Technology Record - Issue 30: Autumn 2023

175 PUBLIC SECTOR instruct LLMs to produce intermediate reasoning. It also utilised few-shot prompting, enabling GPT-4 to learn from a few examples without requiring extensive training data. Pangaea demonstrated GPT-4’s capability to attain performance levels as high as 96 per cent F1 scores for disease classification, and the potential of providing assistance to clinicians in the prospective evaluation of patients. Another company, ScienceIO, is on a mission to develop language models to decipher healthcare data. Its solution transforms medical text into enriched data to enhance patient care. With the application programming interface, users can dissect medical records and pinpoint essential details like medications, treatments and procedures. Additionally, ScienceIO ensures data security by excluding sensitive healthcare information. The foundation behind ScienceIO’s work is its healthcare-specific language models, which are trained using high-quality biomedical and clinical data to avoid potential biases and bad practices that may seep in from web data, posing risks to healthcare applications. ScienceIO recently unveiled its Embeddings API, which helps developers to construct search products that uncover critical patient information. Beyond aiding patient care, startups also leverage large language models to understand and improve patient experiences and engagement. Hyro, a provider of conversational AI healthcare solutions, empowers enterprises to automate workflows and interactions across their most valuable platforms, services, and channels – including call centres, websites and mobile applications. Using a plug-and-play approach alongside conversational intelligence, Hyro delivers omnichannel analytics, including engagement metrics, trending topics and knowledge gaps that offer industry-leading control and optimisation. Its new GPTpowered assistant, Spot, is trusted by the world’s leading health systems. It generates customised responses to queries while navigating patients to relevant pages for deeper exploration. Once clinicians have engaged with patients, CommerceAI can help them to monitor their sentiment via generative AI. CommerceAI’s latest offering, auto-MATE, is a generative AI tool catering to diverse industries, including healthcare, by processing unstructured data such as contact centre calls, Teams meetings or telemedicine recordings to extract structured insights and subsequently automate workflows. This tool streamlines provider operations, elevates patient care and enhances overall healthcare outcomes. CommerceAI even tailors the auto-MATE tool specifically for pharmaceutical companies to support activities ranging from drug discovery to regulatory compliance. While the potential of generative AI in healthcare is immense, there is still much to learn and safeguards to be developed. The key principle in successfully adopting LLMs is to begin with low-risk, high-impact scenarios such as administrative tasks, with all findings undergoing clinician adjudication. As an industry, we should work together to solve the challenges of data privacy, algorithmic biases and interpretability to ensure the safe and responsible deployment of LLMs in patient care. Sally Ann Frank is worldwide lead for health and life sciences at Microsoft for Startups ScienceIO is developing language models to help clinicians decode healthcare data