An Information Visualization Approach to Classification and Assessment of Diabetes Risk in Primary Care

Chronic disease risk assessment is a common information processing task performed by primary care physicians. However, efficiently and effectively integrating information about many risk factors across many patients is cognitively difficult. Methods for visualizing multidimensional data may augment risk assessment by providing reduced-dimensional displays which classify patient data. This study develops a framework which combines medical evidence, statistical dimensionality reduction techniques, and information visualization to develop visual classifiers for the task of diabetes risk assessment in a population of patients. The framework is evaluated in terms of classification accuracy and medical interpretation for two case studies, prediction of type 2 diabetes onset and prediction of heart attacks in adults with type 2 diabetes. These models are instantiated and tested using a unique health information database from the American Diabetes Association and gold standard risk predictions made by the Archimedes model. Results suggest that the visual models approximate the gold standard predictions and are comparable to commonly used classification methods. In addition, the methods provides rich visualizations of a patient population that contextualize the classification problem, giving insight into (i) the relative importance of many individual risk factors, (ii) confidence in individual patient predictions and (iii) overall distributions of risk in the population. The framework is based on computationally efficient methods and its parameters can be modified to meet the needs of individual physicians with different patient populations. These models may be embedded in existing health information systems to provide interactive visual analysis tools that support physician decision making for chronic disease prevention and management.