How we can use data to address the opioid epidemic

President Nixon declared the war on drugs in 1971. Drug overdoses are now the leading cause of death for people under the age of 50, with opioids accounting for most of the devastation. Can data science help? Data can provide hints about underlying truths, but it is not immediately clear how it can be applied to the opioid crisis. Any solution that data could provide would have to be better than across-the-board solutions, such as more careful prescribing, or other approaches that have been tried, such as increased resources.

Exploring how data could be applied to the opioid epidemic was the goal of the recent Opioid Symposium and Code-a-Thon put on by the U.S. Department of Health & Human Services. Chief Technology Officer Bruce Greenstein and Chief Data Officer Dr. Mona Siddiqui and their people brought together 70 data sets to be examined. During the symposium, they discussed how getting departments to share data was difficult, but they found that when you frame it in terms of saving lives, people will do what needs to be done to get the data shared. Indeed, the tagline of the event was “Connecting data to save lives.”

Two themes emerged from the event with respect to how data science could mitigate the negative effects of opioids. The first theme was using data science to allocate scarce resources. If one can predict which places are going to have a spike in overdoses, health officials can get resources to those places ahead of time. These spikes can be predicted by looking at the distribution patterns of illegal drugs or by looking at comments on Twitter, or even just past patterns. For example, by noticing that one county generally spikes a few days after a neighboring county. The second theme was using data to balance the tradeoff between pain management and opioid addiction risk on an individual basis. This was the approach taken by our team.

We built a model to predict whether someone is likely to overdose or become dependent on opioids by applying machine learning to Medicaid claims and the National Drug Code Directory. This model indicates which features are predictive of opioid problems in general, and it also tells a healthcare provider which characteristics of a particular patient put him or her at risk of opioid addiction or overdose. The model enables healthcare providers to interactively evaluate different opioid risk scenarios for an individual patient by allowing them to alter the features of the patient that are changeable, such as the other medicines the patient is taking, and to see the predicted results. In this way, the provider can identify the optimal balance between pain management and risk of opioid dependence.

In summary, data is useful because it provides hints about underlying truths. The trick is to turn those truths into straightforward plans for action. We saw here that for the opioid epidemic, data can help to maximally apply resources and to balance competing needs. Let’s get more of these ideas implemented and operational, and let’s start turning the tide.

Thank you, HHS, for hosting a top-notch event!