-based saliency mapping, and concept-based explanations—built into the model rather than applied post hoc.

3. Biological validation of the explanations themselves: testing whether the model's attention actually tracks real, known biology (e.g., correctly attending to inhibitory GABAergic connections when predicting excitatory/inhibitory rebalancing) rather than picking up spurious correlations. This is the crux of the safety case—an interpretable-looking model that explains itself incorrectly is arguably worse than an honest black box.

4. Open publication of the resulting pipeline, model, and validation methodology so the interpretability approach is usable as a template for other high-stakes graph-structured clinical prediction problems, not just this one disease.

Why This Is Tractable Now

• The mathematical scaffold (graph construction, multi-task loss function, and composite integration scoring) is already written and published.

• The relevant open datasets and GNN tooling (PyTorch Geometric, DGL) are mature and accessible.

• The scope is deliberately narrow: this is not a claim to solve interpretability in general, but to build and validate one rigorous case study with a clear, checkable success criterion (does the model's stated reasoning match known biology?).

Who is on your team? What's your track record on similar projects?

Solo project led by me. Track record includes published preprint  outputs on AI-driven epilepsy research—including the neural circuit mapping/GNN framework this project directly extends, the Aurevia seizure-prediction wearable design, and related work on AI and stem cell therapy modeling

• Aurevia: AI-Driven Seizure Prediction Through Wearable EEG Systems. A technical framework for real-time seizure prediction using wearable EEG systems and edge AI. [zenodo] [pdf]

• Artificial Intelligence and Stem Cell Therapy for Epilepsy. Review and framework exploring the convergence of regenerative medicine and machine learning. [zenodo] [pdf]

• Neural Circuit Mapping and Integration Analysis. Graph neural network approaches for analyzing stem cell integration within epileptic neural circuits. [zenodo] [pdf]

How will this funding be used?

Estimated $20,000–$35,000 total budget over 6 months:

• Compute (GPU/cloud costs for GNN training and experimentation): ~$8,000–$12,000 — training dynamic/heterogeneous GNNs across multiple architecture variants and interpretability configurations is compute-intensive relative to standard ML workloads.

• Personal runway/time (solo researcher, 6 months, part-time-to-full-time equivalent): ~$12,000–$18,000

• Data access/licensing: ~$1,000–$2,000 — most datasets (OpenNeuro, EBRAINS) are open access, but this covers any restricted-access requests or preprocessing tooling.

• Publication and dissemination (open-source release, preprint/journal fees, documentation): ~$1,500–$2,500

What are the most likely causes and outcomes if this project fails?

The most likely failure mode is a data bottleneck: dedicated datasets for neural stem-cell transplantation in epilepsy remain scarce, and open datasets like OpenNeuro IDEAS weren't built for this specific integration-tracking task, so the model may need to be validated on proxy or synthetic data rather than real longitudinal transplant cohorts. A second risk is that the interpretability methods don't cleanly separate "the model is right for the right reasons" from "the model looks interpretable, but its attention doesn't track real biology" — which is itself part of what this project is designed to test but could mean the validation step returns an ambiguous or negative result rather than a clean success. If it fails, the most useful outcome would be an open-source pipeline and a clear negative/methodological result—still valuable for others trying to build interpretable GNNs on sparse multimodal clinical data—rather than a working clinical tool.

How much money have you raised in the last 12 months, and from where?

None—this would be the project's first funding.