California Scientific

California Scientific
1000 SW Powell Ct
Oak Grove, MO 64075
Sales@CalSci.com
916-225-5119
California Scientific  *  BrainMaker Neural Network Software  *  Predict Forecast Classify Stocks Bonds Markets Commodities Diagnose Medical

BrainMaker Neural Network Software

Diagnosing Giant Cell Arteritis with BrainMaker

Five doctors have trained a neural network using the American College of Rheumatology (ACR) database of patients with vasculitis. The ACR has developed standards for classifying a number of rheumatic diseases. In addition to traditional classification approaches, other methods have been used such as decision trees, linear discriminant function analysis, logistic regression, and neural networks.

For the classification of Giant Cell Arteritis (GCA) of patients in the ACR database, all approaches have been used and compared. BrainMaker was trained on this set of patients, with the ACR diagnosis standards for comparison reasons. The inputs to the neural network were eight ACR predictor variables: 1) age greater than 50, 2) new localized headache, 3) temporal artery tenderness or decrease in a temporal artery pulse, 4) polymyalgia rheumatica, 5) abnormal artery biopsy, 6) erythrocyte sedimentation rate greater than 50mm/hour, 7) scalp tenderness or nodules, and 8) claudication of the jaw, tongue or on swallowing. If the predictor variable was present, a 1 was input. If the variable was not present, a 0 was input. The output was a 1 if the patient was diagnosed as having GCA or a 0 if not.

There were 807 patients in the database, 214 with GCA and 593 with other forms of vasculitis. The 807 patients were broken into three groups for neural network design and testing. One group of 80 or 81 patients was set aside for testing. A second group of 200 patients was set aside for monitoring the training (testing while training). A third group of 526 patients was set aside for training. Ten sets of these triplet groups were created using a different set of 80 or 81 patients each time. Ten different neural networks were trained on slightly different training groups.

After training, each network was tested on its corresponding testing set. In this way, the networks would test each and every case in the database without having seen the case during training. The trained networks correctly classified 94.4% of the testing cases that had GCA and 91.9% of the cases that did not have GCA.

This work was carried out by Michael L. Astion*, PhD, MD, Mark H. Wener, MD, Ron Thomas, MD, Daniel A. Bloch, PhD, Gene G. Hunder, MD. University of Washington, Dept of Lab Med., Seattle, WA

Inputs:

Age greater than 50
new local headache
temporal artery tenderness
polymyalgia rheumatica
abnormal artery biopsy
erthrtcyte sedimentation rate
scalp tenderness
claudication of jaw or tounge

Output:

arteritis diagnosis