Hurricane Harvey was a category 4 major hurricane that attacked Texas in August 2018. The company I worked for had a state health plan with over 500K members at that time. Though sitting behind a desk at Long Beach, CA, I was proud to have led a part of the hurricane relief efforts. This post is to document what we did, and what we learned from the incident. I particularly like this use case because the results had relatively large social impacts, without the development of sophisticated statistical models or machine learning algorithms.

Harvey was reckoned to be a destructive force at a very early stage. The local state health plan started their preparation days ahead of the arrival of the hurricane, but there were only such much they could do within a limited amount of time. On August 31, we were tasked to find ways to generate some useful information that could be helpful to the recovery mission, and lightweight on the workload in the meantime. At that time we already built a pretty impressive Hadoop data platform. For this particular assignment, we had multiple years’ medical claims, pharmacy claims, and case management data available in the same place, all cleaned up and could be cross-referenced with a set of primary keys. The case management data were especially useful here, and details on the data was included in another post.

The first thing we did was to come up with a list of impacted areas based on the weather forecast because we wanted to know the members that were or would be affected by the hurricane. Besides of including members whose residential addresses were in the “hurricane zone”, we also filtered for members who recently went to a provider or pharmacy that was located in the impacted area because such members may not have a valid permanent address on file, or may be homeless. This filter gave us a list of ~39K members who may have been affected by the hurricane in various ways.

The list of 39K members was simply too long for the state health plan to take any actions on, so we came up with several criteria/flag that could help assess the risk of the member in an emergency. The first flag was chronic conditions that needed regular care, such as diabetes, asthma, dialysis, or any kind of physical disabilities. The second was behavior health conditions that could lead to mental instability without proper medications. And the third was recent use of home care services or skilled nurse facilities (SNF), which indicated certain degree of physical disabilities or inability to taking care themselves by the patients. All these information was derived from medical claims with the diagnosis codes, procedure codes, place of services, or revenue codes. Some grouping mechanisms were used since it could be hard to define a certain condition with the onset/filing of only one episode. We also obtained a list of critical medications from CVS. This list ranked medications from high to low by that severe health consequences would occur if access or compliance was cut off for any amount of time, for days or weeks, or for some time. A list of members who were on critical medications was created then because we were afraid that these prescriptions could be flooded or absent when the members evacuated, and refills may not be immediately available. We also looked for members with an active/open management case, which was a good indicator that there was some health issue, particularly high risk pregnancy with the member. We came up with a list of ~300 members after applying these filters to the 39K impacted, we sent the list to local health plan, along with the conditions and medications each member had, so they could prepare accordingly before contacting the members.

The next day after coming in to the office, I was told about a webpage that Google set up with FEMA on which people could request help for evacuation. We matched the addresses and the cell phone numbers from the webpage with our database and identified ~200 members whose residencies were flooded and requested rescue. We sent this list to the state health plan, including details about these members, such as existing conditions, medications, or medical equipments. We also attached the notes the members left on the webpage that described their current situations, about the damages to their living places, and potential new contact information. We cross-referenced the two lists we sent to state health plan, and found quite a few matches. These members became our top priorities and staffs from the local office contacted them imminently for any urgent health care services.

Some thoughts: </br> This is a good example that a use case can be of high social impact with relatively low to moderate technical challenges. There were no sophisticated machine learning algorithms or statistical models involved, nor did we had time for the tuning of such things (the total turnaround time was less than 24 hours). Some business logics or clinical definitions were used to generate those flags, but all of them were already created during daily operations before the event. It took some time to build the data platform, and we were fortunate that it was mature enough at that time for such task. I was deeply moved by the moment when I found one of the high risk pregnant mom I identified requested rescue because she and her two daughters were trapped on the roof of their house. I realized suddenly that all the dots on my screen were some real people in distress, and my work might actually alleviate some pain from their life.