“Let’s start building a framework now to hold algorithms accountable for the long term. Let’s base it on evidence that the algorithms are legal, fair, and grounded in fact. And let’s keep evolving what those things mean depending on the context.” - Cathy O’Neil, Weapons of Math Destruction
For us at Delfina, the context is the worsening maternal health crisis in the United States: where rates of pregnancy-related hypertension have nearly doubled over the past decade. Where access to basic reproductive healthcare depends on your employer and state of residence. Where Black women have a three times higher rate of pregnancy-related death than white women. Where over one third of US counties have been classified as maternity care deserts. So, in the context of increasing adverse pregnancy outcomes, growing racial and geographic disparities in access to care, and the politicization of pregnancy and childbirth – how do we responsibly build algorithms?
The Delfina Data Science team has worked extensively with de-identified clinical data from Mayo Clinic Platform_Discover as part of the Mayo Clinic Platform_Accelerate program. As a result, we have created a dataset that tells the story of 30,000 pregnancy journeys. Or, more precisely, the documented pieces of each pregnancy journey. Here are my reflections on how to responsibly build algorithms from this rich data source.
.png)
Obstetricians are an integral part of the data science team. As a biostatistician, my first question is “where did this data come from?” To find the answer, I engage with the individuals responsible for data collection (e.g., care providers) and, if possible, witness data collection firsthand (e.g., shadow patient visits). Who better to engage in our data endeavor than the Mayo obstetricians who wrote the data story of each pregnancy journey? The obstetricians became our sanity checkers – “does it make sense that the prevalence of gestational diabetes is double that of the general US population?”, our cryptographers – “what does G2P1001 mean?”, and our cartographers – “where do you typically enter information on family medical history?” These insights have been crucial for our ability to construct each pregnancy journey.
.png)
The pregnancy care experience at Mayo Clinic may not be transportable to other settings. The pregnant population consists of individuals receiving care at a highly specialized medical system in Rochester, Minnesota. As a result, their prenatal, birth, and postnatal care experiences may differ from pregnant individuals seeking care at smaller community-based obstetric clinics throughout the United States. The population cared for at Mayo Clinic is also predominantly white and non-Hispanic such that racial and ethnic groups of smaller sizes will be underrepresented in our algorithms. This is problematic because it is well known that Black and Brown patients face unique challenges in our healthcare system, and capturing the experiences of these groups is essential to creating equitable AI algorithms.
The electronic health record is far from the whole story. Many patients have sought care at Mayo Clinic since they were born, giving us an extremely detailed history of prior diagnoses, medications, and care seeking behavior. Despite the richness of this dataset, there are several data gaps we encountered. First, lack of documentation on past health events or pre-existing conditions does not mean absence of an event or condition, as this information may be missing (e.g., patient received care outside of Mayo Clinic). Second, biometric data such as blood pressure are only captured during in-office visits – snapshots that will not accurately capture patterns or variation. Third, studies have shown that air pollution, weather, abortion restrictions, experiences of racism and sexism, food access, and neighborhood greenspace can impact pregnant patients’ care experiences and outcomes. None of these factors are currently captured in standard electronic health records.
Interacting with this rich data source has been a journey for our team. We have leaned on each other’s strengths to identify and address data gaps, enabling us to build algorithms that can be reliably deployed in our context. This experience also highlighted how Delfina Care is uniquely positioned to close these data gaps. Specifically, Delfina Care engages patients early in pregnancy, collects longitudinal biometric data via remote patient monitoring, and integrates external data sources that capture an individual’s social and physical environment. We are grateful for all we learned during Mayo Clinic Platform_Accelerate. We look forward to continuing to work with our colleagues at Mayo Clinic on data-driven approaches to improve maternal health outcomes.
