Integrating multiple-domain rules for disease classification

In psychiatry, clinicians use criteria sets from the Diagnostic and Statistical Manual of Mental Disorders to diagnose mental disorders. Most criteria sets have several symptom domains, and in order to be diagnosed, an individual must meet the minimum number of symptoms required by each domain. Some efforts are now focused on adding biomarkers to these symptom domains to facilitate the detection of and highlight the neurobiological basis of psychiatric disorders. Thus, a new criteria set may consist of both clinical symptom counts in several domains and continuous biomarkers. In this paper, we propose a method to integrate classification rules from multiple data sources to estimate an optimal criteria set. Each domain-specific rule can be counts of symptoms, a linear function of symptoms, or even nonparametric. The overall classification rule is the intersection of these domain-specific rules. Based on examining the expected population loss function, we propose two iterative algorithms using either support vector machines or logistic regression to fit intersection rules consistent with the Diagnostic and Statistical Manual of Mental Disorders. In simulation studies, these proposed methods are comparable with the true decision rule. The methods are applied to the motivating study to construct a criteria set for complicated grief. The developed criteria set shows a substantial improvement in sensitivity and specificity compared to the current standards on an independent validation study.