Validation of Preoperative Lesion Identification Algorithm in Pediatric Focal Cortical Dysplasia-related Epilepsy
Kara Hom1, Venkata Illapani2, Hua Xie2, William Gaillard2, Taha Gholipour3, Nathan Cohen2
1George Washington University School of Medicine, 2Children's National Hospital - Center for Neuroscience Research, 3UC San Diego Health - Comprehensive Epilepsy Center
Objective:

We evaluated MELD’s clinical utility using an independent sample.

Background:

Focal cortical dysplasia (FCD) is the commonest etiology of pediatric pharmacoresistant epilepsy (PRE). Surgery can be curative, but lesions can be poorly demarcated on MRI. The Multi-centre Epilepsy Lesion Detection (MELD) project designed an automated, machine-learning algorithm to segment FCD on preoperative MRIs.

Design/Methods:

Images from PRE, healthy control (HC), and “MRI-negative” subjects were segmented by MELD. PRE patients with lesional preoperative MRIs were also segmented manually; overlap between manual/MELD masks was measured by Dice Similarity Coefficient. Outcomes measured: Sensitivity: % patients with MELD/manual mask overlap; Specificity: % HC yielding no MELD clusters. In MRI-negative patients, resulting MELD clusters were compared to estimated seizure onset zones (SOZ) from presurgical data.

Results:

90 PRE, 19 HC and 17 MRI-negative patients were included. Twenty-two patients were excluded due to inadequate preprocessing. Sensitivity was 38.2%. Specificity was 36.8%. MELD accuracy by lobe (%): frontal, 46.2%; temporal, 26.9%; parietal, 11.5%; multilobar, 15.38%. MELD accuracy by FCD pathology: Type I, 23.1%; IIa, 23.1%; IIb, 42.3%; III, 7.7%; NA, 3.9%. Frontal and Type IIb FCDs were the most accurately masked. Temporal and Type I FCDs were the most inaccurately masked. Engel I outcomes were similar between accurate (65.4%) and inaccurate masks (72.4%). MELD localized lesions in eight MRI-negative patients (42%); four had resections with Engel I outcome.

Conclusions:

Our sensitivity, specificity, and MRI-negative detection were lower than the original MELD study (65%, 52%, 62.9%, respectively). Pathology influenced accuracy: well-defined Type IIb were detected more often than the more nebulous Type I FCDs. Compared to the original MELD study, our population included more temporal lesions and Type I FCD, which are harder to mask. Importantly, FCD detection in “MRI-negative” cases demonstrates potential for real-world SOZ prediction in nonlesional patients, improving diagnostic confidence, surgical decision making and rates of seizure free outcomes.

10.1212/WNL.0000000000204723