A primary care provider’s perspective on health tech
Want to make your well-meaning primary care doc chuckle at your checkup? Tell them about how the latest app, electronic medical record software, or risk scoring algorithm will simultaneously lower healthcare costs, improve health outcomes, and improve patient experience. As primary care physicians ourselves, we have PTSD from learning about new technologies that claim to solve healthcare’s biggest problems, yet somehow fail to engage with the kernel of its operating system: the financially-entrenched workflows that constrain what primary care can be or how people access it.
The challenge of building modern healthcare technology is to communicate with the existing healthcare system we are a part of. We need to meet patients, providers, and health plans where they are, because that is how we can catalyze change.
A technology stack for community-based care teams
At Waymark, most of our tech is focused on enabling community-based care teams to know whom to reach out to, what to recommend, and what to expect next in the care pathway. We train and deploy local teams within communities to expand the capabilities of primary care providers serving Medicaid patients. Our care teams use technologies that enable 1:1 SMS texting; app-free video visits that use low bandwidth for limited data plans; more traditional secure email, telephone, and in-person visits; and care management workflows. And while customizing and deploying those technologies is not easy, it is 100x easier (and cheaper) to stand them up today than it was a decade ago, due to advances in technology.
We do not need to reinvent HIPAA compliant software. Rather, we are focused on technologies that enable our teams to effectively care for patients in the communities we serve, allow a new health worker to rapidly gain the pattern-matching skills of someone with years of experience, and help us efficiently scale our services to patients covered by Medicaid across the country.
Learning what works requires moving beyond healthcare data
Take a routine workflow our care teams experience: should a patient with high blood pressure have a conversation today with a pharmacist to focus on how to titrate their prescriptions, a video visit with a social worker to help enroll in a local healthy food program (such as the Vouchers 4 Veggies program we’ve published several studies about), or an in-person visit from a community health worker to talk about how to use a home blood pressure monitor?
We’ve focused our healthcare technology and data science research over the past decade–research that ultimately contributed to the formation of Waymark–on problems like these that are specifically encountered by primary care providers working among underserved patients. Many healthcare entities want to deliver the right intervention, at the right time, for a patient. Unfortunately, applying modern machine learning methods–particularly reinforcement learning, which trains software to suggest the best action to take at a given time–is usually ineffective when applied only to traditional electronic health records or claims data. Health records and insurance claims are created for financial reimbursement purposes, not to identify whether a text from a pharmacist, or a phone call from a social worker, or a door knock from a community health worker will be most effective for engaging a patient; whether that patient will benefit from that engagement; and what to do when a patient has been engaged.
Our approach at Waymark is to enrich traditional medical records and insurance claims data with new information generated from our own interactions with patients–information that keeps track of which patients respond best to different staff members, to different forms of outreach (from SMS to in-person visits, depending on context, location, and patient personas), and how well patient health outcomes improved (or not) after intervention. By adding information on social determinants of health, we have been able to improve how well we can identify “rising risk” populations who are not yet “high cost claimants” but are on the pathway to potentially experiencing healthcare catastrophes. Many of our conversations with patients will be dictated by the patient’s circumstances during that particular day, but also by numerous other factors–like whether they’ve availed of unemployment benefits–a factor we’ve found to be important for behavioral health management.
Keeping technology hidden in our workflows
Problems like social determinants of health or financial misalignment between fee-for-service payments and patient needs are not solved simply by creating richer datasets. Strong evidence supports our focus on deploying technology that is hidden behind a human-centered service offering: to train and deploy local teams within communities to augment the capabilities of primary care providers, paid through value-based (capitated) contracts with health plans that seek to meet state quality metrics for their funding. Our local teams include community health workers who we train and provide salaries, benefits, and equity in our company. These community health workers are supported by licensed clinical social workers for behavioral healthcare delivery (using the Collaborative Care Model) and referrals to social services; by advanced practice pharmacists who can titrate medications, address prior authorizations, and provide patient support for medication adherence; and by staff members embedded into primary care providers’ offices to work within the providers’ existing medical record systems, coordinate referrals, and communicate between the brick-and-mortar clinics and the rest of our team who can visit patients at home (if there is a home) or community sites like food banks, churches, or homeless shelters.
Addressing bias in machine learning
As machine learning approaches are subject to reinforcing biases in our datasets–such as confusing low access to healthcare among Black Americans for lower risk of a healthcare event–we have chosen to adopt measures to keep our algorithms in check and proactively develop de-biasing algorithms. One approach we’ve taken is the most obvious: to increase the gathering of information from under-studied populations who are typically excluded from research or have disproportionately low access to healthcare. For example, when we studied current standard algorithms for heart disease prevention services, we found that Black Americans often received inaccurate estimates of their risk for a heart attack or stroke. Only by collecting and synthesizing large datasets from Black Americans, and developing a data science approach that avoided over-fitting algorithms to White Americans’ data, were we able to correct the problem and deploy a new alternative risk score.
Our motivation in conducting our technology and data science work is to move beyond the use of machine learning methods that just predict the “risk” of a bad outcome to actually predicting which intervention is most beneficial for a patient. As we presented at the National Academy of Medicine and more recently at NeurIPS (the major international machine learning conference), we have found that new algorithms can be created that point to actions: what our staff should do about such risk, and which of many options should be chosen by those staff members to lower the risk of a bad outcome for a patient. As we published previously (in work that won one of three prizes from The New England Journal of Medicine’s SPRINT Data Challenge), our work on these so-called “heterogeneous treatment effect” algorithms can help us move away from one-size-fits-all workflows. These algorithms help us refocus not just on whether a patient will be readmitted to the hospital or experience a heart attack, and instead on whether the best choice for that patient is a new medication, a home nutrition support service, or some other action that may be seen as a reasonable alternative. In this setting, our data science approach can help identify similar patients and compute multiple alternative risks and benefits based on a huge volume of studies that can be synthesized for our teams into personalized risk/benefit dashboards for discussion with patients.
As we continue to address healthcare technology and data science problems, we hope to collaborate with humble, passionate, and smart people who care deeply about improving the experience of Medicaid patients. If these technology or data science frameworks are appealing to you–or, better yet, if you feel that you have a better approach we should learn from–please let us know and consider joining our team.