Authors: Edsel Ing, Neil Miller, Angela Nguyen, Wanhua Su, Lulu Bursztyn, Meredith Poole, Vinay Kansal, Andrew Toren, Dana Albreiki, Jack Mouhanna, Alla Muladzanov, Mark Gans, Mikael Bernier, Dongho Lee, Colten Wendel, Claire Sheldon, Marc D. Shields, Lorne Bellan, Matthew Lee-Wing, Yasaman Mohadjer, Navdeep Nijhawan, Felix Tyndel, Arun Sundaram, John Chen, Amadeo Rodriguez, Angela Hu, Nader Khalidi, Royce Ing, Samuel W. K. Wong, Martin ten Hove, Nurhan Torun

Author Disclosure Block: E. Ing: None. N. Miller, Angela Nguyen: None. W. Su: None. L. Bursztyn: None. M. Poole: None. V. Kansal: None. A. Toren: None. D. Albreiki: None. J. Mouhanna: None. A. Muladzanov:None. Mark Gans: None. M. Bernier: None. D. Lee: None. C. Wendel: None. C. Sheldon: None. M.D. Shields: None. L. Bellan: None. Matthew Lee-Wing: None. Y. Mohadjer: None. N. Nijhawan: None. Felix Tyndel: None. Arun Sundaram: None. J. Chen: None. A. Rodriguez:None. Angela Hu:None. Nader Khalidi: None. R. Ing: None. S.W.K. Wong: None. M. ten Hove: None. N. Torun: None.

Abstract Body:

Purpose: Oculoplastic surgeons are frequently called upon to perform temporal artery biopsy (TABx) for patients with suspected giant cell arteritis (GCA). However, TABx is an invasive procedure with a median utility rate of 25%. We aimed to develop and validate neural network (NN) and logistic regression (LR) diagnostic prediction models that might aid in the triage of patients with suspected GCA.
Study Design: Multicenter retrospective chart review
Methods: An audit of consecutive patients undergoing TABx was conducted at 14 international medical centers. The outcome variable was biopsy-proven GCA. The predictor variables were age, gender, headache (HA), clinical temporal artery abnormality (TAabn), jaw claudication (JC), vision loss (VL), diplopia, erythrocyte sedimentation rate (ESR), C-reactive protein (CRP), and platelet level. The data were divided into three groups to train, validate and test the models. The dataset split ratio was approximately 64:18:18. Geographic external validation was performed using the test set.
Results: Of 1,833 patients who underwent TABx, there was complete information on 1,201 patients, 300 (25%) of whom had a positive TABx. On multivariable, cluster LR age, platelets, JC, VL, log CRP, logESR, HA and TAabn were statistically significant predictors of a positive TABx (p<=.05). The area under the receiver operating characteristic curve (AUC)/Hosmer-Lemeshow p for LR was 0.867 (95% CI, 0.794, 0.917)/0.119 versus NN 0.860 (95% CI, 0.786, 0.911)/0.805, with no statistically significant difference of the AUC curves (p=0.316). The misclassification rate/false negative rate of LR was 20.6%/47.5% versus 18.1%/30.5 for NN. The cut-off values for 95% and 99% sensitivity are provided. Missing data analysis did not significantly change the results.
Conclusions: Statistical models can aid in the triage of patients with suspected GCA, and potentially increase the positive yield of TABx. Misclassification remains a concern for both models, but the NN had fewer false negatives than LR.