Temporal lobe epilepsy lateralization and surgical outcome prediction using diffusion imaging

Abstract

Objective:

We sought to augment the presurgical workup of medically refractory temporal lobe epilepsy by creating a supervised machine learning technique that utilizes diffusion weighted imaging to classify patient-specific seizure onset laterality and surgical outcome.

Methods:

151 subjects were included in this analysis: 62 patients (aged 18-68 years, 36 female) and 89 healthy controls (aged 18-71 years, 47 female). We created a supervised machine-learning technique that utilizes diffusion-weighted metrics to classify subject groups. Specifically, we sought to classify patients vs. healthy controls, unilateral vs. bilateral temporal lobe epilepsy, left vs. right temporal lobe epilepsy, and seizure free vs. not seizure free surgical outcome. We then reduced the dimensionality of derived features with community detection for ease of interpretation.

Results:

We classified the subject groups in withheld testing datasets with a cross-fold average testing areas under the receiver operating characteristic curve of 0.745 for patients vs. healthy controls, 1.000 for unilateral vs. bilateral seizure onset, 0.662 for left vs. right seizure onset, 0.800 for left-sided seizure free vs. not seizure free surgical outcome, and 0.775 for right-sided seizure free vs. not seizure free surgical outcome.

Conclusions:

This technique classifies important clinical decisions in the presurgical workup of temporal lobe epilepsy by generating discerning white matter features. We believe this work augments existing network connectivity findings in the field by further elucidating important white matter pathology in temporal lobe epilepsy. We hope this work contributes to recent efforts aimed at using diffusion imaging as an augmentation to the presurgical workup of this devastating neurological disorder.

INTRODUCTION

Epilepsy is one of the most common and disabling neurological disorders suffered by over 50 million people worldwide.[1] Temporal lobe epilepsy (TLE) is the most common form of focal epilepsy and approximately 30-40% of these patients continue to have debilitating seizures despite maximal therapy with anti-seizure medications.[2] Epilepsy surgery can eliminate or markedly reduce seizures in many TLE patients using resection,[3] ablation,[4] or neurostimulation [5] of regions that generate seizures. Unfortunately, about 33-50% of patients who undergo epilepsy surgery have persistent disabling seizures post-operatively.[6] One important factor contributing to epilepsy surgery failures is inaccurate or incomplete lateralization and localization of seizure onset zones ahead of treatment.[7]

Standard non-invasive presurgical techniques to evaluate TLE rely on a combination of scalp-electroencephalography (EEG), T1/T2 weighted magnetic resonance imaging (MRI), positron emission tomography, single-photon emission computerized tomography, functional MRI, and magnetoencephalography.[8] However, these modalities are often not sufficient to confidently lateralize (i.e. which brain hemisphere is responsible for seizure generation) and localize epileptogenicity. Thus, invasive intracranial-EEG (iEEG) monitoring is often pursued.[9] Adverse events may occur in up to 10% of these long and uncomfortable inpatient iEEG monitoring sessions, and seizures may not occur at all.[10, 11] Thus, many research groups have sought other non-invasive modalities to help reduce the need for invasive monitoring and improve presurgical evaluation of TLE.

Many groups have leveraged machine learning (a.k.a. artificial intelligence) to maximize the information gleaned from non-invasive modalities to address existing problems in epileptology.[12] Generally, machine learning is the process of inputting large amounts of raw data into a fine-tuned algorithm to provide a specific clinically relevant result. Some popular applications include predicting seizure occurrence on EEG or detecting lesions on MRI T1 imaging. Developers of machine learning models must take great caution to avoid over-fitting their own datasets. An overfit model will only give accurate predictions on a single dataset and will give erroneous results when applied to a new dataset. When carefully applied to properly posed questions, machine learning has the potential to meaningfully augment clinical decision making.

With these considerations in mind, there has been a recent effort to use diffusion weighted imaging (DWI) metrics and machine learning to predict presence of TLE[13, 14], left or right-sided seizure onset,[13, 15, 16] and surgical outcome.[17-19] DWI models may enhance previously proposed models that rely solely on categorical clinical variables.[20, 21] More work is needed to improve model accuracy and simultaneously enhance our biological understanding of TLE. Additionally, a unified technique that incorporates a wide scope of important clinical classifications in the presurgical evaluation of TLE would be very useful to the clinician and could help avoid unnecessary intracranial monitoring. We provide evidence here that a supervised machine-learning technique utilizing joint independent component analysis (joint-ICA) feature extraction on DWI-derived whole-brain structural connectivity (SC) and fractional anisotropy (FA) metrics can be used to identify the presence of TLE, classify TLE laterality, and predict surgical outcome. Furthermore, we sought to increase the interpretability of our technique by reducing the dimensionality of our derived features and re-evaluating model performance. Through this work, we aim to contribute to recent efforts using DWI-derived predictive modelling to augment the existing presurgical evaluation of TLE.

METHODS

Participants

We included 151 subjects (62 TLE patients and 89 healthy controls) in these analyses. We have outlined the demographics and clinical information of the subjects in Table 1. This study received approval by Vanderbilt University IRB (170560), and we obtained informed consent from each subject. Only patients with surgical, drug resistant TLE were included in this study. Specifically, the diagnosis of TLE was determined by our medical center’s standard multidisciplinary process, including epileptologists, neurosurgeons, and neuropsychologists. This process included analyzing patient history, seizure semiology, video EEG, MRI, positron emission tomography, memory/language localization by functional MRI or Wada, and neuropsychological testing. We included patients with and without mesial temporal sclerosis (MTS) but excluded patients with any other diagnosed lesions observed on neuroimaging by a neuroradiologist or epileptologist, such as encephalocele, cavernous malformation, and tumor. All unilateral TLE patients had resective surgery or laser ablation, and surgical outcome was evaluated at the most recent time point available (mean 1.95 years, min 1.1 years, max 3.1 years) according to the Engel Surgical Outcome Scale.[22]

Table 1:

Subject characteristics

	Right TLE (n=37)	Left TLE (n=18)	Bilateral TLE (n=7)	Controls (n=89)
Sex
Female	22 (59.5%)	8 (44.4%)	6 (85.7%)	47 (52.8%)
Male	15 (40.5%)	10 (55.6%)	1 (14.3%)	42 (47.2%)
Age (years, mean ± STD)	40.0 ± 11.3	43.1 ± 21.7	33.6 ± 8.9	38.3 ± 13.2
Epilepsy duration (years, mean ± STD)	21.3 ± 14.2	19.1 ± 15.4	16.0 ± 4.3	-
Preoperative seizure frequency (per month, mean ± STD)	17.2 ± 40.2	14.7 ± 20.2	18.3 ± 23.7	-
FBTC seizures
Yes	16 (43.2%)	13 (72.2%)	3 (42.9%)	-
No	21 (56.8%)	5 (27.8%)	4 (57.1%)	-
MRI
No Diagnosed Lesions	14 (37.8%)	7 (38.8%)	4 (57.1%)	-
Unilateral MTS	17 (45.9%)	11 (61.1%)	3 (42.9%)	-
Bilateral MTS	6 (16.2%)	0 (0%)	0 (0%)	-
Surgery type
Resective SAH	27 (73.0%)	15 (83.3%)	0 (0%)	-
TL	10 (27.0%)	1 (5.6%)	0 (0%)	-
Laser SAH	0 (0%)	2 (11.1%)	0 (0%)	-
B/L RNS	0 (0%)	0 (0%)	6 (85.7%)	-
DBS	0 (0%)	0 (0%)	1 (14.3%)	-
Follow-up duration (years, mean ± STD)	1.8 ± 0.9	2.0 ± 0.9	0.78 ± 0.59	-

B/L RNS = bilateral responsive neurostimulation in hippocampi and amygdalae; DBS = deep brain stimulation of bilateral anterior thalamic nuclei; FBTC = focal to bilateral tonic-clonic seizures; MTS = mesial temporal sclerosis; SAH = selective amygdalohippocampectomy; STD = standard deviation; TL = temporal lobectomy.

For the healthy controls vs. all TLE patients, there was no significant difference for age (two-sample t-test statistic −0.319, 149 df, P=0.750) and sex (chi-squared statistic 0.408, 1 df, P=0.523). For unilateral vs. bilateral TLE there was no significant difference between age (two-sample t-test statistic 1.190, 60 df, P=0.239), sex (Fisher’s exact, 1 df, P=0.224), epilepsy duration (two-sample t-test statistic 0.807, 60 df, P=0.4227), preoperative seizure frequency (two-sample t-test statistic −0.136, 60 df, P=0.892), and focal to bilateral tonic-clonic (FBTC) seizures (Fisher’s exact, 1 df, P=0.703). For right and left TLE, there was no significant difference between age (two-sample t-test statistic −0.718, 53 df, P=0.476), sex (chi-squared statistic 1.101, 1 df, P=0.294), epilepsy duration (two-sample t-test statistic −0.525, 53 df, P=0.602), and preoperative seizure frequency (two-sample t-test statistic −0.240, 53 df, P=0.811). For left vs. right TLE there was a significant difference between patients that experienced FBTC seizures (chi-squared statistic 4.080, 1 df, P=0.043), with left TLE patients experiencing FBTC seizures more frequently.

DWI data collection and preprocessing

We acquired high angular resolution diffusion imaging with 92 b-vector directions at a b-value of 1600 s/mm² and a single b0 volume, 50 slices, 2.5 mm isotropic, collected on a Phillips 3T MRI scanner with a 32-channel head coil. We preprocessed the DWI data with an integrated pipeline called PreQual.[23] We used a T1 image to generate a synthetic susceptibility-corrected b0 volume using SYNB0-DISCO, a deep learning framework by Schilling et al.[24] We used this synthetic b0 image in conjunction with FSL's topup to correct for susceptibility-induced artifacts in the diffusion data. We then utilized FSL's eddy algorithm to correct for motion artifacts and eddy currents, and to remove outlier slices.[25]

Probabilistic tractography

We used MRTrix3 to conduct probabilistic tractography.[26] Briefly, we used FreeSurfer to parcellate each T1 image into 84 region of interest (ROI) segmentation based on the Desikan-Killiany (DK) atlas.[27] We then performed spherical deconvolution on the preprocessed DWI to get direct fiber orientation density function for every voxel.[28] We ran probabilistic tractography [29] and weighted the tractogram with spherical-deconvolution informed filtering of tractograms (SIFT2) for improved biological accuracy of white-matter representation.[30] Finally, we generated the structural connectomes by assigning SIFT2-weighted streamline counts between ROIs (SC connectome), and average fractional anisotropy (FA) between ROIs in the DK segmentation (FA connectome). We merged the SC and FA metrics into a joint SC-FA matrix to be used for machine learning.

Diagnostic and prognostic clinical decisions (binary decision groups)

We sought to characterize the following clinically relevant decisions, outlined in Fig. 1, using the DWI-derived SC-FA matrices: Decision 1) First, we classified healthy controls [n=89] vs. TLE patients [n=62]. Decision 2) Next, we classified unilateral seizure free (“SF”, Engel 1a) TLE [n=26] vs. intracranial EEG confirmed bilateral TLE [n=7]. Decision 3) Then, we classified unilateral SF right TLE [n=16] vs. unilateral SF left TLE [n=10]. We then brought the not seizure free (“not-SF”, Engel 1b-4) surgical outcome patients back into the analysis and classified the patients based on surgical outcome as follows: Decision 4) Left TLE SF [n=10] vs. not-SF [n=8] surgical outcomes. Decision 5) Right TLE SF [n=16] vs. not-SF [n=21] surgical outcomes.

Figure 1: Binary decision tree for TLE diagnosis and prognosis.

This diagram outlines the clinically relevant diagnostic and prognostic decisions on which we chose to focus. The color of each stage indicates the proposed clinical suggestion: orange, more information required to provide clinical suggestion; blue, suggest to proceed directly to unilateral surgical resection; red, suggest to localize further with SEEG.

*Focal non-temporal neocortical onset excluded due to inadequate sample size

Supervised machine learning

We used the MATLAB (MathWorks inc., Natick, MA, USA) built-in Bayesian optimization function [31] to train a supervised learning algorithm to produce a “decoding network” that classified each of our binary groups of interest (Fig. 2B). Specifically, within each iteration of the optimization we extracted features with MATLAB’s reconstruction independent components analysis (r-ICA) [32] algorithm. Using r-ICA to extract features from the raw SC/FA matrices allows us to generate independent predictor variables while minimizing the potential for overfitting that could occur if we had allowed all 6,972 individual SC/FA values to be fit individually. We then ran elastic net regression on the r-ICA components using the binary decision labels as output values. To classify the groups, we then assembled a “decoding network” by multiplying each r-ICA component by its elastic net regression coefficient and calculating the sum. Each subject will have a unique r-ICA loading that represents how similar or dissimilar his or her original input SC/FA matrix is to the decoding network. To quantify this classification, we calculated an area under the curve (AUC) for a receiver operating characteristic curve (Fig 2C).

Figure 2: Methods overview.

(A) We used MRTrix3 to conduct anatomically constrained probabilistic tractography on DWI images for each subject. Structural connectivity (SC) and fractional anisotropy (FA) were combined into a single matrix and the SC was corrected using spherical deconvolution informed filtering of tractograms 2 (SIFT2). (B) We implemented a supervised learning approach to extract features with reconstruction independent component analysis (r-ICA) and fit component loadings to an elastic net regression to perform binary classification for each decision of interest. The “decoding network” is the summation of the networks generated by r-ICA weighted by the elastic net regression coefficients. We optimized the number of r-ICA components “q”, regression regularization coefficient “lambda”, and elastic net regression coefficient “alpha” by using Bayesian optimization. (C) We then used the decoding network loadings to classify the binary decision of interest. The blue columns represent a label of 0 and the red columns represent a label of 1. We can then calculate the area under the receiver operating characteristic curve (AUC) to determine the model’s efficacy of classifying the two groups.

We repeated this process with 250 iterations to determine the optimum hyper-parameters to maximize the AUC for each binary decision. We optimized the following hyper-parameters: 1) the number of independent components 2) the r-ICA lambda regularization coefficient and 3) elastic net regression alpha value. We calculated the loss function for the optimization according to Eq. 1 with the training dataset AUC represented as AUC_Tr, and the AUC of the validation dataset represented as AUC_V. We completely withheld the testing set from training and validating within each individual cross fold.

$$(1-AUC\_Tr)+0.5(1-AUC\_V)$$ (1)

Cross validation and summary decoding network

We used five-fold cross validation with 70% of the dataset used for model training, 20% used as a validation set for model training termination to prevent overfitting, and 10% completely withheld data used as a testing set for each cross fold. We report the results of the training, validation, and testing datasets across the five cross folds for each binary decision group. For ease of interpretation and presentation, we combined all cross fold decoding networks into a single summary decoding network. This single decoding network for each decision allows us to give a summary of how well we can classify all patients in the cohort. We reported the average AUC across the testing sets for the summary decoding networks as “Combined Folds” AUC.

Existing clinical prediction tools

As a baseline prediction and motivation for this work, we sought to evaluate the performance of existing surgical outcome prediction tools. We combined left and right TLE patient groups for this analysis (n=55), and bilateral onset patients were excluded (n=7). We predicted SF outcome at one year with Dugan et al.’s Epilepsy Surgery Grading Scale (ESGS),[20] and SF outcome at two years with Jehi et al.’s Nomogram.[21] We calculated the Jehi et al. Nomogram for every patient that had two-year surgical outcomes (n=30) and the Dugan et al. ESGS and clinical concordance for all unilateral TLE patients with 1-year surgical outcomes (n=55).

Community detection to form decoding subnetwork

To further ease interpretation, we sought to reduce the dimensionality of the decoding network. We focused on the SC portion of the matrices because it consistently demonstrated higher absolute value contributions (weights) compared to FA. For dimensionality reduction we used the community-Louvain [33] algorithm as implemented by the Brain Connectivity Toolbox in MATLAB.[34] We ran community-Louvain 100 times with resolution parameters ranging from 0.5 to 2.0 at an increment of 0.01. We selected the resolution with the highest mutual information as the most stable community detection. For each community we calculated how well it could individually classify the binary decision of interest. We designated the community with the highest AUC as the “Decoding Subnetwork” that preserves the greatest amount of classification ability compared to the original decoding network.

RESULTS

Decoding networks can classify TLE laterality and surgical outcome

Using a completely withheld testing dataset, this technique classified with an AUC of 0.745 for TLE patients vs. healthy controls, 1.000 for unilateral vs. bilateral seizure onset (small sample size), 0.662 for left vs. right seizure onset, 0.800 for left-sided SF vs. not-SF surgical outcome, and 0.775 for right-sided SF vs. not-SF surgical outcome (Figs. 3 & 4). We then merged the decoding networks from each cross fold to ease interpretation. The AUCs for these summary decoding networks are outlined in Figs. 3 & 4. We will further outline the interpretation of these decoding networks in the “Decoding subnetworks maintain predictive power” section and Discussion. To assess the affects of MTS on the surgical outcome predictions, we conducted a post-hoc sub analysis that showed there was no significant difference in loadings (i.e. raw prediction values) for patients with MTS vs. patients without MTS for left TLE surgery (t-test p=0.064) or right TLE surgery (t-test p=0.3134). These post-hoc results suggest that a heterogenous training set has led to high classification accuracy that generalizes across the identification of MTS on MRI during standard presurgical workup.

Figure 3: Decoding networks to predict TLE vs. controls and TLE laterality.

This figure outlines the DWI-derived decoding network, network loadings, and binary ROC curve for the first three decisions of the binary decision tree outlined in “Diagnostic and prognostic clinical decisions (binary decision groups)”. The decoding network for each decision (e.g. controls vs. TLE) is the average decoding network across the five cross folds. The decoding network in (A) has labels to represent which DK ROI connections are being represented in each part of the matrix – i.e. upper right is bilateral (B/L) structural connectivity (SC) connections, the bottom left quadrant is B/L fractional anisotropy (FA) values, and right (R)/left (L) SC and FA values are in the remaining triangular sections of the matrix. Please see Supplementary Table 1 for a full list of the DK ROIs. One can interpret the decoding network with reference to the network loadings. For example, if a specific subject group has higher network loadings than another group, then that group has higher SC or FA connections outlined by the positive (red) weights in the decoding network and lower SC or FA connections outlined by the negative (blue) weights in the decoding networks. We generated the binary ROC curve by comparing the network loadings between the groups of interest for each binary decision. The two groups of interest for each decision are indicated by the blue background and pink background in the network loading violin plots. Results: (A) The controls vs TLE patients were classified with a cross fold average testing area under the receive operating characteristic curve (AUC) of 0.745, and a combined fold AUC of 0.721. (B) Unilateral vs. bilateral TLE patients were classified with a cross fold average testing AUC of 1.000 and a combined fold AUC of 0.877. (C) Next, we classified left vs. right TLE patients with a cross fold average testing AUC of 0.662 and a combined fold AUC of 0.970. Each of the cross-fold average training and validation AUCs can be seen in the legend of the binary ROC curves.

Figure 4: Decoding networks to predict surgical outcome for left and right TLE patients.

This figure outlines the DWI-derived deciding networks’ ability to predict surgical outcome at greater than one year for left and right unilateral TLE. For an explanation of how to interpret the sections of this figure refer to Fig. 2 caption. (A) For left TLE, we classified SF against not-SF with a cross fold average testing AUC of 0.800 and a combined fold AUC of 0.975. (B) For the right TLE, we classified SF against not-SF with a cross fold average testing AUC of 0.775 and a combined fold AUC of 0.923. Each of the cross fold average training and validation AUCs can be seen in the legend of the binary ROC curves.

Surgical outcome prediction tools could benefit from advanced neuroimaging analyses

As motivation for this work, we previously had used the Dugan et al. Epilepsy Surgery Grading Scale (ESGS) [21] to predict surgical outcome at one year and the Jehi et al. Nomogram [20] to predict surgical outcome at 2 years. Both the ESGS and Nomogram gave comparably uncertain predictions of surgical success for our patient cohort. Specifically, 15 (56%) of our known 27 SF patients at one year were given an ESGS of Grade 1, the highest chance of achieving SF, and 14 (50%) of our 28 known not-SF patients were also given Grade 1 (chi-squared statistic 0.170, 1 df, P=0.680 – combined ESGS Grade 2&3). For the Jehi Nomogram, nine (64%) of our known 14 SF at two years being predicted by the Nomogram to be SF and 11 (69%) of our 16 known not-SF at two years also being predicted to be SF (chi-squared statistic 0.0670, 1 df, P=0.796). Due to the single-institution development of our DWI-derived methods, it is not fair to directly compare our DWI model results to these surgical outcome prediction tools. Rather, we wish to outline that there is the potential for prediction improvement by using patient-specific advanced neuroimaging features.

Decoding subnetworks maintain predictive power

After dimensionality reduction, we were still able to classify the TLE patients vs. healthy controls by the decoding subnetwork with an AUC of 0.623 (Fig. 5A). We classified unilateral vs. bilateral patients with a decoding subnetwork AUC of 0.830, which represents a loss of only 0.047 AUC compared to the full decoding network (Fig. 5B). Next, we classified left vs. right TLE with a decoding subnetwork AUC of 0.892 (Fig. 5C). The decoding subnetworks also maintained predictive power for classifying surgical outcome with an AUC of 0.913 for predicting SF at one year in left TLE patients (Fig. 6A) and 0.875 for right TLE patients (Fig. 6B). This reduction in AUC, as compared to the full decoding network, is the compromise for more easily interpretable subnetworks.

Figure 5: Decoding subnetworks to predict TLE vs. controls and TLE laterality.

This figure outlines the structural connectivity (SC) decoding subnetworks for the first three binary decisions present in “Diagnostic and prognostic clinical decisions (binary decision groups).” The decoding subnetwork is a subnetwork of nodes that maintains a comparable ability to classify the subject groups of interest compared to the full decoding matrix. These subnetworks are much easier to interpret and provide insight into the SC connections that are most important to classify the subject groups. (A) TLE patients classified against healthy controls by the decoding subnetwork with an AUC of 0.623, whereas the full decoding network had a combined fold AUC of 0.721. (B) Unilateral vs. bilateral patients were classifed by the decoding subnetwork with an AUC of 0.830, whereas the full decoding network had a combined fold AUC of 0.877. (C) Left vs. right TLE were classified by the decoding subnetwork with an AUC of 0.892, whereas the full decoding network had a combined fold AUC of 0.970.

Figure 6: Decoding subnetworks to predict surgical outcome for left and right TLE patients.

This figure outlines the structural connectivity (SC) decoding subnetworks for the surgical outcome binary “Decision 4” & “Decision 5” present in “Diagnostic and prognostic clinical decisions (binary decision groups).” The decoding subnetworks is a subnetwork of nodes that maintains a comparable ability to classify the subject groups of interest compared to the full decoding matrix. These subnetworks contain fewer ROIs compared to the full decoding networks, and thus provide simpler insight into the specific SC connections that are most important to classify the subject groups. (A) Left TLE patients with SF outcome were classified against those with not-SF outcome by the decoding subnetwork with an AUC of 0.913, whereas the full decoding network had a combined fold AUC of 0.975. (B) Right TLE patients with SF outcome were classified against those with not-SF outcome by the decoding subnetworks with an AUC of 0.875, whereas the full decoding network had a combined fold AUC of 0.923.

DISCUSSION

We intended to create a machine-learning technique that generated interpretable white-matter features that accurately classify clinically important diagnostic and prognostic decisions in the presurgical workup of medically refractory TLE. We suggest that advanced neuroimaging analyses can be effectively used to enhance current clinically driven prediction models. Incorporation of a DWI-enhanced clinical decision tree could reduce the need for intracranial monitoring by suggesting patients proceed directly to surgery if they are predicted to have a high chance of SF with unilateral surgical resection.

Due to the large size of the full decoding networks, we sought to further aid the biological interpretation by reducing the dimensionality with community detection. This resulted in smaller “decoding subnetworks” that maintained an adequate predictive power. In general, our methods identified TLE patients vs. controls with moderate accuracy, laterality and surgical outcome with medium-high accuracy, and may supplement clinical decision-making and help guide patient counselling.

Interpretation of key regions in decoding subnetworks

1). Healthy Controls vs. TLE Patients

The important SC connections to discern controls from TLE patients were a combination of left and right brain structures (Fig. 5A). Notably included was the left insula, many left deep brain structures, the right cingulate, and the right superior, middle, and inferior frontal gyri. An interpretation of this asymmetric finding is that left, right, and bilateral TLE patients likely do not harbor symmetric disease.[15, 16] The strongest SC predictors in the decoding subnetwork were between the right caudal middle frontal gyrus to the left basal ganglia and left insula. Also included in patients were augmented SC connections from the left thalamus. Patients experienced an average of much higher SC between these regions. A possible synthesis of these findings is that patients could have an augmented structural connection in known seizure propagation pathways involving the basal ganglia and thalamo-cortical networks.[35, 36] Notably absent from this decoding subnetwork are limbic structures like the hippocampi, amygdale, and orbitofrontal regions that have well documented pathology in TLE.[37] A possible interpretation of this finding is that there is likely strong pathological heterogeneity amongst the patients’ limbic structures because the left, right, and bilateral TLE patients were all grouped together. The predictive ability of this technique could have possibly been improved by subdividing this decision into controls vs. left, controls vs. right, and controls vs. bilateral TLE separately.

2). Unilateral vs. Bilateral TLE:

The most important SC connections for this classification were bilateral temporal and orbitofrontal structures, and the left thalamus (Fig. 5B). The single most important connections were those from the right temporal pole, with bilateral TLE patients experiencing much higher connections from the right temporal pole to the left and right lateral orbitofrontal regions. Perhaps these patients have enhanced involvement of the temporal pole in their seizure generation and/or propagation networks - possibly indicating they could have a higher incidence of the difficult to diagnose temporal-pole subtype of TLE.[38]

3). Left vs. Right TLE:

The discernment of right from left TLE in known SF patients was driven exclusively by connections between left-sided ROIs (Fig. 5C). We surmise this is concordant with previous findings that left TLE exhibits a more “severe” white matter pathology.[15, 16] The most important decoding subnetwork weights were SC connections between the left superior and transverse temporal regions to the inferior frontal gyrus – with left TLE having much weaker SC connections between these regions. These findings are in direct alignment with known decreases in functional connectivity of the superior temporal and inferior frontal gyrus in left TLE.[39] The decreased SC of the left superior temporal gyrus in left TLE is also in alignment with previous findings of altered superior temporal nodal degree laterality in left mesial TLE.[40]

4). Left TLE – SF vs. not-SF:

The decoding subnetwork to classify left TLE SF from not-SF patients only contained five ROIs and had all left sided structures except for the right caudal anterior cingulate (Fig. 6A). The single most important connection was between the left thalamus and the left superior parietal lobe, with not-SF patients experiencing significantly higher SC. This can be interpreted in alignment with previous findings that high functional connectivity of the thalamus was a predictor of poor surgical outcome.[41]

5). Right TLE – SF vs. not-SF:

Interestingly, not a single ipsilateral or contralateral ROI is shared between the left and right surgical outcome decoding subnetworks (Fig. 6B), also supporting the idea that left and right TLE exhibit very different white matter pathologies,[15, 16] – and further suggests that predicting seizure freedom with white-matter connectivity is different for left and right TLE. The most important weights were high bilateral SC connections stemming from the left rostral anterior cingulate and left caudate – specifically with not-SF patients experiencing much higher SC from these regions to the right inferior frontal gyrus and right accumbens area. One interpretation of this finding is that right sided TLE patients with not-SF outcome are experiencing enhanced seizure propagation through the contralateral basal ganglia.[35] Notably absent from this decoding subnetwork is the right thalamus, suggesting that unlike left TLE, ipsilateral structural thalamic connections are not as important for predicting surgical outcome in right TLE.

Study limitations

Potential for overfitting

This technique allows TLE patient groups to be classified with five-fold cross validated interpretable white matter connectivity patterns. However, there still exists the ever-present potential to overfit our dataset and thus reduce generalizability of these results. A further concern for overfitting was the relatively small sample size for certain subject groups, specifically bilateral TLE. This small sample resulted in model instability for validation and testing AUCs of 0.513 and 1.000, respectively. The reduction in AUC to 0.877 when combining folds reflects this instability. Overall, this potential for overfitting means we must be cautious in the detailed interpretation of the decoding subnetworks. The best mitigation for these concerns is to incorporate more patients from other medical centers and from a multitude of different MRI scanners and DWI acquisition protocols.

Patient heterogeneity

Another potential concern in this study is the heterogeneity of the patient population. We did restrict the patient population to just TLE due to relatively few of our patients with DWI scans having non-temporal focal neocortical epilepsy – Thus, the clinical decision tree outlined in Fig. 1 is incomplete and future work should incorporate datasets with an adequate sample size of patients with focal neocortical onset. Furthermore, even amongst the TLE patients there is heterogeneity in important factors like presence of MTS, presence of FBTC seizures, and type of surgery received. Specifically, 94% of left temporal cases received selective procedures compared to 73% of right temporal cases. This imbalance in the training set could lead to different structural connectivity features incorporated in the respective decoding networks. A multi-institutional dataset with sufficient subject numbers for each subcategory would be needed to address these concerns. Thus, we grouped all of these patient subtypes together for two main reasons: 1) to maintain an adequate number of patients for machine learning and 2) to hopefully make our findings more generalizable to all TLE patients and not just certain subtypes - e.g. just patients with right MTS who received a right selective amygdalohippocampectomy. Regardless, it would still be worthwhile to aggregate data across multiple sites so that we could evaluate these subtypes separately with adequate numbers for machine learning.

Conclusions

This technique accurately classifies TLE laterality and SF surgical outcome. Furthermore, by reducing the dimensionality of the predictive features with community detection we obtain more easily interpreted subnetworks that maintain adequate classification ability. This work builds upon existing network connectivity findings in TLE and complements other proposed techniques using DWI to classify subject groups. With validation in a larger multi-site dataset, this technique may serve as a source of insight into the biological underpinnings of TLE, and a potential enhancement to the surgical workup of this complex neurological disorder.

Key messages:

Diffusion MRI has been proposed as an augmentation to the presurgical workup of drug resistant focal epilepsy, potentially improving clinical workflow and providing important prognostic information. However, the specifics of translating this modality to the complicated presurgical workflow remain unclear. In this work, we propose a practical machine-learning diffusion MRI model that could augment the clinical decision-making process at many of the important presurgical stages. Furthermore, due to the model design, the white matter features important for classification at each stage can be interpreted directly.

Abbreviations:

AUC

area under the receiver operating characteristic curve

df

degree of freedom

DK

Desikan-Killiany atlas of brain regions

DWI

diffusion-weighted imaging

ESGS

epilepsy surgery grading scale

FA

fractional anisotropy

FBTC

focal to bilateral tonic-clonic seizures

MTS

mesial temporal sclerosis

r-ICA

reconstruction independent components analysis

ROC

receiver operating characteristic

ROI

region of interest

SF

seizure free

SC

structural connectivity

SIFT2

spherical-deconvolution informed filtering of tractograms 2

TLE

temporal lobe epilepsy

Data availability

Data is available upon request to the corresponding author.