CAPTURING FUNCTIONAL INFORMATION

Mahmoudi said as far as he can tell, information about functional status is not currently gathered anywhere. Given that, he wondered how current technology—machine learning, natural language processing (NLP), and artificial intelligence (AI)—could collect this information and make it readily and easily available to providers. Vydiswaran answered that collecting information is not what AI and NLP can do. Rather, somebody must capture that information and document it in the EHR, which is the unsolved problem with its own biases. However, if the information is captured and it is not in a standardized field, that is where NLP, and particularly AI, can help. NLP systems typically start with a keyword-based approach to define what the model needs to capture, but these do not cover all the variations in the English language. This is where AI-based approaches that use context are more efficient and effective, and it is where generative AI approaches are getting better at summarizing information that might be in the clinical documentation. The key is training the AI models with good-quality data, which requires human input to identify factors more representative of the conditions of interest and those that are not, which is equally important.

Morris commented that disability and functional status can be separate concepts. She noted that in the standards the Office of the National Coordinator for Health Information Technology is promulgating, disability status and functional status appear among the health status identifiers, not in the demographics section. However, if the goal is to use the data in the disability determination process, it is imperative to think about who is at risk, and that means identifying the demographics of people who are at risk.

Mahmoudi asked Laddha how the community can advocate to add functional status and functionality to the EHR. Laddha said that when national standards are available, it is easy for a technology vendor to implement a change in a way that is available to every consumer out of the box. Absent standards, the EHR vendor can add flexibility in the software to enable the user to define a version or variation of a standard in a way that is most meaningful to the user’s demographic area, health system, or target population. This, she said, is where the work Morris is doing is helpful for technology partners as it serves as a starting point for health systems, absent a national standard, to make their own modifications to their EHR system.

ENABLING BIDIRECTIONAL COMMUNICATION WITH COMMUNITY PARTNERS

Laddha also discussed how health care systems can communicate with community-based services in a format compliant with the Health Insurance Portability and Accountability Act when the community-based organization is not using the Epic system. Michigan Medicine, for example, had to use Dropbox to serve as an intermediary between the health system and community organizations. Noting there is a great deal of interest in addressing this issue, Laddha said several things need to happen when bringing in community partners and engaging them to help patients. The first is understanding the target population and what assistance they need. “Identify the target population and have the analytics in place that can identify inequities and needs,” she said.

The second step is to for the health care system to name a community liaison who will engage with the community. “I do not want to underestimate how difficult that task is, so having a named community liaison that can help you through that process is extremely beneficial,” said Laddha. The community liaison, working with others in the health care system, then identifies the appropriate community partners. The third step is to involve the technology partner and take advantage of their open-source application interfaces that community-based organizations can use to create a closed-loop referral system, though Laddha acknowledged that taking this step is still in its infancy. “For [community-benefit organizations] to set up a process where someone is going in electronically and responding to referrals is still a hard task,” she said. This is where financial incentives for these community-benefit organizations to do this work can help.

Laddha said her organization has developed a light-weight platform for its Epic system to create closed-loop systems with community partners. In Wisconsin, for example, the company developed this platform for doulas because the health care system thought it beneficial to share more of a patient’s history with the doulas to provide the right care.

PROVIDING EVIDENCE-BASED, DISABILITY-FRIENDLY CARE

Following Laddha’s comment about incentives for community-based organizations, Mahmoudi wondered if data showed that addressing the social needs of patients, particularly those with disabilities, reduced hospital readmissions, then might it be possible to provide incentives to these community resources? Laddha replied that value-based contracts with providers could work, given payers have shown some interest in funding community-based organizations. Adler-Milstein said this question ties into the challenge of making a case for health systems to invest in disability-friendly care. Unaware of whether the data to make that case exists, she said this would be a good area for study to generate such data. She wondered if it would be possible to articulate a national-scale transformation effort around what disability-friendly health systems would look like that would engage community partners and then design incentives based on that effort. To go along with the incentives, it would be imperative to provide evidence-based care practices that, if followed, would earn those incentives.

Lagu said there is definitely a need for more research to identify evidence-based practices that improve care for people with disabilities. However, there are some practices that work because they have face validity, because patients report they are happier when those practices occur, and because they are best for quality and safety. The simplest practice, she added, is providing accommodations for people who need them.

D’Sena’ Warren said the question should not be how to keep people with a disability out of the hospital, but why are they going to the hospital in the first place. Usually, she said, it is because going to the emergency department to receive treatment is better than waiting for six months to two years to get an appointment with a specialist. Lagu said she has data to back up Warren’s comment. “We have qualitative data for patients with disabilities who report that they cannot access care, they cannot get appointments with some specialists, and they cannot get the testing they need,” said Lagu. “In some cases, it is their physicians who tell them to go to the emergency room because they say, ‘You are not going to be able to get this care anywhere but in the hospital.’”

Mahmoudi, agreeing with Laddha, Lagu, and Warren, said moving forward requires providing evidence-based data to provide disability-friendly care. In his opinion, it may be productive to provide data showing that if a health system provides that type of care, it will reduce the costs of readmissions and preventable hospitalizations or emergency department visits in addition to helping their patients have a better quality of life and fewer adverse health events. Morris said she would love to do that research if anyone wants to fund it, something that has been difficult so far. The National Institutes of Health’s (NIH) decision to consider individuals with disabilities a disparity population may be a good first step to getting this type of research funded.

Z CODES AND DOCUMENTING SOCIAL DETERMINANTS OF HEALTH

Changing subjects, Mahmoudi asked the panelists for their thoughts on how to use Z codes more efficiently to identify social determinants in a more standardized manner. Adler-Milstein said her health system does not use them, even when clinicians know there is a social need. “I think the issue is some combination of awareness and our clinicians not seeing the value of documenting the codes,” she said. “I think it is a helpful step that they exist, but I think we have to be sure that there is some structure to show the value of using them.” She also blamed a reluctance to use Z codes for this purpose because of the messiness of the problem list tied to the Z codes and the complexity of deciding what code to use for which social need. “Unfortunately, I just do not think there is a strong case for their use right now,” said Adler-Milstein.

Lagu argued the question is not about what diagnostic codes should be in the EHR, but what is the reason for thinking they are even needed. “I think the reasons are that we do not know who has a disability, who needs accommodation, and who has health-related social needs,” said Lagu. Her view is to collect the data to identify who has needs and then inform the health systems to provide the services and accommodations to meet those needs.

Laddha noted that Z codes help with interoperability, and Warren added that what she does not like about them is that the names and diagnoses associated with them are always changing. Vydiswaran said Z codes are not meant to be an end-to-end solution. “It is a very specific solution people have come up with that could help with interoperability,” he explained. To him, the critical first step is not about documenting disability, but documenting the need for accommodations early, often, and continuously so changes in a person’s functional status that occur over time are also recorded. In addition, he emphasized the need to develop a continuous, standardized way of capturing information before thinking that Z codes could be a solution.

IDENTIFYING BIAS

Vincent Nibali asked if there are any indicators to look for in existing EHRs to identify when the information contains biases to adjudicate. Lagu replied that since the EHR is not providing information about who has a disability, spotting biases is difficult. Spotting biased language may be one hint, she said, as could being discharged from a practice, though that can be difficult to determine. Morris noted she and her colleagues completed a randomized controlled trial for which they recruited people with communication-related disabilities. “We came up with 300 ICD10 codes that might have been relevant, and those patients with scheduled appointments with those ICD10 codes were called,” said Morris. “About 50 percent of the people who we were able to get ahold of actually denied having a communication-related disability.”

Though this study could not answer the question of why people did not identify as having a disability, what it indicated was that ICD10 codes have problems. “I think the message is coming across: We need to ask,” said Morris. Once people are asked and that information is in the EHR, it might be possible to find indications in the medical notes that could identify these individuals. “There might not be. I think we have to be prepared for that,” Morris said. Lagu added that it is important to ask about accommodations too.

Warren, who said she has dealt with physician bias throughout her journey of applying for disability, had a note in her chart from her neurologist that contradicted what he told her. She appealed and was told that changing the EHR was up to her neurologist, and he would change his notes. However, her EHR noted her appeal and that, together with a wealth of other information she supplied to Social Security, led to her disability determination being approved.

Lagu said that there are certain health systems that do better at avoiding bias toward people with disabilities, including rehabilitation hospitals and stroke rehabilitation facilities. Children’s hospitals have advanced methods of dealing with people with developmental disabilities. These systems could provide valuable lessons about best practices, she said. Vydiswaran commented that research has shown the doctor–patient dyad has a significant role in determining whether information in the EHR is biased.

Amy J. Houtrow, noting that medical students and residents are now being trained not to mention someone’s race or ethnicity in their notes, asked the panelists for their thoughts about how to appropriately include information in the EHR in a manner that best serves people with disabilities and does not further perpetuate discrimination against them, especially for those with an invisible disability that might not be readily apparent at a first encounter. Morris replied that even if clinicians are being instructed not to include that information in their notes, it is still in the EHR as part of the demographic information. To her, disability status needs to be in a prominent place in the EHR to ensure the necessary accommodations are in place for a person’s clinical encounter. Mahmoudi noted that when he and his colleagues contacted individuals in wheelchairs because of a spinal cord injury, the individuals did not want to be labeled as someone with a disability because they felt they were not disabled. Instead, they wanted to be identified as a person with a spinal cord injury.

ABOVE ALL, ASK

Morris recounted the findings of a study she and her collaborators conducted in which they had medical students come into a standardized clinical scenario after randomizing them to either have or not have a person with a disability as their patient. The student noticed the disability but did not ask the patient about it because they assumed the patient would be ashamed of their disability and bringing it up would make them feel bad. However, what she has heard repeatedly from patients with disabilities is that the best thing a provider can do is ask, about both the disability and any accommodations they need to have a more productive appointment. Lagu noted that providers now ask about the pronouns people use, so in the same vein, they should ask about disabilities. She added there are web pages with information on basic disability etiquette, and that all clinicians would benefit from learning basic disability etiquette so they could speak to people with disabilities in the way they want to be addressed. Mahmoudi agreed there is a need for more education for all medical professionals on how to talk about sensitive subjects with their patients, and Lagu said the accreditation bodies that oversee medical education need to take this up as an important issue.

IMPROVING THE DISABILITY DETERMINATION PROCESS

Lagu raised the issue of how SSA could improve the disability determination process to make it more accessible to the community it serves. She said things she took away from the workshop were the importance of collecting data and the need to educate clinicians on how to document their patients’ disabilities in the EHR. She applauded Yvonne M. Perret for her approach of compiling a holistic picture of her patients that incorporates social determinants, medical issues, and other crises they face and putting that in the EHR so it can be considered in the determination process. “I just wonder why there are not more Yvonnes out there and why our system is not supporting more people who do this for vulnerable patients,” said Lagu.

Adler-Milstein said what struck her is that everyone feels they provide a great deal of information and fill out many forms. “It is not like there are no opportunities to collect information from people,” she said. She wondered if there is a need to step back and look for the touch points at which to best collect the most information most efficiently in a manner that would engender trust. She noted that Medicaid enrollment is taking this approach to identify natural opportunities to catch people when they first move.

Vydiswaran said that educating future clinicians early in their training about documenting disabilities and social needs is important, but SSA should also recognize that biases exist and that they are probably playing a role in how clinicians are documenting disabilities in the medical record. Recognizing that and looking at the entire package of what patients are saying and significantly enhancing their voices in the process would help too. This is where training those making these determinations about bias would help.

Warren said learning to speak respectfully to individuals with disabilities would help improve what clinicians are entering into the EHR, and Morris highlighted the need to think about access to the disability determination process given the challenge people with disabilities have with getting to a disability determination hearing or appointment. She also noted that this process can hurt the patient–provider relationship because of the onerous amount of paperwork a provider must complete for their patient.

Amanda Price, from NIH, said one issue for individuals going through the disability determination process is the amount of outdated information in an EHR that follows a patient as the years go by and that, while no longer true, still affects both how other clinicians view the individual and how the disability determination process considers an application.

THE ROLE OF AI

Mahmoudi asked the panelists for their thoughts on the use of AI in the health care system. Vydiswaran replied that AI is not a solution, just a means to an end. In his view, AI should not be used to create a persona of what a patient is, and that persona should not be put in the EHR to serve as a summary of what the health system thinks about the patient. He noted Warren’s issue with contesting a statement in her EHR from her neurologist and wondered if the situation would be the same for an AI-generated statement. “I would caution against using AI as a replacement for what the health system thinks about the patient,” said Vydiswaran. It would be important, too, to use the appropriate training data to avoid biased results from an AI-powered process.

Adler-Milstein said AI is not ready to make disability determinations, but it could help humans be more efficient so they have more time to have the necessary discussions. One thing AI can do well is access a much larger pool of information to see relevant signals in the EHR that are now overlooked, though as Price noted, there can be misleading and outdated information in the EHR. Joy Amaryllis Johnson added that AI might be a problem for a community that does not know about it because her clients are suspicious of new technology.

Jonathan Platt commented that people are now aware that AI can be biased, and he wondered if that knowledge can be used to train clinicians to use language that is less biased. Kenrick Cato said that AI does a good job of capturing the demographic-level bias in the workflow of health care, so AI systems today will give answers that are biased. Adler-Milstein said the question for AI systems is what the acceptable level of precision is versus the efficiency gains that might result from using an AI model to extract information from the EHR.

THE IMPORTANCE OF TRUST

Speaking of trust, Adler-Milstein, responding to Johnson’s question about the role of insurance companies in this process, said she is inclined to say that insurers have a significant role to play in solving the problems raised at the workshop, but they are the least trusted entity in the health care system, according to consumers. That is why she has been hesitant to involve insurers.

Morris said there is not enough conversation about the mistrust the disability community has in the medical establishment. “There is a long history of abuses in the United States against people with disabilities, especially in medicine, and we have not recognized that and acknowledged how entering into the health care system can be retraumatizing for these individuals,” she said. “We need to think about building our trust with our community of individuals with disabilities.” As a final comment to conclude the workshop, Houtrow said the big message from the discussions is to include people with disabilities at all steps of the process.