Human-AI Collaboration in Healthcare: AI-Enabled Adaptive Learning Systems

Human-AI collaboration (HAIC) describes systems whereby humans and artificial intelligence (AI) systems work in tandem to produce outcomes superior to independent solutions. Today, AI systems are beginning to be deployed in healthcare systems, for workflow optimisation, and potential economical and productivity benefits. They have been shown to encroach or even outperform the performance of trained experts in some clinical settings ([1]).

However, these benefits do not come without drawbacks. Specific to the healthcare domain, AI such as deep learning models are susceptible to bias, can be prone to poor generalisation, and can produce uninterpretable predictions ([2]). Humans are, of course, also not without weaknesses.

A study approximated that medical errors from radiologists rank as the third most significant cause of death, with an annual occurrence rate of up to 9.5% ([3]). The error rate can be attributed to several factors, such as high concentration, large workload and quick turnover, which contributes to fatigue of the radiologists ([4]). HAIC seeks to mitigate the individual weaknesses of humans and AI while leveraging their respective strengths, ultimately developing an enhanced system.

HAIC encompasses a wide range of topics, including out-of-distribution generalisation, deferral-based systems, explainable AI, audio-visual computer vision, and multimodal AI models. In this DPhil project, our initial goal is to develop large multimodal language models for creating AI-enabled adaptive learning systems. For instance, sonographers face the demanding profession of maintaining high diagnostic precision in stressful situations with time constraints, which requires a high level of skill. Transferring expert knowledge and expertise to new trainees presents a significant challenge.

Streamlining this time- and cost-intensive process can be achieved through HAIC by developing AI systems that convey task-specific expert knowledge to novice trainees while adapting to their evolving expertise levels over time. This project falls under the EPSRC's research area of human-computer interaction and will be carried out in collaboration with the OxSTaR (Oxford Simulation, Teaching, and Research) team. HAIC is a relatively new field in healthcare; therefore, there are many novel problems to address.

Although there have been some preliminary developments in various types of AI-enabled adaptive learning systems, these have primarily been implemented in the education domain. There are no publications of similar work in the healthcare domain. References [1] Babak Ehteshami Bejnordi, Mitko Veta, Paul Johannes Van Diest, Bram Van Ginneken, Nico

Karssemeijer, Geert Litjens, Jeroen AWM Van Der Laak, Meyke Hermsen, Quirine F Manson, Maschenka Balkenhol, et al. Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer. Jama, 318(22):2199-2210, 2017. [2] Milena A. Gianfrancesco, Suzanne Tamang, Jinoos Yazdany, and Gabriela Schmajuk. Potential

Biases in Machine Learning Algorithms Using Electronic Health Record Data. JAMA Internal Medicine, 178(11):1544-1547, 11 2018. [3] Martin A Makary and Michael Daniel. Medical error-the third leading cause of death in the us. Bmj, 353, 2016.

[4] Stephen Waite, Srinivas Kolla, Jean Jeudy, Alan Legasto, Stephen L Macknik, Susana Martinez-Conde, Elizabeth A Krupinski, and Deborah L Reede. Tired in the reading room: the influence of fatigue in radiology. Journal of the American College of Radiology, 14(2):191-197, 2017.