Last year’s release of Large Language Model (LLM) applications, most prominently the ChatGPT, has captured public imagination and adoption like no other technology in recent times. The ability to play around with the latest shiny toy in an easy-to-use interface, with no or low cost of usage has led to a predictable hype cycle.
My first interaction with ChatGPT was also a fun exercise - to convert the prior authorization requirements for another “new” flavor of the year – GLP-1 receptor agonist drugs for weight loss, to iambic pentameter. The results were impressive.
“The sought drug gains coverage by request,
With prior nod, specific terms invest.
Three months in, therapy's grasp secure,
A stable dose, foundation's strength endure.
When body weight retreats, a 5 percent loss,
Or if the initial loss, maintenance gloss,
Prove this with documents, a requisite plea,
For approval sought, a documented decree.
Alternatively, in chronic weight's demand,
The drug finds purpose, part of a grand plan.
A diet's cut, and physical ascent,
In an adult's realm, chronic weight's intent.
A program comprehensive, weight in its fold,
Behavior altered, calories controlled.
For half a year or more, this path endured,
Follow-up continued, commitment assured.
The patient's BMI, a threshold laid,
Thirty or more, in this realm displayed.
Or BMI of twenty-seven, hold it true,
With comorbids in tow, approval's cue.
Hypertension's grip, or diabetes near,
Dyslipidemia's shadow drawing near.
With one in tow, the BMI must stand,
Approval granted, as per the plan.”
- Courtesy ChatGPT
And then there was the hype – Is it conscious? Whose job will it take first? What will humans do when AI does everything? While answers to these questions will be known in due course, Amara’s law should be our guide today –
“We tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run.”
In this article, we will focus on how:
1) Physician practices can benefit from Generative AI,
2) Guardrails that would be needed, and
3) To get started.
(In the next article, we will delve into how Generative AI can significantly benefit physicians in their clinical practice)
Physician Practices Automation
Artificial Intelligence, including ML predictive models and Gen AI-mediated content generation, is set to modernize physician practices. For clinical usage clinical validity and reproducibility of outputs will need to be ironed out before LLMs can be safely utilized to aid clinical decision making. However, the current state of Generative AI is mature enough to automate several administrative processes such as automated charge capture and revenue cycle management to give back to physicians several hours of their day.
Charge capture and coding:
Current charge capture software is dependent on manual review of medical records for capturing billable services. LLMs can “read” the medical records and do entity recognition for diagnoses, medications, and procedures. These entities then can be matched to appropriate billing codes. Generative AI can be used to extract structured information from the medical records and provide accurate coded information for billing purposes.
Payer contract management:
LLMs can compare and contrast payer contracts using semantic similarities of line items and identify areas of improvement in negotiations. Accurately extracting payment information about a service when combined with a patient’s insurance plan can help provide accurate payment transparency to the patients as well.
Clinical policies review:
“Chat” with clinical policies published by various insurance payers to identify expectations on clinical documentation, coding, and billing. Understanding the clinical policies is important to know payer specific coverage rules for example for prescription drug step therapy, or quantity limits. Knowing this beforehand can prevent claim rejections at the pharmacy and help patients get their prescription medications quickly.
Prior authorization:
Let the machine identify payer specific requirements for prior authorization of a procedure or a prescription, extract relevant data from the medical records, and automatically generate the prior authorization request.
Quality scores:
Keep a track of quality score numerators and denominators required for end of the year submission to get credit for the quality of services provided. LLMs can fill in the gaps by extracting data from unstructured records without a need for manual effort.
Value-based care:
Improve clinical documentation and processes to establish the value of services provided for value-based contracts. Contract specific quality metrics and cost performance can be deciphered directly from medical records, without needing another system to keep a track of these.
Follow-ups:
Automatically identify visit, diagnostic, or procedure follow-ups in text reports from medical records, lab/pathology reports or radiology reports. Payer-specific covered well-care visits can be derived and scheduled accurately from contract documents.
Guardrails for using Generative AI in your Practice
Generative AI is another tool that has the potential to increase your productivity by taking on mundane tasks so that you can focus on patient care without interruptions. As like any other tool, Gen AI also needs to be adapted to your systems to effectively do the tasks you assign to it.
Fine-tuning:
Fine-tuning the foundation models on your own data is an optional task that can make the models work more accurately by learning your practice patterns, record keeping styles, patient types, etc. This is the most effective method to reduce confabulations and fabrications that the models produce, if left alone.
Retrieval augmented generation:
RAG is a much faster and cost-effective technique to improve outcome accuracy. Using vector embeddings and semantic search to provide an accurate context to the LLMs increases the accuracy of information extraction from organization’s unstructured data.
Human in the loop:
While integrating Generative AI in the practice workflow for the first time it is important to include humans in the loop to assess quality of generated responses. Domain experts working together with technologists can fine-tune the prompts that generate accurate responses and potentially identify branching prompts for more sophisticated use cases. Iteration and experimentation will be the quickest methodology to achieve results.
Bias reduction:
As foundation models are trained on data that comes from across the internet, it is more likely that the existing biases will be preserved and potentially accentuated by Gen AI even in production. Having compliance and governance oversight in place will be essential where the outputs are likely to include age, race, gender, or economic biases. This is especially true when utilizing Generative AI to aid in clinical use cases.
A Winning Generative AI Strategy
Leaders and organizations across industries are trying to understand the phenomenon of Generative AI and how to harness this technology’s “weird” potential to improve their productivity and profitability. While healthcare has been behind other industries in adopting new technologies, with Gen AI it might just be different. With a heavy dependence on accurate record keeping – both for clinical purposes as well as reimbursement – regulatory and compliance needs, and a unique consumer, provider and payer relationship, Generative AI is more suited for the healthcare industry than any other technology before.
Harvest the Low-hanging fruit:
Start with the low-risk, high reward areas of your practice where the output of Generative AIcan reduce manual effort of reading and summarizing information for other downstream processes. Generative AI can accurately read medical records, clinical guidelines and policies, payer contracts, prior authorization requirements, etc and extract structured information for the administrative processes.
Record once and retrieve:
Recording a clinical interaction is a crucial part of any practice. EMRs and EHRs are great tools for getting the data in. However, most of the time information retrieval from these systems is a complex and technologically time-consuming process. With Generative AI – guided by domain expertise and prompt engineering – extracting information from unstructured EMR records can be made easy. For example – Find me cohort of 40- to 50-year-old females with Diabetic Retinopathy – and other such ad hoc queries, is much easier to answer using Generative AI on medical records than current processes of raising a report request and waiting for a database engineer to get you that report.
Create your own knowledge bank:
Let Generative AI learn from the best. As machine learning models, including LLMs, are a projection of the past, the models can be trained/fine-tuned on new data from your practice. Your own knowledge bank should include the medical records, the care plan, the expected outcomes, and the actual outcomes. Models trained on such a knowledge bank can then help you build your own clinical decision support as well as help guide the newer clinical staff to best practices for future.
Experiment and iterate:
Rapid experimentation and hypothesis testing will make your Generative AI work for you much more quickly than long-term project planning and goal setting. Once the experiment is producing desired results, step back and think of other areas in your practice which can be automated or enhanced by this technology. Usage of Generative AI should be a fun journey.
—----------
Aman has recently joined the Datamede team and is focused on unlocking Automation and Analytics value to customers.
%20(1).webp)