Artificial intelligence in healthcare is no longer a theoretical discussion. Major hospital systems across the US are using AI to improve the way they interact with patients and their families.

Solutions are being implemented not just to improve clinical decision making, but to impact nearly every touch point a patent has with healthcare organizations. Scheduling, communication, follow-up are all being impacted. And these solutions are faster, smoother, and perhaps paradoxically more human.

In this article we highlight some examples of how leading healthcare orgs are taking advantage of this new technology, and abstract some common patterns that you can leverage for your own implementations.

How Systems are Leveraging AI

Lifting the weight of documentation

At UCSF Health and Cleveland Clinic, AI scribes now record and summarize doctor-patient conversations. Historically clinicians have had to spend the end of hteir day (and often into the evening) typing their notes from the day into their organization’s EHRs or other software. Now, clinicians are able to review the notes before they are entered into the record, cutting the time spent with data entry.

Both hospital systems are reporting higher completion rates for same-day documentation. They also report stronger patient feedback, as physicians are able to maintain eye contact and demonstrate improved bedside manner during visits.

Making access simpler

Mass General Brigham’s Care Connect platform is using conversational AI to triage patients who lack a primary-care physician. The initiative was designed to ease the shortage of primary care physicians within their system. They report patients are now able to schedule a telehealth visit within 30 minutes.

Similarly, Kaiser Permanente’s “Intelligent Navigator” has rolled out in Southern California, and allows patients to book appointments and request care by describing their needs in plain language vs. navigating complicated menus or phone trees. A study published by Nature Magazine found that the system was able to suggest appropriate care pathways with 89% accuracy, and reduced abandonment rate before completing an appointment to just 3%.

Reaching patients before problems escalate

UC Davis Health created a predictive model to identify people at high risk of hospitalization. The goal was to proactively identify patients that might benefit from care management services, before their health problems led to emergency department visits or hospitalization.

During the process, they learned how important it is to train AI tools on your own data. When they first rolled the platform out, they found that it under-predicted the potential for these kinds of events when evaluating African American and Hispanic populations. They found that if you aren’t tuning your model based on your specific patient population, you risk inaccurate recommendations.

Improving communication at scale

At NYU Langone Health and Penn Medicine, clinicians are testing LLMs to draft replies to patient messages. NYU found the number of messages physicians are receiving daily has increased by nearly 30% per year, with physicians reporting regularly seeing more than 150 messages per day.

These tools seem to be helping alleviate this burden - in the NYU study, the AI-written responses were just as accurate as those written by physicians. And surprisingly, they were rated as more empathetic as well. While the physicians still review the messages before sending to ensure there are no issues, the time savings were still considerable.

Guiding treatment decisions

Cedars-Sinai deployed a chatbot that collects symptom data, compares it with records from similar patients, and produces a treatment plan. The physicians review the treatment plan, can ask clarifying questions if necessary, and are free to disagree with the AI’s recommendations.

A study on the new system found the AI’s recommendations were judged “optimal” 77% of the time, versus 67% for the physicians themselves. And while it focused on a relatively narrow set of medical conditions, they believe these systems can help standardize evidence-based care in the future.

What These Examples Demonstrate

Across institutions, some common patterns are evident:

• Emphasis on reducing friction. Because AI tools are largely based on natural language, patients and clinicians are able to engage with these tools more efficiently than previous solutions. This lets them save time and increases the likelihood of task completion.
• Reducing workload for clinicians. This reinforces our belief that the biggest opportunities for AI right now are in streamlining cumbersome or manual workflows. But this requires an understanding of the current state at the ground level. It’s imperative to pilot these tools with a consistent feedback loop from the clinicians to maximize adoption.
• The data matters. UC Davis’s experience with model bias demonstrates that predictive models are only as useful as the data behind them. It’s important to plan for a process that includes regular auditing, retraining, and a fair amount of human oversight to make sure you are getting accurate results and not introducing unintentional biases. And in the case of healthcare, that probably includes community health experts and data ethicists, not just data scientists.
• Integration before scale. The most effective deployments started with limited use cases (albeit a high volume ones). They integrated it tightly with existing EHR or portal workflows. And only after they knew it was working did they worry about expanding.
• Data quality as the foundation. Each example underscores the importance of clean, connected data. Hospital systems with fragmented patient records or inconsistent coding would likely find it difficult to embrace any of these solutions. It’s imperative to get your data foundation in order first.

How Your Organization Can Follow Their Lead

We’re going to see more of these types of tools in the coming years. And the organizations that will benefit the most will be the ones who do the hard work now of getting their data house in order. Who embrace a posture of rapid experimentation, working collaboratively with the clinicians and team members who interact with patients on a daily basis. Who believe AI-powered patient engagement needs to become part of their operating fabric to deliver better patient outcomes and reduce the burden on their teams.

If that sounds like you, we’d love to talk.

‍