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

Aizierjiang Aiersilan; Mohamad Koubeissi

cs.MMApr 10, 2026

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

Authors: Aizierjiang Aiersilan, Mohamad Koubeissi

Organizations: The George Washington University

Abstract

Predicting post-surgical seizure outcomes in pharmacoresistant epilepsy is a clinical challenge. Conventional deep-learning approaches operate on static, single-timepoint pre-operative scans, omitting longitudinal morphological changes. We propose Neuro-Oracle, a three-stage framework that: (i) distils pre-to-post-operative MRI changes into a compact 512-dimensional trajectory vector using a 3D Siamese contrastive encoder; (ii) retrieves historically similar surgical trajectories from a population archive via nearest-neighbour search; and (iii) synthesises a natural-language prognosis grounded in the retrieved evidence using a quantized Llama-3-8B reasoning agent. Evaluations are conducted on the public EPISURG dataset ( $N{=}268$ longitudinally paired cases) using five-fold stratified cross-validation. Since ground-truth seizure-freedom scores are unavailable, we utilize a clinical proxy label based on the resection type. We acknowledge that the network representations may potentially learn the anatomical features of the resection cavities (i.e., temporal versus non-temporal locations) rather than true prognostic morphometry. Our current evaluation thus serves mainly as a proof-of-concept for the trajectory-aware retrieval architecture. Trajectory-based classifiers achieve AUC values between 0.834 and 0.905, compared with 0.793 for a single-timepoint ResNet-50 baseline. The Neuro-Oracle agent (M5) matches the AUC of purely discriminative trajectory classifiers (0.867) while producing structured justifications with zero observed hallucinations under our audit protocol. A Siamese Diversity Ensemble (M6) of trajectory-space classifiers attains an AUC of 0.905 without language-model overhead.

Figures & tables

Figure 1: Overview of the proposed Neuro-Oracle framework. (i) Pre- and post-operative MRI pairs are registered and encoded through a shared 3D Siamese ResNet-50 backbone. (ii) The morphological difference is projected into a compact trajectory vector, which is used to retrieve historically similar cases from a population archive. (iii) A quantized LLM reasoning agent evaluates the retrieved evidence to synthesize a transparent, natural-language surgical prognosis.

Category	Method	AUC $\uparrow$	F1 $\uparrow$	Sensitivity $\uparrow$	Specificity $\uparrow$	Balanced Acc. $\uparrow$
Static encoders	ResNet-50 Static	$0.793$	0.775	$0.396$	$0.921$	$0.659$
Static encoders	3D-ViT Static	$0.867$	0.775	$0.811$	0.930	0.871
Siamese-based models	Siamese + Logistic Regression	$0.883$	$0.629$	0.849	$0.791$	$0.820$
	Siamese + 2-Layer MLP	$0.834$	$0.628$	$0.811$	$0.809$	$0.810$
	Siamese + $k$ -NN	$0.868$	$0.717$	$0.717$	0.930	$0.824$
	Siamese Diversity Ensemble (M6)	0.905	$0.741$	$0.755$	0.930	$0.843$

Table 1: Epilepsy surgical prognosis on EPISURG (

N{=}268

, 5-fold stratified CV). BalAcc = (Sensitivity + Specificity) / 2. Best value per column in bold .

Figure 2: ROC curves for six baseline and proposed methods under 5-fold stratified CV on EPISURG ( $N{=}268$ ); Siamese Diversity Ensemble (M6) is omitted for visual clarity. AUC values are reported in Table 1 .

Figure 3: Aggregated confusion matrix for Neuro-Oracle (M5), 5-fold stratified CV. Rows: ground-truth proxy label ( $0{=}$ temporal resection, corresponding to Success / Engel I; $1{=}$ non-temporal or complex resection, corresponding to Failure / Engel II-IV); columns: predicted label. M5 achieves specificity of 0.921 and sensitivity of 0.566 at the default threshold.

ID	Method	$\tau$	AUC $\,\uparrow$	Sensitivity $\,\uparrow$	Specificity $\,\uparrow$	Balanced Acc. $\,\uparrow$
M3	Siamese + Logistic Regression	0.50	0.883	0.849	0.791	0.820
M4	Siamese + $k$ -NN	0.42	0.868	0.717	0.930	0.824
M5	Neuro-Oracle (ours)	0.34	0.867	0.698	0.916	0.807

Table 2: Threshold-optimised performance. Threshold selected via grid search over

[0.30,0.50]

(step 0.02), maximising balanced accuracy on aggregate held-out predictions. M3 and M3b retain

\tau=0.50

as optimal.

Model	Configuration	AUC $\uparrow$	Sensitivity $\uparrow$	Balanced Acc. $\uparrow$
Random encoder	Gaussian init., no training	$0.488$	$0.038$	$0.486$
Retrieval baseline	$k=1$ (greedy 1-NN, no vote smoothing)	$0.741$	$0.585$	$0.741$
Neuro-Oracle	w/o age filter ( $\delta\rightarrow\infty$ )	$0.867$	$0.566$	$0.744$
Neuro-Oracle M5	Full model ( $k=5$ , $\delta=15$ yr)	$0.867$	$0.566$	$0.744$
Siamese Diversity Ensemble (M6)	Ensemble of diverse models	0.905	0.755	0.843

Table 3: Ablation study of Neuro-Oracle pipeline components (

N=268

, 5-fold stratified cross-validation).

Figure 4: AUC-ROC across input modality and reasoning engine combinations. Single scan with $k$ -NN achieves 0.793; switching to longitudinal trajectory $\Delta v$ (0.868) or using a Llama-3 reasoning agent (0.867) individually increase performance. The full Neuro-Oracle pipeline (Trajectory $\Delta v$ + Llama-3) preserves the high performance of the geometric embedding and generates interpretative clinical reasoning. The chance level AUC=0.5 is shown as a dashed red line.

Figure 5: t-SNE projection [ 28 ] (perplexity = 30, 2000 iterations) of the 512-dim $\ell_{2}$ -normalised trajectory vectors for all 268 EPISURG subjects. Green circles show proxy label $y{=}0$ (temporal resection, corresponding to Success / Engel I in legend); Red circles show $y{=}1$ (non-temporal or complex resection, corresponding to Failure / Engel II-IV).

Metric	Value
Top-5 retrieval fidelity (%) $\uparrow$	83.0
Mean cosine similarity (top-5) $\uparrow$	0.919
Calibration MAE (label-weighted) $\downarrow$	0.189

Table 4: Embedding space and retrieval quality metrics (

N=268

subjects; 512-dimensional FAISS exact inner-product index).

Clinical Outcome Prediction

Clinical outcome prediction uses patient records, imaging, signals, and biomarkers to estimate risks, disease trajectories, treatment response, and survival.

See all 265 papers

May 21, 2026

CLSP-REQA: A Real-Time Quality-Aware Closed-Loop Seizure Prediction Framework with Mamba-BiLSTM and Confidence-Gated Intervention

Reliable seizure prediction is a prerequisite for closed-loop neurostimulation therapy, yet existing methods rarely account for the variability in EEG signal quality encountered in real-world deployment, and the overwhelming majority adopt non-strict evaluation protocols that overestimate generalisation performance. We propose CLSP-REQA (Closed-Loop Seizure Prediction with Real-time EEG Quality Assessment), a unified framework that embeds a lightweight signal quality estimator directly within the prediction pipeline. A Real-time EEG Quality Assessment (REQA) module runs in parallel with a Mamba-BiLSTM backbone, producing a scalar quality score q in [0,1] that modulates output confidence through a tiered non-linear fusion function (ECLO). Under strict cross-patient evaluation on the CHB-MIT Scalp EEG Database (n = 23 subjects, 198 seizures), CLSP-REQA achieves an AUC-ROC of 0.7426 +- 0.0199, outperforming the unadapted cross-patient baseline of 0.69 reported by Jemal et al., using only 16 EEG channels compared to 23 in prior work, and without requiring any target-patient data or domain adaptation. On the SIENA Scalp EEG Database (n = 14 subjects, 47 seizures), CLSP-REQA achieves AUC 0.7012 +- 0.0249, substantially surpassing the best domain-adapted cross-patient result of 0.61 on the same dataset, demonstrating strong cross-dataset generalisation. The framework outputs a structured four-tuple (p, q, c, Phi_SHAP) directly compatible with closed-loop neurostimulator interfaces.

Mufeng Chen, Qi Wu, Bingchao Huang +4

Department of Engineering Science, University of Oxford, Oxford OX1 3PJ, United Kingdom · Mathematical Institute, University of Oxford, Oxford OX2 6GG, United Kingdom · School of Computer Science and Engineering, Beihang University, Beijing 100191, China +4

Electroencephalography EEG Seizure Detection Clinical Prediction

Ask PDF

Sep 26, 2026

Hybrid Ensemble Learning for EEG-Based Epileptic Seizure Forecasting

Epileptic seizure forecasting aims to provide actionable warnings before seizure onset, yet patient-independent generalization and false-alarm control remain major challenges. We propose a calibrated hybrid ensemble for EEG-based seizure forecasting that combines five deep learning models and three classical machine learning models through a logistic regression stacking meta-learner. The proposed pipeline integrates signal preprocessing, handcrafted feature extraction, class-imbalance handling, probability calibration, and clinically motivated post-processing. We evaluate the framework on CHB-MIT using strict Leave-One-Patient-Out (LOPO) cross-validation, with threshold and post-processing parameters selected only on held-out meta data. On the filtered cohort, excluding patients with anomalous preictal rates below 1% or above 15%, the model achieves 74.2% seizure-level sensitivity at 1.24 false alarms per hour, with an average warning time of 16.9 minutes. A test-tuned oracle constrained to the target false-alarm budget achieves 60.9% sensitivity at 0.951 false alarms per hour, highlighting the importance of reporting sensitivity together with realized false-alarm rates. Our code is available at: https://github.com/DanaMason/IEEE-CARS-Hybrid-Ensemble-Learning-for-EEG-Based-Epileptic-Seizure-Forecasting

Mason Dana, Khandaker Mamun Ahmed

The Beacom College of Computer and Cyber Sciences, Dakota State University Madison, SD, USA

Ensemble Learning Electroencephalography

Ask PDF

May 10, 2026

EpiGraph: Building Generalists for Evidence-Intensive Epilepsy Reasoning in the Wild

Epilepsy diagnosis and treatment require evidence-intensive reasoning across heterogeneous clinical knowledge, including biosignal patterns, genetic mechanisms, pharmacogenomics, treatment strategies, and patient outcomes. In this work, we present EpiGraph, a large-scale epilepsy knowledge graph and benchmark for evaluating knowledge-augmented clinical reasoning. EpiGraph integrates 48,166 peer-reviewed papers and seven clinical resources into a heterogeneous graph containing 24,324 entities and 32,009 evidence-grounded triplets across five clinical layers. Built upon this graph, EpiBench defines five clinically motivated tasks spanning clinical decision-making, EEG report generation, pharmacogenomic precision medicine, treatment recommendation, and deep research planning. We evaluate six LLMs under both standard and Graph-RAG settings. Results show that integrating EpiGraph consistently improves performance across all tasks, with the largest gains observed in pharmacogenomic reasoning (+30--41%). Our findings demonstrate that structured epilepsy knowledge substantially enhances evidence-grounded clinical reasoning and provides a practical benchmark framework for evaluating knowledge-augmented LLMs in real-world neurological settings. Our code is available at: https://github.com/LabRAI/EEG-KG.

Yuyang Dai, Zheng Chen, Jathurshan Pradeepkumar +4

Florida State University · The University of Osaka · University of Illinois Urbana-Champaign

KG Construction Retrieval-Augmented Generation Clinical Decision-Making Clinical Reasoning Biomedical KGs

Ask PDF

Abstract

Clinical Outcome Prediction

Explore similar work

CLSP-REQA: A Real-Time Quality-Aware Closed-Loop Seizure Prediction Framework with Mamba-BiLSTM and Confidence-Gated Intervention

Hybrid Ensemble Learning for EEG-Based Epileptic Seizure Forecasting

EpiGraph: Building Generalists for Evidence-Intensive Epilepsy Reasoning in the Wild