Multicenter validation of [18F]-FDG PET and support-vector machine discriminant analysis in automatically classifying patients with Amyotrophic Lateral Sclerosis versus controls

Objective: 18F-FDG PET single-center studies using support vector machine (SVM) approach to differentiate ALS from controls have shown high overall accuracy on an individual patient basis using local a priori defined classifiers. The study aim is to validate the SVM accuracy on a multicentric level. Methods: A previously defined Belgian (BE) group of 175 ALS patients (61.9±12.2 years, 120M/55F) and 20 screened healthy controls (62.4±6.4 years, 12M/8F) was used to classify another large dataset from Italy (IT), consisting of 195 patients (63.2±11.6 years, 117M/78F) and 40 control subjects (62±14.4 years; 29M/11F) free of any neurological and psychiatric disorder who underwent whole-body 18F-FDG PET-CT and that were found negative for lung cancer and for any evidence of paraneoplastic symptoms. 18F-FDG within-center group comparisons based on SPM (Statistical Parametric Mapping) were performed and SVM classifiers based on the local training sets were applied to differentiate ALS from controls from the other center. Results: SPM group analysis showed only minor differences between both ALS groups, indicating pattern consistency. SVM using BE data set as training, classified 183/193 IT ALS correctly (accuracy of 94.8%). However, 35/40 IT CON were misclassified as ALS (accuracy 12.5%). Furthermore, using IT data set as training, BE ALS could not be distinguished from BE CON. Within-center SPM group analysis confirmed prefrontal hypometabolism in IT versus BE CON, indicating subclinical brain changes in patients undergoing oncological scanning. Conclusion: This multicenter study confirms that the 18F-FDG ALS pattern is stable across centers. Furthermore, it highlights the importance of a carefully selected control group, as subclinical frontal changes might be present in patients in an oncological setting.