Prediction is the act of forecasting what will happen in the future. Prediction is central to medicine as preventive and therapeutic interventions are prescribed or recommended on implicit or explicit expectations about future health outcomes. Prediction models aim to quantify the probability of these future health outcomes based on a set of predictors. They are used to score the health status of newborn babies (APGAR), for cardiovascular disease prevention, and for stratifying cancer screening programs. They are also increasingly offered outside health care such as the personal genome tests, wearables, and risk calculators.
The field of prediction research is gaining importance because of the increasing interest in personalized medicine. Personalized medicine is a medical model in which interventions are targeted to the individual or to risk groups rather than given to populations at large. Prerequisites for personalized medicine are that there is a test or prediction model that can divide a population into subgroups and that there are alternative interventions available to assign to each of the subgroups.
The focus of the course is on the methodology of prediction research. Examples will come from clinical and public health practice as well as from public health and clinical research. Well-known practical examples are the Framingham risk score for the prediction and stratification of cardiovascular disease risk, the Gail model for breast cancer risk and the APACHE score for mortality at the intensive care unit.
The lectures broadly cover the following topics:
- Introduction and examples. A broad overview of the course, what characterizes prediction research, difference prediction and epidemiological research
- The health care scenario: what is to be predicted in whom for what purpose, decision making, risk stratification, screening programs,
- Development of a prediction model (1): selection of outcome, diagnostic criteria, misclassification, risk period, population selection, selection of predictors, types of predictors.
- Development of a prediction model (2): statistical model, missing values, outliers, multicollinearity, interaction effects, variable reduction methods.
- Assessment of performance of prediction model (1): model fit, calibration, discrimination, predictive value, reclassification.
- Assessment of performance of prediction model (2): net benefit, internal and external validation, generalizability, transferability.
- Advanced issues: precision medicine, big data, machine learning.
The manual of the course can be downloaded here.