Identification of four biotypes in temporal lobe epilepsy via machine learning on brain images

Artificial intelligence provides an opportunity to try to redefine disease subtypes based on similar pathobiology. Using a machine-learning algorithm (Subtype and Stage Inference) with cross-sectional MRI from 296 individuals with focal epilepsy originating from the temporal lobe (TLE) and 91 healthy controls, we show phenotypic heterogeneity in the pathophysiological progression of TLE. This study was registered in the Chinese Clinical Trials Registry (number: ChiCTR2200062562). We identify two hippocampus-predominant phenotypes, characterized by atrophy beginning in the left or right hippocampus; a third cortex-predominant phenotype, characterized by hippocampus atrophy after the neocortex; and a fourth phenotype without atrophy but amygdala enlargement. These four subtypes are replicated in the independent validation cohort (109 individuals). These subtypes show differences in neuroanatomical signature, disease progression and epilepsy characteristics. Five-year follow-up observations of these individuals reveal differential seizure outcomes among subtypes, indicating that specific subtypes may benefit from temporal surgery or pharmacological treatment. These findings suggest a diverse pathobiological basis underlying focal epilepsy that potentially yields to stratification and prognostication – a necessary step for precise medicine.

To examine whether the prediction performance is significantly better than random predictions, we used a permutation test to evaluate significance by random permutation of predictive label.Specifically, the label (seizure-free or not) of subject were randomly permuted across all subjects.Subsequently, we re-conducted machine learning procedures (Supplementary Figure 10), and predicted the label of each subject on test set using leave-one subject-out cross-validation (LOOCV).Prediction performance was evaluated using Youden Index.The above random permutation was repeated 1000 times, yielding a distribution of random prediction performance.The significance of true prediction performance was estimated by its location within the distribution of random prediction performance.One-side P<0.05 (i.e., better than 95% of random predictions) was considered as a significance that rejects the null hypothesis.
We follow up brain MRI data of the part of individuals with TLE (n=23, without surgery until follow-up).The average of interval time between the baseline scanning and follow-up scanning is 39.0 months (SD=16.8months), range from 10.5 to 76.7 months.Using this subsample, we re-estimated the SuStaIn subtype labels of these individuals using their follow-up MRI data.We examined whether the subtype label at follow-up keeps consistent with the baseline label.We found that subtype labels remained consistent for almost all patients at baseline and follow-up (Supplementary Figure 11), except for two patients with subtype 4 (i.e., stage=0) who converted to subtype 1 and 3 at follow-up, respectively.This result suggests that once certain initial brain injury is established, it is less likely to shift from one trajectory pattern (i.e., subtype) to another.This assumption is also supported by Supplementary Figure 9, which shows that the probability of maximum likelihood subtype is high across all SuStaIn stage, indicating that there was no "cross-over events" in the subtype sequence.
We divided all of individuals with TLE (n=296) into two disease subgroup according to their disease durations (cutoff = median value (i.e., 9.5 years)), yielding a short-term subgroup (n=148, mean disease duration=4.8±2.7 years) and a long-term subgroup (n=148, mean disease duration=17.5±7.4 years).Such a subgrouping rule take into account the same size of subsamples.We re-estimated the 'spatiotemporal patterns of brain atrophy' (i.e., SuStaIn trajectory) in each subgroup, separately.We found there was a similar pattern of the three trajectories (left hippocampus-led, right hippocampus-led and cortex-led) in the two disease subgroups (Supplementary Figure 12).This result suggests that the distinct spatiotemporal patterns of brain atrophy may not be affected by disease progress.
In our data, we examined the associations between HS and other clinical variables (age of onset, illness duration, medication outcome and surgery outcome) using a linear regress or logistic model analysis.We observed that patients with HS+ show younger age of onset compared to those p=0.001).In addition, we found that patients with HS-experience worse surgical outcomes compared to those HS+ patients (χ²=5.99,p=0.014) (see Supplementary Table 9).To examine whether the clinical differences among SuStaIn subtypes are affected by HS, we re-analyzed the correlations between clinical features and subtype with HS effect as a covariate using a linear regress or logistic regress model (Y~X+C+ε).Here, clinical variable is dependent variable (Y); SuStaIn subtype is independent variable (X); and HS is covariate variable (C).After controlling HS effect, we still found significant associations of SuStaIn subtype with age of onset, illness duration and medication outcomes (see Supplementary Table 10).This suggests that the clinical differences between SuStaIn subtypes remain significant after controlling HS effect.

Supplementary Table 1. ROI-wise correlation coefficients between SuStaIn stages and regional z scores.
**p<0.001, *p<0.05,two-sided.Spearman correlation test is conducted for data analysis.Multiple comparisons were corrected by FDR.Supplementary Table2.Regional z score for each subtype.