Lead Investigator: Christian Dansereau, Perceiv AI
Title of Proposal Research: Validation of prognostic machine learning tools for forecasting clinical progression to enrich Alzheimer’s disease trials
Vivli Data Request: 9236
Funding Source: Perceiv Research Inc. will fund the proposed research.
Potential Conflicts of Interest: Christian Dansereau is a full-time employee and holds stocks/stock options with Perceiv Research Inc. We will declare all conflicts of interest in any subsequent publications. Angela Tam is a full-time employee and holds stocks/stock options with Perceiv Research Inc. We will declare all conflicts of interest in any subsequent publications. César Laurent is a full-time employee and holds stocks/stock options with Perceiv Research Inc. We will declare all conflicts of interest in any subsequent publications.
Summary of the Proposed Research:
It is estimated that around 50 million people currently live with dementia worldwide and this number will reach 152 million by 2050. Alzheimer’s disease is a neurodegenerative disorder and a leading cause of dementia as it is believed to account for 60-80% of cases. Over 90% of trials in Alzheimer’s disease have failed to demonstrate the effectiveness of the tested treatments. Key contributing factors to the failures of trials include individual heterogeneity (differences) in clinical progression and suboptimal patient selection. Clinical progression (progress of the disease) across Alzheimer’s disease patients is highly variable as some patients will decline very quickly and others will remain relatively stable for a long time. The inadvertent enrollment of patients who will remain stable into a trial is problematic because not only does it expose patients who are unlikely to benefit from treatment to unnecessary risks, but it also reduces a trial’s potential to detect treatment effects. Alzheimer’s disease trials therefore need enrichment strategies to select the most appropriate participants, individuals who will likely decline within the trials’ durations.
We propose to use prognostic machine learning models that forecast disease progression to select participants who will likely decline within the typical trial duration as an enrichment strategy. We have previously trained and validated models on individuals with mild cognitive impairment (decline in mental abilities) and Alzheimer’s disease dementia who participated in observational studies (i.e. research that did not test potential treatments). With these models, we demonstrated enrichment with likely decliners can allow for substantial reductions in the number of patients that need to be recruited compared to standard clinical trial recruitment while preserving the same potential to detect treatment effects. We would like to validate these models on clinical trial subject groups to assess their generalizability, to quantify the impact of enriching trial samples with likely decliners, and to compare treatment efficacy in subgroups of likely decliners against the total patient populations.
Better patient selection is important for accelerating therapeutic development in Alzheimer’s disease. Prognostic models for enrichment can give trials higher chances for success by enhancing their power to detect treatment effects in the most appropriate patient populations.
Statistical Analysis Plan:
We have previously trained models to forecast clinical progression in early Alzheimer’s disease (AD) populations. These models predict whether an individual will decline or remain stable in a given time window on the CDR sum of boxes (CDR-SB), a common outcome in AD clinical trials. We are requesting clinical trial studies in AD populations to validate our models on unseen data and to assess the utility of enriching trials with individuals who are at high risk of clinical decline. The requested studies were chosen because they represent phase 2 and 3 placebo-controlled clinical trials in AD populations and populations at high risk for AD dementia that measured change on CDR-SB as an outcome over a time period of at least 12 months.
In each requested study’s placebo arm, we will assess the generalizability of our models to predict decliners and non-decliners with area under the curve (AUC). We plan to impute missing data with various imputation strategies. We will measure the mean change from baseline on the CDR-SB in all individuals in the placebo groups and then in individuals predicted as decliners in the placebo groups. Based on the mean progression on the CDR-SB in the overall placebo groups and the subgroups of decliners within the placebo groups, we will perform power analyses to estimate required sample sizes for trials that would enrich for decliners and for trials without such an enrichment strategy.
Then, in each requested study, we plan to assess treatment efficacy as defined by the difference between placebo and treatment arms on change from baseline on the CDR-SB with either mixed models for repeated measures (MMRM) or analyses of covariance (ANCOVA). For studies that repeatedly measured CDR-SB across scheduled visits, MMRM will be used, where we will include terms for: baseline score, site, treatment, visit, treatment-by-visit interaction, concomitant use of a cholinesterase inhibitor/memantine at baseline, gender, and age at baseline. For studies that measured CDR-SB only at baseline and end of study, ANCOVA models will be used, and we will include terms for baseline score, site, treatment, concomitant use of a cholinesterase inhibitor/memantine at baseline, gender, and age at baseline. For each requested study, the null hypothesis is that the difference in least squares mean between the treatment and placebo arms equals zero.
Based on the predictions of our pre-trained models, we will stratify each requested study’s population into likely decliners and likely non-decliners in all placebo and treatment arms. We will then perform subgroup analyses separately on the likely decliners and the likely non-decliners. We will assess treatment efficacy separately on each of these subgroups in each requested trial with MMRM and/or ANCOVA as previously described to see if treatment efficacy differed between the subgroups.
Requested Studies:
Effect of Passive Immunization on the Progression of Mild Alzheimer’s Disease: Solanezumab (LY2062430) Versus Placebo
Data Contributor: Lilly
Study ID: NCT01900665
Sponsor ID: 15136
Continued Efficacy and Safety Monitoring of Solanezumab, an Anti-Amyloid β Antibody in Patients With Alzheimer’s Disease
Data Contributor: Lilly
Study ID: NCT01127633
Sponsor ID: 11935
Effect of Passive Immunization on the Progression of Alzheimer’s Disease: LY2062430 Versus Placebo
Data Contributor: Lilly
Study ID: NCT00904683
Sponsor ID: 11934
Effect of LY2062430, an Anti-Amyloid Beta Monoclonal Antibody, on the Progression of Alzheimer’s Disease as Compared With Placebo
Data Contributor: Lilly
Study ID: NCT00905372
Sponsor ID: 6747
Effect of LY450139 a y-Secretase Inhibitor, on the Progression of Alzheimer’s Disease as Compared With Placebo
Data Contributor: Lilly
Study ID: NCT00762411
Sponsor ID: 11271
Effect of γ-Secretase Inhibition on the Progression of Alzheimer’s Disease: LY450139 Versus Placebo
Data Contributor: Lilly
Study ID: NCT00594568
Sponsor ID: 7666
A Phase 2 Multiple Dose, Multicenter, Randomized, Double-Blind, Placebo-Controlled Study to Evaluate the Efficacy and Safety of ABBV-8E12 in Subjects With Early Alzheimer’s Disease
Data Contributor: AbbVie
Study ID: NCT02880956
Sponsor ID: M15-566
A Double Blind, Randomized, Placebo Controlled, Parallel Group Study to Simultaneously Qualify a Biomarker Algorithm for Prognosis of Risk of Developing Mild Cognitive Impairment Due to Alzheimer’s Disease (MCI Due to AD) and to Test the Safety and Efficacy of Pioglitazone (AD-4833 SR 0.8 mg QD) to Delay the Onset of MCI Due to AD in Cognitively Normal Subjects
Data Contributor: Takeda
Study ID: NCT01931566
Sponsor ID: AD-4833/TOMM40_301
Summary of results:
We did not have conclusive results for neither our power analyses to compare a typical trial population to an enriched sample of predicted progressors, nor our planned analyses to assess treatment efficacy on subgroups that were stratified by the model’s predictions. We had hypothesized that while each of the requested trials had negative results at the group level, it may have been possible that there was a subgroup of individuals in the treatment arm who responded particularly well, and we had hoped to identify these super responders. However, it is possible that the investigated drug in each trial had no impact on cognition and function. As a consequence, this may have confounded our analyses to identify potential super responders with the predictive model, and our results are therefore inconclusive. These inconclusive results might also have been caused by differences in data preparation, and our models were trained and evaluated in Perceiv AI’s infrastructure, while external evaluations were performed on the Vivli platform. We will take some time to further analyze these differences. While the results of our analysis are unlikely to yield publishable results, we greatly appreciate the work of Vivli and trial sponsors for sharing their data.