A prospective cohort study was performed on two cardiac care units of the University Hospitals Leuven, Belgium, between September 2016 and June 2017. Patients aged 75 years or older were recruited on admission if they were admitted for non-surgical treatment of an acute cardiovascular disease, had an expected length of stay of 3 days or longer, consented to participate and were able to complete the assessment. Patients who were admitted from another hospital or from the intensive care unit were excluded. The Medical Ethics Committee of the Leuven University Hospitals approved this study and written informed consent was obtained for all patients who consented to participate. The study adhered to the Declaration of Helsinki.
All data were assessed by three researchers using standardised assessment forms. Assessments on admission were performed within 72 h: 88% within 24 h and 99% within 48 h. Researchers were trained by performing paired bedside assessments and having case discussions until a 100% inter-rater agreement was observed. A meeting was organised every 3 months to confirm agreement.
Baseline characteristics
Sample characteristics included age, gender, living situation (home, retirement home or nursing home) and medical diagnosis (heart failure, valvular heart disease, ischemic heart disease, arrhythmia or other).
Outcome
Hospitalisation-associated functional decline was defined as the development of new or worse dependency in Activities of Daily Living (ADL) according to the Katz index when patients were discharged home. The Katz index was measured on admission to the unit and on the day that the patient was discharged home. Patients were asked whether they needed assistance for bathing, dressing, walking, continence, toileting, or feeding [13]. Assistance was scored on a three-point scale: independent, partially dependent, completely dependent. A change of one point on the scale was considered clinically relevant. Nurses and informal caregivers were asked to confirm the ADL status in patients with cognitive impairment.
Development and validation of a prediction model
Predictors were a priori selected based on their association with functional decline in previous studies, [1, 4,5,6,7,8,9,10,11] and their clinical utility (i.e., easy to assess with no additional costs or resources than those available during routine care). The model complexity was a priori restricted to five predictors based on the assumption that we would observe a minimal of 50 events (assuming a conservative incidence of 27% based on the review of McCusker et al. [7]). This allows for a more precise estimate of the coefficients for the risk score and prevents against overfitting of the model.
The predictors were assessed within the first 72 h of hospital admission. Mobility impairment was defined as the use of a walking aid before hospital admission as reported by the patient. Cognitive impairment was defined as a Mini-Cog score < 3 out of 5 points. The Mini-Cog assesses performance on two cognitive tasks, i.e., three-item word recall and clock drawing test. A higher score indicates a better cognitive performance [14]. The presence of depressive symptoms was defined as a score > 3 on the 10-item version of the geriatric Depression Scale. Patients were asked whether they experienced any of the 10 symptoms in the past week. Sum scores vary between 0 and 10 with higher scores indicating more depressive symptoms [15]. Loss of appetite was defined as self-reported loss of appetite in the past 3 months and was used as a proxy for risk for malnutrition. Use of restraints was defined as the use of physical restraints (e.g., vests, limb ties or chairs with restraints) or an indwelling urinary catheter between admission to the unit and assessment of the predictors [16]. Bed rails were not considered a restraint.
A multivariate logistic regression model was built using a full model approach. Unadjusted and adjusted Odds Ratios (OR) and regression coefficients were calculated with 95% Confidence Intervals (CI). Coefficients were shrinked to compensate for potential overfitting of the model. A uniform shrinkage factor was calculated based on the Chi2 and the degrees of freedom (df) of the model ((model Chi2– df)/model Chi2) [17]. Multicollinearity was evaluated using the Variance Inflation Factor and Tolerance values. A score chart for the prediction model was developed by multiplying the shrinked coefficients with 10 and rounding the score [18].
Discrimination was assessed using the C-index. We performed an internal validation of the model using nonparametric bootstrapping samples (n = 1000). This procedure estimates the optimism of developing and validating the prediction model in the same sample of patients, and provides a bias-corrected measure for model discrimination.
Comparison with existing prediction models
Three clinical prediction models for hospitalisation-associated functional decline were previously developed in patients admitted to a general medical ward (see Table S1 and S2 in the supplementary material). Each model was scored within the first 72 h of hospital admission (see Table S3 in the supplementary material for definitions of all predictors). The model by Inouye et al. defines the presence of a decubitus ulcer, cognitive impairment, functional impairment and low social activity level as predictors [9]. The model by Mehta et al. defines age, premorbid dependency on instrumental activities of daily living and basic ADL, inability to run a short distance, inability to walk stairs, metastatic cancer or stroke, cognitive impairment and albumin level as predictors [10]. The model by Sager et al. defines age, cognitive impairment and premorbid dependency on instrumental ADL as predictors [11].
The discrimination of the models was compared using the chi2 test for equality for two or more Receiver Operating Characteristic areas. Calibration was assessed using the Hosmer-Lemeshow Goodness of Fit test and by constructing calibration plots. The clinical usefulness was assessed using classification statistics (sensitivity, specificity, predicted values). For all models, the cut-off value, that identifies if a patient is considered at risk, was based on the Youden index to allow for equal comparison between the models [19]. The Youden index corresponds to the cut-off value with the optimal combination of sensitivity and specificity.
Missing data
There were 53 cases (28%) with missing data for the Mehta et al. model because of missing albumin levels, which were not routinely assessed on hospital admission. There was no significant relationship between missing data and hospitalisation-associated functional decline (OR 1.8, 95% CI (0.9–3.5)). We therefore assumed that data were missing completely at random [20]. Multiple imputation (M = 5) with all predictors, the outcome and auxiliary variables was performed to allow comparison between models with an equal number of subjects. A parametric multiple linear regression model was used as the data were normally distributed. A sensitivity analysis with a complete case analysis was performed to evaluate the influence of data imputation.
Post-hoc analysis
After the model was developed and validated, we used a nested model approach to investigate if adding the predictor ‘age’ improved the discrimination. We also performed a chi-squared test to compare the C-index estimates per age group (75 to 80, 81 to 84, or > 84).