As the availability of big data increases and statisticians assist with predicting outcomes or understanding patterns in an ever-wider variety of scenarios then supervised and unsupervised learning methods become increasing called upon. Such machine learning algorithms offer the opportunity to understand potential predictors or clusters amongst large datasets but are also subject to the risks of overfitting or over-interpretation. This Webinar seeks to introduce ideas and share experiences in this field. The talks will introduce several supervised and unsupervised learning methods and cover data-driven subgroup identification in clinical trials, and case studies of implementation clustering algorithms. 

Not all patients are created equal, but are there subgroups that are more homogenous? Alexander Schacht, Lilly

Can I divide my overall patient population into meaningful segments? Do patients follow different patterns over time? We should ask these questions more often and techniques of unsupervised learning, where the classification of a patient into a group is unknown, answers these questions. We differentiate these approaches from supervised learning techniques in which classification of the patients is known. Typical questions for supervised learnings algorithms include: Can I predict patients outcomes given his/her baseline characteristics? Cluster analysis represents a class of approaches in unsupervised learning. It helps to answer the above questions. Cluster analysis stands on the determination of metrics, which measure the distances between patients in terms of their many different characteristics. In this presentation, I will present and discuss different approaches available in SAS. The determination of the number of clusters represents a classical problem of bias-variance trade-off. The presentation will discuss various heuristics but also practical considerations to determine a reasonable choice of clusters.The practical implementation of cluster analyses comes with various challenges. I will discuss standardization of variables, weighting of variables, correlated data, outliers, finding spurious small clusters, and identification of relevant clusters.  Finally, the communication of cluster analyses has its unique challenges and I will mention various approaches based on real case studies.

Overview of methods for subgroup and biomarker identification from clinical data, Ilya Lipkovich, Lilly

In this talk I will provide a high-level description of a broad class of statistical methods for subgroup/biomarker identification in early and late-phase clinical trials. First, I contrast “data-driven” subgroup analysis with a traditional “guideline-driven” approach and describe key elements of principled data-driven subgroup analysis. Then I review 4 classes of methods for subgroup identification that had emerged recently as a result of cross-pollination across machine learning, causal inference and multiple testing (global outcome modelling, global treatment effect modelling, modelling individual treatment regimes, and local treatment effect modelling). I also briefly review available software and key features of subgroup identification methods.

Using the SIDES algorithm to the identify patient phenotypes that have the potential to benefit most from switching to Relvar, Andy Nicholls, GSK

In 2016 GSK successfully completed the Salford Lung Study, a 12-month, open label, randomised, effectiveness study to evaluate fluticasone furoate (FF, GW685698)/vilanterol (VI, GW642444) Inhalation Powder delivered once daily via a Novel Dry Powder Inhaler (NDPI) compared with the existing COPD maintenance therapy alone in subjects with Chronic Obstructive Pulmonary Disease (COPD).  Upon completion of the study, the Scientific Committee expressed an interest in using a data-driven approach in order to identify patient subgroups for which the treatment effect was strongest.  In this presentation we will look at why SIDES was chosen for this analysis, the design parameters, and how it fared.