From modelling disease risk to helping with diagnosis, artificial intelligence is starting to make itself felt in the medical field. Abi Millar looks at the latest studies and technological developments.
Machine learning has long been touted as the next big thing for healthcare. With countless startups investing in that promise, applications are emerging across everything from diagnostics to drug discovery.
Of course, artificial intelligence (AI) within healthcare has some way to go before it realises its potential. Since the costs of getting it wrong are so high, the medical profession has tended to approach this field with caution. However, machine-learning techniques are poised to hit the mainstream over the next few years.
There is a political momentum here as well as a scientific one. In May, UK Prime Minster Theresa May set targets for a “whole new industry around AI in healthcare”, stating that artificial intelligence could prevent up to 22,000 annual deaths from cancer by 2033.
Her plans were described as “pioneering” by Cancer Research UK, which added that advances in detection technology “have the potential to save hundreds of thousands of lives every year”.
So how are machine learning techniques being used currently, and what kind of opportunities could be on the cards?
While human radiologists won’t lose their jobs any time soon, deep learning computers are already beginning to outpace them in diagnosing certain cancers. This, at any rate, was the verdict of several recent papers, which explored machine-learning applications within oncology.
In one paper, published in the Journal of Medical Imaging, a team at Case Western Reserve demonstrated that machine learning might have the edge when it comes to picking out malignant lung nodules. Human radiologists have a tough job here – of all the ‘suspicious’ or ‘indeterminate’ nodules that are flagged up on a CAT scan, around 98% turn out to be benign. The study found that a computational imaging technique was between 5% and 8% more accurate.
Another study, published in the Journal of Magnetic Resonance Imaging, found that machines were better than humans at diagnosing prostate cancer from an MRI scan. When a machine analysed the images, rates of false positives were around 50% lower, and rates of false negatives around 70% lower. (Israel-based startup Ibex has also developed an AI system for diagnosing prostate cancer.)
In a third study, published in Annals of Oncology, researchers showed that a deep learning conventional neural network could diagnose skin cancer better than dermatologists. The network was trained to tell the difference between benign moles and malignant melanomas. Compared to the control group of dermatologists, it missed fewer melanomas and misdiagnosed fewer moles as cancerous.
“When I came across recent reports on deep-learning algorithms that outperform human experts in specific tasks, I immediately knew that we had to explore these artificial intelligence algorithms for diagnosing melanoma,” said study author Professor Holger Haenssle, of the University of Heidelberg in Germany.
AI is also being used to analyse molecular information from cancer patients, identifying people who may respond to certain therapies despite falling outside the target demographic. A recent study used machine-learning techniques to find ‘hidden responders’ – i.e., patients who slip through the net when using conventional sequencing strategies.
Assessing the risk of stroke and dementia
In May, scientists from Imperial College London and the University of Edinburgh announced they had used machine learning to detect a common cause of strokes and dementia. Their software could identify cerebral small vessel disease (SVD) with unparalleled accuracy, as well as estimating its severity.
SVD is a progressive neurological condition, in which blood flow to the white matter regions of the brain is reduced. Over time, brain cells die, which can lead to dementia or stroke.
While CT scans can be used to diagnose the condition, it can be hard to estimate exactly how far the disease has spread. Knowing that could be hugely advantageous, as it could help you flag up patients who are unsuited to certain medications, as well as assessing their dementia risk.
The software learnt to detect SVD by analysing more than 1,000 CT scans from stroke patients. According to results published in the journal Radiology, it was 85% more accurate than an MRI scan (the ‘gold standard’ of diagnosis) in predicting the severity of the condition.
Professor Joanna Wardlaw, head of neuroimaging sciences at the University of Edinburgh, said: “This is a first step in making a scan reading tool that could be useful in mining large routine scan datasets and, after more testing, might aid patient assessment at hospital admission with stroke.”
It’s not the first time this year that machine learning has been used in the service of stroke patients. In March, a team from the Medical University of South Carolina and the University of Tennessee Health Sciences Center, studied a device called the Cerebrotech Visor that can detect stroke within seconds. Since the device can tell the difference between minor and severe strokes, it could enable a personalised approach to treatment.
Meanwhile, team at the Boston University School of Medicine has used machine-learning techniques to identify risk factors for dementia. Through mining retrospective data, they found new combinations of factors that could make people more susceptible in later life.
Machine learning has also been used to predict epileptic seizures. Researchers at the University of Sydney are working towards a portable, affordable device that could be used by people with treatment-resistant epilepsy.
This device uses an algorithm that reads a patient’s electroencephalogram data. It predicts if they will have a seizure to 81.4% accuracy, giving them a 30-minute warning and time to find somewhere safe. The system learns as brain patterns change, and will become more sensitive over time.
Meanwhile, researchers in Massachusetts have developed machine learning models that can predict patients’ risk of contracting C-difficile. This widespread infection is a huge problem for hospitals, and kills around 30,000 Americans every year. The model was able to generate daily risk scores for each patient, classing certain people as high-risk well before their actual diagnosis.
Machine learning also has its uses within mental health, with a number of startups developing apps that detect symptoms of depression. Cogito, for example, merges a number of AI techniques to analyse communication patterns. Its Cogito Companionapp, trialled on patients at Massachusetts General Hospital, uses voice analysis tools as a gauge of wellbeing.
A further use for machine learning lies in monitoring heart disease. Most recently, researchers at the University of Southern California developed a new predictive model for this condition. The person uses an app on their smartphone to measure their pulse, and a machine-learning model detects their arterial stiffness.
“That’s how you go from an $18,000 tonometry device and intrusive procedure to an iPhone app,” said Niema Pahlevan, one of the inventors.
As should be clear, this rundown is far from an exhaustive list. New ideas and applications are emerging all the time, with powerful implications across the med-tech sector. As machine-learning techniques grow more sophisticated, we will surely start to move beyond speculation and into the realm of everyday applications.
This article appears in the July 2018 edition of Medical Technology