The use of AI in real-world healthcare decision making

This article originally appeared in the April '22 magazine issue of Electronic Specifier Design – see ES's Magazine Archives for more featured publications.

TurinTech was conceived by four co-founders who met at university during their PhD research. They felt growing frustration at the manual machine learning developing and manual code optimising process. Such processes were time-consuming, resource-intensive and required deep domain expertise. To overcome these challenges, they built an AI optimisation platform, EvoML, to automate the process of creating, deploying and optimising AI.

Benefiting drug discovery and patient diagnosis

AI can be used to generate new models to predict or model scenarios, a key use case being in drug discovery. The technology could be used to discover how a drug may affect certain candidates – without the need for testing. Similarly, it can compare samples against each other, which would usually be performed using statistical approaches.

EvoML can assist in medical diagnoses, providing the tools to identify cancers or diseases affecting neurology and cardiology, which are the most difficult and sensitive areas for a lot of medical professionals. A particular dataset can be uploaded to the platform, and then models can be generated and a healthcare professional can choose the models to make a diagnosis.

"If you’re looking at the lungs, trying to identify something that is causing discomfort to the patient, the process of identification through different consultants can be lengthy,” said Kanthan.

"Let’s assume the patient has a life-threatening condition like pulmonary fibrosis, and the GP is not necessarily able to make the full diagnosis.”

The GP might not be able to make a diagnosis for many reasons, including a lack of experience or expertise to recognise the condition straight away from standard MRI scans. While the scans would then be passed along, and eventually seen by a senior expert, this process can take months to happen before the patient receives a final diagnosis.

TurinTech’s AI can help with this problem, allowing diagnoses to take place straight from the MRI using the correct models. The model can notify a GP that it has discovered something, speeding up the process and assisting experts at a system level.

Explainable AI

AI explainability refers to making the predictions made by AI more understandable to humans. In healthcare, if you are relying on such models, there needs to be a level of explainability to allow for transparency and understanding of how the model came to a specific diagnosis. A lack of understanding could have major consequences, especially if something goes wrong.

“It’s one of those use cases where if you’re looking at diagnosing somebody who has something as serious as cancer, if the machine learning model is predicting that they might have cancer, that’s quite insightful. At the same time, the explainability of that prediction needs to be understood,” said Leslie.

AI informing healthcare trust and regulations

If a healthcare provider uses TurinTech’s technology, they can see how the diagnosis for life-threatening diseases has been made and determine how that is compliant with regulations.

Currently, TurinTech’s model doesn’t make the decision alone: it assists and offers its conclusion to an expert for review. As is so often the case in AI operations, a human being is still involved in regulating the decision-making process.

Energy and memory constraints in wearable devices

Wearables are used massively throughout the healthcare industry, but they have some problems as Kanthan explains: “The more often you’re using your device, the more it is going to consume energy. Even when you’re not using the device, it’s consuming energy.”

With wearable devices that measure heart rate, for example, an algorithm uses an AI model to predict what the patient’s heart rate is going to be, based on sensor data and the interval of collection. This makes it “as accurate as it could be” according to Kanthan. This model consumes a lot of data to provide that level of accuracy, draining the battery even more. Battery life is what matters more to consumers, continued Kanthan: “The longer the battery lasts, the longer you can make real value of the use of a product.”

TurinTech’s platform uses code optimisation, which can reduce the memory and CPU and the energy consumption of any device that the given code is running on.

Strengths and challenges of working with AI

Kanthan said: “The greatest strength is in the automation of the work, the trusting of AI, explainability of AI, and making things more efficient.

“If you have a repetitive process using AI, it can predict the process – at the same time, augmenting the capability of existing people and technology, improving that efficiency and making it more robust.”

Kanthan explained that the biggest challenge to overcome is AI making the wrong decisions. It requires adoption and getting people to collaborate. AI needs to make mistakes to improve and learn and adapt to the user. He said: “The challenge is: how do you improve the AI and accuracy by making minimal or no mistakes in an ideal world?”

How AI will benefit the future of the healthcare industry

Having variants modelled and understood before the population is affected by them is the future of healthcare, Kanthan explained. This has been demonstrated by the COVID-19 pandemic, showing how a virus can appear suddenly and disrupt the world.

Advances in MRI technology will also benefit the industry. AI will begin to be applied straight at the source, allowing an MRI to pick up something that perhaps the healthcare professional wasn’t initially looking for, wasn’t related to the initial industry, or wasn’t a medical concern.

In addition, the use of AI in healthcare data will be a huge benefit to the industry. The idea of having shared data, so people’s experiences and history are amalgamated into one large dataset, means that predictions can be used for your own healthcare. Privacy is a challenge here, as healthcare data is a well-known sensitive field, and care must be taken around the handling and anonymisation of such data.

In addition, the use of AI in healthcare data will be a huge benefit to the industry. The idea of having shared data, so people’s experiences and history are amalgamated into one large dataset, means that predictions can be used for your own healthcare. Privacy is a challenge here, as healthcare data is a well-known sensitive field, and care must be taken around the handling and anonymisation of such data.

The future of AI

Kanthan explained that, currently, it seems as though TurinTech is the only one working on this specific area of code optimisation, thus pioneering the technology.

Whilst he considers that people are building AI to solve problems, TurinTech has focused on every model being based on code, allowing code optimisation to efficiently solve problems within healthcare.