This article was originally published in STAT+ on December 16, 2024. View the full article here.
When I worked as a social worker at a community clinic in Detroit, I often saw James (not his real name), a patient receiving Medicaid benefits who was grappling with chronic substance use. Every appointment revealed a deeper, more disturbing layer of systemic disarray impacting James’ life and care: fragmented health records, conflicting medication lists, incomplete therapy session notes, and scattered commentary on detox treatments. This disorganized care record frequently delayed his treatment, given how long it took me to piece together his medical history and, as a result, how little time was left to address his immediate care needs.
This experience was by no means abnormal. Many patients came to me only after an avoidable visit to the hospital. Access to proactive preventive care, however, is hard to come by when care records are fragmented and disorganized or when patients like James’ chaotic lives make keeping regular primary care appointments difficult, even impossible, without adequate support.
It was this desire to solve the data discoordination problem in Medicaid that led me to pivot from social work to data science. Our research, recently published in the American Journal of Managed Care, demonstrates that 39% of acute care visits among Medicaid recipients are for nonemergent conditions, indicating that many low-income patients often don’t receive the proactive outreach and early interventions they need to manage their health and stay out of the hospital or emergency department.
What makes this data even more troubling is that the technology exists to identify and engage these patients — a study published in Nature Scientific Reports in January 2024 shows new machine learning algorithms can predict avoidable acute care utilization with greater than 90% accuracy. These algorithms, calibrated to help providers identify patients at rising risk for acute health events, are proven to reduce care costs and enhance patient quality of life –– so why aren’t more Medicaid programs using tools like these to keep their beneficiaries healthy and out of hospital/ED settings?
For starters, many underserved communities quite reasonably associate the technology industry with mistreatment and systemic inequalities. The poor reputation of technology companies — often perceived as filled with billionaire tech bros who violate privacy practices and social norms with impunity — goes a long way to distance technology innovations from historically underserved people and those in charge of managing their care. The net effect of the bad technology of prior years has rendered many Medicaid providers and their patients skeptical of newer technology that may actually streamline workflows and improve outcomes.
Additionally, because these tools are not tailored to meet the specific social and clinical needs of low-income patients, adoption is also a major barrier. In my experience and in those of my peers, patients receiving Medicaid benefits often lack stable access to modern technology, reliable phone service or the internet, which exacerbates the digital divide and limits the potential impact of patient-facing AI-based solutions.
On the other side of the exam table, healthcare providers serving Medicaid patients frequently operate within under-resourced environments (such as in Federally Qualified Health Centers) and lack the necessary infrastructure to support advanced technological implementations.
Finally, I’ve heard that many tech tools touted as being helpful for providers have instead created additional layers of bureaucracy, which has, in turn, created resistance to new tools that are meant to streamline tasks. The incentives to adopt new technologies via traditional payment models (like fee-for-service, for example) are limited — research demonstrates that these approaches don’t incentivize proactive, preventative patient care. When you add that to the historical barriers presented to technological adoption, the answer to the question of, “We have these solutions, so why aren’t we using them?” becomes clearer.
Addressing these barriers requires a participatory approach to software design — one that actively involves and seeks input from patients and the people who serve them: PCPs, community health workers (CHWs), pharmacists, therapists, care coordinators, and community organizations. Involving patients receiving Medicaid and the care workers who serve them in the AI design process ensures these tools are relevant and sensitive to the unique needs of these populations. This approach fosters trust and acceptance of new technologies, as stakeholders feel their perspectives and needs are being considered.
At Waymark, we learned that CHWs had difficulty identifying patients who urgently needed assistance. In response, my data science team and I developed “rising risk” algorithms to help the team find patients who were likely to use the ED for primary care needs and hence would be likely to benefit from proactive outreach to connect them with early interventions. Our ML algorithm also accounts for social and environmental factors — for example, we combined wildfire smoke data with the list of patients who had asthma so that they could receive early inhaler refills before smoke came into their neighborhood.
Similarly, a persistent challenge in scaling community-based care programs is limited technology to enable remote, multidisciplinary care teams operating on the frontlines of the community. To solve this problem, we identified which workflows CHWs found most difficult to remember during their routine care, and had it automatically populate for them inside their software systems, saving time and energy having to search through manuals and documents. We also deployed a robot that can write cumbersome “prior authorization” letters to help us overcome obstacles to getting patients the correct medications.
By conducting user research and integrating feedback from patients and care workers, data science tools can be designed to meet the specific needs of people who have been historically excluded from the technology development process, enhancing the effectiveness and acceptance of these technologies. This participatory approach ensures that AI solutions are not only technologically advanced but also practically applicable and personally tailored, ultimately leading to better health outcomes for Medicaid recipients. By embracing participatory approaches to developing and refining data science algorithms, we can transform Medicaid into a model of innovation and excellence in health care delivery.
