Healthcare and generative AI: What are the emerging use cases?

Machine learning has been widely adopted in healthcare, with predictive AI algorithms being used for a variety of functions ranging from image-based diagnosis in radiology to genome interpretation. Generative AI – which uses algorithms (such as large language models (LLMs)) to create rather than simply analyse – has captivated the tech world, but brings with it both risk and opportunity. On one hand, the risk of bias and inaccuracy calls into question the ethics of using it in healthcare, but on the other hand, research suggests that it also has the potential to vastly streamline and improve services.

According to research by Accenture, 40% of working hours across all industries could be impacted by LLMs. The firm looked at work time distribution and potential AI impact by identifying 200 tasks related to language and how these were distributed throughout industry (based on employment levels in the US in 2021). Language tasks accounted for 62% of total worked time, with 65% of those tasks having high potential to be automated or augmented by LLMs.

In health specifically, 28% of working hours were defined as tasks with higher potential for automation (with the potential to be transformed by LLMs and requiring reduced involvement from a human worker). Eleven percent of tasks had higher potential for augmentation (requiring more human involvement).

So, how exactly can generative AI impact healthcare? Here’s a look at some recent examples.

Microsoft designs clinical tools to enhance productivity and ease the burden on HCPs

Rich Birhanzel, Accenture’s global health industry lead, suggests that, as in areas such as customer support, the immediate benefits of generative AI may lie in summarisation. He told HealthcareITNews:

“Instead of a nurse or doctor recording information – from vitals to treatment plans – gen AI can listen to the conversation during the appointment and create a summary that can be added to an electronic health record. Additionally, the technology can also simplify complex medical language into summaries that patients can understand, and that can be easily translated into any language.”

We are already seeing solutions for some of these sorts of use cases. At the end of March, Microsoft’s Nuance Communication announced a new clinical documentation tool powered by GPT-4. The tool, called Dragon Ambient eXperience (DAX) will enable healthcare workers to automate clinical documentation simply by ‘listening’ to physician-patient consultations.

“Nuance and Microsoft came together with the goal of helping to digitally transform healthcare, and today we are marking the next step forward in the ongoing evolution of AI-powered solutions for overburdened care providers,” Mark Benjamin, CEO of Nuance, explained in a statement.

Microsoft is also expanding its AI capabilities in healthcare with its partnership with electronic health vendor Epic. The partnership will see Microsoft integrate Azure OpenAI Service technology added to Epic’s electronic health record (EHR) software, which the company says will “increase productivity, enhance patient care and improve financial integrity of health systems globally.” One of the first initiatives is using generative AI to automatically draft message responses, currently being piloted by UC San Diego Health, UW Health and Stanford Health Care.

Google aims to make strides towards medical diagnosis with Med-Palm 2, but gaps in accuracy remain

As well as automating tasks like note-taking, pharma and healthcare companies are experimenting with generative AI for greater efficiency in other areas of medicine, such as decision-making and diagnosis.

Recently, Google unveiled its latest generation large language model, Palm-2, which now has improved multilingual, reasoning, and coding capabilities. On the back of this has also come Med-Palm 2 – an AI that has been specifically developed for the healthcare industry and is trained to answer medical questions. According to Google, Med-Palm 2 achieved 85%+ accuracy on the MedQA dataset of US Medical Licensing Examination (USMLE)-style questions and scored 72.3% on the MedMCQA dataset comprising Indian AIIMS and NEET medical examination questions.

While this means Med-Palm 2 has improved on the original version, Google stated that there are still “significant gaps when it comes to answering medical questions and meeting our product excellence standards,” suggesting that there is a way to go before the technology can be implemented into real-life clinical settings. However, the tool will be available to a select group of Google Cloud customers for testing, which Google says is “to explore use cases and share feedback as we investigate safe, responsible, and meaningful ways to use this technology.”

Elsewhere, start-up Glass Health is implementing generative AI with the aim to assist clinicians in drafting clinical plans and generating differential diagnosis (DDx) – which means distinguishing a particular disease or condition from others that present as similar – based on patient symptoms.

According to co-founder Paul Dereck, in just two days, 14.6K people used Glass AI to submit 25.7K queries, with users rating 84% of DDx and 78% of clinical plan outputs as helpful. Like Google, however, Dereck also highlighted a lack of total accuracy, with users citing an accuracy rate of 71% for DDx outputs and 68% for clinical plans. Consequently, Dereck adds that the focus should be on Glass AI’s ‘helpfulness’ rather than precision, suggesting that this type of technology should merely be used as an aid (and to sit alongside) human judgement.

Even so, there has been suggestion that, regardless of how much of an impact AI has on an end diagnosis, generative AI should be used with caution in healthcare settings. NPR cites research by MIT student, Marzyeh Ghassemi, which found that AI systems can be biased, particularly when they have been trained by humans (as they will reflect existing human biases). “It has the sheen of objectivity: ‘ChatGPT says you shouldn’t have this medication. It’s not me – a model, an algorithm made this choice,'” Ghassemi explained, bringing into question the accountability for decisions made by this tech.

Biopharmaceutical companies make advances in drug discovery and development

Generative AI is also starting to having an impact on drug development, both in terms of revealing new therapies and the speed at which they can be discovered.

Researchers at the University of Toronto, for example, have developed an AI system using generative diffusion – the same technology as image creation tools like DALL-E – to develop new proteins that are not found in nature.

“Our model learns from image representations to generate fully new proteins, at a very high rate,” explains Philip M. Kim, a professor at the University of Toronto. “All our proteins appear to be biophysically real, meaning they fold into configurations that enable them to carry out specific functions within cells.”

Bio-tech company Insilico Medicine has also published findings on how it is using generative AI to design new molecular structures that can target proteins which contribute to disease progression.  To do so, it uses Chemistry42, its machine learning platform that connects generative AI algorithms with medicinal and computational chemistry methodologies. Chemistry42 has also enabled Insilico Medicine to discover a small molecule inhibitor of CDK8 for cancer treatment using a structure-based generative chemistry approach.

Elsewhere, German biotechnology company Evotec has recently invested in UK-based Exscientia, to accelerate AI-powered drug development. The partnership recently resulted in a phase one clinical trial on a new anticancer molecule, which Nature states was found in just eight months using Exscientia’s ‘Centaur Chemist’ platform. Traditional drug discovery processes typically take four to five years.

Zepp Health integrates generative AI into wearables for personalised health management

Another way generative AI is being utilised is to collect and analyse data from smartwatches and wearables. In doing so, the technology can assist companies in providing personalised care, such as health or weight management recommendations tailored to suit the individual user’s needs.

One example is Zepp Health, which recently launched several AI-powered products to integrate with its existing smart wearables. Zepp Aura, for instance, is a sleep and relaxation platform that offers users personalised sleep coaching, sleep analysis, as well as AI-generated sleep music which adjusts in real-time based on the user’s heart rate. Premium subscribers of Zepp Aura can also access an AI chat service, which answers questions about sleep and general wellness.

Meanwhile, Zepp’s smartwatch brand Amazfit also announced back in March that it would be integrating ChatGPT into its GTR4 watch, to enable users to ask ‘ChatGenius’ questions using natural language.

“By applying large language model (LLM) and Generative AI technology to our smart wearables, we are empowering users to make informed decisions and achieve their wellness and fitness goals intelligently,” explained Wang Huang, Chairman and CEO of Zepp Health.

Going beyond general wellness, metabolic health company January.ai is using generative AI to hone in on one specific issue – glucose levels. January.ai estimates and predicts individuals’ glucose response to various foods even when members are not wearing a continuous glucose monitor (CGM), thereby empowering them to make smart food choices. Co-founder Noosheen Hashemi explained:

“With January AI, members enter foods they are considering eating, and the predictive AI model will tell them what it will do to their blood sugar. It helps them make better decisions about what they’re about to put into their body, kind of like having an AI nutritionist in your pocket to let you know whether you should order a smoothie or an arugula salad, and what happens if you take a 25-minute walk afterwards.”

With all of these use cases, there is obviously a crucial role for trials and having professionals ‘in the loop’, and a debate to be had about the potential for bias, accuracy, privacy and overall patient experience. But with big investments being made, these debates about generative AI and its implications for the healthcare industry are set to continue.