Mihai Lazarescu | PubMed Health

PubMed2006

Dynamic knowledge validation and verification for CBR teledermatology system.

OBJECTIVE: Case-based reasoning has been of great importance in the development of many decision support applications. However, relatively little effort has gone into investigating how new knowledge can be validated. Knowledge validation is important in dealing with imperfect data collected over time, because inconsistencies in data do occur and adversely affect the performance of a diagnostic system. METHODS: This paper consists of two parts. First, it describes methods that enable the domain expert, who may not be familiar with machine learning, to interactively validate knowledge base of a Web-based teledermatology system. The validation techniques involve decision tree classification and formal concept analysis. Second, it describes techniques to discover unusual relationships hidden in the dataset for building and updating a comprehensive knowledge base, because the diagnostic performance of the system is highly dependent on the content thereof. Therefore, in order to classify different kinds of diseases, it is desirable to have a knowledge base that covers common as well as uncommon diagnoses. RESULTS AND CONCLUSION: Evaluation results show that the knowledge validation techniques are effective in keeping the knowledge base consistent, and that the query refinement techniques are useful in improving the comprehensiveness of the case base.

Dermatology↗

PubMed2006

A comparison of algorithms for calculating glaucoma change probability confidence intervals.

PURPOSE: To evaluate the ability to detect change in standard automated perimetry data using 4 different methods for calculating the glaucoma change probability (GCP). METHODS: A database of stable visual fields, collected within 1 week from 35 glaucoma patients and within 6 months from 15 normal patients, was used to determine confidence intervals for GCP using 4 different methods. The methods classified visual field locations on the basis of either defect or mean threshold, and used test-retest data or baseline-less-follow-up data to determine values for the confidence intervals. The specificity of the 4 methods was measured using 3700 locations artificially generated to simulate stable visual field data. The sensitivity of the methods was measured using 3330 artificially generated locations that decreased in either a linear, curvilinear, or bi-linear fashion by 2, 3, or 4 dB per year on average. RESULTS: Using GCP with confidence intervals built using the methods described in the literature (on the basis of defect and test-retest differences) resulted in a higher specificity than techniques based on mean threshold. However, the mean-based methods were more sensitive at detecting a decrease in a location. Building confidence intervals using the difference between a baseline and the current measurement (baseline-less-follow-up), rather than test-retest differences, also improved the detection of visual field progression. CONCLUSIONS: Stratifying baseline visual field measurements based on defect and eccentricity as described in the literature results in an unusually high specificity: 98% accuracy in classifying the same stable data that generated the 95% confidence intervals, rather than the expected 95% accuracy. By stratifying measurements based on mean threshold, and using baseline-less-follow-up rather than test-retest differences to build 95% confidence intervals, sensitivity is increased by 14.1%. This increase in sensitivity comes with a corresponding 2.2% decrease in specificity.

Adult↗