Guest Article by Dr Ron Granot.
For quite some time, I’ve been reading a variety of eye-catching headlines about the impact of Artificial Intelligence (AI) on the field of medicine. Some are alarmist and others seem fanciful.
Sceptics may point out the great promise and terrible delivery that technology has brought to medicine in the last decades. However, in areas such as imaging and pathology, successful advances have been made, and we have better access to more powerful tests than ever before.
So, let’s take a deeper dive into the role that machine learning and AI will take into our medical future.
Medicine as the flow of information
One way to fully grasp the power of AI is to look at one primary concern in medicine – managing information flow – and see how machine learning can have an impact.
To begin, let’s examine medicine in terms of data. Doctors have a series of tasks we need to achieve:
1. Diagnosis – The patient presents with a problem, which the patient describes to the doctor as a series of symptoms.
2. Hypothesis – The symptoms are combined by the doctor to create a hypothesis as to what the underlying problem or mechanism of the problem is.
3. Examination – The patient is examined physically, which assists in clarifying and translating some of these symptoms into more specific hypotheses from the data points.
4. Presumptive diagnosis – This is converted to a presumptive diagnosis or list of the possible problems that the patient could have.
5. Testing – This is further confirmed (or refuted) by different tests: blood tests and imaging – CT, MRI – as well as anatomical pathology.
6. Treatment – A course of treatment is prescribed, which can be pharmaceutical, non-medication (physical, mental or emotional) or surgical.
Looking at this progression in terms of information flow, this is a process of obtaining, sorting and reformatting information to best allow the doctor to help the patient feel better. But it is certainly heavily based on information manipulation – the patient history is about obtaining what the patient feels or experiences, the examination is extracting more information from the patient’s body directly, and tests take this to the next level, examining the tissues themselves. Each additional piece of information modifies our prior hypothesis, and this can be correlated to the likelihood of treatment being effective.
Computers and information
Let’s consider what computers can do to information, particularly in the realms of AI and machine learning.
AI takes independent, entirely digital information and transforms it into different types of information (you can see this as attaching different meaning to the information). For example, a digital photograph can be interpreted by a machine learning algorithm and the scene described in words or acted on immediately.
Where AI is proving hugely beneficial, is where it is beginning to assist medical teams at all phases of the patient lifecycle, from diagnosis through to treatment.
Computers and machine learning, in medicine
Now to combine the two aspects:
1. The information content of medicine
2. What machine learning algorithms can use in processing information
To see where AI can be most helpful, the most obvious part of this equation is to look for what is most digital.
Medical imaging, such as CT and MRI, is completely digital from beginning to end. More recently, ophthalmology, through digital photographs, and dermatology, through photographs of lesions of the skin, have become digital too. These are therefore a great place to start machine learning algorithms in medicine – and is why such great machine learning advancements have been made in this area.
Machine learning algorithms can examine and decide if:
· There is a blocked blood vessel to explain a new stroke on a CAT scan
· The heart is beating normally and is normal in structure on a cardiac MRI
· If a patient’s Multiple Sclerosis is getting worse from scan to scan
AI is already amongst us
AI is being increasingly used in medicine to generate ‘decision support algorithms’ for routine patient care. Generally, these ‘decision support algorithms’ are expert written – usually by a group of doctors with reference to the medical literature and in a format that can help immediately, such as a flow diagram or decision tree, to assist the doctor making the right treatment decision.
In the US, a company called Viz.ai is using artificial intelligence to scan CT images for indicators associated with strokes, and then sending text notifications to Neurovascular Specialists if potential large vessel occlusions are identified. Stroke teams can then consult in real time via secure mobile interfaces, driving fast treatment decisions that save lives.
MaxQ is using AI to automatically retrieve and process non-contrast CT images to provide case-level indicators for possible acute intracerebral haemorrhages.
Imagen is using AI to scan X-ray images for distal radius fractures.
This technology is designed to enable doctors to make faster, more accurate decisions when diagnosing and treating patients, and it is having a significant impact.
Another example is IBM’s Watson, which has become involved in assisting with making treatment decisions for cancer patients – in medical oncology. Here, the information about the patient and the specifics of the tumour, such as gene changes and tumour grade, are used to help select the best chemotherapy treatment that should be used on that patient. They do this by finding the most appropriate clinical evidence to support that decision. The more specifics about a patient’s background and other medical problems, the more accurate such a decision support system becomes.
The reason that oncology is favoured is the significant importance of binary and clear-cut data in making decisions. A mutation is present or absent (or at least graded) in a tumour and the degree of spread is categorised. Hence, lots of data points to split patients into definite categories to assist with grouping outcomes of different therapies in different groups to make a good decision ahead of time.
However, this has not been as successful a deployment of AI thus far, and it has not lived up to the hype it originally had. More is to come, I’m sure.
The future of AI in medicine
As researchers obtain greater access to appropriate datasets, they can develop algorithms and machine learning programs to interpret these narrow ranges of clinical circumstances, such as a specific tumour or a specific disease. These algorithms can be served to practising clinicians through digital marketplaces, where they can use the functionality, as appropriate.
This is not science fiction, either! These marketplaces are already beginning, with both the DragonDictate software maker Nuance, as well as Philips – in its role as maker of medical imaging systems and pathology systems – opening up such algorithms for their clients’ use.
As digital marketplaces get bigger, and algorithms get more accomplished, they can be offered as product suites – with their use dictated by both individual clinicians and larger health organisations.
How can AI complement the daily medical routine of doctors?
Gradually, clinicians will likely be able to handle greater patient workloads, but the amount of work required will likely diminish with the productivity improvement that AI will be able to offer.
When we can use machine learning to create technology that can reprioritise the work of the medical community and save patients’ lives, the opportunities are endless.
Thomas M Maddox, Professor of Medicine and Director of the Health Systems Innovation Lab at Washington University sees the future of AI in medicine like this: “Doctors’ roles may shift from being data collectors and analysers, to being interpreters and counsellors for patients as they try to navigate their health.” I couldn’t agree more.
About Dr Ron Granot
Dr Ron Granot is an experienced Neurologist, Clinical Advisor and Conjoint Lecturer for the University of New South Wales. Ron graduated from UNSW in 1999, with a Bachelor of Medicine and Bachelor of Surgery and Medical Science. He achieved First Class Honours and the University Medal. He has produced several publications and journal articles, as well as book chapters on endozepines, and computer technology in medical practice. He is also the creator of BetterConsult, a pre-consultation tool that captures patients’ presenting symptoms, medication and other relevant clinical information and then translates the data into concise medical notes, ready for doctor review before the consult.