Premium
The use of statistical modeling to predict temporal artery biopsy outcome from presenting symptoms and laboratory results
Author(s) -
LEE M,
DE SMIT E,
WONG TEN YUEN A,
SAROSSY M
Publication year - 2014
Publication title -
acta ophthalmologica
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.534
H-Index - 87
eISSN - 1755-3768
pISSN - 1755-375X
DOI - 10.1111/j.1755-3768.2014.t097.x
Subject(s) - outcome (game theory) , medicine , temporal artery , biopsy , radiology , giant cell arteritis , disease , vasculitis , mathematics , mathematical economics
Purpose Giant cell arteritis (GCA) is a vasculitis affecting the elderly that can cause blindness if untreated. The current diagnostic gold standard is a temporal artery biopsy (TAB) which requires theatre time and histopathology analysis, causing a delay in formal diagnosis. In this study we investigated whether various predictive statistical models could accurately predict the TAB result from presenting symptoms and baseline laboratory tests. Methods Clinical characteristics of 182 patients who underwent TABs performed at the Royal Victorian Eye & Ear Hospital (RVEEH) were analysed. Clinical data was extracted from patient files including demographics (age, gender), presence of typical GCA symptoms (headache, scalp tenderness, jaw claudication, anorexia, fatigue, malaise, fever, weight loss, visual disturbance and previous diagnosis of polymyalgia rheumatica), and laboratory tests (CRP, ESR and platelet count). Various predictive statistical models were fitted in R – Random Forest (RF) , Adaboost, Regression Tree (rpart) and Support Vector Machine (SVM) to fit the dataset which was segregated into test and train parts for cross‐validation. We used a binary classification where equivocal results were considered positive. Results Of the TABs, 65 were negative, 117 were “positive” (94 positive, 23 equivocal). SVM achieved the best classification accuracy at 100% (40/40 predicted correctly). Rpart, Adaboost and RF achieved 65, 65 and 67.5% respectively. Conclusion Our statistical model using open source software has shown a high accuracy in predicting the TAB outcome based on presenting features. This could inform the clinician of the likelihood of TAB positive GCA and hence guide immediate management.