Diagnosis and Subtyping of Autoimmune Encephalitis Using an Attention-Based Multi-Instance Learning Model: A...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Medical Diagnostics

Background:

• Autoimmune encephalitis (AE) presents diagnostic challenges, necessitating advanced tools for differentiation.
• ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) PET imaging offers potential for characterizing AE.
• Developing AI models can improve the accuracy of AE diagnosis and subtyping.

Purpose of the Study:

• To develop an attention-based model using ¹⁸F-FDG PET imaging for AE patient differentiation.
• To discriminate between different AE subtypes using the developed model.
• To compare the performance of the attention-based model against traditional algorithms.

Main Methods:

• A multi-center retrospective study included 390 participants (222 AE patients, 122 healthy controls, 33 antibody-negative AE, 13 viral encephalitis).
• An attention-based multi-instance learning (MIL) model was trained and externally validated.
• A multi-modal MIL (m-MIL) model integrating imaging, age, and sex was evaluated against logistic regression and random forest models.

Main Results:

• The m-MIL model achieved high accuracy in binary classification (AE vs. controls): 84.00% internal, 67.38% external.
• For AE subtype classification, the MIL-based model reached 95.05% internal and 77.97% external accuracy.
• Heatmap analysis revealed distinct brain region involvement patterns for NMDAR-AE, LGI1-AE, GABAB-AE, and GAD65-AE.

Conclusions:

• The m-MIL model effectively distinguishes AE patients from controls.
• The model enables accurate subtyping of different AE subtypes.
• This AI-driven approach serves as a valuable diagnostic tool for AE assessment and classification.