Empirically derived decision trees for the treatment of late-life depression.

OBJECTIVE Several predictors of treatment response in late-life depression have been reported in the literature. The aim of this analysis was to develop a clinically useful algorithm that would allow clinicians to predict which patients will likely respond to treatment and thereby guide clinical decision making. METHOD A total of 461 patients with late-life depression were treated under structured conditions for up to 12 weeks and assessed weekly with the 17-item Hamilton Rating Scale for Depression (HAM-D-17). The authors developed a hierarchy of predictors of treatment response using signal-detection theory. The authors developed two models, one minimizing false predictions of future response and one minimizing false predictions of future nonresponse, to offer clinicians two clinically useful treatment algorithms. RESULTS In the first model, early symptom improvement (defined by the relative change in HAM-D-17 total score from baseline to week 4), lower baseline anxiety, and an older age of onset predict response at 12 weeks. In the second model, early symptom improvement represents the principal guide in tailoring treatment, followed by baseline anxiety level, baseline sleep disturbance, and--for a minority of patients--the adequacy of previous antidepressant treatment. CONCLUSIONS Our two models, developed to help clinicians in different clinical circumstances, illustrate the possibility of tailoring the treatment of late-life depression based on clinical characteristics and confirm the importance of early observed changes in clinical status.