Abstracts

Rationale:
New neuroimaging methods are required to improve outcomes for drug-resistant epilepsy (DRE) patients. We are evaluating the efficacy of normative modelling (NM) at different sample sizes applied to the epilepsy population, using hippocampal segmentation and reconstruction.

Methods:
Hippocampal thickness was extracted from T1w Magnetic Resonance Imaging (MRI) with Hippunfold which segments and reconstructs hippocampal surfaces (DeKraker et al., 2022). Bayesian linear models were trained with the PCNtoolkit Python package on control subjects to predict normal hippocampal thickness at each vertex based on inputs of age, sex, and acquisition site (Rutherford et al., 2022). Z-scores were computed to compare ground truth and predicted hippocampal thickness. For this study, we used multiple datasets, including T1w 3T MRIs from 581 subjects sourced from the HCP-Aging dataset (0.8mm³), 641 subjects from the Cam-Can dataset (1mm³), and T1w 7T MRIs from 62 subjects (29 controls, 33 DRE) obtained from an in-house dataset (0.8mm³) (Bookheimer et al., 2019; Shafto et al., 2014; Taha et al., 2022). To study sample size, a holdout set of 251 subjects was taken for testing and sampling between 25-1000 subjects for training. Applying NM in epilepsy, the model trained on 1000 subjects from our sample size analysis was applied to the 33 DRE subjects.

Results:
As we increased the number of subjects in our training set, we observed a decrease in bias as evidenced by decreasing error and, with Z-scores and residuals appearing to converge for models trained on greater than 100 subjects. We used our 1000-subject model to predict thickness and to compute Z-scores on a number of known DRE subjects. Currently, we have epilepsy localization information for a subset of eight subjects. For these eight subjects, z-score values appear to range between zero and two regardless of epilepsy localization. A cluster of Z-score values greater than three standard deviations is apparent in the left head between CA2-4 for one of the left temporal DRE patients.

Conclusions:
This analysis shows differences between smaller and larger training set size models and possibly relevant differences between epilepsy and control subjects for thickness z-scores. Further research requires larger datasets of epilepsy subjects with more detailed clinical information to enable an understanding of the possible clinical utility of normative modeling in diseases like epilepsy. These findings can help to inform future applications of normative modeling to epilepsy and guide the development of personalized treatment strategies for individuals with epilepsy based on their unique neuroanatomical profiles.

Funding: NSERC CRSNG, Digital Research Alliance of Canada, Western University BrainsCAN.