Roberta Vasta (Italy)
Magna Graecia University Neuroscience Research Center, Department of Medical and Surgical SciencesAuthor Of 1 Presentation
RANDOM-FOREST CLASSIFICATION OF PSYCHOGENIC NON-EPILEPTIC SEIZURES AND TEMPORAL LOBE EPILEPSY.
- Maria Eugenia Caligiuri (Italy)
Abstract
Background and Aims:
Psychogenic nonepileptic seizures (PNES) represent a multifactorial psychopathology, which makes diagnosis particularly challenging: PNES can be misdiagnosed as pharmaco-resistant temporal lobe epilepsy (TLE), and approximately 80% of subjects actually undergo anti-epileptic drug (AED) at the time of correct diagnosis. In this study, we used machine learning (ML) to differentiate PNES and TLE patients.
Methods:
Thirty-six PNES subjects and 43 demographically-matched TLE patients underwent neuropsychiatric/neuroimaging assessment. A 10,000-trees random forest (RF) algorithm, considered more robust to overfitting compared to other ML algorithms, was trained on T1-weighted MRI, i.e., on the entire set of morphological metrics obtained through FreeSurfer (cortical thickness, surface, volume, curvature, gyrification index). All features with a mean decrease in Gini index ≥ 0.30 were selected to construct a new classifier with the lowest out-of-bag error (OOB; accuracy = 100 - OOB).
Results:
Figure 1 shows the most important features discriminating PNES from TLE, ranked according to mean Gini Index decrease. This discriminant network included regions across all lobes of the brain, from parietal-occipital regions to frontal regions, as well as the anterior portion of the corpus callosum. Based on these selected features, the RF algorithm was able to distinguish PNES from TLE with an average accuracy of 77.2%.
Conclusions:
This work provides evidence that ML techniques could aid the differential diagnosis of PNES. Involvement of cingulate and orbitofrontal cortices, a frequent finding when comparing PNES to controls, also represented a distinctive feature from TLE patients. This finding supports the hypothesis that PNES subjects experience disrupted processing of emotional information, which might ultimately lead to the insurgence of seizure-like episodes.