Page 116 - Risk quantification and modification in older patients with colorectal cancer

Chapter 6
to the outcome using univariable logistic regression analysis to estimate Odds Ratio (OR) with corresponding 95% Confidence Interval (CI) and p-value.
To investigate the added value of a geriatric predictors, two models were created. A “demographic model” included only preoperative demographic predictors, comorbidity, tumour location and stage, and ASA score. For a “geriatric” model, the geriatric predictors from the EMR were added to all candidate predictors from the demographic model.
The questions of the Katz ADL, self-reported need for ADL assistance, previous delirium and self-reported cognitive impairments (classified as a risk for delirium) were added as a categorical predictor on an individual level and dichotomised (Katz ADL ≥2 and risk for delirium ≥1). Because of expected co-linearity between Katz ADL questions and the self-reported need for ADL assistance, either the Katz ADL or self-reported ADL assistance were used as candidate predictors.
In both the demographic and geriatric model, the final model selection was obtained using the Least Absolute Shrinkage and Selection Operator (LASSO) method. LASSO applies a penalty on the absolute value of the regression coefficients, such that some are set to zero whereas others are shrunk towards smaller (absolute) values. Variables that are shrunk to zero are omitted from the model. The goal of this process is to minimize the prediction error. Compared to backward selection, the addition of shrinkage may improve model performance by avoiding overfitting and miscalibration.30
The validity of both models was tested by performing bootstrap validation with 500 replications and optimism correction. The discriminative predictive performance of the models was demonstrated with the Area Under the Curve (AUC). For the optimism corrected model, no valid 95% CI can be calculated. The final shrunk coefficients from the LASSO were used to generate a score chart which is intended as a clinical tool. The shrunk β coefficients from the geriatric model were rounded for selection in the simplified clinical tool. Predictors with a β of less than 0.1 were therefore not selected for the clinical tool to increase the robustness of the model.30 At least 1 point was given to each predictor included. Subsequent risk groups were created based on at least 70 observations in each risk category.
114