General discussion 177 To ensure that patients continue to exercise, and for modifications to the rehabilitation program, follow-up is necessary [29]. Based on shared decision making, physiotherapists and patients can determine how unsupervised exercise will take place. If the physiotherapy practice has the facility, the patient may choose to exercise independently at the physiotherapy practice. The patient may also choose to exercise independently at home or use regular sports and exercise services [30]. PATCH was developed in a cohort of patients with COPD already following PR (70% of the patients followed PR for one or more years). Therefore, the prediction model should be applied with some caution to a different population. Chapter 5 described that adherence was constant over a period of 12 months in this cohort, therefore, when using the prediction model for a different time period, the results should also be interpreted with some caution. Methodological considerations To develop the prediction model, a probabilistic approach was adopted. The current era of evidence-based medicine asks for an individualized approach to medical decision-making [31]. Healthcare providers need to make predictions on the likelihood of an underlying disease (or in this case the probability of adherence) in their decision-making on choice of therapy. Evidence-based medicine applies this scientific method to medical practice [31]. Another development is that healthcare is moving towards “shared decisionmaking,” where healthcare providers and patients both actively participate in deciding on choices for diagnostic tests and therapeutic interventions [32]. Clinical prediction models may provide the evidence-based input for shared decisionmaking, by providing estimates of the individual probabilities of risks and benefits [31]. In developing the model, uncertainty about the assumptions was accepted; the continuous data of the predictors were considered linear. This choice was based on a sample size that was too small to add additional determinants to the model as this would take too many degrees of freedom [33]. This way of handling linear data is not optimal. However, it is a better alternative than categorizing the data; this results in too much information loss, reduced predictive accuracy, and also in too many degrees of freedom (each response category has its own degree of freedom). If a sample size is limited, assuming a linear relationship between continuous variables may make a model less sensitive to extreme observations [33], making the solution in this study the better of the two least optimal alternatives. However, research shows that assuming linearity is not necessarily a bad choice; applying more flexible alternatives such as fractional polynomials and restricted cubic splines does not always provide a better calibrated and predictive model [34]. If follow-up
RkJQdWJsaXNoZXIy MjY0ODMw