A-Z of AI in Healthcare
Validation
Want to see how rigorous validation moves from research to clinical deployment?
What is validation?
Validation refers to the process of testing whether an algorithm "works," or simply put, whether it can do the job it was trained to do well. This typically involves giving the algorithm new data it hasn't seen before and checking whether it can still correctly perform its intended task, for example, classifying images (telling a healthy scan from an unhealthy one) or predicting a patient's future risk (identifying a high risk of developing diabetes) based on electronic health record (EHR) data.
Validation is similar to sitting an exam at the end of a school or university course: an exam is a way of testing whether you've genuinely learned and understood a topic.
What does the typical validation process look like?
A typical process for testing a healthcare AI algorithm involves:
- The algorithm is trained on data from Hospital A
- It achieves high levels of predictive accuracy
- It's then given data from Hospital B to see if it can still make accurate predictions
To ensure generalisability, the "unseen" dataset used for validation should come from a different location (e.g., a different hospital) and/or a different time period. The exact metric used to test the algorithm depends on the context, what the algorithm is designed to do and what type of algorithm it is (see the C-index for one common example).
Why does validation matter?
After an algorithm has been trained, it's important to test the model to ensure its "learned knowledge" is accurate, complete, and consistently capable of producing a suitable output, whether that's advice (via clinical decision support), classification, or prediction.
This matters because validation can surface issues like bias, overfitting (and the resulting poor generalisability), and problems with overall accuracy or performance, such as a tendency to generate excessive false positives. If these issues aren't caught early in development, a deployed algorithm could cause harm through misdiagnosis or missed diagnosis, or undermine research by producing results that rely on spurious correlations.
What are the general guidelines for validating a model?
Exactly how a model is validated depends on the type of model, what it's designed to do, and the clinical outcome it's targeting. That said, some general guidelines apply:
Validation should involve multiple stages:
- Internal validation — testing performance against a subset of the training dataset
- Temporal validation — testing against data from the same clinical centre (e.g., the same hospital) but from a different time period
- External validation — testing against a dataset of patients from an entirely different clinical setting (e.g., a different hospital)
Validation should test for robustness — assessing how well the model copes with different patient case mixes, variations in missing data and uncertainty, and differences in instrumentation and care protocols across settings.
Validation should include benchmarking — comparing the model's performance to another model, a similar intervention, or human performance.
Validation should use a range of statistical performance tests and metrics to identify the model's strengths and weaknesses in detail. These might include ROC AUC, mean, absolute mean, mean squared error, rand error, warping error, positive predictive value, the concordance index, and others, plus metrics used for testing recommender systems. The exact metrics chosen depend on the type of model (supervised or unsupervised) and the clinical task (classification or prediction).
Validation should be transparently and openly published, following relevant reporting guidelines such as MINIMAR (MINimum Information for Medical AI Reporting), with statistical metrics reported alongside confidence intervals or measures of statistical significance wherever feasible.
Validation studies should involve multi-disciplinary teams — at a minimum, statisticians, clinical informaticians, data scientists, and clinicians.
Is technical validation the same as clinical efficacy?
No. Rigorous technical validation, testing for accuracy, reliability, and robustness, confirms that a model's "knowledge" is accurate and performs well across a variety of conditions. But technical accuracy is not the same as clinical efficacy.
Clinical efficacy can only be established by testing a model in a real-life clinical setting, known as clinical evaluation. See how Owkin validates its own models against real patient data and in state-of-the-art laboratory systems before deployment.
Clinical evaluation might involve a randomised controlled trial, a cluster trial, an observational study, or a time series analysis. Generating this kind of evidence, using an appropriate method based on the model's associated level of risk, is a legal requirement if the model is intended for use as a medical device (i.e., deployed in a clinical setting rather than used solely for research).
Curious how these principles are being put into practice?
Owkin is building agentic AI and biological reasoning models to better understand biology and advance biological superintelligence.
Patient Validation
Owkin's approach to testing AI predictions against real patient data and real-world clinical outcomes, guided by a medical oncologist co-founder to keep patient impact at the centre of every model.
