scientific concept · BEGINNER LESSON
model prediction
Model prediction is the use of known inputs, also called predictors or features, to estimate an outcome that has not yet been observed. In health and other data-driven settings, a useful prediction model must use only information available before the outcome is known and should still work on new, unseen data.
What it is and what is measured
A prediction model is a construct that takes known variables, such as patient factors or other measured inputs, and estimates an outcome before that outcome is observed. The thing being measured is the future or not-yet-seen value of the outcome, not just the relationship among variables inside one dataset. In health research, prediction is mainly used for clinical practice, but it also appears in research and administration. Related terms include independent variables or inputs for the known factors, and dependent variable or output for the outcome being estimated.
[1] Development and Reporting of Prediction Models: Guidance ... - PMC [2] Table 1. Prevalence Of...
How prediction works
Prediction modeling begins by choosing variables that are known at the time of prediction. The model is then developed to estimate the outcome from those inputs, and it should remain accurate when applied to new observations. For a binary outcome, meaning one with two possible states such as present or absent, the model needs a classifier function. In machine learning (ML), prediction models are usually built with generalization in mind, meaning they are intended to perform well on new data, not only on the training data. Some models may include causal factors, but causation is not required for prediction.
[1] Development and Reporting of Prediction Models: Guidance ... - PMC [2] Table 1. Prevalence Of...
Why validation matters
After a model is built, it must be validated, which means checked on data not used for development. Internal validation asks whether performance is reproducible in the same underlying population, while external validation asks whether the model works in a different population. This step is necessary because a model can fit the original data well yet fail on new data. In other words, a prediction claim is strongest only when the model has been tested on previously unseen observations.
[1] Development and Reporting of Prediction Models: Guidance ... - PMC [2] Table 1. Prevalence Of...
Limitations and common errors
Prediction is not the same as causal inference, which tries to learn what happens when an exposure is changed. A model can predict without explaining cause. Another limitation is overfitting, where too many variables are used for the amount of data available; this can produce a close fit to the training data but poor performance elsewhere. Feature selection based only on p values, a statistical measure of how surprising data are under a null hypothesis, is strongly discouraged in most prediction settings. Models in complex systems may also need careful mechanistic detail and can still be constrained by data limits and human modeling limits.
[1] Development and Reporting of Prediction Models: Guidance ... - PMC [3] pmc.ncbi.nlm.nih.gov › articles › PMC5758710Modeling complexity: cognitive constraints and computational ...
Simple takeaway
A prediction model is judged by whether it estimates an unseen outcome from known inputs, uses only information available before the outcome, and keeps working when tested on new data. If it only summarizes patterns in the original dataset, it is not yet a validated prediction model.
[1] Development and Reporting of Prediction Models: Guidance ... - PMC [2] Table 1. Prevalence Of...
SEE THE IDEA
How a prediction model is used
Variables measured before the outcome is known, such as features or predictors.
A mathematical or computational rule links inputs to an outcome estimate.
The model is checked on new, unseen data to test generalization.
The estimated future or not-yet-observed result.
Known inputs go into model building; validation checks whether the resulting prediction still works on unseen data before the outcome is observed.
Prediction depends on pre-outcome inputs and is only trustworthy when it generalizes beyond the data used to build it.
TRY IT YOURSELF
Test your understanding
Choose an answer, then check the explanation.
PUT IT TOGETHER
Match each concept to its meaning
Drag a label onto its description, or select a label and then select a description. This also works with a keyboard.
4 pairs to match
Sources and further reading
This general lesson was generated from public source material. It explains a concept; it does not interpret a private paper or provide medical advice.