Decomposing MRI phenotypic heterogeneity in epilepsy: a step towards personalized classification

Abstract

In drug-resistant temporal lobe epilepsy, precise predictions of drug response, surgical outcome and cognitive dysfunction at an individual level remain challenging. A possible explanation may lie in the dominant 'one-size-fits-all' group-level analytical approaches that does not allow parsing interindividual variations along the disease spectrum. Conversely, analysing inter-patient heterogeneity is increasingly recognized as a step towards person-centred care. Here, we used unsupervised machine learning to estimate latent relations (or disease factors) from 3 T multimodal MRI features [cortical thickness, hippocampal volume, fluid-attenuated inversion recovery (FLAIR), T1/FLAIR, diffusion parameters] representing whole-brain patterns of structural pathology in 82 patients with temporal lobe epilepsy. We assessed the specificity of our approach against age- and sex-matched healthy individuals and a cohort of frontal lobe epilepsy patients with histologically verified focal cortical dysplasia. We identified four latent disease factors variably co-expressed within each patient and characterized by ipsilateral hippocampal microstructural alterations, loss of myelin and atrophy (Factor 1), bilateral paralimbic and hippocampal gliosis (Factor 2), bilateral neocortical atrophy (Factor 3) and bilateral white matter microstructural alterations (Factor 4). Bootstrap analysis and parameter variations supported high stability and robustness of these factors. Moreover, they were not expressed in healthy controls and only negligibly in disease controls, supporting specificity. Supervised classifiers trained on latent disease factors could predict patient-specific drug response in 76 ± 3% and postsurgical seizure outcome in 88 ± 2%, outperforming classifiers that did not operate on latent factor information. Latent factor models predicted inter-patient variability in cognitive dysfunction (verbal IQ: r = 0.40 ± 0.03; memory: r = 0.35 ± 0.03; sequential motor tapping: r = 0.36 ± 0.04), again outperforming baseline learners. Data-driven analysis of disease factors provides a novel appraisal of the continuum of interindividual variability, which is probably determined by multiple interacting pathological processes. Incorporating interindividual variability is likely to improve clinical prognostics.

Keywords: MRI; machine learning; phenotypic heterogeneity; precision medicine; temporal lobe epilepsy.

Figure 1

Surface-based feature extraction. To model prevalent features of TLE pathology (atrophy, gliosis, demyelination and microstructural damage), we carried out surface-based sampling of morphological (MOR; cortical thickness, hippocampal volume) and intensity features (FLAIR, T₁-weighted/FLAIR), as well as diffusion-derived fractional anisotropy (FA) and mean diffusivity (MD), across grey and white matter (GM, WM) and hippocampal surface points (or ‘vertices’). T1w = T₁-weighted.

Figure 2

Analysis of latent disease factors. (A) Schematic representation of the LDA. LDA uncovers latent relations from MRI data representing distinct patterns of alterations (or disease factors) and quantifies their coexpression within each patient. Factors are extracted across grey matter (GM) and white matter (WM) surface points (or vertices) and are expressed as posterior probability P(Vertex | Factor). Here, Factor 1 (blue), 2 (red) and 3 (green) are localized to the somatomotor, temporal lobe and perisylvian grey matter, respectively; Factor 4 (yellow) is localized in the white matter of the posterior quadrant. Factors may be partially overlapping (3 with 1 and 2) or non-overlapping (1 and 2). This schematic example illustrates only one pathological process. By allowing patients to express varying degrees of disease factors [P(Factor | Patient)], instead of assigning patients to a single disease subtype, LDA captures interindividual variability. (B) Factor modelling. For each modality, features were z-normalized with respect to the analogous vertices of healthy controls ipsi- and contralateral to the seizure focus. Scores were transformed into counts, multiplied by 10/−10 and rounded to the nearest integer such that the larger counts indicated more severe pathology. We then applied Latent Dirichlet Allocation to uncover latent relations (namely, disease factors) from these features [expressed as posterior probability P(Vertex|Factor), or disease load] and quantify their coexpression (namely factor composition) within each patient [P(Factor|Patient)].

Figure 3

Mapping whole-brain latent disease factors. Each latent factor is modelled as a weighted combination of neocortical atrophy, hippocampal atrophy, FLAIR hyperintensity, T₁-weighted/FLAIR decrease, FA decrease and MD increase across the grey and white matter (GM, WM), as well as hippocampus. Disease factors are expressed as posterior probabilities [P(Vertex|Factor)] across ipsi/contralateral MRI vertices; higher probability (brighter colour) signifies greater contribution of a given feature to the factor or disease load (P_FDR < 0.05). For each feature, inset bar graphs indicate mean and standard deviation of the disease load in ipsi/contralateral cortical grey matter, white matter and hippocampus

Figure 4

Factor composition and specificity. (A) Factor composition in TLE. In the tetrahedron, each patient is a dot and its barycentric coordinate the factor composition expressed as posterior probability [P(Factor|Patient)]. Patients located close to the corners predominantly express a given factor (F), whereas those located towards the centroid express various combinations of all factors. The scale represents the kernel density, with yellow/blue indicating similar/dissimilar composition among patients. (B) Specificity of factors. The bar graphs compare the severity of factor expression [z-score weighted by factor maps P(Vertex| Factor)] in TLE, healthy controls (HC) and disease controls composed of patients with FLE. The error bars indicate standard deviation. The matrices show subject-wise severity of each factor.

Figure 5

Individualized predictions. Drug response (A), seizure outcome (B), verbal IQ (C), memory index (D) and motor index (E) are more accurately predicted when using latent disease factors than when relying on conventional group-level features (P_FDR < 0.001). Data-points indicate mean balanced accuracy for categorical data (drug response, seizure outcome) and Pearson correlation coefficients for numerical data (cognitive scores) evaluated based on 100 repetitions of 10-fold cross-validation. Inset bar graphs represent the magnitude and direction of contribution from each selected feature. Note that the magnitude of bars adds to one and thus reflect relative feature importance. Upwards bars indicate contribution towards drug-control, Engel I outcome or normal cognition, whereas downwards bars indicate contribution towards drug-resistance, Engel II–IV and impaired cognition.