Machine Learning for MRI Classification of Systemic Lupus Erythematous Patients with and without Neuropsychiatric Events.
Authors
Affiliations (11)
Affiliations (11)
- School of Medicine and Surgery, University of Cagliari, Cagliari, Italy.
- Department of Radiology, Azienda Ospedaliero-Universitaria (A.O.U.), di Cagliari-Polo di Monserrato, Cagliari, Italy.
- Department of Medical Science and Public Health, University of Cagliari, Cagliari, Italy.
- Rheumatology Unit, Azienda Ospedaliero Universitaria Di Cagliari, 09042, Monserrato, Italy.
- IRCSS SDN, Naples, Italy.
- Stroke Monitoring and Diagnostic Division, AtheroPointâ„¢, Roseville, CA, 95661, USA.
- Department of Neuroscience, Imaging and Clinical Sciences, University of Chieti-Pescara, Chieti-Pescara, Italy.
- Division of Neuroimaging and Neurointervention, Stanford University Medical Center, Stanford, CA, USA.
- Department of Electrical and Electronic Engineering (DIEE), University of Cagliari, Piazza d'armi, Edificio M, Cagliari, Italy.
- Department of Radiology, Azienda Ospedaliero-Universitaria (A.O.U.), di Cagliari-Polo di Monserrato, Cagliari, Italy. [email protected].
- Department of Medical Science and Public Health, University of Cagliari, Cagliari, Italy. [email protected].
Abstract
To provide a useful and practical Machine Learning framework to facilitate the diagnosis of Neuropsychiatric Systemic Lupus Erythematosus (NPSLE) and Systemic Lupus Erythematosus (SLE) from Magnetic Resonance Imaging (MRI) derived features. Twenty-seven SLE patients (14 NPSLE, 13 SLE; 24 females and 3 males; average age:Â 43Â years, age range: 21 to 62) and 20 healthy controls (17 females and 3 males; average age 41, age range: 21 to 56), were included in this cross-sectional study. VolBrain online platform was used to quantitatively assess brain structural features (regional cortical thickness) which were used as input for the Machine Learning models. Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF) and XGBoost were trained and tested, using a fivefold cross validation in the process. The Random Forest model demonstrated superior performance with an accuracy of 90% on the test, distinguishing itself as the most effective classifier between the three classes without the need for L1 or L2 regularization. The most relevant volumetric features for the classification were Precentral gyrus right thickness norm, angular gyrus thickness asymmetry and Parietal thickness asymmetry. The Random Forest algorithm was the most effective model for classifying between NPSLE, SLE and controls patients in the context of this study, highlighting a potential path for clinical diagnosis support through machine learning.