A VALIDATION STUDY OF COMPUTER-BASED DIAGNOSTIC ALGORITHMS FOR CHRONIC DISEASE SURVEILLANCE
MetadataShow full item record
Background: Chronic conditions comprise a significant amount of healthcare utilization. For example, people with chronic diseases account for 51% of family physician encounters. Therefore, diagnostic algorithms based on comprehensive clinical records could be a rich resource for clinicians, researchers and policy-makers. However, limitations such as misclassification warrant the need for examining the accuracy of these algorithms. Purpose: To investigate and enhance the accuracy of the diagnostic algorithms for five chronic diseases in the Canadian Primary Care Sentinel Surveillance Network. Methods: DESIGN: A validation study using primary chart abstraction. SETTING: A stratified random sample of 350 patient charts from Kingston practice-based research network. OUTCOME MEASURES: Sensitivity and specificity for the diagnostic algorithms. ANALYSIS: A multiple logistic regression model along with the receiver operating characteristic curve was employed to identify the algorithm that maximized accuracy measures. Results: The sensitivities for diagnostic algorithms were 100% (diabetes), 83% (hypertension), 45% (Osteoarthritis), 41% (COPD), and 39% (Depression). The lowest specificity was 97% for depression. A data-driven logistic model and receiver-operating characteristic curve improved sensitivity for identifying hypertension patients from 83% to 88% and for osteoarthritis patients from 45% to 81% with areas under the curve of 92.8% and 89.8% for hypertension and osteoarthritis, respectively. Conclusion: The diagnostic algorithms for diabetes and hypertension demonstrate adequate accuracy, thus allowing their use for research and policy-making purposes. A multivariate logistic model for predicting osteoarthritis diagnosis enhanced sensitivity while maintaining high specificity. This approach can be used towards further refining the diagnostic algorithms for other chronic conditions.