Page 115 - Risk quantification and modification in older patients with colorectal cancer

GerCRC risk prediction model for severe complications
Predictors
A systematic review of prediction models for adverse outcomes of CRC was used to identify commonly used predictors in younger and older patients.17 Candidate predictors that were available from the DCRA database included demographic information (age, gender, body mass index (BMI)), tumour stage and location, ASA (American Society of Anesthesiologists) score and comorbidity. Comorbidity included previous abdominal surgery, cardiac comorbidity (including arrhythmias, myocardial infarction, cardiac surgery and cardiomyopathy), pulmonary comorbidity (COPD/Asthma/Emphysema and other), and previous thrombo- embolic such as Pulmonary Embolism (PE) or Deep Venous Thrombosis (DVT). From the comorbidity data, a Charlson Comorbidity Index (CCI) was calculated.26
From the EMR the following preoperative additional candidate predictors were extracted: undernutrition (or at risk of becoming undernourished), functional impairment, the use of a mobility aid (the use of a cane, crutches, a walking frame or wheelchair), the risk of delirium and falls in the past six months. Risk for undernutrition was assessed with either the SNAQ27 or MUST28 screening tool. Functional impairment was assessed with the six-item Katz ADL29 consisting of questions regarding bathing, dressing, using the toilet, eating, transferring from bed to chair and the use of incontinence materials. Risk for delirium was assessed using three yes or no questions concerning previous delirium during hospitalization, self-reported need for ADL assistance (in the past 24 hours) and self-reported cognitive impairment. Polypharmacy (using five or more prescribed medications) was based on preoperative medication/prescriptive data from the EMR. All predictors from the EMR had been registered at the day of hospital admission or in the weeks before surgery (up to 6 weeks).
Statistical analysis
Data were inspected for missing variables. Missing predictor data were estimated in a regression model using all other predictor variables and outcomes as independent variables. Missing data on candidate predictors were subsequently imputed with a single imputation technique and used for final predictor selection and model development.
Baseline characteristics were reported as means with standard deviation (SD) or as frequencies and percentages. Before imputation, candidate predictors were related
113
6