General Information
Conflict of Interest
- Laska-conflict-of-Interest-.pdf
- Siegel-conflict-of-interest.pdf
- Kim-conflict-of-interest.pdf
- Wanderling-conflict-of-interest.pdf
- Marmar-conflict-of-interest.pdf
Request Clinical Trials
Associated Trial(s): What type of data are you looking for?: Individual Participant-Level Data, which includes Full CSR and all supporting documentationRequest Clinical Trials
Data Request Status
Status: Approved Pending DUA SignatureResearch Proposal
Project Title: Prediction of Likely Responders to Topiramate to further personalized treatment for Alcohole Use Disorder
Scientific Abstract:
Background: Based on the goal of precision medicine to use “the right drug for the right patient at the right dose,” we seek to identify the characteristics of persons likely to respond (LR) to treatment for alcohol use disorder (AUD) with Topiramate. Topiramate is known to be efficacious in treating AUD, but there is considerable heterogeneity in treatment response. To understand the nature of the variability a large sample size is required.
Objective: to obtain a function that produces predictions of post-treatment outcomes based on demographic and other patient specific baseline and pretreatment phenotypic variables, and to test whether it is causally superior to placebo among this subgroup of participants in a RCT.
Study Design: This is a post-hoc reanalysis of three randomized clinical trials of topiramate versus placebo. The drinking outcomes from participants in three similar randomized clinical trials will be used to obtain the prediction function and perform the tests.
Primary and Secondary Outcome Measure(s): The primary outcome variable is the percentage of drinking days and the secondary measure is percentage of heavy drinking days.
Statistical Analysis: The data from the three RCTs of topiramate for treating AUD will be input to a random forest (RF) regression analysis to obtain a predictive model. The expected or predicted response to topiramate will be computed for patients and counterfactually for placebo patients. Those meeting prespecified criteria on the predictive outcomes are LRs. We will perform tests of causal superiority comparing the actual outcomes for the LRs who received topiramate to LRs who received placebo in groups matched on their expected response. We will fit the model on two of the trials and test the predictive accuracy in the third study for the three possible pairs of studies. The model will be refined as we learn about the variability of response in the sample population on which the model is fit. The final model will utilize all three studies. As the models are fit, we will identify the variables that are most influential in the prediction functions. The analysis, if there is good fit, will increas understanding of the heterogeneity of outcomes of treatment with topiramate and potentially enable a targeted approach to treating AUD promoting personalized medicine.
Brief Project Background and Statement of Project Significance:
The introduction of the LR method of analysis (Laska et al 2021, 2023) has increased the value of reanalyzing RCTs to determine whether predictive models can be built so that targeted treatment approaches can be identified. Laska and Siegel are the principal investigators of an NIAAA grant to study the statistical properties of the LR methodology and to analyze six RCTs, (that were funded by the Institute) of potential AUD treatments that have participant-level data available. Except for placebo, no two treatments in the six studies are the same, so validation of findings cannot be performed in an independent sample. Separately, we have obtained the individual participant data of two topiramate studies (N=278) conducted by Kranzler and colleagues (2014, 2021). The Johnson et al study (2007), which has a similar design as the two topiramate studies has 371 participants. This is a substantial increase in sample size and at the same time enables predictive accuracy to be evaluated on independent data sets. It is clear that before a predictive function can be accepted into clinical practice, testing in diverse samples is required. Successfully achieving our aims can begin to change the standard practice of treatment with topiramate and the method of analysis of RCTs. The reanalysis of the three topiramate studies can uncover clinically meaningful relationships between patient characteristics and likely outcomes furthering the goals of personalized medicine for AUD .
Specific Aims of the Project:
Aim 1. Utilize LR statistical and machine learning methodology to analyze three clinical trials of topiramate to obtain a predictive function of drinking outcomes based on pre-treatment patient characteristics
Aim 2. To identify individualized pre-treatment patient characteristics that are the most influential predictors of outcome.
Aim 3. Refine the statistical and machine learning methodology and investigate the performance of methods for appraising model misspecification.
Study Design: Meta-analysis (analysis of multiple trials together)
What is the purpose of the analysis being proposed? Please select all that apply.: New research question to examine treatment effectiveness on secondary endpoints and/or within subgroup populations Confirm or validate previously conducted research on treatment effectiveness Participant-level data meta-analysis Meta-analysis using data from the YODA Project and other data sources Develop or refine statistical methods Research on clinical trial methods Research on clinical prediction or risk prediction
Software Used: RStudio
Data Source and Inclusion/Exclusion Criteria to be used to define the patient sample for your study: The two topiramate studies are at NYU Langone School of Medicine under a data use agreement with Dr. Henry Kranzler. The participant data from the Johnson study will be shared with us by YODA. All individuals in each of the three trials will be included in the analysis in an intent to treat approach.
Primary and Secondary Outcome Measure(s) and how they will be categorized/defined for your study:
We will consider six drinking outcomes for participants in the trials. The outcomes included
1. measures of quantity of drinking (drinks per day and percentage change in drinks per day),
2. measures of frequency of drinking (percentage of drinking days and change in percentage of drinking days),
3. measures of frequency of heavy drinking days (percentage of heavy drinking days and percentage change in heavy drinking days).
Although we will apply the methodology to all measures, the primary measures are percentage of heavy drinking days and percentage change in heavy drinking days.
Main Predictor/Independent Variable and how it will be categorized/defined for your study:
Available in the Kransler studies: Sex, Age, Marital Status, Employment Status full-time or not, Number of years in school, Income range, Normalized depression score on PHQ-9, Cigarette smoker status, AUQ score, Alcohol Expectancy Measure (Less.Tense), Alcohol Expectancy Measure (Good.Time), Count (sum) of DSM-5 AUD symptoms, Age of onset of AUD, Number of standard drinks per day in 30 days prior to screening, Percentage of abstinent days in 30 days prior to screening, Percentage of Heavy Drinking days in the 30 days prior to screening, Baseline Short Index of Problems scale score, Baseline Short Form 12 score.
Available in the Johnson study (from the manuscript): assessment of (1) physical health (via physical examination, electrocardiogram), vital signs (blood pressure, pulse, and BMI), hematological
and biochemical screenings, urine tests (including urine drug screening), and a urine pregnancy test for women with childbearing potential); (2) withdrawal symptoms (CIWA-Ar); (3) depressed mood (MADRS); (4), personal and psychosocial harm from alcohol (Alcohol Use Disorders Identification Test)
We will need to harmonize the predictors across studies. Note that random forest has no difficulty "shaving" the number of predictors even when they number in the tens of thousands, without fear of overfitting.
Other Variables of Interest that will be used in your analysis and how they will be categorized/defined for your study: Will depend on what we find in the details but none are contemplated now.
Statistical Analysis Plan:
A random forest in regression mode will be used to develop a model to predict the outcome measures for patients treated with topiramate utilizing only patients who received the drug. The model will be based on baseline features that include demographic and pretreatment drinking characteristics. These variables need to be harmonized across studies. The task will be informed by colleagues who are experts in the field of AUD. Important predictors will be identified in the RF model based on the mean square error. Tuning parameters will be adjusted to maximize the goodness of fit. Outcome measures will be calculated using drinking data from the last 4 to 6 weeks of treatment. The resulting RF model will then be "scored" for all participants, including those receiving placebo. This is accomplished by running their baseline predictors down the tree in trees in which they were “out of bag”. Thus, no tree in which an individual was part of the bootstrap sample used to grow the tree is used in obtaining an estimate of the expected response to topiramate. This yields an estimated or predicted outcome for each subject on each outcome measure. Next, for each outcome, participants will be ranked by their predicted outcomes and grouped into quintiles. Within each quintile, patients in the topiramate and placebo groups are matched in that they have similar multivariate baseline feature distributions based on their predicted response to topiramate. This is analogous to matching on propensity scores in an observational study. As demonstrated in Laska et al. [18] this permits a causal treatment comparison of topiramate and placebo. An analysis of variance model will be used to test for overall treatment differences across the quantiles. The models included terms for LRs, treatment assignment, and quantile as factors and treatment-by-quantile interactions. Participants are deemed to be likely responders if their predicted outcome exceeds a clinically determined threshold that identifies a lower (or upper as appropriate) limit of a clinically meaningful response. For heavy drinking measures patients with a predicted reduction during treatment of greater than, for example, 50% in percentage reduction in HDDs (%ΔHDD) may be deemed likely responders. For the LR analysis, all patients who are LRs (topiramate or placebo) are ranked by their predicted outcomes and grouped into quantiles. As in the whole population analysis, in each quantile, participants in the topiramate and placebo groups are matched in that they have similar predicted outcomes and therefore baseline multivariate feature distributions. ANOVA for the pooled test for treatment differences will be performed in this LR group. We will fit the model on two of the trials and test the predictive accuracy in the third study for the three possible pairs of studies. The model will be refined as we learn about the variability of response in the sample population on which the model is fit. The final model will utilize all three studies. As the models are fit, we will identify the variables that are most influential in the prediction functions.
Narrative Summary: Most randomized clinical trials (RCTs) are designed to determine whether the mean responses to a test treatment, T, and a control treatment, P, are the same in a sample from a target population. Prognostic features are usually ignored unless there is an imbalance between the treatment groups as a chance consequence of the particular randomization. Based on the goal of precision medicine to use “the right drug for the right patient at the right dose,” in our analyses, we seek to identify the characteristics or features of persons likely to respond to T, and to test whether T is causally superior to P among this subgroup of participants in the RCT. Topiramate is known to be efficacious in treating alcohol use disorder (AUD), but there is considerable heterogeneity in treatment response. To understand the nature of the variability a large sample size is required. We will combine participant data from three similar clinical trials, two conducted by Kranzler et al (2014, 2021) and the third by Johnson et al (2007) to identify individuals most likely to benefit from topiramate, (likely responders, LRs) based on pretreatment features. We will use a recently developed LR causal inference method to test topiramate’s clinical efficacy in this group. The fact that there are three studies available means that the accuracy of the predictive models can be checked in independent data samples by fitting in two of the studies and validated in the third. The LR analysis has the potential to enable the use of an individualized approach to the treatment of AUD with Topiramate.
Project Timeline: This project will commence as soon as the Johnson et al study data is made available to us. The following timeline is an approximation. The first three months will be spent studying all of the documents, "cleaning" the data, handling missing data, and harmonizing the coding structure for items present in all three studies (e.g., males =1). In about the fourth month, when the level of missingness is known, we will harmonize scales that measure the same construct. For example, the Kranzler et al studies utilized the Normalized Depression Score on the PHQ-9, whereas the Johnson et al study utilized the MADRS for measuring depression. In months five through seven, the major LR analyses and the model validity checks will be carried out. Depending on the complexity of the findings, manuscript preparation should take about 3 months to be ready for submission for publication and results reported to the YODA Project.
Dissemination Plan: Clinical members of the team will present at professional AUD meetings. Because there is considerable interest in this area of research, particularly because it is focused on individualized therapeutics, there should be no problem publishing the results in high-impact journals. These include such possibilities as JAMA, JAMA Psychiatry and Alcoholism: Clinical and Experimental Research.
Bibliography:
Johnson BA, Rosenthal N, Capece JA, Wiegand F, Mao L, Beyers K, et al. Topiramate for treating alcohol dependence: a randomized controlled trial. Jama. 2007;298(14):1641-51.
Kranzler HR, Covault J, Feinn R, Armeli S, Tennen H, Arias AJ, et al. Topiramate treatment for heavy drinkers: moderation by a GRIK1 polymorphism. Am J Psychiatry. 2014;171(4):445-52.
Kranzler HR, Morris PE, Pond T, Crist RC, Kampman KM, Hartwell EE, et al. Prospective randomized pharmacogenetic study of topiramate for treating alcohol use disorder. Neuropsychopharmacology. 2021;46(8):1407-13.
Kranzler HR, Hartwell EE, Feinn R, Pond T, Witkiewitz K, Gelernter J, et al. Combined analysis of the moderating effect of a GRIK1 polymorphism on the effects of topiramate for treating alcohol use disorder. Drug Alcohol Depend. 2021;225:108762.
Laska EM, Siegel CE, Lin Z, Bogenschutz M, Marmar CR. Gabapentin Enacarbil Extended-Release Versus Placebo: A Likely Responder Reanalysis of a Randomized Clinical Trial. Alcohol Clin Exp Res. 2020;44(9):1875-84.
Laska E, Siegel C, Lin Z. A likely responder approach for the analysis of randomized controlled trials. Contemp Clin Trials. 2022;114:106688.