Neuro-Oracle: A Trajectory-Aware Agentic RAG Framework for Interpretable Epilepsy Surgical Prognosis

1. Problem Statement

Predicting post-surgical seizure outcomes for patients with pharmacoresistant epilepsy is a significant clinical challenge. Current deep learning approaches suffer from three primary limitations:

• Static Inputs: They rely on single-timepoint pre-operative scans, ignoring longitudinal morphological changes (e.g., progressive cortical thinning) that occur between pre- and post-operative states.
• Data Scarcity & Overfitting: Surgical cohorts are small (often <300 patients), making high-capacity end-to-end classifiers prone to overfitting.
• Lack of Interpretability: Standard classifiers output probability scores without verifiable clinical rationale, limiting their adoption in surgical decision support.
• Data Limitation: The study utilizes the public EPISURG dataset, which lacks ground-truth seizure-freedom scores (Engel Class). Consequently, the authors use a clinical proxy label based on surgery type (Temporal resections = "Favorable"; Non-temporal/Hemispherectomy = "Less Favorable"), acknowledging this introduces a risk of the model learning anatomical resection patterns rather than true prognostic pathology.

2. Methodology: The Neuro-Oracle Framework

Neuro-Oracle is a three-stage framework designed to encode longitudinal changes, retrieve similar historical cases, and generate interpretable prognoses using a Retrieval-Augmented Generation (RAG) paradigm.

Phase 1: Contrastive Trajectory Encoding

• Architecture: A 3D Siamese contrastive encoder (based on MONAI 3D ResNet-50) processes paired pre-operative ($X_{t0}$) and post-operative ($X_{t1}$) T1-weighted MRIs.
• Preprocessing: Scans are rigidly registered (6-DOF) to align coordinate spaces, avoiding deformable registration to preserve localized volume changes.
• Encoding: The network extracts feature vectors ($v_{t0}, v_{t1}$). The difference vector is projected via an MLP into a compact 512-dimensional trajectory vector ($\Delta v$).
• Training Objective: The model is trained using a combination of Supervised Contrastive Loss (to cluster similar outcomes) and Focal Loss (to address the 4:1 class imbalance).

Phase 2: Population Archive & Trajectory Retrieval

• Archive Construction: All training subjects are encoded into a FAISS index using their 512-dim trajectory vectors.
• Retrieval: For a new query patient, the system retrieves the top-$k$ (k=5) historically similar trajectories based on cosine similarity (inner product of normalized vectors).
• Filtering: An age-gap heuristic ($\delta = 15$ years) is applied to filter out retrieved neighbors with significant age discrepancies, ensuring demographic relevance.

Phase 3: Calibrated LLM Reasoning Agent

• Agent: A quantized Llama-3-8B-Instruct model (4-bit NF4) acts as the reasoning engine.
• Prompting: The agent receives the query patient's metadata and the retrieved neighbor profiles (age, sex, outcome). It is instructed to output a verdict token (SUCCESS or FAILURE) followed by a single justification sentence.
• Probability Calibration: The final prediction probability ($p_{Oracle}$) is a weighted blend of the neighbor majority vote ($p_{neighbor}$) and the LLM's binary verdict ($p_{LLM}$):
  $$p_{Oracle} = 0.60 \times p_{neighbor} + 0.40 \times p_{LLM}$$
  This balances geometric evidence with semantic reasoning.

3. Key Contributions

1. Trajectory-Aware Architecture: First application of a 3D Siamese contrastive encoder to distill pre-to-post-operative MRI changes into a retrieval-compatible trajectory vector for epilepsy prognosis.
2. Agentic RAG for Clinical Prognosis: Introduction of a visual RAG system where an LLM agent synthesizes natural-language prognoses grounded in retrieved 3D medical imaging evidence, rather than relying on a fixed classifier.
3. Interpretability without Sacrifice: The framework achieves performance comparable to purely discriminative models while providing auditable, natural-language justifications with zero observed hallucinations in the audit protocol.
4. Open-Source Reproducibility: The entire pipeline relies on public datasets (EPISURG) and open-source tooling (MONAI, FAISS, HuggingFace).

4. Experimental Results

Evaluated on the EPISURG dataset (N=268) using 5-fold stratified cross-validation.

• Performance vs. Baselines:
  • Static Baselines: A single-timepoint ResNet-50 achieved an AUC of 0.793.
  • Trajectory Classifiers: Methods using the trajectory vector ($\Delta v$) significantly outperformed static baselines. The Siamese Diversity Ensemble (M6) achieved the highest AUC of 0.905.
  • Neuro-Oracle (M5): Achieved an AUC of 0.867, matching the 3D-ViT static baseline and the k-NN trajectory classifier, while providing interpretability.
• Sensitivity: Trajectory-based methods showed substantially higher sensitivity for the minority class (less favorable outcomes) compared to the static ResNet-50 (0.396 vs. ~0.80+).
• Ablation Studies:
  • Removing the trained encoder (random initialization) dropped AUC to 0.488 (chance level), confirming the trajectory vector carries discriminative signal.
  • Reducing retrieval to $k=1$ dropped AUC to 0.741, highlighting the benefit of multi-neighbour voting.
• Audit Metrics:
  • Zero Hallucinations: The LLM produced no outputs containing information outside the provided context window.
  • Age-Filter Adherence: 73.9% adherence to the age-gap filtering rule in generated reasoning.

5. Significance and Limitations

• Clinical Impact: Neuro-Oracle bridges the gap between high-performance AI and clinical trust. By providing a "traceable evidence chain" (retrieved similar cases + LLM reasoning), it addresses the "black box" problem in surgical decision support.
• Trade-off: While the purely discriminative ensemble (M6) has a slightly higher AUC (0.905 vs. 0.867), Neuro-Oracle is deemed more deployable because its reasoning is transparent and auditable.
• Limitations:
  • Proxy Labels: The model predicts surgery type as a proxy for seizure freedom, meaning it may learn anatomical resection features rather than true epileptogenic pathology.
  • Cohort Size: The small dataset (N=268) from a single center limits generalizability claims and statistical power for pairwise comparisons.
  • Future Work: Requires linking to true Engel outcome registries and testing on multi-center, multi-scanner cohorts.

Conclusion: Neuro-Oracle demonstrates that encoding longitudinal morphological changes into a trajectory vector, combined with RAG-based reasoning, offers a robust, interpretable, and high-performing approach for epilepsy surgical prognosis, setting a new standard for explainable medical AI.