Intelligent Reasoning Cues: A Framework and Case Study of the Roles of AI Information in Complex Decisions

Paper Title

Intelligent Reasoning Cues: A Framework and Case Study of the Roles of AI Information in Complex Decisions

Publication Info

• Topic area: AI-based decision support systems in high-stakes expert domains
• Keywords: AI decision support, intelligent reasoning cues, clinical decision-making, sepsis treatment, human-AI interaction, reasoning patterns, contextual inquiry, think-aloud study, ICU, medical AI

Background and Problem

• Problem / challenge: Existing AI decision support systems often fail to effectively support expert users, particularly in complex, high-stakes decisions. Conventional models focus on calibrating reliance on AI advice but do not address how AI information influences reasoning processes.
• Significance: Enhancing AI decision support systems could improve decision quality and outcomes in critical domains like healthcare, where errors or inefficiencies can have life-threatening consequences.
• Motivation and related work: Prior research has explored AI explanations, uncertainty estimates, and user control to promote reliance on AI. However, these approaches yield mixed results, suggesting a need for a broader framework to understand how AI information interacts with human reasoning. This paper introduces a new framework, "intelligent reasoning cues," to address this gap.

Solution

• Proposed approach: The "intelligent reasoning cues" framework conceptualizes AI interfaces as collections of discrete pieces of AI-derived information (reasoning cues) that influence decision-making processes.
• Novelty:
  1. Introduces the concept of reasoning cues to expand beyond the conventional "second-opinion" AI model.
  2. Identifies specific patterns of influence that reasoning cues exert on human reasoning.
  3. Develops and evaluates reasoning cues in the context of sepsis treatment in intensive care units (ICUs).
  4. Provides design principles for creating adaptive, context-sensitive AI decision support systems.
• Procedure and key techniques:
  1. Developed eight reasoning cues (e.g., unusual features, risk predictions, peer actions) and embedded them in AI interfaces.
  2. Conducted two studies: contextual inquiries with six ICU teams and a think-aloud study with 25 physicians using AI interfaces.
  3. Analyzed reasoning processes, patterns of influence, and perceptions of reasoning cues through qualitative coding and thematic analysis.

Results

• Concrete findings:
  • Identified 11 distinct patterns of influence, including "Considering Alternatives," "Resolving Contradictions," "Second-Guessing," and "Plan Preference."
  • Reasoning cues like unusual features (R2) and plan-dependent risk differences (R4) were particularly effective in supporting complex decisions.
  • Clinicians valued reasoning cues that provided "true data" and complemented existing guidelines.
• Advantage over baselines:
  • Demonstrated that reasoning cues can influence reasoning beyond simple reliance on AI recommendations.
  • Showed that clinicians are more likely to engage with AI when it supports discretionary, complex decisions rather than protocolized tasks.
• Experiments / evaluation:
  • Study 1: Contextual inquiries with six ICU teams to understand real-world reasoning processes.
  • Study 2: Think-aloud study with 25 ICU fellows using AI interfaces to make decisions on four complex sepsis cases.
  • Analysis included 544 coded reasoning processes and thematic mapping of reasoning patterns and AI influence.
• Limitations and future work:
  • Focused on sepsis treatment; findings may not generalize to other domains without further study.
  • Participants were ICU fellows, whose reasoning patterns may differ from more experienced clinicians.
  • Future work should explore reasoning cues in other decision contexts and test quantitative impacts on decision quality.

Summary

This paper introduces the "intelligent reasoning cues" framework to better understand how AI-derived information influences expert decision-making. Through a case study on sepsis treatment in ICUs, the authors identify specific reasoning patterns influenced by AI cues and propose design principles for adaptive decision support systems. The findings suggest that reasoning cues are most effective when they support discretionary decisions, align with clinicians' goals, and complement existing knowledge. Future research should explore the generalizability of this framework and evaluate its quantitative impact on decision outcomes.