Extracting Temporal Knowledge

C. Cimino and S. A. Turken

We have attempted to examine the value of temporal knowledge for two physicians making predictions about blood glucose levels in diabetic patients. Our hypothesis is that physicians evaluating sequential data about a patient derive an abstract description of the patient that is not temporally related (atemporal knowledge). For example, a patient might be identified as a "brittle diabetic". Atemporal knowledge shouM allow a physician to make isolated predictions about a patient that are better than random chance. A combination of temporal and atemporal knowledge shouM allow a physician to make better predictions than atemporal knowledge alone. We used the AIM-94 diabetes datasets supplied by Dr. Michael Kahn. These datasets consist predominately of blood glucose measurements and insulin dosages.We failed to show any different in predictive ability between temporal and atemporal data, between the two physicians, and between the individual physicians and three different controls. We did show that by grouping all 20 physician cases (temporal and atemporal) their predictions had an average error of 69 mg/dl of blood glucose. These predictions were better than the average of one of the controls by 30 mg/dl (standard deviation of 6.5). While this is statistically significant, we do not feel it is clinically significant. It is our conclusion that physicians may not make accurate predictions based solely on insulin dosages and glucose measurements.

This page is copyrighted by AAAI. All rights reserved. Your use of this site constitutes acceptance of all of AAAI's terms and conditions and privacy policy.