Page 44 - Risk quantification and modification in older patients with colorectal cancer

Chapter 2
include selecting candidate predictors based on meta-analysis or literature15,27, or more modern techniques such as least absolute shrinkage and selection operator (LASSO).31,52 These methods still need a sufficient sample size to provide reliable estimations.
Concerns in generalisability exist, when data-driven models are not internally or externally validate.22,24,25,27,30 Furthermore, a split-sample validation does not assess the external validity of a model in the development study.10,11,18,26,32,34 For more recently published models, it was more easy to judge bias and applicability, because these were more often reported in line with the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD)53. That does not disqualify the validity of the earlier developed models, however it hampers the formal assessment of the quality and performance and applicability for older patients as used by the CHARMS checklist.
This review summarised the information available on the included predictors and performance of the different models. By selecting 25 studies out of almost 1900 publications, it is unlikely that we missed any unknown prediction models, which adds to the strength of this review. The assessment of the risk of bias aids in the critical appraisal of a prediction model for clinical practice. Albeit, the various prediction models did not prove to be specifically useful for older CRC patients.
There are some limitations to our review. First, we focused on clinical prediction models, excluding studies only describing logistic models without further analysis of their model performance. Second, with 25 studies included in this review, we decided not to assess the individual predictors on their association with the outcomes. Therefore, no information is provided on the weight of predictors, although the CHARMS checklist suggests providing these details.7 For these details as well as for the definition of outcomes such as anastomotic leakage and severe morbidity were refer to the individual studies.
42