The first time I heard of the ERAS tour, I texted our perioperative nurse manager to ask why the Enhanced Recovery After Surgery program had a billboard during a soccer match in Singapore (this is a true story). Suffice it to say that I’m not a Swiftie, though my daughter plays Taylor Swift often enough that I at least recognize her music now. However, I like the concept of eras in life and in technology, and when I saw the title of this recent paper by Google Health, I was excited. I wanted to know what they thought the future would hold, and how we would know we were transitioning from one epoch (or era) to another.
Unfortunately, the paper just takes the three commonly accepted eras of AI and talks a little about healthcare with each one. Healthcare AI is becoming such a huge field that it’s hard to say anything interesting while generalizing. So I decided to write this post with what I was hoping to see in that paper.
Let’s take three of the main innovations in healthcare AI right now: EHR integration including ambient scribes, clinical decision support, and patient-facing tools.
Clinician-facing EHR Integration
Phase 1 will be defined by AI tools that are available but not well-integrated into the EHR. Hallmarks include having to copy and paste the information (ambient scribe) or click on a link that brings the user outside the EHR (risk calculator). Patient data won’t be automatically pulled in, and there will still be duplicative EHR tasks like checkboxes for information that’s already written.
In Phase 2, AI assistants will populate fields in the background and customize to a physician’s workflow without explicit rules-based instruction. More complex tasks like inbox sorting solutions will become common at large medical centers. However, physicians will still be interacting regularly with the EHR and maintain frequent oversight. Monitoring and validation will begin but will be a manual process.
By Phase 3, seamlessly integrated AI will perform EHR activities proactively without needing physician input. Physician interaction with the EHR will be hands-free through conversational voice commands and require no typing. The AI assistant will be able to anticipate a specific physician’s needs. Behind the scenes, AI will constantly optimize system performance via validity checks, monitoring usage patterns, and processing user feedback.
Clinical Decision Support
In Phase 1, AI for clinical decision support will be limited to simple interventions like automating rules-based order sets. The actual suggestions and decision trees will be largely unchanged, and there will be few validation studies.
In Phase 2, decision support AI will gain adoption and become more sophisticated and responsive. AI will incorporate more nuanced characteristics from a population health standpoint and proactively respond to changes. Validation of the tools will happen prior to implementation but there won’t be ongoing monitoring yet.
Phase 3 will see clinical decision support AI become integral to routine practice. Drawing on individually tailored patient factors, comprehensive population data, and the latest evidence, AI can construct highly personalized care plans for clinician review. Validation and monitoring will be routine and vendors will compete to demonstrate their ability to continuously monitor the effectiveness of their tools.
Patient-Facing Tools
In Phase 1, patient-facing AI will see very limited adoption, with early experiments primarily focused on using chatbots to deliver basic medical advice or mental health counseling. AI tools will be rules-based, similar to current phone triage decision-trees. The level of helpfulness will be similar to that of an automated voice for a 1-800 number.
Phase 2 will make AI-powered patient apps more mainstream, with leading hospitals deploying smart symptom checkers and voice assistants to offer basic triaging services or pre-visit guidance. Yet even as adoption grows, the depth of AI will remain narrow—zeroing in on immediate concerns without managing underlying health trajectories or coordinating broader care. Gaps in validity and oversight will persist at this middle stage.
In Phase 3, patient-centered AI applications will aspire to deliver coordinated care that addresses both acute and chronic needs. A hospital level of care can be provided by the assistant via integration of wearable monitors and proactive responses to changes in patient conditions. Chronic conditions will be monitored without patient involvement and the information will be synthesized for the treating team to review.
Summary
We’re at the beginning of a long process of integrating AI into healthcare. While we don’t know exactly what the future will look like, these are my best guesses as to what the eras will look like for these three big buckets of healthcare AI adoption.
What did I get wrong? Are there other indicators for the different eras that I missed?
Join ML for MDs for a special virtual event on February 20th, 2024 at 5PT/8ET. Four of our talented physician healthcare AI experts will hold an Ask Me Anything style discussion about career transitions. Register and submit your questions today!