Introduction:
ALS drug development is plagued by high clinical trial failure rates. Subgroup analysis is a key tool used to account for patient heterogeneity, but current methods fall short. DEC analysis systematically groups and analyzes patients based on predicted disease path, creating more homogeneous patient subgroups with reduced noise around the endpoint.
Methods:
A multivariate machine-learning model trained using PRO-ACT was used to rank order trial participants by predicted disease progression. Fifty initial subgroups were expanded by adjusting prediction thresholds in 2% increments until the FAS was reached. A matrix was plotted in which each block had distinct upper and lower thresholds. A series of analyses were performed to assess variance (RMSE), treatment effect (TE), effect size, and P-value, thus developing a series of heat maps that revealed subgroups with favorable conditions for detecting a significant effect. The method was applied to the Ceftriaxone-ALS and Topiramate-ALS data sets available from the US NINDS.
Results:
We used the 285 patients in the Ceftriaxone-ALS dataset who remained on study for one year, which included 190 treated and 95 placebo patients. We randomly separated the 190 treated patients into two groups, one for the exploratory analysis and a second for hypothesis testing. One-year predictions using our validated percent expected vital capacity model were made for all patients in the dataset. A broad central region, a hot-spot, where moderately progressing patients localized, was detected and a subgroup, representing predicted 15% to 25% one-year decline in percent expected vital capacity was selected to determine whether the subgroup could be detected in the test set. Examination of the test group confirmed the results of the exploratory analysis.
The topiramate trial reported a negative TE for the primary endpoint. Similarly, DEC analysis showed broad zones of negative TEs. This experiment provides a strong negative control for DEC analysis.
Conclusions:
DEC analysis organizes trial participants in an unbiased way into homogeneous subgroups:
•Reveals “hot-spots” of detectable TEs that could form the basis for a subsequent successful trial
Numerous ways to implement this approach are envisioned:
•“Rescue” of drugs that failed late-stage clinical development
•All-comers trials to identify patients with detectable effects that can be seamlessly expanded into a fully powered trial