A Novel Approach to Developing Comprehensive Profile of Adverse Event (AE) Data from Clinical Trials

Proposal
1595

Title of Proposed Research

Lead Researcher

Sharayu Paranjpe

Affiliation

Cytel Statistical Software and Services Pvt. Ltd

Funding Source

The funding will be provided by Cytel Statistical Software & Services Pvt. Ltd. ?“Cytel”?, an Indian company with whom I am a consultant on contract. The team members 2 to 7 are full time employees of Cytel.

Potential Conflicts of Interest

I have requested for data on four trials. The sponsor companies are Eli Lilly and Company, Hoffmann-La Roche, Boehringer Ingelheim and Glaxo-Smith-Kline. There is no conflict of interest. But as a matter of disclosure, I would like to state that all four sponsor companies are customers of Cytel's software products (East and StatExact) The sponsor company Hoffmann-La Roche is an existing client of Cytel for statistical and programming services. That relationship will not influence the conduct of the project or the outcome that we will document. None of the team members (including me) are part of the team working for Roche and we do not have direct/ indirect access to client data.

Data Sharing Agreement Date

20 September 2017

Lay Summary

Summary and analysis of AE data is crucial for judging suitability of a drug and for comparison between drugs. AE data from clinical trials is typically reported mostly as counts and percentages. In line with recommendation from ICH3 Guideline, only a few events of special interest are further analyzed and used for comparison of treatments. Clearly this is incomplete use of AE information. This happens because of the complex nature of data. First of all, variety of AEs is immense. Some are minor while others are severe. Some are clearly drug related while others are not. Some are instantaneous while others linger for some duration. Secondly the Same AE may occur in multiple subjects and a single subject may experience multiple AEs. There may be repeat episodes of same AE to same patient. There may be associations between occurrence/non-occurrence of two or more AEs. Total number of AEs experienced by group of patients in a trial is generally very large. All these issues make AE data complicated to comprehend as a whole. It would be very useful to devise statistical methods to extract more information from the complex AE data. The problem could use some 'out of the box' thinking. One way of getting hold of such ideas is to borrow from a different domain of application of statistics. Towards this my collaboration with ecologists during my tenure as University teacher has an advantage. I believe that some statistical methods widely used in analysis of ecological data can be fruitful in analysis of AE. The biodiversity measurement issue in ecology has complexity level of data comparable to AE data. My idea is to apply some of these methods to AE data to build a comprehensive picture.
Objectives: The main objective is to treat all AEs as a single entity and explore patterns in AE occurrence as a whole instead of one AE type in isolation. The emergent AE profile will help in comparison between drugs. I believe that with this approach some new features of the AE profile will emerge. Some of the indices available in ecological literature for measurement of species diversity can be directly used for AE data. Some modifications, if needed will be attempted. The interpretation of these indices and possible comparison of treatments will also be part of the objective. AE accumulation curve will be generated. A suitable model will be fitted. This will help estimating ‘Possible number of maximum AEs we may observe' and what fraction we have recorded thus far.
Key question is whether some new insights emerge from the proposed analysis.
In particular, it is of interest to see whether inferences drawn by the trial team can be supplemented/strengthened further. The work is intended as a prototype. If results are encouraging, biostatistics community will need to be informed. The results will be published in a peer reviewed journal.

Study Data Provided

[{ "PostingID": 1309, "Title": "ROCHE-BO17708", "Description": "A randomized, double-blind study of the effect of first line treatment with Avastin (bevacizumab) in combination with docetaxel on progression-free survival and disease response in patients with HER2 negative metastatic breast cancer" },{ "PostingID": 1664, "Title": "GSK-GLP114179", "Description": "A Randomized, Open-Label, Parallel-Group, Multicenter Study to Determine the Efficacy and Safety of Albiglutide as Compared With Liraglutide in Subjects With Type 2 Diabetes Mellitus" },{ "PostingID": 2510, "Title": "BI-1245.23", "Description": "Efficacy and Safety Study With Empagliflozin (BI 10773) vs. Placebo as add-on to Metformin or Metformin Plus Sulfonylurea Over 24 Weeks in Patients With Type 2 Diabetes" },{ "PostingID": 3179, "Title": "LILLY-B9E-MC-JHQG", "Description": "Phase III Study of Gemcitabine Plus Paclitaxel Versus Paclitaxel in Patients With Unresectable, Locally Recurrent or Metastatic Breast Cancer" }]

Statistical Analysis Plan

I plan to analyze AE data for two different therapeutic areas (breast cancer and Dibetes Melitus Type 2), and two trials for each TA. The above discussed statistical methods will be applied and tabular and graphical summaries will be generated. Wherever appropriate treatment comparison will be carried out with suitable statistical tools. This testing will have the only purpose of checking feasibility of suggested approach. No confirmatory conclusions will be drawn. The AEs data have variables of different nature. Occurrence/ non-occurrence is binary, number of episodes of same AE by same subject is a count variable, duration is continuous, severity may be an ordered category. Hence the tools needed will vary from one variable type to another. As a general strategy, logistic regression will be used for binary variables. Count variables will be modeled by a discrete distribution like Poisson. Duration will be modeled by suitable continuous distribution. Categorical (nominal or ranks) will be analysed using contingency table approach or non- parametric methods. For all tests of hypothesis a 5% level of significance will be used. It will be explored if different AE groups are associated with some demographic parameters or other baseline conditions. Towards this, depending on nature of variable, correlations, measures of associations will be used. In fact all possible statistical tools commonly used in exploratory data analysis will be attempted. For example, Association between AEs may be explored using cluster analysis. The routinely used tools practiced in AE data analysis may not be pursued. As stated earlier, we hope to use indices well known in ecological literature and to modify them if needed. A regression model may be attempted for data on accumulation of AE. This will help estimating maximum number of AEs likely to occur. Subgroup analysis may be attempted if improved results seem possible.The patterns which look interesting will be further analyzed by body system and/ or severity grades. Intersection sets of AEs (AEs common to different treatment groups) and exclusive set (occurring in only one of the treatment groups) will be formed. The questions of interest may be different for each of these sets.In ecology frequency of frequencies concept is used and a suitable probability distribution is fitted to these data. This concept will also be used for AE data. It may provide insightful summary of AE data.Here number of occurrences (0, 1, 2...) is on X axis and number of AE types with exactly X occurrences is on Y axis.This is not an inferential study. The treatment comparison will have the only objective of demonstrating that new measures of safety of a treatment better facilitate comparison. There is no intention of using the conclusions for any kind of claims.This is a frequentist approach and not Bayesian.

Publication Citation

Sharayu Paranjpe & Anil Gore. Adversity Index for Clinical Trials: An Inclusive Approach for Analysis of Safety Data.

https://arxiv.org/abs/1806.00204