Large language models in healthcare: transforming the industry with AI-driven solutions

Despite millions of dollars of investment, healthcare’s digital transformation is still very much in its infancy.

“In 2008, fewer than one in ten American hospitals had an electronic health record,” said Dr. Robert Wachter, Professor and Chair of the Department of Medicine at UCSF. “By 2017, fewer than one in ten did not. So it was really over the past ten to fifteen years that medicine went from fundamentally an analog industry, to mostly a digital industry.”

While significant progress has been made, the fact remains that the electronic health record (EHR) was created to be a system of record, not a system of intelligence. 

“We have better access to information,” said Dr. Manish Naik, Chief Medical Officer and CMIO at Austin Regional Clinic (ARC). “But, my own assessment is that in the current state of things, in no way shape or form has the electronic health record made us more efficient. And that's the challenge that we deal with today.”

To deal with these challenges, leading health systems are focused on capitalizing on remarkable advancements in artificial intelligence (AI) and machine learning (ML). 

Decoding large language models

Previous AI tools were a step in the right direction from what existed before them, however, they were limited by the fact that they were largely unable to read and understand things like medical records, research papers, and even conversations between doctors and patients.

The AI community has been buzzing about large language models (LLMs) and their unique ability to understand and produce human-like text. And for good reason.

“There is no doubt,” said Pranay Kapadia, Notable co-founder and CEO. “Large language models will revolutionize the way we access and process data across virtually every industry.”

But what exactly are these models and how do they work? 

Large language models are essentially machine learning algorithms that are designed to comprehend, generate, and manipulate human language. The models are considered “large” due to their enormous number of parameters – ranging from five to tens of billions – which are refined throughout the training process. Think of parameters like neurons in a brain. The more neurons there are, the larger the brain that can represent and process more complex information.

The training process involves exposing these models to an extensive collection of text data, sourced from diverse materials such as websites, books, documents, articles, and more. Exposing the models to this data enables them to learn to recognize patterns in the data, including aspects of grammar, syntax, and context. The end result is that the models are able to produce coherent and contextually appropriate text. 

How do LLMs understand context? 

Look no further than the use of transformers, which are an architecture that facilitates parallel processing and allows models to attend to different parts of the input text simultaneously. While all deep learning models can be configured to have a lot of parameters, transformers have continued to perform better, learning more and more complex patterns as they scale up.

More specifically, transformers that have been scaled up to billions of parameters, such as OpenAI's GPT-3, have shown remarkable capabilities in a wide range of language tasks like reading comprehension and entity extraction.

Once trained, large language models can be fine-tuned for specific tasks or industries, such as healthcare. Fine-tuning happens by supplying the model with additional domain-specific training data, which allows it to generate more accurate and relevant insights tailored to that particular context. In healthcare, this can include clinical notes, medical journals, medical imaging records, and more.

In this article, we discuss the significance of large language models in healthcare, dive into how they’re being used today, and explore their potential to rapidly impact the industry moving forward.

Why do large language models make sense for healthcare?

The healthcare industry is inundated with various types of text-based documentation, such as patient medical records, claims documents, clinical notes, consent forms, and waivers just to name a few. The ability of LLMs to rapidly read, interpret, and act on massive amounts of text means they are instantly a competitive differentiator for any health system that puts them to work safely, securely, and in a unique manner.

“Healthcare is ideally suited to leverage large language models because of the sheer volume of latent data that exists across a given health system’s medical records,” said Kapadia. 

Despite the massive volume of data, capitalizing on it to personalize the patient experience and improve care has remained elusive for most health systems. The labor-intensive nature of mining, reading, and drawing conclusions from these datasets, combined with staffing shortages, has made it infeasible to personalize care for all patients.

Large language models: unique capabilities

In such a document-heavy industry, there is no shortage of applications for large language models in healthcare. The following are examples of large language model capabilities that, when applied in a healthcare setting, streamline processes and improve overall workflows.

• Enhanced clinical documentation: Healthcare professionals use LLMs to create accurate and comprehensive clinical notes, using an LLM to analyze patient data and generate relevant summaries. This helps reduce the burden on providers while ensuring that essential information is accurately documented.
• Improved information extraction: Identifying and extracting pertinent information from unstructured text, such as patient records or research articles, is another natural application of LLMs. This ability allows healthcare organizations to unlock valuable insights hidden in vast data repositories, ultimately supporting better decision-making.
• Advanced communication capabilities: Given that they excel at understanding natural language, LLMs are well-suited for tasks such as patient-provider communication. These models can facilitate more effective interactions by interpreting and responding to patient inquiries, concerns, or feedback in a contextually relevant manner.
• Document generation and management: LLMs can automate the generation of essential healthcare documents, such as consent forms, waivers, and discharge summaries. By extracting relevant information from patient records and pre-populating these documents, LLMs save time, minimize the risk of errors, and enable providers to make patients feel like they know them by not asking for repeat information. When incorporated into an intelligent intake and registration workflow, these tools are able to capture accurate patient information with ease to maximize clinical and financial outcomes.

Real-world use cases in healthcare

Moving beyond the LLM capabilities detailed above, let’s explore a few real-world use cases that illustrate how LLMs can help improve overall efficiency in healthcare. 

• Appointment scheduling and management: today, LLMs are being used to develop intelligent appointment scheduling systems that consider factors such as patient preferences, provider availability, and optimal appointment durations. These models automatically send appointment reminders, reschedule appointments in case of cancellations, and manage waitlists to optimize clinic workflows.
• Billing and revenue cycle management: LLMs can be employed to optimize billing and revenue cycle management processes by identifying discrepancies in billing codes, flagging potential issues, and suggesting corrective actions. This can help healthcare organizations ensure accurate billing and improve revenue collection.
• Insurance claims processing: LLMs can analyze insurance claim documents, identify relevant information, and automatically populate the required fields in claims processing systems. This automation can speed up the claims process, minimize manual data entry errors, and reduce the administrative burden on healthcare staff.
• Patient education and health literacy: LLMs can generate easy-to-understand explanations of complex medical concepts, helping patients better comprehend their health conditions and treatment options. By promoting health literacy, these models can empower patients to make informed decisions about their care and take a more active role in managing their health.
• Personalized care recommendations at scale: LLMs are pivotal in delivering personalized care recommendations at scale, which they are able to do by analyzing individual patient data and generating contextually relevant recommendations. Notable’s Patient AI is one such example and is already being deployed by health systems across the country. The process of customizing care using LLMs involves several steps:

• Data analysis: Once the relevant data is collected, LLMs can analyze it to identify patterns, trends, and correlations. This analysis helps the model recognize potential health risks, treatment response patterns, and patient preferences.
• Care recommendations: Based on the analysis, LLMs can generate personalized care plans, interventions, and recommendations. These may include tailored medication regimens, clinical trial matching, medical programs offered by the health system, or follow-up appointments. By considering factors such as patient demographics, medical history, and individual preferences, LLMs can power care recommendations that are uniquely suited to each patient, without staff intervention.
• Continuous monitoring and adaptation: LLMs continuously monitor patient data and adjust care plans as needed. By integrating real-time data, such as changes in vitals or new test results, LLMs are able to dynamically adapt recommendations to ensure that patients receive the most appropriate care at all times.

There is little doubt, the integration of LLMs into healthcare workflows holds immense potential. As they evaluate new technology solutions, health systems must ensure that potential partners not only deeply understand LLMs but have integrated these advanced AI technologies into the very fabric of their offerings.

Additional benefits of large language models

As we have seen, there are myriad applications for large language models throughout the healthcare ecosystem. Health systems of all shapes and sizes – urban or rural, actuate or ambulatory – stand to benefit from the continued innovation around these advanced AI technologies.  

Here are examples of key areas where LLMs can deliver value to all health systems.

Advanced analytics and reporting

LLMs process and analyze large volumes of healthcare data, generating valuable insights and reports that support decision-making at various levels within a health system. By providing actionable intelligence, LLMs help health systems identify areas for improvement, optimize resource allocation, and monitor the effectiveness of interventions. 

What’s more, rather than providing static data visualizations, ChatGPT-type interfaces allow leaders to query the data by asking questions, drill into specific areas of interest, and receive real-time, data-driven insights into things like patient feedback.

Continuous learning and adaptation

As discussed, large language models learn and adapt over time, allowing them to stay current with the latest knowledge and best practices. This continuous learning capability ensures that health systems are able to leverage the most up-to-date insights and recommendations, leading to better patient care and outcomes.

Scalability and flexibility

LLMs can easily be integrated into existing healthcare workflows and can be scaled up or down to meet the specific needs of a health system. This allows health systems of all sizes to benefit from the advanced capabilities of LLMs without significant disruption or investment in new infrastructure.

Patient safety and privacy considerations

As we move forward, ensuring responsible, safe integration of large language models in healthcare must remain a top priority. As we’ve established, these models require vast amounts of data in order to generate accurate and relevant insights. In healthcare, this data often includes sensitive patient information. 

Anyone working to incorporate large language models in a healthcare context must ensure that patient safety, security, and privacy remain paramount. This can be accomplished through the implementation of robust security measures, and data anonymization techniques to protect patient information from unauthorized access and potential misuse. 

What’s next for AI and large language models?

“I don’t think AI has hit its tipping point in healthcare just yet,” said Dr. Wachter. “There’s certainly applications of it that are impressive, and the potential, obviously, is enormous. I think that has to happen in the next 10 years.”

By embracing advanced AI technologies, health systems can not only optimize their current operations but also pave the way for future advancements that promise to further elevate patient care. The potential of large language models in healthcare is immense, and the applications are only just beginning to be realized.