Clinical Prediction Models and Machine Learning
The purpose of a prediction model is to estimate the probability of the presence of a particular outcome as accurately as possible. Prediction models are often developed with clinical practice in mind, and involve combining information about individual patients to calculate an individual’s probability of illness or recovery. The model can then be presented in the form of a clinical predictive rule. General applicability – i.e. the accuracy of the prediction model when applied to new patients in the future – is another very important aspect.
Nowadays, access to data is becoming easier and easier and therefore the data sets are getting bigger and bigger. The problem when developing prediction models in these data sets include the difficulty of selecting the most important predictors from a large number of variables. If this is not done carefully, the quality of the prediction model can be adversely affected. Machine learning methods can be used to develop prediction models in these large data sets. Also, prediction models may be adjusted before they can be applied to new persons. All these issues are frequently overlooked or underestimated by clinicians and researchers.
The aim of the course is to provide better knowledge and understanding of the development of prediction models in smaller and larger data sets that are relevant to real-life practice. We will focus on common methods for selecting variables as backward selection but also more advanced Machine learning procedures as lasso regression and tree based methods as well as their pros and cons. Once prediction models have been developed, it is important to assess the quality of the prediction model. For example, we will look at whether the predictions of the model are accurate and will consider various ways of measuring performance by using measures for overall quality, discrimination and calibration. The question of applying the model to new (future) patients will also be addressed. An important element of this is investigating whether the performance of the prediction model deteriorates when it is applied to new patients. This component is entitled the validation of the prediction model and we will cover various techniques for internal and external validation of prediction models and ways to train and test the model by using bootstrapping and cross-validation.
The course consists of an intensive programme of partly interactive lectures, combined with computer-based practical work. Examples taken from clinical practice will be used for the computer-based work.