AI in Healthcare: How Can It Be Applied to Medical Records & Interoperability?

Overview

Artificial Intelligence and Machine Learning have been all the buzz recently. Rightfully so, as it seems like a very promising revolution in technology and our lives. Just as the agricultural and industrial revolutions changed how previous generations of our ancestors lived and worked, the AI/ML revolution will likely influence our generation similarly. AI in Healthcare is a trend that is just beginning, yet likely to continue for years.

In this blog, we’ll talk about what AI/ML is and what capabilities it has, what it looks like in healthcare today, where AI/ML still needs work, and how we see artificial intelligence and machine learning playing a role in healthcare interoperability.

What is AI/ML and What Can It Do?

As the name implies, Machine Learning models take inputs of information to learn, predict and produce outputs similarly to humans. AI/ML models bring together analytics, data science, and automation. Machine learning algorithms are trained with data and can continuously improve the quality and accuracy of their outputs. One unique characteristic of these learning models and algorithms is that they improve exponentially.

Some examples of what AI/ML can do today:

1. Write a long text output (e.g., write a blog, essay, or post) from a short text prompt
2. Provide support to engineers and developers by writing basic programming scripts
3. Provide general answers with explanations to specific questions
4. Fraud detection, spam filtering

GPT-3 (Generative Pre-trained Transformer 3), a language model that produces human-like text released in 2020, uses 175 billion parameters. GPT-4 (Generative Pre-trained Transformer 4), set to be released in 2023, will support 170 trillion parameters. It also will expand from natural language processing capabilities and support tasks such as image and video creation.

Those numbers suggest that GPT-4 will be 100x more powerful than the release three years earlier. That speed of capability development is unheard of anywhere except maybe in young human brains. The breadth of capability development is unheard of anywhere. Machine Learning and Artificial Intelligence will permeate every industry, including healthcare.

What Does AI Look Like in Healthcare Today?

AI/ML began to enter the healthcare tech market within the last five years or so. Adopting AI/ML is risky and takes time, like other innovation-induced industry shifts. Machine Learning solutions are new and have limits in their application. By removing the human element, we rely on the accuracy and quality of technology and the logic and thinking the algorithm applies. Since healthcare deals directly with the well-being and life of humans, caution and careful evaluation are essential.

Healthcare is a human capital and resource-intensive business. The size and volume of the global healthcare market are large, the stakes of delivering patient care are high, and the profit margins for healthcare organizations are low. Healthcare may be the prime industry candidate for adopting Machine Learning and Artificial Intelligence solutions and automation.

There are several areas and specialties that are currently implementing AI/ML applications into their workflows. The main benefits that can be realized today are improved clinician efficiency, increased diagnosis speed and accuracy, and better treatment outcomes.

Here are a few examples of AI/ML employed in healthcare today:

1. Clinical Note Generation – solutions now use NLP (Natural Language Processing) and guided questionnaires to generate clinical notes for provider and patient visits. These applications significantly improve a provider’s efficiency by decreasing the documentation and administrative task time. An example of this solution is SOAP Health, which developed a conversational-AI Medical Assistant.
2. Diagnosis Decision Support – based on patient responses to questions, current symptoms, demographics, and medical history, programs can identify likely conditions and diagnoses for providers to investigate.
3. Sepsis Detection – real-time predictive analysis algorithms can monitor and detect patients at risk or with early signs of sepsis.
4. Radiology & Image Reading – AI and deep learning methods can recognize, categorize, and sometimes make recommendations from medical imaging. Applications can identify patterns that may not be apparent to human physicians and provide quantitative analysis and guidance rather than a qualitative reading.

Where Does AI Still Fall Short

Artificial Intelligence and Machine Learning are proficient in specific, complex tasks and modalities but not as effective for requests requiring emotional or interactive capability. For example, a model can deduct a cancer diagnosis by parsing through a high volume of unstructured data from multiple sources, something a human may miss. But the model cannot deliver an empathetic and encouraging diagnosis to a patient in the same manner as a physician.

AI/ML has excellent potential to be a supporting tool and, in some cases, a “Swiss army knife” for providers and clinical teams. However, it won’t replace them. Not anytime soon, at least. And we should not let it.

The 2008 financial crisis was devastating because risky debt was collateralized and leveraged repeatedly. In essence, the mortgage lending business created the debt bubble and then built on exponentially by the Mortgage-Backed Securities, CDOs, and synthetic CDOs created by the banking industry. When defaults on loans began, these financial practices created uncontrollable subprime debt exposure that exacerbated the economy’s downturn.

Like the situation in 2008, if AI/ML permeates all areas and eliminates human monitoring and intervention, a bad or “rogue” model could have devastating repercussions. Deep Learning and neural networks have hidden layers and build upon themselves with the information they take as inputs. By its nature, it runs autonomously, and errors in the program compound far beyond the immediate output. Caution and care must be kept at the forefront when developing and running AI/ML applications in healthcare.

Luckily, AI/ML is more of a marketing and sales play for many digital health and healthcare technology companies. The vision is to be an AI/ML solution that algorithmically learns to improve the outputs it produces. Right now, many programs take inputs and build conjunctive concepts using AND and OR statements and still rely on conditional programming, using IF, ELSE, THEN type logic to produce outputs.

How Intely Sees Using AI/ML For Healthcare Interoperability?

Data Interoperability is a hot topic in healthcare, and AI/ML models most certainly have a place to improve data sharing and communication between systems and technology solutions. Healthcare interoperability solutions generally sit behind the scenes yet significantly improve provider efficiency and patient outcomes.

The product team’s vision at Intely is to incorporate AI/ML models in our platform, adding value and efficiency to interfacing and standardizing data between multiple systems and applications. We believe AI/ML can assist interoperability for point solutions and other digital health applications.

Here is an example of how it may look in the future:

1. Determining Integration Requirements – using AI/ML models take workflow needs between two or more systems and use business needs to define what interfaces and connection needs.
   
   
   Input: “Vendor application needs to send form links to patients before the visit, collect the responses, and send to the EMR for the provider to review. The EMR system is Epic.”
   
   
   Output: “You need a scheduling and patient demographic feed from Epic. Epic does not support webhook events. An Outbound SIU HL7 feed will provide the information about the appointment and patient.
   
   You will also need an integration or method to deliver the secure links to the patient using the information provided by the EMR.
   
   After completing the form or questionnaire, send the responses back using an HL7 ORU Inbound or FHIR Request to create a document. Use the patient ID in the original message to match the patient in the EMR system.
   
   This is primarily an HL7 v2 workflow and can be supported by FHIR interfaces.”
   
   
   
2. Building Interfaces – training an AI/ML model to use interface standards and system specifications to make connections and an interoperability workflow.
   
   
   Input: Using the parameters of the customer Epic EMR system (IP Address, Ports, Device Info) build an HL7 interface and VPN connection to receive SIU messages and send ORU messages. Use the client FHIR server (URL, Authorization URL, API endpoint specification) to establish an API interface to query patient data.
   
   
   Output: Configured and tested Interface Channels and Instances
   
   
   
3. Standardizing Data – standardize, convert, and map data from multiple sources using system specifications such as API specs and standard shared vocabularies.
   
   
   Input: “Using the responses to the patient questionnaires, format outputs that conform to the HL7 v2 standard for Observation and FHIR Resource standard for Questionnaire and QuestionnaireResponse.”
   
   
   Output: “Formatted HL7 ORU messages and FHIR Questionnaire and QuestionnaireResponse resources that can be accepted by client EMR system.”
   

The above examples and scenarios are full of nuance and the need for human intervention. However, they roughly match the work required by a human to connect and share healthcare data. It is a pie-in-the-sky vision, but the AI revolution is coming, and it’s coming to healthcare. Fast!