Accurately predicting bipolar disorder mood outcomes: implications for the use of electronic databases

Monitoring mental health treatment outcomes for populations requires an understanding as to which patient information is needed in electronic format and is feasible to obtain in routine care.

To examine whether bipolar disorder outcomes can be accurately predicted and how much clinical detail is needed to do so. RESEARCH DESIGN, DATA SOURCES, AND PARTICIPANTS: Longitudinal study of bipolar disorder patients treated during 2000 to 2004 in the 19-site Systematic Treatment Enhancement Program for Bipolar Disorder observational study arm (N=3168). Clinical data were obtained at baseline and quarterly for over 1 year. We fit a "gold standard" longitudinal random-effects regression model using a detailed clinical information and estimated the area under the receiver operating characteristic curve (AUC) to predict accuracy using a validation sample. The model was then modified to include patient characteristics feasible in routinely collected electronic data (eg, administrative data). We compared the AUCs for the "limited-detail" and gold standard models, testing for differences between the AUCs using the validation sample.

Remission, defined as Montgomery-Asberg Depression Rating Scale score <5 and Young Mania Rating Scale score <4.

The gold standard models had baseline AUC=0.80 (95% confidence interval=0.74 to 0.86) and 0.75(0.64 to 0.86) at 1-year follow-up. The predicted accuracies of the limited-detail model were lower at baseline [AUC=0.67(0.60 to 0.75)]; correlated test χ=14.25, P=0.002] and not statistically different from the gold standard model at 1 year [AUC=0.67(0.54-0.80); correlated test χ=2.88, P=0.090].

Future work is needed to develop clinically accurate and feasible models to predict bipolar disorder outcomes. Clinically detailed and limited models performed similarly for shorter-term prediction at 1-year; however, there is room for improvement in prediction accuracy.