The promise and reality of ambient AI in health care settings

What ambient AI in health care is capable of doing: ©Catalin - stock.adobe.com

As health care organizations seek solutions to address physician burnout and documentation burden, ambient artificial intelligence is emerging as a promising technology. By automatically capturing and processing clinical conversations, ambient AI aims to free providers from manual documentation tasks. However, the technology's implementation comes with both opportunities and challenges that health care organizations must carefully navigate.

How ambient AI works in clinical settings

At its core, ambient AI deployment relies on listening devices –– often tablets or phones –– to capture clinical conversations during patient encounters. While earlier iterations of this technology relied heavily on human transcriptionists overseas to convert recordings into structured notes, recent advances in speech-to-text capabilities and AI have transformed the process.

Modern ambient AI systems can now convert spoken conversations directly into text and use artificial intelligence to structure that information into appropriate clinical documentation sections, such as physical exams, review of systems, and history of present illness. Some systems allow for customization of how this output is structured though the level of structure can impact accuracy.

Current implementation and results

The technology has shown the most promise in outpatient settings, particularly in primary care. These environments are well-suited for ambient AI because they typically involve focused, 15-minute encounters where all relevant clinical information is discussed in a single conversation. In contrast, inpatient and surgical settings have proved more challenging due to the continuous nature of care and the presence of multiple care providers.

Most current implementations are small-scale pilots, with individual providers or small groups testing the technology. Organizations are measuring success through various metrics, with a particular focus on indicators that reflect reduced administrative burden. One creative approach tracks the time lag between patient encounters and note closure – a proxy for "pajama time" spent on documentation after hours. When providers can complete their documentation within scheduled hours, it suggests the technology is successfully reducing their administrative burden.

Technical challenges and limitations

The relationship between specificity and accuracy presents a significant technical challenge. As systems attempt to capture more specific medical terminology or structural elements, the risk of errors increases. This creates a tension between generating highly structured data useful for billing and quality reporting and ensuring accuracy in clinical documentation.

Current solutions show promise in distinguishing between relevant clinical conversation and background noise or casual banter. However, challenges remain with accurately capturing specific medical terminology, especially when dealing with different accents or dialects. The more specific the required output, the higher the risk of misinterpretation or error.

Integration and workflow considerations

Integration with existing EHR systems varies significantly across different ambient AI solutions. Many current implementations rely on basic copy-and-paste functionality, with providers manually transferring the AI-generated notes into their EHR systems. While some vendors have established partnerships that enable more seamless integration, the industry still largely struggles with translating semi-structured ambient AI outputs into the fully structured data needed for billing and quality reporting.

This integration challenge can create downstream impacts on workflow, potentially shifting burden from documentation to review and validation. Health care organizations implementing ambient AI must carefully consider how semi-structured data affects various stakeholders, from billing staff to quality reporting teams.

Privacy and security safeguards

Privacy and security considerations are paramount in ambient AI implementation. Vendors typically provide assurances that information passed through their large language models won't be used for model training. Systems must maintain robust security certifications (such as SOC 2 or HITRUST) and operate in hardened environments due to the sensitive nature of the information being processed.

Future outlook

The future development of ambient AI in healthcare settings depends on several factors. The emergence of medical-specialized large language models, such as Google's MedLM, shows promise but remains experimental. EHR integration remains a significant hurdle that requires continued investment and development.

As health care organizations begin to see concrete returns on investment through improved quality metrics and satisfaction scores, investment in the space is bound to accelerate rapidly. The technology's evolution will likely be gradual, driven by increasing provider comfort with the systems and continued improvements in EHR integration and usability.

Success will ultimately hinge on striking the optimal balance between automation and human oversight, ensuring that the technology enhances, rather than compromises, the quality of clinical documentation while meaningfully reducing provider burden.

Ben Scharfe is executive vice president of AI Initiatives for Altera Digital Health.