Authors :
Presenting Author: Francois Okoroafor, MD – Institue of Child Health, University College London
Ghassan Makhoul, BS – Vanderbilt University Medical Center
Anas Reda, MS – Vanderbilt University Medical Center
Bruno Hidalgo, BS – Vanderbilt University Medical Center
Emily Liao, BS – Vanderbilt University Medical Center
Kate Wang, BS – Vanderbilt University Medical Center
Addison Cavender, BS – Vanderbilt University Medical Center
Victoria Morgan, PhD – Vanderbilt University Medical Center
Dario Englot, MD PhD – Vanderbilt University Medical Center
Martin Tisdall, MD PhD – UCL
Rory Piper, MD PhD – UCL
Aswin Chari, MD PhD – UCL
torsten Baldeweg, PhD – UCL
Rationale:
Accurate presurgical biomarkers that predict 12‑month seizure outcomes after responsive neurostimulation (RNS) or focal resection could improve patient selection and counselling for drug‑resistant epilepsy. We evaluated whether neural‑avalanche–derived transition matrices (Fig. 1), analyzed with Random Forest, Decision Tree and a matrix‑adapted LeNet‑5, predict one‑year outcomes and generalize across therapy types.
Methods:
Retrospective resting‑state SEEG (20 min, bipolar montage) from RNS (n=24) and resection (n=39) cohorts were preprocessed (0.5–119 Hz band‑pass, 60 Hz notch), thresholded to detect neural avalanches and converted into per‑patient avalanche transition matrices. Models were trained and optimized via normalization, seed‑grid search and hyperparameter optimization. Our performance metrics included accuracy, sensitivity, specificity and ROC AUC. Engel class stratification exploratory analyses were performed to determine their influence on model performance. External validation of the trained RNS models allowed comparison of the RNS‑trained models applied to the resection dataset with the resection‑trained models on the same dataset.
Results:
For the RNS cohort the top model was LeNet‑5 (seed‑grid) achieving accuracy 91.7%, sensitivity 87.5%, specificity 94.4% and ROC AUC 0.75; Random Forest reached 75% accuracy, sensitivity 68.8%, specificity 72.2% and ROC AUC 0.5.
Engel group stratification analysis demonstrated heterogeneous behavior: e.g., "1a vs rest" showed RF accuracy 55.0% with sensitivity ~22% and specificity ~82%; broader groupings such as "1+2 vs rest" and "1+2+3 vs 4" yielded high accuracy driven by sensitivity ≈1.0 and specificity ≈0.0 for several models, indicating class‑imbalance or tendency to predict only the responder class. “1a+1b vs rest” grouping was used for downstream resection analysis.
In the resection cohort, the best single-model performances were modest: Random Forest (seed‑grid) accuracy 65.0% and LeNet‑5 peaked at 60.0% with high sensitivity but low specificity; many resection ROC AUC metrics were not computable in the supplied output.
The external validation analysis confirmed that the top RNS‑trained model demonstrated better accuracy, specificity and AUC ROC scores, when tested on resection data, compared to resection-trained models tested on resection data (Fig. 2).
Conclusions:
Neural‑avalanche transition matrices combined with machine learning models can achieve strong within‑cohort predictive performance for RNS outcomes, but cross‑modality generalizability is limited. Further work should prioritize larger, balanced multi‑center datasets, explicit calibration to improve specificity, and prospective external validation before clinical translation.
References
< ![if !supportLists] >1. < ![endif] >Sancetta BM, Matarrese MAG, Ricci L, et al. Altered neural avalanche spreading in people with drug-resistant epilepsy. Neuroimage. 2025;311:121188. doi:10.1016/j.neuroimage.2025.121188
< ![if !supportLists] >2. < ![endif] >Ghatan S. Pediatric Neurostimulation and Practice Evolution. Neurosurg Clin N Am. 2024;35(1):1-15. doi:10.1016/j.nec.2023.09.006
Funding: University College London (UCL) 2025 Bogue Fellowship