HOUSTON – A machine learning interpretation of presurgical MRI studies did a better job of predicting which patients would have a successful outcome after anterior temporal lobectomy for temporal lobe epilepsy than did commonly-used clinical indicators in a prospective cohort study.
Xiaosong He, PhD, and his associates used two different machine learning classification methods to find two markers for thalamocortical connectedness that best predicted a good surgical outcome for temporal lobe epilepsy (TLE) in a small sample of patients. They presented their findings during a poster session of the annual meeting of the American Epilepsy Society.
copyright graphicsdunia4you/Thinkstock
After selecting a variety of possible predictors and building a model using resting state functional MRI (rsfMRI) data from 48 patients, the investigators then validated the prediction accuracies with rsfMRI data from 8 patients.
In predicting which TLE patients would have a good surgical outcome, models built with machine learning techniques using rsfMRI functional connectivity values had sensitivity ranging from 80% to 89% and specificity ranging from 52% to 57%. By comparison, models using clinical predictors only had sensitivity of 66% to 83% and specificity of 29% to 33%.
Dr. He and his coauthors dichotomized the surgical outcome for 56 patients who underwent TLE surgery into good outcome (n = 35) for those achieving and Engel class I and poor outcome (n = 21, class II-IV) at 1 year post surgery. All patients had a 5-minute rsfMRI scan before surgery.
MRI has been helpful in elucidating the importance of thalamocortical network pathology in TLE. Dr. He and his associates had previously used rsfMRI to examine the strength of functional connectivity between thalamic regions and their corresponding cortical regions in patients with TLE. Analysis of rsfMRI data of “both the left and right TLE groups showed that compared to controls there was a pattern of decreased thalamocortical [functional connectivity] in multiple thalamic segments,” wrote Dr. He and his collaborators (Epilepsia. 2015;56[10]:1571-9).
For the validation cohort, the two measures of connectedness found most predictive of a good surgical outcome were degree centrality and eigenvector centrality. In the graph theory and network analysis used in mapping functional connectivity, centrality refers to how highly connected one node, or data point, is to other data.
In the present study, the investigators used the Automated Anatomical Labeling cortical parcellation map to identify 45 cortical regions of interest per hemisphere, for a total of 90 cortical regions. They built a matrix with five topological parameters (global efficiency, global clustering coefficient, degree centrality, betweenness centrality, and eigenvector centrality) and the 90 cortical regions, yielding 272 variables. When nine commonly-used clinical predictors of surgical outcome (age, gender, handedness, laterality of TLE, epilepsy onset age and duration, seizure focality, interictal-spike type, and the presence of hippocampal sclerosis) were included, the model was made up of 281 variables.
The investigators used two different machine learning classification methods, called support vector machine and random forest, to build models that included various combinations of the 281 variables based on data from the initial 48 patients. The models were then tested with data from the remaining 8 patients.
Of the 35 patients with a good outcome, 18 had a left-sided epileptogenic temporal lobe; for the 21 patients with a poor outcome, the left temporal lobe was epileptogenic in 8. The mean age was similar for both groups: 41.25 years in those with good outcome, and 38.58 years in those with a poor outcome. Age at epilepsy onset also was similar, with each group having had epilepsy for about 17 years at the time of surgery. A total of 15 of the 20 patients with good outcome had seizure focality, compared with 10 of the 11 with poor outcome. Of those with a good outcome, 29 had an ipsilateral interactive spike, while 15 of those with poor outcomes had an ipsilateral interactive spike.
Since the random forest model best predicted surgical outcomes in the small sample size tested, the investigators plan to further fine-tune the random forest parameters to increase the robustness of their model.
Dr. He reported no conflicts of interest.
On Twitter @karioakes