Distinct pathophysiological progressions of brain atrophyDistinct patterns of spatiotemporal development of brain atrophy have been recognized utilizing SuStaIn, primarily based on cross-sectional MRI information from 296 people with TLE. Three distinct trajectories of atrophy, labeled as ‘trajectory’ 1, ‘trajectory’ 2, and ‘trajectory’ 3, had been noticed (Fig. 1a-c). In ‘trajectory’ 1, the preliminary regional quantity loss was noticed in the left hippocampus, adopted by the left thalamus, after which prolonged to the appropriate thalamus and eventually to the left entorhinal cortex and cerebral cortex (Fig. 1a). Conversely, in ‘trajectory’ 2, quantity loss started in the appropriate hippocampus, adopted by the appropriate thalamus, after which unfold to the left thalamus and left hippocampus earlier than affecting the cerebral cortex (Fig. 1b). Lastly, ‘trajectory’ 3 displayed a cortical-predominant phenotype. It was characterised by preliminary discount in the cortex, particularly involving the bilateral center and superior frontal lobes. Subsequent cortical atrophy was extra extreme and expanded to different lobes, together with the bilateral parietal, occipital, and temporal lobes. Finally, the subcortical areas, the hippocampus and thalamus, had been affected (Fig. 1c). The noticed variations in the atrophy trajectories throughout particular brain areas point out potential phenotypic heterogeneity in the pathophysiological progressions of TLE. We additionally estimated spatiotemporal trajectories of brain atrophy in a short-term subsample (n = 148, imply illness period = 4.8 ± 2.7 years) and a long-term subsample (n = 148, imply illness period=17.5 ± 7.4 years), individually (Supplementary Materials). There was an analogous sample of the three trajectories in the 2 illness subsample (Supplementary Materials). This means that the distinct spatiotemporal patterns of brain atrophy might not be affected by illness progress.Fig. 1: Spatiotemporal patterns of development of brain atrophy via SuStaIn.Trajectory exhibits that cortical thickness or quantity loss is firstly noticed in the left hippocampus (a), the appropriate hippocampus (b) and cortex (c) in folks with temporal lobe epilepsy relative to wholesome controls. The shade of brain area reveals the severity of gray matter loss; white: unaffected areas (z < 1); light blue: mildly affected areas (z = 1–2); dark blue: severely affected areas (z > 2). d Individual subtyping based on the utmost chance of belonging to which ‘trajectory’ (crimson, left hippocampus-predominant trajectory; blue, right-hippocampus-predominant trajectory; inexperienced, cortex-predominant trajectory). e–g Correlation between SuStaIn levels and z scores (i.e., the diploma of thickness/quantity lower in sufferers relative to wholesome inhabitants) of common cortical thickness, the quantity of left and proper hippocampus individually in every subgroup (crimson, left hippocampus-predominant trajectory; blue, right-hippocampus-predominant trajectory; inexperienced, cortex-predominant trajectory). Spearman correlation check is performed for information evaluation in figures e-g. It exhibits a big correlation between SuStaIn levels and common cortical thickness (trajectory 1: r = 0.599, p = 1.4 × 10-9; trajectory 2: r = 0.791, p = 1.8 × 10-25; trajectory 3: r = 0.847, p = 3.0 × 10-12), in addition to the quantity of the left hippocampus (trajectory 1: r = 0.627, p = 1.3 × 10-10; trajectory 2: r = 0.577, p = 2.3 × 10-11; trajectory 3: r = 0.431, p = 0.005). The important correlation between SuStaIn levels and proper hippocampus quantity was solely discovered in the ‘trajectory’ 3 (trajectory 1: r = 0.269, p = 0.013; trajectory 2: r = 0.157, p = 0.097; trajectory 3: r = -0.006, p = 0.973). The error bands in figures (e, f, and g) symbolize 95% confidence interval. n = 85, 113, and 41 biologically impartial samples in left hippocampus-predominant trajectory, right-hippocampus-predominant trajectory, and cortex-predominant trajectory. **p < 0.001, *p < 0.05, two-sided. Multiple comparisons had been corrected by FDR.Subtype-specific neuroanatomical signaturesThe SuStaIn method calculated the chance of every affected person belonging to a particular ‘trajectory’ (Fig. 1d) and additional assigned them to a sub-stage inside that trajectory. It is vital to notice that sufferers who didn't exhibit apparent reductions in any ROI had been assigned a ‘stage=0’ by SuStaIn, indicating a ‘regular’ neuroanatomical signature. The SuStaIn levels confirmed correlations with z scores, which symbolize the diploma of thickness/quantity lower in sufferers relative to a wholesome inhabitants. Specifically, there was a big correlation between SuStaIn levels and common cortical thickness (Fig. 1e, trajectory 1: r = 0.599, p < 0.001; trajectory 2: r = 0.791, p < 0.001; trajectory 3: r = 0.847, p < 0.001), in addition to the quantity of the left hippocampus (Fig. 1f, trajectory 1: r = 0.627, p < 0.001; trajectory 2: r = 0.577, p < 0.001; trajectory 3: r = 0.431, p = 0.005). The important correlation between SuStaIn levels and proper hippocampus quantity was solely discovered in the ‘trajectory’ 1 (Fig. 1g, r = 0.269, p = 0.013). These findings counsel that the SuStaIn stage could mirror the underlying neurophysiological and pathological processes. Supplementary Table 1 offers ROI-wise correlation coefficients between SuStaIn levels and regional z scores.By evaluating ROI-wise z scores (Fig. 2) between every subtype and the wholesome management group, four distinct neuroanatomical signatures had been recognized because the left hippocampus-predominant signature (subtype 1), the appropriate hippocampus-predominant signature (subtype 2), the cortex-predominant signature (subtype 3), and the ‘regular’ signature (subtype 4). Compared to the wholesome management group, subtypes 1 and a pair of exhibited essentially the most extreme atrophy in the ipsilateral hippocampus. In subtype 3, grey matter loss was primarily noticed in the neocortices. Conversely, subtype 4 confirmed elevated grey matter quantity, with essentially the most pronounced enlargement noticed in the amygdala (Supplementary Table 2). Furthermore, the bilateral amygdala quantity in subtype 4 was bigger than in the opposite subtypes (left, t = 6.39, p < 0.000001; proper, t = 7.53, p < 0.000001) in addition to the wholesome management group (left, t = 7.63, p < 0.000001; proper, t = 7.40, p < 0.000001) (Supplementary Fig. 1). In addition, comparisons of ROI-wise z score between any two subtypes are visualized in Supplementary Fig. 2. Results of inter-subtype comparison that includes all ROIs across the brain are described in Supplementary Table 3.Fig. 2: Four distinct neuroanatomical signatures of brain atrophy patterning in people with temporal lobe epilepsy.Subtype-specific signature in neuroanatomical pathology includes (1) the left hippocampus-predominant signature (subtype 1), (2) the right hippocampus-predominant signature (subtype 2), (3) the cortex-predominant signature (subtype 3) and (4) the ‘normal’ signature (subtype 4). ROI-wise z-scores are mapped to a brain template using visualization tools implemented in ENIGMA Toolbox (https://enigma-toolbox.readthedocs.io/en/latest/index.html). Color bar indicates z-scores (i.e., normative deviations) relative to the healthy control group. Note that a higher z-score represents a larger gray matter loss. Data in violin plot are presented as mean values +/− SD. Asterisk indicates significant regional volume reduction in subtype group compared to healthy control group using two-sided two sample t-test following FDR multiple comparisons correction. n = 85, 113, 41, and 57 biologically independent samples in the subtype 1, subtype 2, subtype 3 and subtype 4. In subtype 1, significant reductions are observed in left hippocampus (p = 2.9 × 10−43), left thalamus (p = 2.1 × 10−15) and right thalamus (p = 3.0 × 10−5). In subtype 2, significant reductions are found in right hippocampus (p = 7.2 × 10−62), left hippocampus (p = 4.7 × 10-5), left thalamus (p = 3.5 × 10−9) and right thalamus (p = 9.2 × 10−28). In subtype 3, significant reductions are found in right thalamus (p = 7.2 × 10−62), right caudalmiddlefrontal (p = 1.8 × 10−13), right paracentral (p = 1.3 × 10−12), right parsopercularis (p = 4.5 × 10−13), right parstriangularis (p = 3.5 × 10−15), right precentral (p = 1.2 × 10−14), right precuneus (p = 2.5 × 10−11), right superiorfrontal (p = 4.2 × 10−17), left caudal middle frontal (p = 6.6 × 10−15, left entorhinal (p = 1.2 × 10−3), left fusiform (p = 2.3 × 10−9), left paracentral (p = 2.0 × 10−12), left precentral (p = 3.3 × 10−14), left precuneus (p = 2.8 × 10−11), left superiorfrontal (p = 2.7 × 10-17), left temporalpole (p = 3.9 × 10-6), and left transversetemporal (p = 1.0 × 10-5) regions.Reproducibility of SuStaIn subtypesWe examined the reproducibility of the SuStaIn trajectories in another independent validation sample including 109 patients diagnosed with temporal lobe epilepsy (61 females, age=33.1 ± 10.4 years). The SuStaIn trajectory was re-estimated based on the validation data. The spatiotemporal trajectory can be mathematically characterized as a sequence of ranked biomarkers (here n = 23), which is shown in the (Supplementary Table 4). We observed again that the three trajectories from the validation data were began at the left hippocampus, right hippocampus and cortex separately, consistent with the findings from the discovery cohort. In addition, SuStaIn assigned each patient into a subtype, which allowed us to calculate average of z score map across individuals within the same subtype as a representation of subtype-specific atrophy signature. Four distinct signatures of brain atrophy patterning were replicated in the validation dataset (Supplementary Fig. 3). We observed a high consistency of z score map between discovery dataset and validation dataset (r > 0.7, p < 10e−10). These outcomes recommended the reproducibility of SuStaIn subtypes.In addition, we evaluated the soundness of SuStaIn utilizing totally different quantity of options. We noticed that 93.9% of people had been constant for subtype label (Supplementary Table 5), indicating a excessive stability for particular person subtyping even at comparatively fewer spatial options for SuStaIn mannequin.To look at whether or not the subtype label retains consistency as illness progresses, we adopted up brain MRI information of a subsample (n = 23, common of interval time = 39.0 ± 16.8 months). The labels of subtype at follow-up remained per baseline for nearly all sufferers (Supplementary Materials), suggesting that since sure preliminary brain harm is established, it's much less prone to shift from one trajectory sample (i.e., subtype) to a different.To look at the generalization of SuStaIn subtype to unseen information, we performed a generalization evaluation with ten-fold cross-validation. For every fold, a brand new SuStaIn mannequin was skilled on 90% of the information, and was used to deduce particular person subtype and stage on the left-out 10% information. We in contrast whether or not the subtype and stage assignments of unseen information are per unique mannequin that has been skilled on all information. We noticed that 98.6% of people preserve constant subtype assignments with the unique subtype (Supplementary Table 6). Spearman correlation check exhibits a excessive consistency between levels of unseen information and unique outcome (r = 0.986, p < 0.001) (Supplementary Materials). These counsel a excessive generalizability of SuStaIn subtype to unseen information.Clinical characterization of subtypesAmong all sufferers, 28.7%, 38.2%, 13.9%, and 19.2% had been categorized into subtypes 1 to 4, respectively. Significant variations had been noticed in varied medical variables among the many four subtypes, together with age of onset, sickness period, seizure lateralization, MRI hippocampal sclerosis (HS) charge, historical past of febrile seizures, aura, and therapy outcomes (Table 1). Each subtype exhibited distinct medical traits.Table 1 Demographic and medical characterization of subtypesConsistent with expectations, a better proportion (χ² = 102.4, p < 0.0001) of people with TLE who had constructive findings of HS on their MRI had been assigned to the left hippocampus-predominant subtype 1 (95.3%) and the appropriate hippocampus-predominant subtype 2 (92.9%) in comparison with the cortex-predominant subtype 3 (39.0%) and the ‘regular’ signature subtype 4 (42.1%) (Fig. 3a). Furthermore, the left/proper hippocampus-predominant subtypes included sufferers with TLE whose seizure lateralization was positioned in the corresponding left or proper hemisphere (Fig. 3b), indicating that the preliminary atrophy occurred primarily in the ipsilateral hippocampus. Patients assigned to the left hippocampus-predominant subtype had the youngest age of onset, with a imply of 12.3 ± 7.7 years, in comparison with the opposite three subtypes (t = −4.34, p < 0.0001) (Fig. 3c). The left/proper hippocampus-predominant subtypes additionally had an extended sickness period in comparison with the cortex-predominant subtype 3 and the ‘regular’ subtype 4 (t = 3.70, p = 0.0003) (Fig. 3d). We additionally discovered a subtype impact on complete intracranial quantity (TIV) — people with the cortical subtype 3 had considerably bigger intracranial quantity than the opposite three subtypes; an exploratory evaluation was used to look at the affiliation of TIV with subtypes and medical options.Fig. 3: Clinical characterization of subtypes.a Proportion of TLE people with a visual hippocampal sclerosis on their magnetic resonance imaging (MRI) in every subtype. b Proportion of people with TLE whose seizure lateralization positioned on the corresponding left or proper hemisphere. Red asterisk represents important distinction between a particular subtype vs. all different subtypes (subtype 1, p = 3.8 × 10−22; subtype 2, p = 2.5 × 10−29; subtype 3, p = 0.723; subtype 4, p = 0.015). c Differences of age of onset amongst four subtypes. d Differences of sickness period amongst four subtypes. e Proportion of people with seizure-free (i.e., efficient), not seizure-free (i.e., ineffective) or misplaced follow-up in 144 medicated people (MG) on the follow-up (imply interval is 56.3 months). f Proportion of people with seizure-free (i.e., efficient), not seizure-free (i.e., ineffective) or misplaced follow-up in 152 anterior temporal lobe operative people (OG) at follow-up (imply interval is 64.1 months). The white dotted line (a, b, e, and f) exhibits the common of the four subtypes. Data in figures (c and d) are introduced utilizing a box-plot (middle line, median; field limits, higher and decrease quartiles; whiskers, 1.5×interquartile vary [IQR]; factors, outliers). n = 85, 113, 41, and 57 biologically impartial samples in the subtype 1, subtype 2, subtype 3 and subtype 4. Pearson’s Chi-square check is performed for information evaluation in figures a, b, e and f. Two-sided two-sample t check is used for information evaluation in figures c and d. Multiple comparisons had been thought of with FDR correction. LHIP, left hippocampus-predominant signature (subtype1); RHIP, proper hippocampus-predominant signature (subtype2); Cortex, the cortex-predominant signature (subtype3); Normal, the ‘regular’ signature (subtype4).We investigated whether or not the neuroanatomical subtype classification primarily based on baseline MRI was associated to differential therapy outcomes with drugs or anterior temporal lobe surgical procedure. In the drugs group (MG, baseline n = 144, follow-up n = 107), 21 sufferers reported seizure freedom on the follow-up (imply interval of 56.3 months). In the anterior temporal lobe operative group (OG, baseline n = 152, follow-up n = 145), 96 people reported seizure freedom following the operative therapy on the follow-up (imply interval of 64.1 months). Interestingly, in the medications-treated sufferers, a considerably greater follow-up seizure freedom charge was noticed in the ‘regular’ signature subtype 4 (39.29%) in comparison with the opposite three subtypes (12.66%) (χ² = 9.29, p = 0.0023) (Fig. 3e). However, for sufferers handled with anterior temporal lobe surgical procedure, the follow-up seizure freedom charge in subtype 4 was 45.00%, which was considerably worse than the opposite three subtypes (69.60%) (χ²=4.66, p = 0.031) (Fig. 3f).We additionally noticed that sufferers with MRI proof of HS (HS+) present youthful age of onset in comparison with these with regular MRI outcomes (HS-) upon visible examination (t = −3.49, p = 0.001). In addition, we discovered that sufferers with HS- expertise worse surgical outcomes in comparison with these HS+ sufferers (χ² = 5.99, p = 0.014). To look at whether or not the medical variations amongst SuStaIn subtypes are affected by HS, we re-analyzed the correlations between medical options and subtype with HS impact as a covariate (Supplementary Materials). We nonetheless discovered important correlations of SuStaIn subtype with age of onset (t = −3.51, p = 0.001), sickness period (t = −3.15, p = 0.002) and medicine outcomes (χ² = 5.64, p = 0.018) after controlling HS impact (Supplementary Materials).Subtype-based classifier predicts surgical procedure prognosisWe evaluated prediction efficiency on classifying the topic who achieves seizure freedom (OG+) or not (OG-) after surgical procedure, utilizing a classical machine learning prediction procedures (see Methods). To look at whether or not the SuStaIn subtype data might assist to enhance prediction, we performed machine learning prediction procedures by means of a framework underneath SuStaIn subtype background (Supplementary Fig. 4). We proposed a perspective that every subtype could require particular options/classifiers to foretell postoperative end result, given that every subtype has particular brain construction and medical traits. Thus, utilizing assist vector machine (SVM), we constructed a particular sub-classifier corresponding to every SuStaIn subtype. By ten-fold cross-validation, we noticed an acceptable-to-good prediction efficiency for every sub-classifier to every SuStaIn subtype (Supplementary Fig. 5); yielding an general accuracy (71.72%), specificity (81.03%) and sensitivity (47.87%) on the check information. As a complete analysis, the Youden Index for the SuStaIn subtype-based classifier (J = 0.289) on check information was considerably greater than randomly predictions by permutation check (p = 0.012) (Supplementary Fig. 6). Details of prediction efficiency of every subtype classifier are described in Supplementary Table 7.As a reference, we additionally performed a predictive check with none SuStaIn subtype data as prior. Specifically, SVM classifier was skilled utilizing medical data at baseline as options. By ten-fold cross-validation, we noticed 67.59% accuracy, 89.58% specificity and really low sensitivity (24.49%) on the check information; whereas Younden Index ( J = 0.141) didn't present important distinction in comparison with randomly predictions by permutation check (p = 0.307). This means that these (OG-) sufferers weren't efficiently recognized if solely medical data was relied upon. In addition, we discovered that even when we added way more options (medical variables + MRI regional measures) to coach classifier, the prediction efficiency didn't enhance (Younden Index = 0.130, accuracy = 66.90%, sensitivity=24.49%, specificity=88.54%).
https://www.nature.com/articles/s41467-024-46629-6
Identification of four biotypes in temporal lobe epilepsy via machine learning on brain images
