Leveraging the potential of Human Centered Artificial Intelligence in healthcare and ensuring the rigor of data generated.

Large Language Models, ChatGPT, and AI and healthcare…we keep hearing these terms but how do they relate to how we use and apply Artificial Intelligence (AI) in cancer R&D? And how do we ensure rigor in the data used for medical decision making?

?“At the end of the day, everything we produce has to be explainable - to study statisticians, to healthcare providers and ultimately to patients.” -? Benjamin Tannenwald ,?Data Science Lead?

A critical component of the Digital Health Oncology R&D team at AstraZeneca is my Human-centered AI (HAI) team led by the amazing Emmette Hutchison that supports our efforts to accelerate data availability for decision making as well as reduce the burden on patients and HCPs. HAI aims to create applications of AI that augment human expertise, often focused on either upskilling those without subject matter expertise or enabling those with subject matter expertise to be far more efficient, typically by automating away the mundane tasks. The HAI team accomplishes this through holistic design of solutions that seek to?augment?rather than?replace and to provide?interpretable?solutions that mitigate concerns around?bias and liability.

?What we do is not only about delivering algorithms; it’s about designing and building technology-enabled capabilities and products that support our patients and partners in healthcare delivery.?This team is directly enabled through our extensive collaboration with scientific, operational, regulatory, clinical, IT and measurement experts throughout the business and industry.

?Acceleration through iterative modular development

Another fundamental principle of our approach is acceleration through iterative, modular development. Frequently, there is a lack of sufficient evidence when building a new solution or process. Our patients are waiting. We can’t afford the luxury of delaying until the perfect data set is available.

To meet the demands of our growing portfolio, the HAI team recently collaborated with scientific and measurement experts within Digital Health Oncology R&D to design a modular,?rules-based algorithm centred on clinical expertise and extensive RWE that included a rapid retrospective longitudinal cohort study. This was created to support the development of a digital therapeutic and the modular design of the algorithms for this new solution allows for the incorporation of new evidence or more complex algorithms as further scientific data becomes available. By taking this approach, the HAI team dramatically accelerated our ability to deliver to our patients.

Advancing with NLP & Machine Learning

The HAI team has also developed innovative new algorithms using advanced?Natural Language Processing and machine learning to extract Patient Reported Outcome (PRO) concepts from the patient’s own language. An explicit design choice was made to enable the patient as the algorithm never tells the patient what symptoms they experience but rather allows the patient to confirm the mapping. This capability can be paired with voice interfaces to both reduce patient burden and enable collection of clinically relevant acoustic data. This innovative combination of modern technology and traditional PRO approaches allows for the?low-burden collection of patient data on more frequent basis and a better understanding of the patient journey.

Overall, to support our patients it’s critical that we leverage as many multi-modal signals around the patient experience of disease and treatment as possible. In this way, we can improve the patient journey and reduce patient burden. Algorithms that rely upon multi-modal and longitudinal data can accommodate missing data associated with the real world setting as well as provide a greater level of certainty around a patient’s clinical status. The downstream implications of this include enhanced symptom management, more timely clinical intervention, accelerated recruitment of patients and augmented support from Health Care Providers. Natural Language Processing (NLP), Computer Vision (CV), Reinforcement Learning and Large Language Models (LLMs) are all technologies that we use to support these use-cases.

In a world of increasing uncertainty, we should not shy away from the exploration of novel technologies, including AI and machine learning.?We also can’t afford to employ black box technology, including generative models such as chat-GPT variants, in an unconstrained manner.?For example, constraints can be placed on such models to provide enhanced patient experience and HCP interface design, but effort must be made to ensure that information surfaced from such interfaces is correct. An imagined outcome or fact from a LLM can negatively impact patients if the implementation of the LLM is unconstrained or the hallucination is not detected. When implementing such complex models care should always be taken to assess risk of potential harm and models should be monitored for bias, drift and hallucinated content. Model interpretability or explainability, in essence the ability to step into a model to understand what features result in a particular outcome, is another tool that we utilize to address these issues. We cannot afford to deprive patients of the benefits of these approaches and the efficiencies they bring but we must also implement them in a holistic framework that can monitor and mitigate undesirable consequences.

?