Zielsetzung Our aim is to predict the antibody status of patients with autoimmune encephalitis
in a fully automated, non-invasive way using machine learning. For this purpose, radiomics-based
machine learning models are developed on the basis of the patient's MRI images.
Material und Methoden Our study cohort consists of 98 patients with autoimmune encephalitis and known antibody
status. Antibodies were previously detected in 57 of the 98 patients, and no antibodies
were detected in the remaining 41 patients. We extracted 107 radiomic factors from
the corresponding MRI images and tested a total of 6 different machine learning algorithms.
Specifically, we used the Random forest algorithm, Naive Bayes, linear discriminant
analysis (LDA), lasso regression, ridge regression and a neural network to predict
antibody status. Each model was developed 100 times with new training data and subsequently
tested each time with new independent test data in order to accurately assess the
stability of the model results.
Ergebnisse Our results show that antibody status in patients with autoimmune encephalitis can
be determined with high accuracy using machine learning algorithms based on MRI images.
We obtained our best results with a lasso regression. Using independent test data,
our 6-feature model yielded a mean AUC of 95.0%, a mean accuracy of 89.2%, a mean
sensitivity of 89.2% and a mean specificity of 89.1%. Further algorithms, such as
the neural network we tested, also resulted in high discriminatory power. Thus, our
machine learning models show very high and stable performance in predicting antibody
status in autoimmune encephalitis diseases, i.e. in distinguishing patients with detectable
and undetectable antibodies.
Schlussfolgerungen Antibody status in patients with autoimmune encephalitis can be determined non-invasively
and fully automatically with high accuracy using machine learning algorithms.
