Adiwijaya BS, Lasser R, Sankoh AJ
Sage Therapeutics Inc
Objectives: SAGE-217 is a potent investigational GABAA receptor modulator at both synaptic and extrasynaptic receptor subtypes, and is currently being studied in patients with psychiatric disorders, including major depressive disorder (MDD) and postpartum depression. In a pivotal, double-blind, placebo-controlled clinical trial of 89 patients with moderate to severe MDD, administration for 14 days with SAGE-217 was associated with a statistically significant mean reduction in the Hamilton Rating Scale for Depression (HAMD) 17-Item total score. The objective of this analysis was to evaluate alternative models of HAMD total scores and their predictions of alternative treatment durations, with emphasis on the comparison of prediction accuracy of models that quantify total score directly (TS method) to models that quantify each of the 17 component scores (CS method).
Methods: Baseline (Day 0 [D0]) and on-treatment HAMD assessments on Days 1-7, and 14 from 89 patients (44 placebo, 45 SAGE-217) were used to evaluate alternative methods. Two mixed-effect models were tested: Model 1 (primary): HAMD~ cumulative concentration (log-transformed), and Model 2 (secondary): HAMD ~ time (log-transformed). Regression estimates were corrected by the placebo effects in each timepoint. In the TS method, HAMD total scores were modeled as continuous variables. In the CS method, HAMD component scores were modeled as ordered categorical variables.
Prediction accuracy was assessed by comparing root mean square error (RMSE) of observed and predicted HAMD total score in the test set, from the model estimated from the training set. Three cases of training/test sets were evaluated: 1) training: D0-7, test: D14; 2) training: D0-6, test: D7 & D14; and 3) training: 80%, test: 20% (5-fold cross-validation). Comparisons of TS and CS methods were performed by evaluating the model comparisons from the same training/test sets.
Results: Alternative methods to model HAMD scores were evaluated by comparing prediction accuracy for TS and CS methods using multiple model covariates and training/test conditions. Compared to the TS method, the CS method demonstrated higher accuracy in all conditions. The RMSE difference of HAMD1 total scores between TS and CS method was highest for Case 1 (training: D0-7, test: D14), with Model 1 RMSE of 4.67 and 4.16 for TS method and CS method, respectively. The RMSE difference was not as large for Case 3 (5-fold cross-validation), with Model 1 RMSE of 3.50 and 3.35 for TS and CS method, respectively. Evaluation of the predictions showed that TS method predicted a few scores that were less than zero (zero is the lower bound of the HAMD scores), while CS method predicted all nonnegative scores. Detailed evaluation of the model predictions in a few example subjects showed the benefits of CS method in providing accurate lower bounds to each component of HAMD score, resulting in more accurate predictions. Model predicted that the optimal treatment duration of SAGE-217 in MDD was 14 days.
Conclusions: HAMD scores in patients with MDD were predicted more accurately by quantifying each component of HAMD scores than by quantifying HAMD total scores directly. The results may permit more accurate study planning during clinical development.
Reference: PAGE 28 (2019) Abstr 8912 [www.page-meeting.org/?abstract=8912]
Poster: Drug/Disease Modelling - CNS