This report evaluates whether classification tree algorithms (CTA) may improve the identification of individuals at risk for bipolar spectrum disorders (BPSD). Analyses used the Longitudinal Assessment of Manic Symptoms (LAMS) cohort (629 youth, 148 with BPSD and 481 without BPSD). Parent ratings of mania symptoms, stressful life events, parenting stress, and parental history of mania were included as risk factors. Comparable overall accuracy was observed for CTA (75.4%) relative to logistic regression (77.6%). However, CTA showed increased sensitivity (0.28 vs. 0.18) at the expense of slightly decreased specificity and positive predictive power. The advantage of CTA algorithms for clinical decision making is demonstrated by the combinations of predictors most useful for altering the probability of BPSD. The 24% sample probability of BPSD was substantially decreased in youth with low screening and baseline parent ratings of mania, negative parental history of mania, and low levels of stressful life events (2%). High screening plus high baseline parent-rated mania nearly doubled the BPSD probability (46%). Future work will benefit from examining additional, powerful predictors, such as alternative data sources (e.g., clinician ratings, neurocognitive test data); these may increase the clinical utility of CTA models further.
Keywords: bipolar disorder; children; classification tree analysis; clinical decision making; risk factors.