The Background: Pioneering Virtual Health Care
Our client, a virtual care center — the only facility of its kind — was established to change lives and the health care system. The virtual care center brings care directly to patients by providing remote delivery of key aspects of health care, including monitoring vital signs, consultation with specialists, and prescription evaluation. Its cutting-edge telehealth program provided each patient with a set of telehealth monitoring equipment (i.e., digital scales, blood pressure cuff). The telehealth monitoring equipment is bluetooth enabled and connects patients to the hospital in real-time.
Patients meet regularly (sometimes as often as daily depending on their level of health) with their healthcare provider via a video call. Holding frequent virtual checkup meetings coupled with the daily tracking of their vital signs, reduces the need for them to visit a healthcare professional’s office. These lower barriers to health-improving activities increase the likelihood of patient compliance and result in better health and higher quality of life.
Our Challenge: Making Sense of Data
The stakes for the virtual care center are high and it faces complex challenges to deliver effective care to its patients. Enter 1904labs. We partnered with this client on a software and data engineering project designed to help streamline data flows and work processes. With the vast amount of person-specific data that the client had from their virtual monitors, they needed a better way to understand it and use it to optimize care being delivered. Our focus was to help them identify hidden signals within vast data sets, and use them to improve patient outcomes.
We first evaluated available data, gained a deep understanding of how the client organization impacts patients’ health, considered possible solutions and, ultimately, arrived on an initial goal for our efforts: create a predictive model for hospitalizations expected to occur within seven days.
This goal was important, measurable, and likely to be solvable with the data sets available. Important in that it could help reduce hospitalizations. Measurable because it is known when a patient is hospitalized. And solvable, because the data sets are likely to include leading indicators of hospitalizations, due to the breadth of metrics recorded.
Our Process: Highly Iterative
1904labs went to work. Over several months of continuous feedback in client check-ins every two weeks, our decision science team built and delivered various models, which led the client increasingly closer to an answer. Close, but not quite. Iteration after iteration seemed to be almost there. However, due to data limitations, these intermediate solutions were not instilling confidence that implementation would be able to impact patient health meaningfully. After a few rounds of incremental improvement, the team stopped to take stock. Following our HCDAgile philosophy, we kept the focus on the client’s ultimate goal: better patient outcomes. While the models we had been delivering may have been able to predict hospitalizations in seven days or fewer, there was a revelation that seven days likely wouldn’t be enough time for meaningful intervention.
And defining which specific patients were exactly predicted to be hospitalized proved to be challenging. Could there be a better solution?
What if, instead of creating a binary prediction of hospitalization within the next seven days, the model could predict a tendency for hospitalization? Might there be some classifications or groupings of similar-risk candidates? Groupings which would provide better insights into the problem?
HCDAgile at work again. With these adjustments, the team purposefully pivoted away from the “hospitalization in seven days” question to look at longer time frames, less absolute scale and more relative risk.
The Right Questions Lead to a Better Outcome
The team discovered that there was some clear clustering of likelihoods within the patient population. We were able to identify patients with similar characteristics and a low likelihood of hospitalization within the next three to six months (not days). These were classified as low risk. Others shared moderate or high-risk characteristics. Overall, virtually everyone could be identified within these three categories.
The opportunity was clear: Could the client use those classifications to tailor care plans to individual needs while still maintaining their high standards?
With this new target (not too far from the original target), the team went back to work. 1904labs developed the risk scoring mechanism and a tool for the client to run regularly and evaluate patient risk.
Now, instead of trying to predict actual hospitalization events, we created a method to quantify the relative likelihood of hospitalization in the next few months. Those with high likelihood should probably have additional check-ins, more in-depth conversations, and more fine-tuning of medications and therapies.
Those with lower likelihoods may be given more autonomy and trust that they would follow their provider’s orders, necessitating less frequent check-ins.
With this in mind, the client was able to refine certain standards of care and begin implementing these updated standards.
One change has been to move low risk patients from a 4x / month (once a week) virtual meeting to 2x / month (every two weeks) virtual meeting schedule. This frees up patient time without impacting their health, as well as giving the providers additional capacity to serve the high-risk population. Future developments will allow even greater refinement of care plans to ensure the highest quality outcomes possible for all patients.
Effective Pivot Created the Opportunity for Success
The implementation is still ongoing and the initial results are extremely positive. In our client’s words: “1904labs was a true partner. Thanks to their work, we have a predictive tool that will give us more opportunities to deliver high-quality care and help us achieve our vision of healthy patients and targeted interventions.”
The essential target, of improving patient health, has clearly been met. It was catalyzed by recognizing the need to change the direction in the middle of the project. An attempt to improve on the data model with limited regard for its end use and implementation would not have led to the positive outcome we achieved for the client. The project was rewarding on many levels and led to improved outcomes for our various stakeholders: patients, health care administrators and providers.
1904labs developed a 3-tier risk score for virtual patients and delivered a tool to update regularly. The client implemented differing care plans by risk tier, ensuring the right level of care applied to each patient.
- Python Data Science Stack (scikit-learn, pandas, NumPy, Matplotlib)
- H2O.ai Open Source Data Science Platform
- Google Cloud Platform
- Docker Container Platform
- Qlik Analytics Platform
- Human-Centered Design
Would you like to learn more?
A digital agriculture company needed a way to enable easy adoption and utilization of their data platform for the engineers that use it both internally and externally to access agricultural insights.
Working remotely the last 3 years, here are my lessons learned on time management, responsibility to yourself and your team, and how to avoid cabin fever.