Using AI to Improve Electronic Health Records

Electronic health record systems for large, integrated healthcare delivery networks today are often viewed as monolithic, inflexible, difficult to use and costly to configure. They are almost always obtained from commercial vendors and require considerable time, money, and consulting assistance to implement, support and optimize.

The most popular systems are often built around older underlying technologies, and it often shows in their ease of use. Many healthcare providers (including the surgeon and author Atul Gawande) find these systems complex and difficult to navigate, and it is rare that the EHR system is a good fit with their preferred care delivery processes.

As delivery networks grow and deploy broad enterprise EHR platforms, the challenge of making them help rather than hinder clinicians is increasing. Clinicians’ knowledge extends far beyond their clinical domain—care procedure knowledge, patient context knowledge, administrative process knowledge—and it’s rare that EHRs can capture all of it efficiently or make it easily available. What’s more, in the U.S., regulatory, billing and revenue cycle requirements add additional complexity to the electronic healthcare workflow and further reduce the time clinicians have to engage with patients.

The options for improving this misalignment between systems and processes are limited. One is to design EHR systems to be more integrated and streamlined from the beginning. One Medical, for example, a concierge medical practice across 40 cities in the U.S., developed its own EHR system that is closely aligned with the care and patient relationship practices it employs. Flatiron Health, a data and analytics-driven cancer care service recently acquired by Roche, bought a company with a web-based EHR and tailored it to fit its OncoCloud EHR for community-based oncology. Although these bespoke systems do seem to fit clinician workflows better, they are themselves difficult and time-consuming to develop (One Medical required ten years to build its system) and they are relatively narrow in scope. Building a system from scratch or extensively customizing a commercial one would probably not work for large delivery networks.

Using an open source EHR is a second option. However, most current ones are designed for small medical practices and aren’t easily scalable or need substantial configuration. And even though the software is free, considerable programming and IT infrastructure is required to implement it and tailor it to the individual practice. Further, open source EHRs are less carefully maintained and less frequently updated than commercial ones and so can quickly become obsolete. Finally, regulatory requirements and reimbursement rules change rapidly. Relying on either open source or internally developed systems in keeping up with those requirements creates both compliance risks and financial challenges.

A third and more promising option is to use AI to make existing EHR systems more flexible and intelligent. Some delivery networks, sometimes in collaboration with their EHR platform vendor, are making strides in this direction. AI capabilities for EHRs are currently relatively narrow but we can expect them to rapidly improve. They include:

Data extraction from free text  Providers can already extract data from faxes at OneMedical, or by using Athena Health’s EHR. Flatiron Health’s human “abstractors” review provider notes and pull out structured data, using AI to help them recognize key terms and uncover insights, increasing their productivity. Amazon Web Services recently announced a cloud-based service that uses AI to extract and index data from clinical notes.

Diagnostic and/or predictive algorithms Google is collaborating with delivery networks to build prediction models from big data to warn clinicians of high risk conditions such as sepsis and heart failure. Google, Enlitic, and a variety of other startups are developing AI-derived image interpretation algorithms. Jvion offers a “clinical success machine” that identifies patients most at risk as well as those most likely to respond to treatment protocols. Each of these could be integrated into EHRs to provide decision support.

Clinical documentation and data entry Capturing clinical notes with natural language processing allows clinicians to focus on their patients rather than keyboards and screens. Nuance offers AI-supported tools that integrate with commercial EHRs to support data collection and clinical note composition.

Clinical decision support  Decision support, which recommends treatment strategies, was generic and rule-based in the past. Machine-learning solutions are emerging today from vendors including IBM Watson, Change Healthcare, AllScripts that learn based on new data and enable more personalized care.

While AI is being applied in EHR systems principally to improve data discovery and extraction and personalize treatment recommendations, it has great potential to make EHRs more user friendly. This is a critical goal, as EHRs are complicated and hard to use and are often cited as contributing to clinician burnout. Today, customizing EHRs to make them easier for clinicians is largely a manual process, and the systems’ rigidity is a real obstacle to improvement. AI, and machine learning specifically, could help EHRs continuously adapt to users’ preferences, improving both clinical outcomes and clinicians’ quality of life.

However, all of these capabilities need to be tightly integrated with EHRs to be effective. Most current AI options are “encapsulated” as standalone offerings and don’t provide as much value as integrated ones, and require time-pressed physicians to learn how to use new interfaces. But mainstream EHR vendors are beginning to add AI capabilities to make their systems easier to use. Firms like Epic, Cerner, Allscripts, and Athena are adding capabilities like natural language processing, machine learning for clinical decision support, integration with telehealth technologies and automated imaging analysis. This will provide integrated interfaces, access to data held within the systems, and multiple other benefits—though it will probably happen slowly.

Future EHRs should also be developed with the integration of telehealth technologies in mind (as is the EHR at One Medical). As healthcare costs rise and new healthcare delivery methods are tested, home devices such as glucometers or blood pressure cuffs that automatically measure and send results from the patient’s home to the EHR are gaining momentum. Some companies even have more advanced devices such as the smart t-shirts of Hexoskin, which can measure several cardiovascular metrics and are being used in clinical studies and at-home disease monitoring. Electronic patient reported outcomes and personal health records are also being leveraged more and more as providers emphasize the importance of patient centered care and self disease management; all of these data sources are most useful when they can be integrated into the existing EHR.

Most delivery networks will probably want to use a hybrid strategy—waiting for vendors to produce AI capabilities in some areas and relying on third party or in-house development for AI offerings that improve patient care and the work lives of providers. Starting from scratch, however, is probably not an option for them. However necessary and desirable, it seems likely that the transition to dramatically better and smarter EHRs will require many years to be fully realized.