Matt Lucas FBCS explores the intersection between technology and healthcare with a particular focus on artificial intelligence.
Thanks to a minor but longstanding health issue, I recently used the NHS. After waiting eight months for a referral appointment (due, I’m told, to a lack of staff), it turned out that I had been misdiagnosed over ten years ago, which is why my condition never fully resolved itself. My experience was far less serious than many: the challenges of waiting times and labour shortages within the NHS are well documented, and there are numerous (often heart-breaking) stories of the impacts of misdiagnoses. But, today I want to talk about the technology: specifically, how advancements in AI can help with ongoing healthcare challenges such as these.
AI, or rather machine learning, is particularly good at making sense of complex, unstructured data – that is, data without a clearly defined, machine-parseable format. In healthcare, there are many formal data structure standards, ranging from imaging formats to taxonomies for diseases and treatments. However, an estimated 80 percent of available medical data is unstructured – why is this?
Science and data ranges
The answer is partly because healthcare is an inexact science. For example, think of the inconsistencies in how COVID cases are reported across the world. Or the range of observations that a clinician might use to decide on a particular course of action – say, whether to recommend a caesarean birth. Faced with a range of possible reporting options, doctors often resort to free-text fields when describing diagnoses; it’s easier and allows them to express a more nuanced reasoning.
Applying machine learning algorithms to these kind of datasets can provide insights that would take too long for humans to do.
The first, maybe most obvious application would be to assist in diagnosis. A patient’s medical background can be complex and important details buried away in the health record; AI can be used to extract these details. For example, IBM Watson has been used to quickly extract relevant data from electronic health records to assist radiologists to diagnose health issues, saving the radiologists’ time and improving the confidence of their decisions.
AI can also be used to search health records for related symptoms and comorbidities, which can help to detect non-obvious causes of illness. And by expanding the sample size, it can also assist in the research of new treatments and vaccines.
Similar technologies can also be used when planning treatments. For example, Microsoft’s InnerEye project can allow oncologists to distinguish between healthy organs and cancerous tumours.
Be part of something bigger, join BCS, The Chartered Institute for IT.
These are just a few interesting examples of AI use-cases in healthcare, but there are several hurdles to its more widespread adoption, including valid concerns around privacy, reliability and investment cost. However, I want to focus on the issue of trust: would you trust your diagnosis or treatment to be handled by an AI?
Importantly, the phrase that underpins most existing AI healthcare applications is ‘decision support’. They are not generally designed to replace human involvement, but instead give experts additional insight that allows them to make better decisions faster. This said, research suggests that AIs are now at least on a par with humans when it comes to diagnosis.
Technical challenges and bias
One of the biggest technical challenges to overcome however, is bias – where errors the AI’s training data can propagate through into errors in the AI’s recommendations. For example, using sensor data that does not correctly detect low blood oxygenation in black patients could lead to AIs misdiagnosing hypoxemia (low blood oxygen) cases.
To help, it is important that AIs are transparent – explaining their thought processes clearly, and with references to primary research as appropriate.
The market for AI in healthcare is significantly , and growing. According to Grand View Research, global AI in the healthcare market was valued at $10.4 billion in 2021 and is expected to expand at a yearly rate compound annual growth rate of 38.4 percent from 2022 to 2030.
But, the size of the market merely gives us an indication of the level of interest we’re experiencing; it does not tell the whole story. We already rely increasingly on AI to assist in everyday activities such as driving, shopping and entertainment; embracing AI to support our future healthcare needs has the potential to directly improve our health.