The Detection of Invisible Abnormal Metabolism in the FDG-PET Images of Patients With Anti-LGI1 Encephalitis by Machine Learning.

Objective:This study aims to detect the invisible metabolic abnormality in PET images of patients with anti-leucine-rich glioma-inactivated 1 (LGI1) encephalitis using a multivariate cross-classification method. Methods:Participants were divided into two groups, namely, the training cohort and the testing cohort. The training cohort included 17 healthy participants and 17 patients with anti-LGI1 encephalitis whose metabolic abnormality was able to be visibly detected in both the medial temporal lobe and the basal ganglia in their PET images [completely detectable (CD) patients]. The testing cohort included another 16 healthy participants and 16 patients with anti-LGI1 encephalitis whose metabolic abnormality was not able to be visibly detected in the medial temporal lobe and the basal ganglia in their PET images [non-completely detectable (non-CD) patients]. Independent component analysis (ICA) was used to extract features and reduce dimensions. A logistic regression model was constructed to identify the non-CD patients. Results:For the testing cohort, the accuracy of classification was 90.63% with 13 out of 16 non-CD patients identified and all healthy participants distinguished from non-CD patients. The patterns of PET signal changes resulting from metabolic abnormalities related to anti-LGI1 encephalitis were similar for CD patients and non-CD patients. Conclusion:This study demonstrated that multivariate cross-classification combined with ICA could improve, to some degree, the detection of invisible abnormal metabolism in the PET images of patients with anti-LGI1 encephalitis. More importantly, the invisible metabolic abnormality in the PET images of non-CD patients showed patterns that were similar to those seen in CD patients.