Healthcare AI Treatment Decision Support: Design Principles to Enhance Clinician Adoption and Trust

Title of the Paper

Healthcare AI Treatment Decision Support: Design Principles to Enhance Clinician Adoption and Trust

Paper Information

• Research Area: Application of health informatics and artificial intelligence in medical support
• Keywords: Artificial intelligence, machine learning, type 2 diabetes, clinical decision support, design principles, medical workflow, drug prescription, trust, user research

Research Background and Issues

• Issues and Challenges: Current AI-driven clinical decision support (CDS) technologies can analyze large volumes of patient data, but their practical application in healthcare still faces challenges. Specific issues include whether the insights generated are useful to clinicians, integration with existing workflows, and clinicians' low trust in these systems.
• Significance: Generating effective AI-supported insights is crucial for improving treatment outcomes, reducing trial-and-error time, and enhancing patient health. Type 2 diabetes, due to its high prevalence and associated healthcare costs, has become a key focus for optimizing treatment decisions.
• Research Motivation and Related Work: Although extensive technological development has been undertaken for AI and machine learning-driven medical support, the focus has primarily been on the technical feasibility and accuracy of algorithms rather than on clinicians' acceptance or usage patterns. This study aims to improve the design of AI treatment decision support systems through prototype research and user feedback.

Solution

• Method/Approach: Develop a prototype of an AI-based treatment decision support system focused on optimizing reductions in glycated hemoglobin (A1c) levels, providing clinicians with insights into drug selection. Feedback was collected through interviews with 41 U.S. clinicians, and the design was iteratively refined.
• Innovations:
  • Proposed six design principles to enhance user trust in medical AI support systems.
  • Highlighted the need for balancing population-level data and personalized insights in clinical decision-making.
  • Validated the effectiveness of drug combinations using real-world case data.
• Key Technologies and Implementation Steps:
  • The prototype was constructed using health insurance claims data from over 140,000 patients, combined with biomarkers and drug prescription records.
  • The model ranked different treatment options based on A1c changes and simulated information such as drug costs and side effects.
  • Three versions of the prototype were designed to continuously optimize data presentation and clinician user experience.

Research Outcomes

• Specific Results:
  • Compared user feedback across different prototypes, summarizing core needs such as clinicians' trust in "existing knowledge forms (e.g., randomized clinical trials)" and the demand for transparency in AI-supported insights.
  • Proposed six design principles (e.g., support clinicians in weighing individual patient characteristics; introducing AI tools as a key opportunity to build trust; avoid adding to clinicians' "research task" burden).
• Comparative Advantages Over Existing Solutions:
  • AI models based on real-world data not only provide recommendations for individual patients but also offer predictive insights into the effectiveness of drug combinations.
  • Integrated the shared decision-making framework between clinicians and patients into tool design.
  • Conducted multiple rounds of optimization for clinician workflows, making the tool more suitable for diagnostic and treatment contexts.
• Experimental and Evaluation Results: The applicability of the prototype tool varied across different medical roles; for example, nurse practitioners were more receptive to this technology, while specialists felt they already possessed relevant knowledge and did not require tool assistance.
• Limitations and Future Directions:
  • Limitations include: research data and participants were confined to the U.S. healthcare system; the prototype study did not address patients with multiple conditions (e.g., diabetes and chronic kidney disease).
  • Future directions:
    • Explore tool designs centered on patient-clinician collaboration.
    • Investigate mechanisms for clinician trust and adaptation in environments with multiple models/multiple AI tools.
    • Conduct in-depth analysis of legal and ethical implications (e.g., accountability for incorrect AI recommendations).

Output Format

• The overall structure of the paper is highly comprehensive. This study provides multidisciplinary insights and design approaches, and the principles and findings can be extended to other medical contexts, paving the way for increasingly complex AI-driven healthcare practices.