Overfitting

Overfitting is when a model learns the particulars of its training data, noise and quirks included, instead of the pattern that generalizes. It performs well on the data it was built on and worse on new patients. A statistician's framing is parsimony: an overfit model includes more terms, or uses a more complicated approach, than the problem needs 1.

Why does overfitting matter in healthcare?

Clinical datasets are often small relative to model complexity, and the outcomes that matter are often rare. That combination invites overfitting, and the literature shows it. A systematic review of 152 machine-learning prediction models published in 2018 and 2019 found that 56% were developed with an inadequate number of events per candidate predictor and 39% assessed overfitting improperly. Of 171 analyses, 87% were rated at high risk of bias 2.

The consequence reaches the bedside. An overfit model's reported accuracy describes its development data. At your hospital it will usually do worse, and nobody finds out until someone measures.

What does it look like in a real study?

Zech and colleagues trained pneumonia-detection networks on 158,323 chest radiographs from three hospital systems 3. A model trained and tested on pooled data from two of them reached an AUC of 0.931 internally. At the third system, it scored 0.815.

Part of what the models had learned was the hospital itself. The networks identified the source hospital system for 99.95% and 99.98% of radiographs from the two main sites. Because pneumonia prevalence was 34.2% at one site and 1.2% at the other, sorting images by hospital alone produced an AUC of 0.861. The authors warned that site signals like these, which track large differences in disease burden, can confound predictions 3.

Large datasets offer no automatic protection. This one had more than 150,000 images.

How is it prevented and detected?

• Enough data for the model's complexity. Riley and colleagues set out how to calculate the minimum sample size for developing a clinical prediction model, with limiting overfitting as a central aim 5.
• Honest testing. Test data kept apart from every development decision, then external validation on a different site or period. Only 6% of 516 imaging AI studies in one review performed external validation 4.
• Regulatory expectations. The FDA, Health Canada, and MHRA's good machine learning practice principles expect controls to manage the risks of overfitting, unintended bias, and model degradation 6.

Common misunderstandings

A high test-set score rules it out. Only if the test set was truly held out and resembles the population where the model will run. The ten red flags guide lists the common ways test data leak into development.

More complex models are always better. Extra flexibility fits noise as readily as signal 1.

Overfitting and drift are the same problem. Overfitting is present at launch. Model drift develops later, as care and populations change.

What to ask a vendor

• How many patients and outcome events were used for development, and how many candidate predictors?
• Was the test set split by patient, by site, or by time, and was it touched during tuning?
• What is performance at an external site, and how large is the drop from internal results?
• Is the model calibrated at that external site?

Related terms

• External validation: the test that exposes overfitting.
• Internal vs external validation: why the two numbers differ.
• Sample size and confidence intervals in AI studies: how much data a claim needs.