Prediction of treatment outcome in daily generalized mental healthcare practice: First steps towards personalized treatment by clinical decision support

Abstract

Rationale, aims and objectives: Results from prediction studies are often of limited value because potential predictors are measured with instruments that are not routinely used, the results are presented in terms that are difficult to translate to individual patients and the relations between predictors and outcome are complex. These problems were faced by deriving decision trees for classification using information routinely accessed in generalized mental healthcare for intake purposes and treatment monitoring. Method: Positive treatment outcome was defined as symptom improvement, measured with the symptom distress scale of the Outcome Questionnaire (OQ-45.2). The analysis process consisted of 3 phases, derivation of a possible decision tree, selection of the best 10 trees and assessment of classification performance after integration of these 10 trees. This analysis was performed 3 times, for all patients without any missing data, for the full set of patients and for the patients in which intermediate outcome data were available. Results: The prediction performance of the 3 integrated classifiers varied from poor (AUC 0.68) in the complete sample including patients with missing variables, to good (AUC 0.83) including early response as predictor. Complex interactions between variables were found. Conclusion: The present study shows the need for registration of clinical and sociodemographic variables and outcome monitoring in a systematic way to prevent missing variables and automated decision support systems to use complex interactions between variables for outcome prediction.