Automatic Categorization of Health Indices for Risk Quantification

Abstract Classification of health data into categories is routinely used for the analysis and understanding of health risks; however, the selection of cut-off points of categories is not a simple task, and mistakes can lead to incorrect interpretation of data. Since inappropriate selection of the cut-off points can lead to unreliable and wrong conclusions, it is desirable to have an automatic method that balances the bias and the variance for constructing categories, and which allows the verification if the amount of available data is enough to draw a conclusion. Such a method is also useful in making decisions on next actions in experiment planning. We show here that a better formulation of cut-off point estimation is required for health data involving wide comparability, and demonstrate how a different method for comparing categorizations can be applied to such data. Our method can help in automation of data analysis pipeline and in promotion of scientific discoveries from health data.