Predicting the Absence and Gaps in Recorded Blood Pressure Follow-up Among Patients With Hypertension
This study developed and validated a gradient-boosted tree model using EHR data to predict gaps in blood pressure follow-up for hypertensive patients, demonstrating moderate predictive performance and potential for targeted outreach while noting significant disparities in sensitivity across racial and ethnic groups and finding that contact patterns offered minimal added value for predicting cardiovascular hospitalization.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
High blood pressure is a silent force that, if left unchecked, can damage the heart and brain over time. Managing it is not a one-time event but a continuous cycle of checking, adjusting, and following up. When a patient's blood pressure spikes, doctors recommend a return visit within a month to see if the treatment is working. Yet, in the real world, many people do not return for that crucial check-up. They might feel fine, forget the appointment, or simply get lost in the shuffle of a busy healthcare system. The absence of a follow-up reading in a patient's medical record is a warning sign that care has stalled, but identifying who is likely to disappear from the system is difficult. Clinics have limited staff and time; they cannot call every single patient who missed a visit. The challenge is to find a way to predict, using the information already available in electronic health records, which patients are most likely to vanish from the system so that limited resources can be directed to those who need a nudge the most.
A team of researchers set out to build a tool to solve this specific problem. They looked at data from thousands of adults with high blood pressure who had recently recorded a dangerously high reading. Their goal was to create a model that could look at a patient's history—their age, their past visits, their medication list, and how long it had been since they last saw a doctor—and predict whether that patient would fail to return for a blood pressure check within the next 90 days. They tested this idea using a massive database of hospital records from a single medical center in Boston. The researchers fed their computer model 62 different pieces of information about each patient, ranging from their insurance type to how often they had visited the hospital in the past. They then watched to see if the model could correctly identify the people who actually disappeared from the records.
The results showed that the computer model was quite good at spotting the patients who were likely to be missed. When the researchers tested it on a group of patients it had never seen before, the model successfully identified about 35% of the people who had no recorded blood pressure check in the following 90 days. If a clinic used this model to prioritize their outreach efforts, focusing only on the top 20% of patients the model flagged as highest risk, they would find that nearly 60% of those contacted had indeed missed their follow-up. This is a significant improvement over random guessing, which would only catch about 35% of the missed patients. The most important clues the model used were simple: how long it had been since the patient's last visit and how many times they had visited the hospital in the past. Patients who had not been seen in a while and had a history of fewer visits were the ones most likely to be missing from the system.
However, the study also uncovered a troubling gap in how well the tool worked for different groups of people. While the model was accurate overall, it did not perform equally across all racial and ethnic groups. When the researchers applied the same rule to everyone, the model found about 37% of White patients who had missed their follow-up, but it only found about 21% of Black patients and 23% of Hispanic patients. This happened even though the model was generally good at predicting risk for everyone. The researchers found that this disparity was not just because fewer Black or Hispanic patients were missing their appointments; it was because the model was less sensitive to the specific patterns of care in those communities. This suggests that relying on historical visit data to predict future behavior can sometimes reinforce existing inequalities, as the model learns from patterns that may already be biased by how different groups access care.
The researchers also asked a second, related question: does the pattern of a patient's contact with the healthcare system tell us anything about their future risk of a heart attack, stroke, or heart failure? They looked at whether patients who had long gaps in their medical records were more likely to end up in the hospital for these serious conditions. Surprisingly, the answer was no. Having a long gap in recorded visits did not consistently predict a higher risk of cardiovascular hospitalization. In fact, the data showed that patients who had more frequent and sustained contact with the healthcare system were slightly more likely to be hospitalized for heart failure. The researchers explained this not as a sign that frequent visits cause illness, but rather that patients who are sicker or have more complex health problems naturally see their doctors more often. Their frequent visits are a sign of their underlying health struggles, not the cause of them. The study concluded that while tracking who is missing their blood pressure checks is a useful tool for clinics to improve follow-up, the pattern of those visits does not serve as a reliable crystal ball for predicting future heart attacks or strokes.
Ultimately, this work provides a practical way for healthcare systems to use their existing data to catch patients who are slipping through the cracks. It shows that a computer can look at a patient's history and say, "This person is likely to miss their next appointment," with a reasonable degree of accuracy. But it also serves as a cautionary tale that such tools must be used carefully. If a clinic uses the model without adjusting for its blind spots, it might end up calling more White patients than Black or Hispanic patients who are equally in need of help. The study does not prove that calling these patients will fix their blood pressure or save their lives; it only proves that the model can help clinics decide who to call. To truly know if this approach improves health, future studies would need to test whether acting on these predictions actually leads to better outcomes for everyone, regardless of their background.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.