Elsevier

Neuroscience Letters

Volume 692, 23 January 2019, Pages 115-121
Neuroscience Letters

Research article
A challenge of predicting seizure frequency in temporal lobe epilepsy using neuroanatomical features

https://doi.org/10.1016/j.neulet.2018.11.005Get rights and content

Highlights

  • Neuroanatomical features serve as promising markers for the detection of TLE.

  • But predicting seizure frequency using neuroanatomical features is challenging.

  • It is due to an unclear relationship between seizure frequency and neuroanatomy.

Abstract

The pathological and clinical characteristics of temporal lobe epilepsy (TLE) tend to be affected by epileptic seizures, specifically represented by seizure lateralization and frequency. Although the lateralization of the epileptogenic zone can be clarified in terms of neuroanatomical damage, there have been conflicting findings on the relationship between seizure frequency and neuroanatomical damage. In this study we sought to examine the relationship in the framework of machine learning-based predictive modeling. We acquired 60 grey matter (GM) anatomical features from structural MRI data and 46 white matter (WM) anatomical features from diffusion-weighted MRI data for 42 TLE patients and 45 healthy controls and applied the random forests method to the neuroanatomical features. We demonstrated that, whereas the neuroanatomical features were promising markers for the discrimination of the TLE patients from the healthy controls, the separation between the TLE patients with low and high seizure frequency on the basis of the neuroanatomical features was challenging. When we applied feature selection techniques for the construction of the predictive models, a greater number of features were selected as being relevant to the distinction of the TLE patients from the healthy controls than to the classification of the TLE patients according to seizure frequency. Furthermore, we adopted model-based clustering analysis and showed that seizure frequency-based subgroups were not matched well with neuroanatomy-based subgroups in the TLE patients. We propose that the challenge of predicting seizure frequency using neuroanatomical features could be attributable to considerable inter-individual variability in neuroanatomical damage among seizure frequency-based subgroups.

Introduction

Machine learning has been applied to predict particular disease states, including the presence and severity of a disease, in individuals. In temporal lobe epilepsy (TLE), machine learning has been employed, for instance, to discriminate TLE patients from healthy individuals [15,33] and to predict surgical treatment outcomes in TLE patients [2,6,23,26]. In such machine learning-based predictive models, features from neuroimaging data are frequently employed as predictors, primarily due to their massiveness and superior predictive values compared with those from other clinical data [9]. Features of brain function can be acquired from EEG [11,33] and functional MRI [35] data, whereas features of brain structure can be acquired from structural MRI (sMRI) [6,20,29] and diffusion-weighted MRI (dMRI) [1,26] data. In particular, since TLE often involves neuroanatomical damage beyond the epileptogenic zone [16], neuroanatomical features are likely to be useful as predictors.

In clinical applications of predictive models, although the detection of TLE, that is, the discrimination of TLE patients from healthy individuals appears to be a primary concern, the subdivision of TLE, that is, the classification between different clinical subgroups of TLE patients could be also a significant concern. The lateralization of the epileptogenic side is crucial for pre-surgical evaluation and the lateralization of TLE patients has been determined by adopting grey matter (GM) anatomical features from sMRI data [20] and white matter (WM) anatomical features from dMRI data [1]. Seizure frequency is another critical seizure characteristic that potentially contributes to the clinical severity of TLE, but the prediction of seizure frequency on the basis of neuroanatomical damage has not been thoroughly sought.

In this study, we examined whether subdivisions of TLE according to seizure frequency could be predicted using neuroimaging features of GM and WM anatomical damage in the framework of machine learning-based predictive modelling. Since there have been conflicting findings regarding the relationship between seizure frequency and neuroanatomical damage [21,24], we hypothesized that such neuroanatomical features may not be as promising for the subdivision of TLE as for the detection of TLE. In addition, we checked how closely seizure frequency-based subgroups would be matched with neuroanatomy-based subgroups when we clustered TLE patients on the basis of the neuroanatomical features.

Section snippets

Participants

Forty-two patients with TLE (age: 40.83 ± 10.06 years, 25 females) and 45 age- and sex-matched healthy controls (age: 41.18 ± 13.48 years, 23 females) participated in the study. The clinical diagnosis of TLE was made according to the criteria of the International League Against Epilepsy [13]. Age at epilepsy onset was 23.07 ± 12.15 years and the duration of epilepsy was 17.98 ± 12.82 years. The healthy controls had no history of neurological disorders and no abnormalities in brain sMRI. Written

Clinical characteristics

All the TLE patients underwent anti-epileptic drug (AED) treatment. The TLE patients with high seizure frequency were drug-resistant, and thus they were considered for epilepsy resective surgery although their surgical procedures and outcomes have not been thoroughly followed up. The number of AEDs differed depending on seizure frequency (p <  0.01), but not on seizure lateralization, at the significance level of a p value of 0.05 or less. According to the sMRI findings of the TLE patients,

Discussion

In this study, we employed machine learning, specifically the random forests method, for the subdivision of TLE, that is, the separation between the TLE patients with low and high seizure frequency as well as for the detection of TLE, that is, the discrimination of the TLE patients from the healthy controls, using the neuroanatomical features as predictors. We demonstrated that classification performance in TLE subdivision was not as adequate as in TLE detection. Furthermore, when we determined

Author contributions

S.H.O. was responsible for the study concept and design. C.P. analyzed the data and drafted the manuscript. All authors critically reviewed content and approved final version for publication.

Funding

This work was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science and ICT in Korea [2015R1C1A1A01052438].

Conflict of interest

The authors have no conflict of interest to declare.

References (35)

  • R. Armañanzas et al.

    Machine learning approach for the outcome prediction of temporal lobe epilepsy surgery

    PLoS One

    (2013)
  • J.D. Banfield et al.

    Model-based Gaussian and non-Gaussian clustering

    Biometrics

    (1993)
  • W.B. Barr et al.

    Bilateral reductions in hippocampal volume in adults with epilepsy and a history of febrile seizures

    J. Neurol. Neurosurg. Psychiatry

    (1997)
  • B.C. Bernhardt et al.

    Magnetic resonance imaging pattern learning in temporal lobe epilepsy: classification and prognostics

    Ann. Neurol.

    (2015)
  • R.S. Briellmann et al.

    Hemicranial volume deficits in patients with temporal lobe epilepsy with and without hippocampal sclerosis

    Epilepsia

    (1998)
  • E. Burke Quinlan et al.

    Neural function, injury, and stroke subtype predict treatment gains after stroke

    Ann. Neurol.

    (2015)
  • F. Cendes et al.

    Atrophy of mesial structures in patients with temporal lobe epilepsy: cause or consequence of repeated seizures?

    Ann. Neurol.

    (1993)
  • Cited by (5)

    • Utility of hybrid PET/MRI multiparametric imaging in navigating SEEG placement in refractory epilepsy

      2020, Seizure
      Citation Excerpt :

      Hybrid PET/MRI systematically offers a complementary combination of two modalities that have often proven to be superior to the single modality approach in the diagnostic work-up of many neurologic and psychiatric diseases. Moreover, simultaneous PET/MRI is well-suited for complex brain function studies in which fast brain signal fluctuations need to be monitored on multiple levels in combination with fMRI [21,27–29]. Initial simultaneous studies have already demonstrated that these complementary brain function measures can provide new insights into the functional and structural organization of the brain [6–8,10,30,31].

    View full text