Digital information is rapidly expanding in the healthcare industry. Healthcare data was predicted to increase at a 36% CAGR between 2018 and 2025, according to a white paper released by IDC in collaboration with Seagate. As a result of this rise, healthcare is now one of the world’s top data sources.
About 80% of medical data in the healthcare industry, comprising formats like text, photos, and signals, is left unorganized and untapped once it is created, per a report by Healthcare Informatics Research at Chungnam National University in Daejeon, Korea.
Finding a balance between privacy and innovation is the difficulty in transforming this data into insight. According to a different study by Administrative Data Research UK, which was supported by the Economic & Social Research Council, organizations are under pressure to use their data responsibly because of strict regulatory oversight and eroding public confidence in data.
De-identification lies in the space between these regulatory scrutiny points and maximizing patient value. De-identification enables teams to use and share patient-level details in a secure manner by eliminating or masking personally identifiable information (PII) while maintaining the therapeutic value of the data. It not only makes regulatory compliance possible, but it also establishes a basis of confidence.
Using de-identification, benchmarking, and education to ensure responsible and successful AI adoption in healthcare, Ben Webster, Modeling and Analytics Team Lead at NLP Logix, and Brad Kennedy, Senior Director of Business Solutions Strategy at Orlando Health, will discuss how to strike a balance between data innovation and patient safety in a special series on the “AI in Business” podcast.
Their discussion emphasizes how critical it is to balance clinical results, patient privacy, and trust when integrating AI in healthcare. In order to ensure responsible innovation that does not jeopardize data ethics or trust, it emphasizes the significance of efficient de-identification procedures, transparent patient communication, and outcome-driven benchmarks.
Two important takeaways from these discussions for healthcare executives implementing AI in their companies are examined in this article:
Improving AI effectiveness by striking a balance between change and compliance: putting an emphasis on adaptable compliance measures and making proactive workflow adjustments to hasten the implementation of AI and provide quantifiable benefits.
De-identification procedures being put in place to promote responsible innovation: establishing strong de-identification procedures that protect patient privacy while utilizing patient data for AI and technological advancement.
Increasing AI Efficiency By Juggling Change And Compliance
Episode: NLP Logix’s Ben Webster Discusses De-Identified Data And AI Adoption In Healthcare
Ben Webster, NLP Logix’s Modeling and Analytics Team Lead, is the guest.
Proficiency in Sentiment analysis, predictive modeling, and advanced analytics
Ben has worked at NLP Logix for the past ten years. From 2013 to 2021, he was a data scientist before being elevated to his present role as Modeling and Analytics Team Lead. He graduated from the University of North Florida in 2016 with a master’s degree in mathematics and statistics.
Ben begins his podcast by stating that de-identification usually occurs when teams wish to use real-world data for experimentation and is frequently motivated by legal and regulatory considerations, especially to maintain compliance with HIPAA. He lists the two main techniques used by businesses to de-identify data:
Using the Safe Harbor Method, 18 different kinds of identifiers are eliminated, ranging from apparent ones like names and photographs to less visible ones like IP addresses and birthdates. Although it is simple legally, it frequently removes too much information, making it less valuable for research.
Expert Determination: In this instance, an expert evaluates whether it is reasonable to re-identify people using the data. Although this approach allows for greater flexibility, it can also result in delays; legal reviews can take months at times, putting projects on hold before any testing has even started.
He points out that there is increasing interest in organizing medical data to help with prescription choices, billing accuracy, and patient home care insights—much of which is still concealed in free-text notes. He identifies two significant market trends:
Developing tools to translate and standardize data as patients move between foreign healthcare systems with different languages and standards, as well as using LLMs to construct conversational interfaces that let physicians query data naturally.
Ben also offers a crucial tip for deciding between adopting off-the-shelf software or developing a solution internally:
Performance measures based on generic datasets are frequently promoted by third-party software solutions, but your particular use case may not be reflected in those benchmarks. When used on your own data, a program that claims 90% accuracy can work better or noticeably worse.
Validating performance on representative datasets is crucial since these variations have the power to make or kill a project. Access to de-identified data that can be securely run through the system to evaluate its efficacy in the actual world is necessary to accomplish that. Naturally, privacy issues must also be carefully taken into account, especially when working with PHI and third-party APIs.
—Ben Webster, NLP Logix’s Modeling and Analytics Team Lead
He emphasizes that if users feel forced to manually go over every choice in order to identify infrequent mistakes, even extremely accurate AI is of little use and will just increase their effort. Workflows need to change if AI is to be useful. Projects run the risk of failing and wasting money if users stick to outdated practices.
The most seamless adoption of AI occurs in jobs like as transcribing, where users may revise AI-generated drafts and instantly save time. On the other hand, because of the high stakes and established routines, resistance is greater in difficult fields like clinical diagnostics or claims management. When people analyze AI’s initial output instead of doing all the work by hand, efficiency advantages take place.
Lastly, Ben notes that companies who are suffering with strict service-level agreements (SLAs) or that are constantly having to hire additional employees to meet demand are the ones that are prepared to adapt. However, if a small, competent crew can complete a work effectively at a cost comparable to automation, it might not be worth implementing AI. Instead, true AI opportunities arise when staffing issues or time constraints plainly call for innovative solutions.
Putting De-Identification Procedures In Place To Promote Responsible Innovation
Episode: Orlando Health’s Brad Kennedy Discusses Preserving The Patient Voice In De-Identified Data Models
Brad Kennedy, Orlando Health’s Senior Director of Business Solutions Strategy, is the guest.
Knowledge of Value-Based Care Strategy, Healthcare Operations Transformation, and Enterprise Technology Implementation
Brief Overview: Brad has worked in healthcare strategy and operations for more than 20 years. He presently manages enterprise transformation projects at Orlando Health as Vice President of Strategic Innovation. Texas A&M University awarded him a Master of Healthcare Administration (MHA).
Brad stresses the “least data necessary” strategy in his podcast, gathering only the bare minimum of personal health data required and frequently depending on unique identifiers rather than comprehensive information. It reduces the danger in the event that the data is ever made public. Despite the safeguards, he admits that responsible use of patient data is essential for innovation, especially in AI. Strong de-identification procedures are therefore necessary to advance healthcare technologies in a safe manner.
Brad explains that depending on the particular use case or study being carried out, different levels of patient information are required. For instance, researchers usually don’t need to know the patient’s identity while dealing with medical imaging data. To train AI models over thousands of picture recordings, they just require clinical context, such as the condition under study. Personal identifiers are not required in these situations.
However, certain clinical or demographic information becomes crucial for other research, particularly those that measure outcomes or compare across patient populations. Age, disease type, zip code, and other health characteristics that aid in classifying patients into relevant cohorts are examples of these datapoints. With the help of these specifics, researchers may evaluate if a new gadget or AI solution is effective for certain subgroups and compare “apples to apples.”
Brad stresses that even in these situations, personally identifying information like names, phone numbers, or addresses is hardly ever, if ever, used. Studying pertinent characteristics of the patient—rather than the patient themselves—is the main goal. Teams can preserve privacy while assessing whether innovations are producing significant results by striking a balance between strategies that prioritize personal data when necessary and complete anonymity.
He goes on to give an example of remote care, in which wearable technology enables patients to recuperate at home while still being closely watched, as opposed to being kept in a hospital bed.
The wearables allow for prompt actions in the case of a problem by sending real-time data back to the care team. For instance, the team can proactively contact the patient and refer them to the proper care if the device notices an incident or anomaly.
Brad points out that both parties gain from the model he outlines:
In the convenience of their own homes, patients can recuperate.
Healthcare systems improve operational efficiency and lessen the burden on hospital capacity.
He goes on to emphasize that patient data must be shared, acquired with consent, and protected in order for wearables to function in a safe and morally responsible manner. When properly applied, this kind of technology promotes a more flexible and effective healthcare system in addition to improving patient outcomes.
Brad goes on to say that knowing the baseline—what is the present level of care—is the first step towards innovation. Which metrics do we want to raise? It is difficult to determine whether a novel solution is truly compelling without that.
In order to ascertain whether the technology is providing value, he supports a data-driven strategy that compares results between test and control groups. However, clinical outcomes are not the sole metric. Experience is also important:
Is there a disruption or improvement in the clinical workflow?
Does the patient have a better experience?
Are patients making the proper use of the technology?
Brad cites the example of wearable technology, which is frequently and increasingly classified as a medical equipment under regulations:
Consider something as basic as a wearable, such as a ring that tracks health. Because it is easily removed, patient education is extremely important. We must be very explicit about the need of continuing use, what we’re tracking, and how it fits into their overall treatment plan. Patients must comprehend both the short-term objective and the long-term objectives we are pursuing once they return home.
In the end, it all boils down to regular communication and execution: ensuring that clinicians and patients understand the rationale for the new procedure, that they feel at ease with it, and that we demonstrate our commitment by taking action.
—Brad Kennedy, Orlando Health’s Senior Director of Business Solutions Strategy