Clinical Statistics

Clinical Trials

The specific challenges of clinical trials necessitate the application and development of specific trial designs and evaluation strategies in order to deal with particular data conditions or research questions.

Biomarker-driven trials, such as the individual profile design trials, are used to stratify the population of diseased patients into many different subgroups and treat patients according to their molecular profile. Usually this stratification results in many, rather small subgroups that only allow the assessment of how beneficial a biomarker-driven strategy is in general without providing strong conclusions about specific marker-defined subgroups. For individual marker-defined subgroups, the research situation is similar to that of rare diseases with all its limitations for clinical trials. More advanced testing strategies based on all subgroups might help to gain information.

Flexible trial designs allow for data-driven design adaptation, e.g., sample size recalculation or treatment arm selection, during an ongoing trial. The necessity for modifications occurs especially when study parameters, e.g. the actual treatment effect, can only be determined roughly or with high incertitude. This may be due to small incidence rates in rare cancer types - and thus little information at trial start - or due to the urgent need for new/any therapy options in patient groups with few/no established effective treatment options. We apply, adjust and develop flexible designs for actual trial plans, e.g., group sequential settings for premature decision-making; conditional power approaches allowing futility stops; or adaptive designs including sample size recalculation or allowing adaptive enrichment strategies.

In some cancer types, advances in therapy have made a cure or long-term survival possible. The most common regression model for survival data, the proportional hazards model, is not appropriate for those heterogeneous patient populations that include both “cured” and “uncured” patients. Cure models, such as the proportional hazards cure model, account for this heterogeneity. Applying such models in sample size calculations helps to avoid over- or underpowered trials. We explore what gain (e.g., in sample size and power) is achieved by these models and how time-dependent interventions, such as allogenic transplantation in acute myeloid leukemia (AML) patients, may be included in cure rate models. Multi-state models are examined as alternative modeling of time-dependent interventions in situations with cure.

Further, we develop methods and tools to support investigators in conducting clinical trials. For example, we offer consulting for data management including: (1) design and implementation of study databases or (2) the set-up of electronic data capture systems. In addition, we provide a web-tool for investigators to perform the randomization directly without a centralized study office. To further simplify the randomization procedure, we have developed a method of integrating the randomization into an electronic data capture system. We made use of two open-source applications in order to offer a flexible and freely available solution for various study designs.

Statistics for personalized medicine

Personalized medicine offers the potential to identify subgroups of patients most likely to benefit from a specific treatment. One important approach is to reliably identify molecular biomarkers or sets of biomarkers (‘signatures’) that predict relevant clinical characteristics for individual patients, such as disease subtype, response to a certain therapy, or time to event.

Prediction based on selected low-dimensional demographic, clinical, and genetic covariates often benefit from being complemented by high-dimensional molecular genomic data. Combining different genomic data sources can help to improve the precision of predictions, and systems approaches try to bring together these various diagnostic and prognostic factors. With the increasing amount of information available, there is also a need for new ways of reporting and visualizing findings that help the treating physician to make the most informed decision. The ultimate goal is to apply the results of these analyses toward improved patient care.

However it has to be be recognized that biomarker based subgroup identification usually does not provide an unbiased assessment of the treatment benefit. Valid assessment of the benefit of a specific treatment in such subgroups requires special clinical trial designs and analyses which is a major topic of actual research.

Another important application area implementing personal data is the monitoring of patients’ treatment and the switch to other treatments according to rules based on the patients’ medical history and treatment outcome, such as disease progression. The development of new approaches to this personalized decision-making process during treatment is still ongoing. One possible approach would be the use of adaptive enrichment designs (cf. clinical trials).

Observational Studies

Retrospective studies can provide the initial evidence for a promising hypothesis, such as a potentially prognostic marker, so that therapeutic strategies may vary depending on the predicted risks. Due to the nature of retrospectively collected and analyzed data, the risk of biased or incorrect results is immanent without setting up pre-defined populations and hypotheses. Statistics play an important role in correcting for bias, confounding, and imbalances, or to account for incomplete covariate information. This can be done, for example, with conditional estimates from regression models, adjusting for covariate information, or with rebalancing groups prior to effect estimations. Reliable estimates from observational studies are needed for successfully planning and conducting prospective trials.
Often when using high-dimensional molecular data in the search of prognostic biomarkers a subset of patients must be selected due to budget restrictions. Commonly the selection of patients is based on extreme survival time data,  e.g. selecting only patients with very poor and very good prognosis. We are currently investigating this approach and comparing  it with alternative sampling approaches. This project aims to provide guidance as to which sampling method would be most useful for biomarker identification.


Clinical Trials

  • Schrimpf D, Plotnicki L, Pilz L (2010). Web-based open source application for the randomization process of clinical trials: RANDI2. International Journal of clinical pharmacology and therapeutics 48:465–467.
  • Schrimpf D, Pilz LR (2012). Adaptive randomization procedures for the web-based randomization system RANDI2. . International Journal of clinical pharmacology and therapeutics 50:85–86.
  • Schrimpf D, Manegold C, Pilz LR (2013). Design of clinical studies: adaptive randomization and progression-free survival (PFS) as an endpoint in clinical studies of advanced non-small cell lung cancer (NSCLC). International Journal of clinical pharmacology and therapeutics 51:84–86.
  • Kunz C, Kieser M (2014). Blinded versus Unblinded Covariate Selection in Confirmatory Survival Trials. Journal of Biopharmaceutical Statistics, DOI:10.1080/10543406.2013.860158.
  • Wunder C [Kunz C], Kopp-Schneider A, Edler L (2012). An adaptive group sequential phase II design to compare treatments for survival endpoints in rare patient entities. Journal of Biopharmaceutical Statistics 22: 294-311.

Statistics for personalized medicine 

  • Meißner T, Seckinger A, Rème T, Hielscher T, Möhler T, et al. (2011). Gene Expression Profiling in Multiple Myeloma–Reporting of Entities, Risk, and Targets in Clinical Routine. Clinical Cancer Research 17: 7240–7247.
  • Schlenk RF, Döhner K, Krauter J, Fröhling S, Corbacioglu A, Bullinger L, Habdank M, Späth D, Morgan M, Benner A, Schlegelberger B, Heil G, Ganser A, Döhner H (2008). German-Austrian Acute Myeloid Leukemia Study Group. Mutations and treatment outcome in cytogenetically normal acute myeloid leukemia. New England Journal of Medicine 358:1909-1918.
  • Werft W, Benner A, Kopp-Schneider A. On the identification of predictive biomarkers: Detecting treatment-by-gene interaction in high-dimensional data. Computational Statistics and Data Analysis 2012; 56: 1275-1286.

Observational Studies

  • Dorner H (Betreuer Stadtmüller U, Hielscher T), Bachelorarbeit (2012). Analyse nicht-randomisierter Beobachtungsstudien, Universität Ulm, Fakultät für Mathematik und Wirtschaftswissenschaften.

to top