Machine Learning-Based Directed Network Analysis
for Drug-Resistant Epilepsy Surgery
A Retrospective, Multicenter, Observational Study
across Four U.S. Level-4 Epilepsy Centers
Drug-resistant epilepsy (DRE)[20]Kwan P et al. Definition of drug resistant epilepsy: consensus proposal by the ILAE. Epilepsia. 2010;51(6):1069-1077. DOI affects ~20 million people worldwide[1,2]World Health Organization. Epilepsy Fact Sheet. Geneva: WHO; 2024. LinkKwan P, Brodie MJ. Early identification of refractory epilepsy. N Engl J Med. 2000;342(5):314-319. DOI. Surgical intervention is the only curative option, but success rates range from only 30-70% depending on etiology and localization accuracy[3,4]Wiebe S et al. A randomized, controlled trial of surgery for temporal-lobe epilepsy. N Engl J Med. 2001;345(5):311-318. DOITellez-Zenteno JF et al. Long-term seizure outcomes following epilepsy surgery: a systematic review and meta-analysis. Brain. 2005;128(5):1188-1198. DOI. Failures carry continued seizure burden, elevated SUDEP risk[6]Harden C et al. Practice guideline summary: SUDEP incidence rates and risk factors. Neurology. 2017;88(17):1674-1680. DOI, and multi-million-dollar societal costs per patient[7,8,9]Langfitt JT et al. Health care costs decline after successful epilepsy surgery. Neurology. 2007;68(16):1290-1298. DOIBegley CE, Durgin TL. The direct cost of epilepsy in the United States. Epilepsia. 2015;56(9):1376-1387. DOIChoi H et al. Epilepsy surgery for pharmacoresistant temporal lobe epilepsy: a decision analysis. JAMA. 2008;300(21):2497-2505. DOI.
Contemporary surgical decision-making increasingly views epilepsy as a network disorder[10,21]Kramer MA, Cash SS. Epilepsy as a disorder of cortical network organization. Neuroscientist. 2012;18(4):360-372. DOISpencer SS. Neural networks in human epilepsy: evidence of and implications for treatment. Epilepsia. 2002;43(3):219-227. DOI, and several computational approaches have been developed to augment the interpretation of iEEG data:
High-frequency oscillations (HFOs, 80-500 Hz)[16,17]Jacobs J et al. High-frequency oscillations (HFOs) in clinical epilepsy. Prog Neurobiol. 2012;98(3):302-315. DOIFrauscher B et al. High-frequency oscillations: the state of clinical research. Epilepsia. 2017;58(8):1316-1329. DOI - the most extensively validated iEEG biomarker for epileptogenic zone localization. HFO resection rates correlate with seizure outcomes.
Functional connectivity metrics - eigenvector centrality, degree centrality[22]Bullmore E, Sporns O. Complex brain networks: graph theoretical analysis of structural and functional systems. Nat Rev Neurosci. 2009;10(3):186-198. DOI, and phase-locking values[23]Lachaux JP et al. Measuring phase synchrony in brain signals. Hum Brain Mapp. 1999;8(4):194-208. DOI identify highly connected nodes but cannot infer directionality.
Granger causality[18]Blinowska KJ et al. Granger causality and information flow in multivariate processes. Phys Rev E. 2004;70(5):050902. DOI, partial directed coherence[24]Baccala LA, Sameshima K. Partial directed coherence: a new concept in neural structure determination. Biol Cybern. 2001;84(6):463-474. DOI, and information-theoretic approaches[19]Wilke C et al. Graph analysis of epileptogenic networks in human partial epilepsy. Epilepsia. 2011;52(1):84-93. DOI have been used to infer causal relationships in seizure networks.
EZTrack[13]U.S. FDA. 510(k) Premarket Notification K201910: EZTrack. Silver Spring, MD: FDA; 2021. FDA is the only FDA-cleared computational tool that analyzes iEEG network dynamics. It computes undirected eigenvector centrality from broadband iEEG correlation matrices, identifying highly connected network hubs without inferring causal directionality.
CN-Suite is a Software as a Medical Device (SaMD) that computes quantitative "criticality scores" for each brain region sampled by intracranial electrodes. Instead of relying on visual pattern recognition, it quantifies directed information flow between neural signals to distinguish seizure "drivers" from passive "responders."
Delay-adjusted wavelet-based transfer entropy (dWTE)[11,25]Schreiber T. Measuring information transfer. Phys Rev Lett. 2000;85(2):461-464. DOIGourevitch B, Eggermont JJ. Evaluating information transfer between auditory cortical neurons. J Neurophysiol. 2007;97(3):2533-2543. DOI - a nonlinear, information-theoretic measure that captures directional causality between neural signals across time-frequency scales. Produces a pairwise directed connectivity matrix for each seizure epoch.
XGBoost classifier[12]Chen T, Guestrin C. XGBoost: A scalable tree-boosting system. ACM SIGKDD. 2016:785-794. DOI - ensemble machine learning that maps connectivity features to a criticality score (0-1) per contact per seizure. Scores are rescaled to 0-10 for clinical use; above the threshold (1 on the clinical scale), a contact is classified as a causal "driver." Model weights were locked before validation.
Retrospective, multicenter, observational, single-arm performance study using sEEG recordings[26]Isnard J et al. French guidelines on stereoelectroencephalography (SEEG). Neurophysiol Clin. 2018;48(1):5-13. DOI. The algorithm was trained on an independent cohort (N=37) from 3 centers and hash-locked before validation on 60 patients from 4 different centers. Zero overlap between training and validation data.
| Role | Centers | N |
|---|---|---|
| Training | BCH, NIH, UMMC | 37 |
| Validation | HUP, JHH, UMF, TCH | 60 |
Age ≥ 3 years - Focal/multifocal DRE[20]Kwan P et al. Definition of drug resistant epilepsy: ILAE consensus proposal. Epilepsia. 2010. DOI - Curative-intent resection or ablation - ≥ 3 stereotyped seizures captured during iEEG - ≥ 12-month follow-up with Engel classification[5]Engel J Jr. Outcome with respect to epileptic seizures. In: Surgical Treatment of the Epilepsies. 2nd ed. Raven Press; 1993:609-621. - Detailed operative notes mapping the surgical zone
Standardized effect size (Cohen's d)[28]Cohen J. Statistical Power Analysis for the Behavioral Sciences. 2nd ed. Lawrence Erlbaum Associates; 1988. DOI of the patient-level interpretability ratio between favorable (Engel I-II)[5]Engel J Jr. Outcome with respect to epileptic seizures. In: Surgical Treatment of the Epilepsies. 2nd ed. Raven Press; 1993:609-621. and unfavorable (Engel III-IV) outcome groups.
75 records screened; 60 met all eligibility criteria. 15 excluded (5 by criteria, 10 by incomplete data). All 60 were successfully processed - zero processing failures.
| Site | Favorable | Unfavorable | Total | % |
|---|---|---|---|---|
| HUP (Penn) | 34 | 13 | 47 | 78.3% |
| TCH (Texas Children's) | 7 | 2 | 9 | 15.0% |
| JHH (Johns Hopkins) | 0 | 3 | 3 | 5.0% |
| UMF (Miami) | 1 | 0 | 1 | 1.7% |
The interpretability ratio asks: does the algorithm assign higher criticality to tissue the surgeon actually treated? Favorable-outcome patients showed a mean ratio of 4.64 vs. 1.83 for unfavorable, confirming the algorithm reliably distinguishes driver tissue in successful surgeries. Ratios were winsorized[29]Dixon WJ, Tukey JW. Approximate behavior of the distribution of Winsorized t. Technometrics. 1968;10(1):83-98. DOI at p95 and bootstrapped[30]Efron B, Tibshirani RJ. An Introduction to the Bootstrap. Chapman & Hall/CRC; 1993. DOI with 10,000 resamples.
| Subgroup | Fav. Mean | Unfav. Mean | Boot. d | 95% CI | p |
|---|---|---|---|---|---|
| All Subjects | 4.64 | 1.83 | 0.74 | 0.39, 1.06 | 0.003 |
| Adults | 5.05 | 2.00 | 0.73 | 0.37, 1.08 | 0.006 |
| Pediatrics | 2.58 | 0.54 | 1.87 | 1.17, 3.24 | < 0.0001 |
This figure sweeps the criticality threshold from 0 to 1, showing how patient-level macro-averaged sensitivity, specificity, and Youden's J[31]Youden WJ. Index for rating diagnostic tests. Cancer. 1950;3(1):32-35. DOI change. The pre-locked operating point at 0.15 (red marker) was derived from the independent training cohort by maximizing Youden's J over seizure-free patients.
Note that this threshold was not optimized on the validation set - it was fixed prior to any validation analysis. The validation-set optimum may differ, but using a training-derived threshold prevents overfitting and ensures generalizability.
686 treated and 3,277 untreated contacts were evaluated in the Engel I/II cohort. Three estimation approaches address different aspects of the clustered data structure, using GEE[32]Liang KY, Zeger SL. Longitudinal data analysis using generalized linear models. Biometrika. 1986;73(1):13-22. DOI:
| Group | Sens (GEE) | Spec (GEE) |
|---|---|---|
| All (N=42) | 64.3% | 84.3% |
| Adults (n=35) | 59.6% | 86.4% |
| Pediatrics (n=7) | 91.1% | 76.0% |
| Group | PPV (GEE) | NPV (GEE) |
|---|---|---|
| All (N=42) | 49.0% | 88.0% |
| Adults (n=35) | 56.4% | 85.6% |
| Pediatrics (n=7) | 10.8% | 99.35% |
The strongest finding: per-patient sensitivity is inversely correlated with surgery size (ρ = -0.62, p < 0.0001). The algorithm is most informative for small focal procedures - the clinical scenario most relevant to minimally invasive ablation (LITT)[14,34]Youngerman BE et al. Long-term outcomes of mesial temporal LITT for drug-resistant epilepsy. J Neurol Neurosurg Psychiatry. 2023;94(11):879-886. DOIChen J et al. MR-guided LITT for drug-resistant epilepsy: a systematic review and IPD meta-analysis. Epilepsia. 2023;64(8):1957-1974. DOI.
For each patient (n=34 with coordinates), we computed the mean nearest-neighbor distance among high-critical contacts and compared it to a null distribution of 2,000 random permutations of electrode labels. Points below the diagonal indicate spatial clustering beyond chance. The ~9 mm mean spacing is compatible with a single LITT ablation trajectory[14,34]Youngerman BE et al. Long-term outcomes of mesial temporal LITT. J Neurol Neurosurg Psychiatry. 2023. DOIChen J et al. MR-guided LITT for drug-resistant epilepsy: IPD meta-analysis. Epilepsia. 2023. DOI.
Beyond the binary threshold, do continuous scores carry rank information? Yes - among the 946 high-critical contacts (≥ 0.15, Engel I/II, n=42), the proportion inside the surgical zone increases monotonically with criticality score (Spearman ρ = 0.28, p = 6.8 × 10⁻¹⁹).
Surgical teams can triage flagged contacts by predicted clinical relevance rather than treating all above-threshold contacts as equivalent.
Each patient is plotted along two axes: the delta-mean (mean criticality inside the surgical zone minus outside, x-axis) and the fraction of high-critical contacts left outside the surgical zone (y-axis). Favorable patients (blue) cluster lower-right; unfavorable (red) cluster upper-left. Median delta-mean: 0.24 favorable vs. 0.05 unfavorable (Mann-Whitney p = 0.002).
SHAP analysis (19,257 observations; 4,235 high-critical) reveals a three-tier feature hierarchy. The top four features are all network-topology measures (|SHAP| 0.27-0.39): Δ betweenness centrality (BC), Δ eigenvector centrality, Δ clustering, and time-averaged BC. Waveform features first appear at rank 5 - after a gap twice the within-tier step.
A critical contact becomes a routing bridge (betweenness ↑), connects to influential hubs (eigenvector ↑), and anchors a cohesive subnetwork (clustering ↑). Although the raw measures are negatively correlated (ρ = -0.49), their SHAP contributions are independent (ρ = 0.06). The model recognizes each as a separate predictor of criticality.
Using directed connectivity and lag estimates from the dWTE network, we tested whether higher criticality scores correspond to earlier position in the inferred seizure propagation sequence. The relationship was phase-dependent: pre-ictally, high-critical contacts were not earlier in flow order; around and after seizure onset, the direction reversed, with high-critical contacts arriving a median 10 ms earlier than low-critical contacts post-ictally (vs. 16 ms, p = 2.9 × 10-41).
What distinguishes a driver contact is therefore not that it is chronically upstream, but that it becomes upstream at seizure onset - consistent with the SHAP finding that criticality reflects seizure-induced network reorganization rather than static electrophysiological properties.
Retrospective design - validates correlation with outcomes but cannot measure real-time impact on surgical decision-making. Prospective integration into Epilepsy Surgery Conferences is the logical next step. Data collected from 2015-2024; temporal confounds mitigated by the algorithm's reliance on fundamental time-frequency neural dynamics.
Site imbalance - HUP contributed 78% of subjects. JHH contributed only 3, all unfavorable. An excluding-HUP analysis (n=13) yielded d = 0.68 (CI: 0.12-1.24), directionally consistent but underpowered.
Missing demographic data - Race, ethnicity, and sex unavailable for some patients due to IRB de-identification. No known biological mechanism links demographics to iEEG network connectivity.
Pediatric sample - Only 7 patients from a single center (TCH). The 99.35% NPV and 91.1% sensitivity are promising but preliminary, and require multicenter replication.
No head-to-head comparisons - This study evaluated standalone CN-Suite performance without direct comparison to HFO detectors[16,17]Jacobs J et al. High-frequency oscillations (HFOs) in clinical epilepsy. Prog Neurobiol. 2012;98(3):302-315. DOIFrauscher B et al. High-frequency oscillations: the state of clinical research. Epilepsia. 2017;58(8):1316-1329. DOI, directed connectivity methods[18,19]Blinowska KJ et al. Granger causality and information flow in multivariate processes. Phys Rev E. 2004;70(5):050902. DOIWilke C et al. Graph analysis of epileptogenic networks in human partial epilepsy. Epilepsia. 2011;52(1):84-93. DOI, or EZTrack[13]U.S. FDA. 510(k) K201910: EZTrack. 2021. FDA.
U.S. Level-4 centers only - Applicability to lower-volume centers, non-U.S. settings, subdural grids, or non-resective interventions (neuromodulation, RNS, DBS) requires separate validation.
FIND Neuro - FDA 510(k) Submission - Multicenter Clinical Validation