AI in healthcare – setting the stage

Recognising Patterns in Medicine: Clinical Bioinformatics/Health Informatics 

The idea of “learning from patterns” or “statistical learning” has been applied to healthcare for decades. However, recently the explosion of AI technologies has facilitated the application of AI to healthcare.  

In this section, we will set the stage for the application of AI to healthcare by looking at different ways the health sector already learns from data.

It may be useful to adopt the terminology used by the NHS in delineating the areas: health informatics, genomics, and bioinformatics applied to the physical sciences.

Health Informatics

Health informatics uses large-scale patient data to inform clinical decisions and public health planning. For example, patterns in prescriptions and patient outcomes have long been used to identify risky combinations of medications, helping clinical systems flag potentially harmful interactions before they occur. By consulting the National Institute for Health and Care Excellence’s Drug Interaction Checker, medical professionals can check for drug interactions. Most instances are evidenced by “studies”, or broadly, learning from data.

During the COVID‑19 pandemic, epidemiological models exemplified statistical learning: they continuously fitted parameters such as the reproduction number and infection‑to‑hospitalisation delay to observed UK time‑series of cases and deaths. By identifying shifts in those patterns following school closures, self‑isolation guidance and lockdowns, models could infer how interventions altered disease spread and predict NHS demand weeks in advance. 

Figure by Katapult Magazin, 2020.

The above figure from Katapult Magazin, shows a simulation illustrating infection spread and mortality resulting from overwhelmed hospital systems under two conditions: unrestricted social interaction (left; 200 individuals freely mobile) and social distancing (right; 25 individuals mobile). Colour legend: Green = Healthy, Red = Infected, Blue = Recovered, Black = Deceased.  

Tools like the Framingham Risk Score show how “statistical learning” integrates with patient facing clinical practice. Researchers used mathematical tools to identify which combination of age, blood pressure, cholesterol, smoking and diabetes best predicted 10‑year heart‑disease risk. These models were then converted into simple point‑based criteria such as age brackets, cholesterol thresholds and blood‑pressure categories so that clinicians can apply them at the bedside. You can even use this Framingham Risk Score Calculator yourself.

TASK: Think about “who” you know that would attend a research study, clinical trial, etc. Does this cohort of people represent a true cross-section of society?

Genomics

Genomics is another area where learning from data is applied. Certain genetic variations are known to increase the risk of disease or alter how a person responds to treatment. However, these links aren’t discovered through logical deduction, they are found by identifying recurring patterns between variations in DNA and measurable health outcomes across large groups of people.

The NHS Genomic Medicine Service relies on databases of known gene–disease associations. These relationships were uncovered by comparing the genomes of people with a particular condition to those without it, and identifying patterns within this huge corpus of data, linking and correlating genomic data to help provide new treatments and diagnostic methods.

Figure from Fundamentals of Machine Learning and Deep Learning in Medicine by R. Borhani, S. Borhani & AK Katsaggelos, Springer, 2022.

The above figure shows how clustering of data can reveal two groups of similar patients and two groups of similar genes.  

The 100,000 Genomes Project is yet another UK-based genomics for healthcare initiative worth exploring.

TASK: Visit the NHS webpage about getting involved in clinical trials. Think critically about the accessibility of this page and service. Will this attract a representative cross-section of the UK?

Bioinformatics in the Physical Sciences

This refers to the use of computational models and pattern recognition techniques to study biological molecules such as proteins, DNA, and potential drug compounds. Before AI, early drug discovery relied on chemical pattern-matching: researchers would test thousands of compounds in the lab, looking for similar structures or behaviours that suggested therapeutic value.

This may also include working with aspects of medical equipment, software, computer interfaces, etc. with imaging devices being a notable example.

The following video, “Genomics Specialist Careers: Meet the Clinical Bioinformaticians” by the Genomics Education Programme, highlights how bioinformatics, and “learning from data” fit into the NHS workflow.

“There is no way that anyone could look at that sea of data and make sense of it in a way that would […] result in anything clinically actionable” – Dr. Hesketh (quotation from above video)