P22-145 ITM Ph II Expansion (CCHAT Protocol)

Current artificial intelligence (AI) algorithms are built to align with ground truth or consensus of trusted human decision-makers. This is not effective for difficult decision domains, such as patient triage, where even experts disagree on the best course of action. A key reason human experts disagree is that their individual attributes influence their decision-making. To develop human-off-the-loop triage systems that are trusted by humans, understanding and alignment with these attributes is critical. This will enable effective communication of system decisions, based on these attributes may further support trust in the system.Ā During this effort, the team will use proven interview-based approaches and machine learning tools to identify attributes that predict communication and decision-making styles used during high-stakes situations such as patient triage, and will develop tools that give insights into alignment scores and communicate AI-based decisions to human users. This effort will produce five unique products: 1) KDMAs / KCAs grounded in data gathered from SME interviews and identified using machine learning models; 2) triage scenarios that elicit decisions and corresponding attributes; 3) a model for cue utilization that will inform scenarios; 4) visualizations to give insight into human-AI attribute alignment; 5) tools to help communicate AI-based decisions to humans.