TL;DR
This study explores critical care pharmacists' perceptions of an AI-powered clinical decision support system (CDSS) for vancomycin dosing. While pharmacists see the potential benefits of AI in improving drug dosing, they exhibit a significant reluctance to fully trust and adopt the AI's recommendations, highlighting a crucial gap between perceived utility and actual clinical integration.
What Problem Does This Paper Address?
Vancomycin, a critical antibiotic for serious Gram-positive infections, has a narrow therapeutic window, making precise dosing essential to avoid toxicity (like acute kidney injury) and ensure efficacy. Critically ill patients, in particular, exhibit high variability in how their bodies process vancomycin, making individualized dosing a complex and time-consuming task for pharmacists. Existing dosing methods, while guided by established formulas, often rely on pharmacist judgment and can lead to inconsistent outcomes. Artificial intelligence (AI) offers a promising avenue to enhance precision and timeliness in vancomycin dosing through clinical decision support systems (CDSS). However, the effectiveness of any AI tool hinges on its acceptance and integration into clinical workflows. This paper addresses the critical question of whether pharmacists, the end-users of such a system, are willing to trust and adopt AI-driven recommendations for vancomycin dosing, despite its potential advantages.
Methods
The researchers conducted a qualitative study involving semi-structured interviews with thirteen critical care pharmacists at Mayo Clinic. The interviews were designed to assess their demographics, current vancomycin dosing practices, and attitudes towards implementing an AI-based CDSS. A key part of the methodology involved presenting pharmacists with eight clinical cases and asking them to recommend a vancomycin dosing regimen. Subsequently, they were shown the AI CDSS's recommendations for the same cases. The study then explored their reactions, willingness to accept the AI's suggestions, and the reasons behind their decisions, particularly when their recommendations differed from the AI's. The interviews were audio-recorded, transcribed, and analyzed using thematic analysis to identify recurring patterns and themes.
Datasets
Not discussed in the paper. The study used hypothetical clinical cases inspired by real-world patient experiences for evaluation, rather than pre-existing datasets for training or testing the AI model itself.
Key Insights and Ideas
- AI as a "Second Opinion" vs. "Decision Maker": Pharmacists generally view AI as a valuable tool for prompting double-checks and providing an additional perspective, rather than a definitive authority to blindly follow.
- The "Black Box" Problem: A significant barrier to acceptance is the lack of transparency in how the AI arrives at its recommendations. Pharmacists expressed a need to understand the "why" behind the AI's suggestions, particularly when they conflicted with their own judgment.
- Distrust and Lack of Dynamic Evaluation: General distrust in AI, coupled with concerns about its ability to perform dynamic, in-depth analysis and integrate all relevant clinical data, contributes to reluctance. The absence of a clear "risk index" associated with recommendations also plays a role.
- Time and Effort Investment: Pharmacists already invest considerable time in individualizing vancomycin therapy, and the introduction of a new system needs to demonstrably improve efficiency or outcomes without adding undue burden.
- Importance of Workflow Integration: For AI to be adopted, it needs to be seamlessly integrated into existing electronic health records (EHRs) and workflows, with a user-friendly interface.
Key Results
- Positive Perceptions: A majority of pharmacists agreed or strongly agreed that an AI-based CDSS could positively affect or enhance their vancomycin dosing decisions.
- Significant Discrepancy: In 85% of the 104 discussed cases, pharmacists' initial dosing recommendations differed from the AI CDSS's suggestions.
- Low Acceptance of AI Recommendations: Even when presented with the AI's recommendations, 78% of pharmacists did not adjust their initial dosing decisions in discrepant cases.
- Reasons for Non-Acceptance: The primary reasons cited were general distrust of CDSS, lack of dynamic evaluation and in-depth analysis, inability to integrate all clinical data, and the absence of a risk index.
- Willingness to Change: In the 22% of discrepant cases where pharmacists were willing to change their doses, it was primarily when their initial dose was very close to the AI's recommendation.
Limitations and Future Work
The authors acknowledge several limitations:
- Small Sample Size and Single Center: The findings may not be generalizable due to the small number of participants from a single academic medical center.
- Potential for Bias: Some interview questions might have influenced clinician responses.
- Limited Model Information: Pharmacists were informed of the AI's good overall performance but not detailed model features, which could have impacted their trust, especially when disagreeing with recommendations.
Future work should focus on:
- Enhancing Transparency: Developing AI systems with clearer rationales and risk indices to build trust.
- Seamless EHR Integration: Integrating CDSS into electronic health records and workflows to improve usability and compliance.
- Change Management: Implementing strategies to transform clinical culture and foster acceptance of digital health solutions.
- Extrapolation to Other Drugs: Investigating the applicability of these findings to other renally eliminated and nephrotoxic drugs.
Summary updated