Neural network and logistic regression diagnostic prediction models for giant cell arteritis: development and...

Area of Science:

• Rheumatology
• Medical Informatics
• Biostatistics

Background:

• Giant cell arteritis (GCA) is a vasculitis requiring timely diagnosis.
• Accurate diagnostic prediction models are crucial for patient management.
• Temporal artery biopsy (TABx) is the gold standard but invasive.

Purpose of the Study:

• To develop and validate diagnostic prediction models for suspected GCA.
• To compare the performance of neural network (NN) and logistic regression (LR) models.
• To assess model accuracy in identifying biopsy-proven GCA.

Main Methods:

• Retrospective chart review of 1,201 patients undergoing TABx for suspected GCA across 14 centers.
• Utilized clinical and laboratory variables including age, ESR, CRP, and symptoms.
• Developed and validated NN and LR models using training, validation, and testing datasets.

Main Results:

• Both NN and LR models showed comparable diagnostic performance (AUC: 0.867 for LR vs. 0.860 for NN).
• The NN model achieved a lower misclassification rate (18.1% vs. 20.6%) and false-negative rate (30.5% vs. 47.5%) compared to LR.
• Multivariable LR identified age, platelets, jaw claudication, vision loss, CRP, ESR, headache, and clinical temporal artery abnormality as significant predictors.

Conclusions:

• Statistical models, particularly NN, can assist in triaging patients with suspected GCA.
• While misclassification is a concern, the NN model offers improved accuracy.
• Provided cutoff values for 95% and 99% sensitivities to aid clinical decision-making.