Background: To summarize the treatment sequences information is complex as it is a longitudinal process involving several treatments and combinations of treatments. The TAK® (Time sequence analysis through K-clustering) is a non-supervised clustering algorithm of time sequences which can be used as a pre-processing step whose results (clusters) summarize treatment sequences information.
Objectives: To describe how to characterize the treatment sequence during the follow-up and to take into account history of treatment sequences in analyses.
Methods: The TAK ( Chouaid C. et al., DOI: 10.1200/CCI.21.00108) can be used as a pre-processing step to explanatory and predictive models. Illustrations will be provided for both modeling techniques.
Results: To explain the treatment sequence, the characteristics and hospital settings of patients from each cluster can be described (age, comorbidities/medical history, year of treatment). Models can also be used to quantify the associations between clusters and characteristics/hospital settings of patients, such as a multinomial logistic regression or a multilevel model. For example, is the heterogeneity of treatment sequences associated with age, comorbidities, or hospital setting? As an extension, TAK clusters can be used in a predictive model of the treatment sequence of a patient (eg. tree-based algorithm). Illustrations will be provided. When clusters are used as a covariable in a model (eg. Cox model), the influence of the treatment sequence on overall survival or progression-free survival can be analyzed. In the case of a TAK on the front line, the influence of the latter on the choice of a subsequent treatment line can be estimated. Similarly, we can assess the impact of treatment sequences on the probability of occurrence of an adverse event after the period included in the TAK. Finally, we can estimate the link between the clusters and a profile signature stemming from a set of variables (eg. Care events such as laboratory test or specialists’ visits) with a multivariate dimensionality-reduction tool such as the Partial Least-Squares Discriminant Analysis.
Conclusions: The clusters obtained from the TAK are simple to interpret, robust to noise and summarize the information of the patient’s treatment sequences (treatments, temporality) and can be further used in predictive or associative models.
