AI’s Impact on Healthcare Data Standards

Will AI make data standards obsolete?

Created by Author using MidJourney

Healthcare Data Standards have a long history, and investment in the advancement of these standards is significant. Research conducted by the Federal Office of the National Coordinator for Health Information Technology (ONC) showed seventy-three-percent of surveyed digital health companies use standards-based APIs, primarily the FHIR standard, to integrate with EHRs. The other Healthcare Data Standards, described below, have been broadly adopted and are used around the World.

But now, with the dramatic advances in AI and particularly Large Language Models (LLMs), will data standards become obsolete? Does one AI, communicating with another AI, need human-created data standards to exchange healthcare data?

This article explains why I believe Healthcare Data Standards will be needed for the foreseeable future. But the “foreseeable future” isn’t the same as “forever,” so I conclude by reminding healthcare standards-makers that they need to keep up with the rapid advances in AI and should not become complacent.

Background

With the advent of computers in healthcare it did not take long to realize the need for standardization. While the data represented within one computer and its associated storage subsystems is only constrained by the applications running on that machine, once two or more computers get involved with some means of communication among them, the need for data standards becomes apparent.

Early in computing standards were needed to implement operating systems, keyboards, displays, and printers; all functionality that works behind the scenes in today’s world. In the healthcare industry, advancing needs for data sharing prompted creation of healthcare-specific standards.

• CPT (Current Procedural Terminology): CPT codes are numerical identifiers assigned to various medical tasks and services provided by healthcare providers.
• HL7 (Health Level Seven): HL7® is a standard for exchanging information between medical information systems.
• NCPDP (National Council for Prescription Drug Programs): NCPDP standards are data interchange standards for the pharmacy services sector of the healthcare industry.
• ICD-10 (International Classification of Diseases, 10th Revision): The International Classification of Diseases is a globally used medical classification system used in epidemiology, health management and for clinical purposes. ICD-11 is now the global standard for recording health information and causes of death.
• FHIR (Fast Healthcare Interoperability Resources): FHIR is designed to enable health data, including clinical and administrative data, to be quickly and efficiently exchanged.

But do AIs need standards? In theory, AI systems could develop their own language for data exchange. In fact, there have been instances where AI systems have created their own form of communication. For example, in a negotiation experiment conducted by Facebook AI Research, two AI agents developed a unique language to communicate with each other. However, this language was not interpretable by humans and the experiment was shut down.

As far back as 2017, separate research at Meta (named Facebook at that time), OpenAI, and Google, showed that AIs, given the opportunity, would develop their own language specific to the assigned communication task.

Extrapolating from these experimental results, it seems likely that AIs could develop their own “language” to exchange healthcare data, and that the data encoded in this AI-created data exchange language might not be understandable by humans.

My Experiment with GPT-4 in February 2024

I thought I would put GPT-4 to the test in a test scenario using a synthetic “history and physical” document. This conversation with GPT-4 is presented in full at the end of this article.

First, I uploaded the synthetic patient record and asked GPT-4 to give me a summary of the patient and her condition. It gave me a nice summary.

“The simulated patient, Ms. Pamela Rogers, a 56-year-old woman, presented with a week-long history of chest pain, described as dull and aching, originating in the left para-sternal area and radiating to her neck. The pain began abruptly and was accompanied by shortness of breath but no other symptoms like sweating, nausea, or vomiting. Initially triggered by gardening, subsequent episodes occurred during walking and even at rest, each resolving with rest….”

Next, I asked it to tell me what ICD-10 codes would describe the patient’s condition. It reported five likely codes. Then, I asked the likely treatment path for this patient and the associated CPT codes. It generated six general areas of treatment, and the associated CPT codes or ranges of possible codes.

After it generated the hypothetical treatments, I asked it to generate the FHIR messages that would likely be sent associated with the treatment of this patient. It generated eight likely message types that would be generated for this patient.

Then I asked GPT-4 what NCPDP claims would likely be generated for this patient during her treatment. It generated four types of codes that could be sent on her behalf.

Finally, I decided to see how well GPT-4 could coordinate the care of Pamela Rogers after her hypothetical hospital stay. I uploaded the “Care Management Workbook” and asked GPT-4 about the care coordination steps that would be recommended by the Workbook. The AI recommended four specific care coordination activities for Pamela Rogers.

The process described above, patient assessment, proposing likely treatments, coding, and care planning, all look accurate from a high-level perspective. In an actual workflow, a qualified human would need to review all codes before recording or submission, and a qualified clinician would need to review and sign-off on any reports included in the medical record.

I posed a hypothetical data exchange use-case to GPT-4 regarding this patient to examine its reaction to the necessity of data exchange standards like FHIR. I asked, “How would an AI helping one hospital communicate with an AI in another hospital regarding a patient who was transferred from one to the other?” Would the AIs spontaneously develop their own language to communicate with each other regarding this specific task? Based on similar cases described above, such an AI-created language would probably be specific to the particular task and would not necessarily be universally applicable across different systems or tasks.

Even though it would be possible for two or more AIs to use unique languages and codes to exchange healthcare information, there would be at least four reasons not to do so.

• Accuracy and Consistency — Use of data standards requires use of codes linked to precisely defined, disambiguated terminology.
• Regulatory Compliance — Data Standards are increasingly required by government agencies and by contracts among healthcare entities.
• Interoperability — Healthcare software and intelligent devices are designed to use standardized data. This is critical for communicating with other computer systems including AIs, as well as with humans.
• Explainability — When AIs use Data Standards, the data being transmitted can be reviewed and understood by humans, helping to resolve the AI Black Box Problem.

Furthermore, using AI-generated languages for data exchange would result in data privacy, security, and ethics problems. For instance, if AI-to-AI communications, not understandable by humans, were used to train the next generation of AI models could we trust those models?

Given these reasons, in the immediate future, we can expect AI-enabled systems to use the Healthcare Data Standards described in this article.

Conclusion

After years of effort, worldwide commitment to the use of Healthcare Data Standards is growing. It is likely that the adoption of AI-enabled healthcare software will make data standards even more valuable in the near future. The use of multimodal LLMs to translate voice, text, images, and video into structured data may dramatically increase the adoption of Healthcare Data Standards by reducing the cost of implementation and use. Coding from natural language, and de-coding into natural language, will be almost instantaneous.

However, that does not mean that data standard proponents should grow complacent. The capabilities of AI models are increasing dramatically, and even today’s AI models may not need standardized data formats to exchange or store clinical and administrative data for limited use cases — such as the hospital-to-hospital patient transfer example we used above. Healthcare Data Standard proponents need to be aware of the increasing impact of AI and ensure that data standards remain relevant in this rapidly changing environment.

Following is the transcript of the author’s interaction with GPT-4. Most readers will not need to read the transcript.

Transcript of Conversation with GPT-4 February 2024

You

I do consulting work for a healthcare IT consulting firm with a specialization in healthcare data exchange standards including HL7, FHIR and NCPDP. I was thinking about the AI revolution which will dramatically help healthcare IT systems acquire and analyze and understand “unstructured” data like human language text including physician notes and patient’s subjective input. I wonder how this AI revolution will impact the movement toward healthcare data standards and the people who earn their living implementing and creating and improving those standards. I bet there will be good impacts and not so good impacts. Please help me think that through starting with a statement of the question, and then providing your list of positive impacts, negative impacts, and just interesting observations.

ChatGPT

Certainly, Ed. Your consideration of the impact of AI on healthcare IT, specifically in the context of healthcare data standards like HL7, FHIR, and NCPDP, is quite pertinent. Let’s frame the question and explore the potential impacts:

Question: How will the AI revolution, with its enhanced capabilities in processing and understanding unstructured data such as human language text (including physician notes and patient subjective input), impact the movement towards healthcare data standards and the livelihood of professionals involved in implementing, creating, and improving these standards?

Positive Impacts:

1. Enhanced Data Integration: AI can significantly improve the integration of unstructured data with existing healthcare data standards. This could lead to richer, more comprehensive patient records.
2. Streamlined Compliance: AI algorithms could automate aspects of compliance with data standards, reducing the burden on healthcare professionals and ensuring more consistent adherence.
3. Innovation in Standards Development: The evolution of AI might drive the creation of new standards or the modification of existing ones to better accommodate AI-driven data insights and analytics.
4. Improved Data Quality and Consistency: AI can assist in standardizing the format and quality of unstructured data, making it more uniform and easier to utilize across different healthcare systems.
5. Career Opportunities in AI-Healthcare Intersection: There will likely be an increased demand for professionals who understand both healthcare data standards and AI, creating new career paths and opportunities for training and development.

Negative Impacts:

1. Obsolescence of Some Skills: Certain tasks related to data standard implementation might become automated, leading to a potential decrease in demand for these specific skill sets.
2. Need for Continuous Learning: Professionals will need to constantly update their skills to keep pace with rapidly evolving AI technologies, which could be challenging for some.
3. Potential for Misalignment: If AI development outpaces the evolution of data standards, there might be a period of misalignment where AI capabilities are not fully utilized due to existing standard limitations.
4. Data Privacy Concerns: The increased use of AI in processing patient data might raise new data privacy and security concerns, necessitating revisions in standards and regulations.

Interesting Observations:

1. Shift in Focus: The role of professionals might shift from implementing standards to more strategic activities, such as data governance, ethical considerations of AI use, and long-term planning.
2. Interdisciplinary Collaboration: The integration of AI into healthcare IT necessitates closer collaboration between technologists, clinicians, and standard bodies, potentially leading to more interdisciplinary innovations.
3. Standardization of AI Models: There might be a move towards standardizing AI models and algorithms themselves within healthcare, similar to how data standards are currently established.
4. Global Impact: The advancements in AI and their integration into healthcare IT could have a global impact, potentially reducing disparities in healthcare quality by providing advanced tools to a wider range of healthcare providers.

These are broad strokes of potential impacts. The actual effects will likely depend on how quickly AI technologies are adopted, how healthcare data standards evolve, and the extent to which these technologies are integrated into existing healthcare systems and practices.

You

I’ve uploaded a simulated “Written History and Physical Examination.” This is not a real patient. Would you please read that and summarize the simulated patient’s condition in two paragraphs?

ChatGPT

The simulated patient, Ms. Pamela Rogers, a 56-year-old woman, presented with a week-long history of chest pain, described as dull and aching, originating in the left para-sternal area and radiating to her neck. The pain began abruptly and was accompanied by shortness of breath but no other symptoms like sweating, nausea, or vomiting. Initially triggered by gardening, subsequent episodes occurred during walking and even at rest, each resolving with rest. Ms. Rogers has a history of hypertension diagnosed three years ago, underwent a total abdominal hysterectomy and bilateral oophorectomy six years ago, and has a family history of premature coronary artery disease. She doesn’t smoke or have diabetes, and her lifestyle includes moderate alcohol consumption.

The physical examination revealed high blood pressure and a heart rate of 90 bpm. Notable findings include crackles in the lung bases, a grade 2/6 systolic decrescendo murmur best heard at the right second intercostal space, and a third heart sound at the apex. There were no abnormal findings in the abdomen, extremities, or neurological examination. The differential diagnosis primarily suggests angina pectoris, likely unstable given the increasing frequency and occurrence of pain at rest, with considerations for ischemic heart disease and potential congestive heart failure. Secondary hypertension due to a renal artery condition and valvular heart disease are also considered.

You

Please tell me what ICD-10 codes would describe this patient’s condition.

ChatGPT

Based on the simulated patient’s condition, the appropriate ICD-10 codes would include:

1. Unstable Angina: I20.0 — Unstable angina is characterized by chest pain or discomfort that typically occurs at rest or with minimal exertion, often worsening in severity and frequency, as seen in Ms. Rogers’ case.
2. Hypertension: I10 — Essential (primary) hypertension. Ms. Rogers has a history of hypertension diagnosed three years ago.
3. Heart Murmur, Unspecified: R01.1 — Cardiac murmurs unspecified, to cover the finding of a systolic decrescendo murmur.
4. Congestive Heart Failure: I50.9 — Heart failure, unspecified. This code is considered given the physical examination findings suggestive of left ventricular dysfunction, such as crackles in the lung bases and a third heart sound.
5. Family History of Ischemic Heart Disease: Z82.49 — Family history of ischemic heart disease and other diseases of the circulatory system. Ms. Rogers has a family history of premature coronary artery disease.

It’s important to note that ICD-10 coding can be complex and is often context-dependent. The codes chosen here are based on the information provided in the simulated patient’s history and physical examination. In a real clinical setting, additional details from the patient’s medical history, laboratory results, and other diagnostic findings might lead to different or additional codes.

You

What would likely treatments be for this patient? Please specify the associated CPT code.

ChatGPT

Based on the simulated patient’s condition, the likely treatments and associated CPT codes would be:

1. Hospital Admission for Monitoring (CPT Code 99221–99223): Due to the unstable nature of her angina, Ms. Rogers should be admitted to the hospital for close monitoring of any increased chest pain that might indicate an impending myocardial infarction.
2. Medication Management (CPT Code 99231–99233 for subsequent hospital care):

• Aspirin and Nitrates (CPT Code 99214–99215 for office/outpatient visit): Start platelet inhibitors such as aspirin to decrease the risk of myocardial infarction and nitrates for pain and to help lower blood pressure.
• Analgesic for Prolonged Pain: For pain unresponsive to nitrates, an analgesic like morphine may be required.
• Cholesterol Management: Cholesterol monitoring and potentially starting cholesterol-lowering medication like HMG Co-reductase inhibitors.

1. Cardiac Catheterization (CPT Code 93458): Schedule a cardiac catheterization to evaluate coronary arteries and assess for ischemic heart disease.
2. Diuretics (CPT Code 99214–99215 for office/outpatient visit): Begin diuretics for dyspnea secondary to volume overload, which will also assist in managing high blood pressure.
3. Lifestyle Modifications (CPT Code 99401–99404 for individual preventive medicine counseling): Advise on an appropriate exercise and weight loss program, including a low-fat diet.
4. Laboratory Tests (CPT Code 80053 for comprehensive metabolic panel): Include BUN/Creatinine to assess kidney function, electrolytes, and a baseline EKG.

It’s important to note that the specific CPT codes can vary based on the detailed nature of each service provided and the setting in which they are provided (inpatient vs. outpatient). The codes provided here are general and may need to be adjusted based on the specific circumstances of the patient’s care.

You

What FHIR messages would likely be sent associated with the care of this patient?

ChatGPT

In the context of the patient’s treatment, various FHIR (Fast Healthcare Interoperability Resources) messages would likely be sent. FHIR is a standard for exchanging healthcare information electronically. Here are some examples of FHIR messages that might be associated with this patient’s care:

1. Admission to Hospital:

• FHIR Resource: Encounter
• Purpose: To document the patient’s admission to the hospital for monitoring due to unstable angina.

1. Medication Orders:

• FHIR Resource: MedicationRequest
• Purpose: To order medications such as aspirin, nitrates, cholesterol-lowering drugs, and possibly diuretics.

1. Cardiac Catheterization Scheduling:

• FHIR Resource: ProcedureRequest or ServiceRequest
• Purpose: To schedule a cardiac catheterization to assess the patient’s coronary arteries.

1. Laboratory Test Orders:

• FHIR Resource: DiagnosticRequest
• Purpose: To order lab tests like a comprehensive metabolic panel, including BUN/Creatinine, electrolytes, and a baseline EKG.

1. Clinical Observations and Vital Signs:

• FHIR Resource: Observation
• Purpose: To record clinical observations such as blood pressure, heart rate, and findings from the physical examination (e.g., systolic murmur, lung crackles).

1. Discharge Planning:

• FHIR Resource: CarePlan
• Purpose: To outline the care plan upon discharge, including lifestyle modifications, exercise, and weight loss programs.

1. Patient’s Personal and Family Medical History:

• FHIR Resource: FamilyMemberHistory and Patient
• Purpose: To document the patient’s medical history and family history of premature coronary artery disease.

1. Diagnostic Imaging Reports:

• FHIR Resource: ImagingStudy
• Purpose: To record the results of any imaging studies, such as a ventriculogram performed during cardiac catheterization.

These FHIR messages enable efficient and accurate sharing of patient information among different healthcare systems, ensuring continuity and quality of care. They cover a range of interactions, from initial admission to discharge planning and follow-up care.

You

What are some example NCPDP claims that would likely be submitted for this patient?

ChatGPT

Based on the simulated patient’s condition and the prescribed treatments, the following NCPDP (National Council for Prescription Drug Programs) claims would likely be submitted:

1. Aspirin (for unstable angina and to decrease the risk of myocardial infarction):

• NDC (National Drug Code) for the specific aspirin product.
• Pharmacy Service Type: Retail Pharmacy.
• Prescription Origin Code: Written.

1. Nitrates (for chest pain and hypertension management):

• NDC for the specific nitrate medication (e.g., nitroglycerin).
• Pharmacy Service Type: Retail Pharmacy.
• Prescription Origin Code: Written.

1. Cholesterol-Lowering Medication (if cholesterol is elevated):

• NDC for the specific statin medication (e.g., HMG Co-reductase inhibitors like atorvastatin or simvastatin).
• Pharmacy Service Type: Retail Pharmacy.
• Prescription Origin Code: Written.

1. Diuretics (for dyspnea secondary to volume overload and hypertension):

• NDC for the specific diuretic medication.
• Pharmacy Service Type: Retail Pharmacy.
• Prescription Origin Code: Written.

Each claim would include the patient’s identification, insurance information, and prescriber details. It’s important to note that the specific NDCs would depend on the exact medications and dosages prescribed. The claims would be processed according to the patient’s pharmacy benefits and insurance plan, ensuring appropriate coverage and payment for the medications.

You

[I uploaded the Care_Management_Workbook.pdf into GPT-4's context window.]

Applying the care management guidelines in the attached “Care Management Workbook” what types of care should be given to our hypothetical patient Pamela Rogers?

ChatGPT

Based on the Care Management Workbook’s guidelines, the care for Ms. Pamela Rogers, a 56-year-old patient with unstable angina, hypertension, and signs suggesting potential heart failure, would include several steps:

1. Comprehensive Needs Assessment: This would involve a thorough evaluation of Ms. Rogers’ medical conditions, medications, and overall health status, taking into account her recent hospitalization due to chest pain and her chronic conditions.
2. Care Plan Development and Implementation: A personalized care plan would be developed, focusing on her specific health needs. This plan would likely include medication management, regular monitoring of her cardiac and hypertension status, lifestyle modifications, and possibly coordination for a cardiac catheterization. The care plan should also consider her mental and emotional well-being, as well as her ability to manage her conditions at home.
3. Monitoring and Modification of Care Plan: The care plan would be regularly reviewed and modified as necessary, based on her response to treatment and any changes in her condition.
4. Coordination of Care: Coordination between different healthcare providers is crucial, including her primary care physician, cardiologist, and any other specialists involved in her care.

The workbook emphasizes a holistic, patient-centered approach, ensuring that the care provided is tailored to the individual needs of the patient, considering not only the physical aspects of their health but also their emotional and social well-being.

You

Changing the topic a bit to a higher level hypothetical, if we had an AI like you working at one hospital and a similar AI working at another hospital, would you AIs need data standards to exchange information on a patient (ICD-10, CPT, NCPDP, FHIR), or would you two AIs be able to exchange data using your own definitions without using standards?

ChatGPT

In the hypothetical scenario where two AI systems, such as myself, are operating in different hospitals, the need for data standards like ICD-10, CPT, NCPDP, FHIR, etc., would still be crucial. Here’s why:

1. Interoperability: Data standards ensure interoperability between different systems. They provide a common language and framework, allowing AIs at different hospitals to understand and interpret the data consistently.
2. Accuracy and Consistency: Using established standards helps maintain the accuracy and consistency of patient data. This is critical for effective diagnosis, treatment planning, and patient care management.
3. Regulatory Compliance: Healthcare data exchange is subject to various regulatory requirements, which often mandate the use of specific standards to ensure privacy, security, and accuracy of patient information.
4. Integration with Existing Systems: Hospitals use numerous systems and technologies. Standardized data formats are necessary for seamless integration between AI systems and these existing healthcare technologies.

Even though AI systems could theoretically develop their own methods of communication, adhering to established standards ensures reliable, consistent, and compliant exchange of healthcare information.