Abstracts

Rationale: The Rey Complex Figure Copy Task (RCFT) is the most widely used neuropsychological test of visual construction and memory. Conventional RCFT administration is poorly diagnostic of non-dominant temporal lobe dysfunction in patients with epilepsy (PWE), potentially due to limitations in the scoring rubric and subjectivity in scoring practice. Convolutional Neural Networks (CNNs) are a class of deep learning algorithms applied to visual recognition which can discover ambiguous features in patient drawings, without the variability inherent in human scoring. We developed a novel self-supervised CNN using a retrospective dataset of RCFT drawings to distinguish between focal epilepsy diagnoses.

Methods: RCFT copy and 30-min delay drawings produced by 197 PWE (66 LTLE, 71 RTLE, 60 ETLE) and 45 healthy controls (HC) were analyzed. Epilepsy localization was determined by a board-certified epileptologist (AL), based on seizure semiology, MRI, and EEG. A synthetic training dataset was generated by combining all 18 RCFT scored features (218 = 262,144 images), and augmented through random rotation, perspective shifts, and line thinning and thickening. This synthetic dataset was used to pre-train a CNN, which was then applied to each drawing to generate a 512-dimensional embedding. A principal component analysis (PCA) was performed on the resulting embeddings, then clustered by K-means clustering. We selected the clustering method which maximally differentiated between seizure diagnoses. We calculated descriptive, ANOVA, and Chi-squared statistics on the resulting clusters.

Results: Subjects were 53% F with a mean age of 38.2 years (SD 14.3) and 14.4 years (2.5) of education. Mean IQ was 93.7 (15.2) and mean RVLT score was 14.0 (1.5). The optimal PCA and K-means cluster analysis generated 3 clusters based on the CNN embeddings among copy and delay drawings. The CNN-generated clusters for delay drawings were more heterogeneous than the copy clusters (Figure 1). Among the delay clusters, there was no difference in age, education, or IQ. However, the distribution of diagnoses differed between the 3 clusters of delay drawings (X2= 29.844, p < 0.0001, Figure 1, Table 1). HCs more likely belonged to Cluster A (p < 0.0001), ETLE patients more likely belonged to Cluster B (p < 0.0001), and RTLE patients more likely belonged to Cluster C (p = 0.001). There was a difference in assigned RCFT scores between clusters (χ2(2) = 42.096, p < 0.0001), with Cluster A (19.1 ± 7.7) scoring higher than Cluster B (10.6 ± 5.3) ( p < 0.0001), or Cluster C (10.5 ± 6.2) (p < 0.0001).