From data to diagnosis – how AI is changing the world of medicine

AI is a rapidly developing technology, especially in the world of medicine. Its ability to take over time consuming processes, such as documentation, can allow healthcare workers to focus their attention towards the treatment of their patient to benefit patients and healthcare workers.

Health experts say when AI is used to its full potential, it has the ability to reduce healthcare costs, increase accessibility and improve the overall system. Current uses of AI in medicine offer promising developments.

Use in medicine

AI has many uses in medicine. Currently, it is being used as a “scribe” which documents doctors’ notes; medical imaging analysis for scans; and bio-marker analysis.

AI scribes reduce the time healthcare workers spend on documentation. Scribes can create live transcripts while the doctor is assessing the patient. When using the most popular scribe programs, such as Lyrebird, once the transcript is made, and the file is downloaded, the transcript and audio is deleted.

Generally, the write-up by a health worker for a patient takes just as long as the assessment itself, which makes scribes a useful development in hospitals worldwide.

A study by researchers at the Perelman School of Medicine at the University of Pennsylvania in the US surveyed 46 clinicians and revealed that the use of scribes increased patient face time by 20%, and reduced time spent outside of working hours by 30%.

Medical imaging analysis uses object detection to recognise anomalies in x-rays, MRIs and CT scans. The AI is trained on a series of images – half of which are normal, and half of which are affected by a tumour or a disease. The AI is trained on hundreds of these imaging sets and learns what to identify within the image. The data must be from a variety of scenarios and locations, to ensure the AI is not recognising environmental patterns outside of the scan itself.

An extensive review of the benefits of AI in health by a team from the University of Louvain in Belgium, concluded:  “Data collection and curation are indeed of paramount importance, since errors, biases, or variability in the training databases are often directly reflected in the model behaviour and can have dramatic consequences in the model performances and its clinical outcome. Some examples of these issues include gender imbalance, racial bias, or data heterogeneity due to changes in treatment protocols over time.”

Bio-marker analysis allows large sets of patient data to be compiled and analysed. AI analyses the patient data and identifies patterns. It then uses these patterns, in combination with a predictive AI model, to identify any markers for disease. It can also predict certain reactions or side effects to a specific medication.

So how does this AI work?

AI programs contain machine learning and natural language processing programs. Natural language processing (NLPs) models break big data inputs into smaller, usable data. The NLP program then makes an assumption based on input values, by asking the data: “Will it meet this requirement… yes/no?”

An example of this is the Supportive Weekend Interprofessional Flow Team (SWIFT), a program that compiles a list of patients ready to see doctors for discharge used in nursing and allied-health at the 522 bed Lyell McEwin Hospital in Adelaide, the capital of South Australia.

Toby Gilbert, who helped create SWIFT, described the process to me.

“The AI takes each patient’s most recent ward round note and breaks the text down so that the AI can understand. From there, it uses an algorithm based on the notes to answer the binary question: “Will the patient be out in 2 days? Yes/no?”

“Then it uses numbers like heart rate and blood test results in a second algorithm to give each patient a score that reflects the likelihood of discharge – the Adelaide Score. We then rank the patients who are for discharge from most to least likely.”

In medical imaging analysis programs, the AI system relies on object detection. However, the reliability of object detection relies on the variation of training. There are other applications of AI in medicine that involve a different type of program called a large language model (LLMs).

LLM’s will ‘profoundly impact‘ the future of fields like chemistry and medical health according to researchers.

These models can be used for scribes and electrical medical records (EMRs).

LLMs convert information into numbers, then create a sequence of numbers based on the input and then it tries to predict what number comes next in the sequence.

It assigns this based on probability percentages. The number that has the highest probability of being the most accurate, gets inserted into the sequence. Once repeated, the now completed sequence of numbers gets converted back into words to be read.

What about security?

When discussing AI, especially in medicine, security is a big concern. Gilbert says there are a number of precautions in place to increase security of these programs.

“Data governance laws that exist in South Australia prevent any data from SA to be uploaded to any network that goes outside of SA.

This means that the data used in AI cannot be uploaded or stored to an international network, which would mean all data stays encrypted locally, and does not contribute to the training data for future algorithms, which is what happens when you use services like ChatGPT.”

Similarly with scribes, as mentioned before, once the transcription has been made, the recording is deleted.

While cyber security will always be a risk, as it would be without AI, people like Gilbert argue AI has improved the capability of cybersecurity services. This is because AI can be used to detect any unusual activity on its network and take pre-emptive actions to deal with these threats and de-escalate these threats in real time.

AI can also use behaviour analysis to detect when user behaviour seems out-of-character. This analysis can enable the AI to detect when an account has been logged into by someone other than the primary user. Both of these applications of AI allow these programs to increase protection and security.

Muhammad Waqas and colleagues from Edith Cowan University in Western Australia, conducted what they believe was the first comprehensive survey to review the AI solutions for all possible security types and threats.

In their paper in Artificial Intelligence Review they argue: “AI can be more effectively used to overcome the upcoming advanced security threats.”

According to statistics revealed in March this year by Embroker, companies that used extensive AI features in security were able to discover and contain breaches 108 days sooner than those who did not use AI features.

“These breachers also cost companies that were using AI $2.75m (USD$1.76m) less on average than companies that did not,” the report says.

Gilbert acknowledges that societal scrutiny and caution makes AI more accurate and developed.

“The people behind these programs are pushed to make their programs as perfect and accurate as they can, as they know that once society sees one small mistake, the whole program can be discredited.”

In addition, there is the question of responsibility. If the AI makes a mistake that causes harm, who is at fault?

“For now, most of these AI programs used in medicine are supervised, which have a ‘human-in-the-loop’,” says Gilbert.

“This human is responsible for application of the program, and in the case that a mistake is made, responsibility falls to them.”

What about the future?

In the emerging field of precision medicine AI is being used to predict machine malfunction before it occurs, allowing safe and proper function of these machines, says Kevin Johnson from Vanderbilt University Medical Centre in Tennessee in a report in Clinical and Translational Science.

AI can also create cheaper, and more accessible healthcare. “Accessibility is increased by AI as it is a portable tool that can complete most information-gathering processes before a clinical interaction, allowing the limited availability of clinicians to spend their time on examination and treatment,” says Gilbert.

Not surprisingly these benefits can assist the workforce too. Healthcare workers who successfully integrate AI into their workflow have potential to enhance their efficiency and productivity, a conference was told last year.

This can enable more work to be done with a smaller workforce, while increasing job satisfaction by allowing healthcare workers to focus on their patient relationships.