Prediction

How does AI make these predictions?

In the context of AI, it's not a human doctor making the prediction, but an algorithm embedded in a medical calculator or clinical decision support software. The algorithm follows a process that can be thought of as a flowchart: if a patient has X health problems now, they might develop Y health problems in future.

The more prediction rules (if/then statements) the algorithm uses, the harder it becomes for humans to understand, but the more accurate the prediction is likely to be. This is why more complex, multi-layered models, such as neural networks, are often used for prediction tasks.

Are AI predictions always accurate?

No. While useful, these predictions aren't always right, and they're not guaranteed to come true. That's why most predictions carry a degree of uncertainty (for example, "we are 67% sure that a particular outcome is likely").

Predictions also rely heavily on comparisons between patient profiles. They assume that because past patients with a specific health profile, similar weight, blood pressure, and cholesterol level, developed a specific condition like heart disease, a new patient with the same profile is likely to develop it too. This isn't a guaranteed outcome. Instead, these predictions describe relative risk: compared to people with lower weight, blood pressure, and cholesterol, patients with higher levels are more likely to be at risk of developing heart disease in future.

How do risk prediction models work in healthcare specifically?

In healthcare, prediction typically involves using a predictive model to estimate an individual's risk of a specific outcome, such as having a particular disease now, being readmitted to hospital, or developing a condition in future, by examining the outcomes of "similar" individuals in the past. This predicted risk is not absolute, but probable and relative: a model predicting an X% likelihood of developing a condition doesn't guarantee that outcome will occur.

‍

‍

What are traditional, regression-based risk prediction models?

Traditionally, risk prediction models have been regression-based. They estimate a baseline risk and then build a linear list of covariates, different combinations of which increase or decrease the likelihood of a particular outcome.

For example, a logistic regression model developed early in the COVID-19 pandemic identified that individuals with a higher BMI, or from Black or minority ethnic groups, were at greater risk of adverse outcomes from infection (such as hospitalisation) compared to baseline. Other well-known regression-based risk scores used in medicine include the Framingham Risk Score (10-year risk of cardiovascular disease) and the Kidney Failure Risk Score (10-year risk of end-stage kidney failure).

What are the limitations of traditional regression-based models?

Despite being widely implemented across healthcare systems globally, these models have several distinct disadvantages:

• They can confuse correlation with causation
• They model what increases risk, but not why
• They assume relationships between variables are always straightforward and linear
• They can't easily be combined to identify how different outcomes (e.g., co-morbidities) interact to alter risk
• Each individual outcome of interest requires a new, bespoke-trained model
• The range of variables they can handle (phenotypic, genetic, or clinical) is limited

How are machine learning models improving on this?

Because of these limitations, it's increasingly common for risk-prediction models to be based on machine learning instead, which is less affected by these constraints.

The exact type of model used depends on the kind of data involved:

• Structured data (e.g., electronic health record data) → typically an artificial neural network
• Unstructured data (e.g., free-text data) → typically a convolutional neural network

In both cases, the model is likely to incorporate Bayesian principles during training, becoming a Bayesian neural network. Bayesian networks are graphical models capable of making predictions involving joint probability (the likelihood of two or more events happening at the same time) and conditional probability (what happens if one factor or variable changes). These capabilities allow Bayesian networks to generate individual-level risk predictions that can eventually be built into clinical decision support software, supporting tasks like treatment planning.