Opportunities and Risks of AI in Patient Portal Messaging in Primary Care

Introduction

• Increased adoption of patient portals and telemedicine post-pandemic has resulted in a surge in patient portal messaging.
• Primary care physicians (PCPs) face increased workloads due to electronic health record (EHR) tasks, leading to burnout.
• Strategies to manage inbox overload have had limited success, including message length limits, billing for messages, and limiting message exchanges.
• Generative AI, using large language models (LLMs), is being explored to draft automatic message responses to reduce workload.
• Early results indicate a reduction in cognitive burden but no significant time reductions.
• Risks include LLMs introducing inaccurate or inappropriate information, possibly affecting patient safety.

Study Objectives

• Primary Objective: Determine how PCPs address errors in AI-generated draft responses.
• Secondary Objective: Assess PCP perspectives on AI-generated responses.

Results

• Study involved 20 participants, mostly female attending physicians with an average of 14.75 years post-medical school.
• Participants missed correcting errors in AI drafts frequently, with a high likelihood of errors reaching patients.

Survey Responses

• Participants held favorable views of AI drafts, reporting reduced cognitive workload and trust in AI tools.

Discussion

• AI-generated drafts can pose safety risks due to hallucinations and outdated information.
• Current AI implementations in healthcare outpace understanding of their safety.
• Low rates of error detection by PCPs in AI drafts raise safety concerns.
• Possible cognitive biases affecting error detection include functional fixedness, confirmation bias, automation complacency, and automation bias.
• Recommendations include design improvements, technological advancements, and human review guidelines.

Limitations

• Lack of full patient records for participants.
• Remote study setting may have introduced distractions.
• No control group for error rate comparison without AI assistance.
• Limited diversity of message content covered.
• Use of non-validated survey instruments.

Methods

• Study approved by MedStar Health IRB; informed consent obtained.
• Participants responded to 18 PPMs with AI-generated drafts in a simulated environment.
• AI drafts reviewed by experienced PCPs to identify errors.

Data and Code Availability

• Datasets and analysis code available upon request from the corresponding author.

Conclusions

• AI drafts offer benefits in reducing cognitive load but present significant safety concerns.
• A need for further research into AI error types, cognitive biases, and effective integration into clinical practice.