King County enlists AI to reduce drug overdose deaths

Like many public health agencies across the US, the King County Medical Examiner’s Office tracks drug overdose deaths to target interventions for populations at risk and save lives.

In past years, reporting fatal drug overdoses has largely been a paper- and human-driven process in King County, which shares the information with state and federal agencies. This cumbersome process to gather and report fatal drug overdoses consumed large blocks of time as fentanyl- and other drug-related deaths have skyrocketed in the Seattle area in recent years.

Seattle, the county seat, reported nearly 600 drug overdose deaths in 2022, an increase of 72% from the previous year. In 2023, King County reported about 1,300 fatal drug overdoses.

Previously, the process of reporting a fatal drug overdose involved a medical examiner filing a written report in an unstructured, narrative form, with the examiner noting factors such as needles, pipes, pills, unidentified substances, or other evidence found at the scene. In many cases, additional documents are involved in the process, such as toxicology reports.

A group of human abstractors would then read the multipage documents, pull out the relevant information and re-enter it into forms used by the state and the US Centers for Disease Control and Prevention. The State Unintentional Drug Overdose Reporting System (SUDORS) at the CDC is a national database of fatal drug overdoses, used to better understand the circumstances, including the drugs involved and their origins.

The ultimate reason to collect and categorize the data is to target interventions to areas where there are large numbers of fatal drug overdoses, says Yang Martin, a forensic epidemiologist in the King County Medical Examiner’s Office. Intervention can take many forms, including educating specific populations about fentanyl, opening new drug treatment centers, and targeted distribution of naloxone, a medication used to reduce the effects of opioids.

A time-consuming process

That manual reporting process, however, took far too long, impacting King County’s ability to shape response. It took county employees about 10 to 12 hours to print and scan the documents associated with 200 fatal drug overdoses and another 4,000 hours to extract the data in the documents and fill in the reporting forms, with dozens of fields to populate.

With funding assistance from the CDC and the US Department of Health and Human Resources, the medical examiner’s office worked with the King County Department of Information Technology to develop a suite of tools, using natural language processing (NLP) and machine learning (ML) to automate the data extraction and form completion needed to report drug deaths.

“This project addresses the bottleneck issues in these programs, what’s preventing them from being more efficient,” Martin says.

Step by step

The new process involves three steps. Multipage incident and toxicology reports filed after a fatal drug overdose are first scanned so that information can be extracted using optical character recognition.

During the second phase, NLP and ML models created and trained by the King County Department of IT extract the pertinent information from these digitized reports. The ML models include classic ML and deep learning to predict category labels from the narrative text in reports.

King County’s NLP model was based on BERT, an advanced large language model (LLM). The IT department also used the Hugging Face online AI service and PyTorch, a Python framework for building deep learning models. Azure Databricks is also employed for data analytics as part of the solution.

“This project focuses on this very classic problem of taking a highly manual process, applying some readily available kind of classic automation techniques,” says Grace Preyapongpisan, director of data strategy and operations for the King County Department of IT. “And then what’s really cool is adding on this more experimental aspect with the machine learning capability.”

The machine learning piece of the project started as a proof of concept to add more automation to the report digitization efforts, she adds.

“We asked, ‘Is there an opportunity to use data that we already have to make this already efficient process even more efficient using some new tooling and new technology we have available to us?’” she says.

For the project’s third step, the IT department created two front-end user interfaces that enable state abstractors to autofill the reporting forms, while also allowing them to review the work done by NLP and machine learning. The first interface allows users to input a text narrative, with the NLP and ML models providing a predicted field label and a confidence estimate of the prediction.

The second interface focuses on automating data entry into the SUDORS database, using JavaScript code. The Washington State Department of Health and a dozen other counties and states across the US have expressed interest in using this interface, Martin says.

In the next phase of the project, the IT team plans to use secure AI models to predict category labels for over 1,000 data entry fields, compared to the initial pilot of seven fields. Another goal is to use the technology to report drug overdoes to the National Violent Death Reporting System (NVDRS).

The project has earned King County a 2024 CIO Award for IT leadership and innovation.

Targeted interventions

As all the pieces came together, King County saw a huge time savings from digitizing the narrative reports and automating the data entry. It now takes 20 minutes to digitize 200 overdose reports, instead of 10 to 12 hours, Martin says. And the time to fill out SUDORS and other forms has dropped by about 30%.

The decreased time to deal with the reports means faster responses when drug overdoses spike, Preyapongpisan says. Paired with geospatial mapping, the more current information enables public health professionals to focus on areas where drug overdoses are increasing.

“It allows us to improve health outcomes with preventative interventions,” she adds. “It allows us to really layer this data in real-time to make sure that not only are we placing the correct resources, but also targeting it very specifically.”