标题：Approches multidimensionnelles pour l’analyse de la relation toxicité-efficacité d’un traitement pour le cancer de l’estomac. Application dans le cadre d’une méta-analyse
摘要：The relationship between toxicity and efficacy are analysed for 2 chemotherapies
in randomised studies using exploratory analyses and a meta-analysis in order to validate the
results.
The first step consists in a series of multidimensional analyses performed on a randomized
study including 124 patients with gastric cancer. The second step is a validation phase of the
results, based on two other clinical trials, one including 221 patients and the second 41
patients treated with the same chemotherapies. In the meta-analysis, in order to decide which
data bases can be pooled, tests are performed to test the homogeneity of the toxicity variables
frequencies and of the baselines variables. Data pre-processing implies the search of a final
coding for each variable in order to run a Multiple Correspondence Analysis on the pooled
datasets. The factorial plans suggest a grouping of the individuals in three clusters. The
corresponding partitions are built by an ascending hierarchical clustering approach, with an
aggregation index associated to the variance one and initial distances compatible with the
ACM ones.
This methodology is infrequently used in the context of clinical trials: Correspondence
Analysis shows a relationship between efficacy and toxicity. The analysis provides also a mapping which allows to define a concept of “useful” toxicities i.e. correlated to the efficacy
as opposed to a concept of “non-useful” toxicities i.e. non related to the treatment efficacy.
The technique of supplementary variables suggests the hypothesis that patients with a low
Karnofsky when entering the trial develop the ‘non useful’ toxicities. At that point one may
consider that somehow a descriptive proof has been obtained thus drafting a solution to the
initial problem. The obtained clusters are projected on to the first factorial plane. Tests for the
significance of the centroids positions are performed in order to get a first step toward the
partition validation.
