Supporting Proactive Behavioral Health Outreach Programs to Improve Mental and Behavioral Health Outcomes

Partner(s): Douglas County Kansas, Johnson County Kansas
Status: In progress
Team: Mihir Bhaskar, Abhilash Biswas, Jameson Carter, Vidisha Chowdhury, Shannon Dutchie, Rayid Ghani, Liliana Millán, and Sarah Nance. This work was initiated at the DSSG 2022 program (Team: Nuria Adell Raventós, Fabian Dablander,  Juan Luque, and Victoria Ritvo).
Github repo: https://github.com/dssg/dojo_mh_public

Summary
Behavioral health crises are a pernicious issue in the United States. Only in 2020, 12 million adults reported suicidal thoughts, and 1.2 million people attempted suicide resulting in 45,000 deaths by suicide. Also, 9.5 million people misused opioids, resulting in 107,000 lives lost; an increase of 15% from the year prior.

Currently, Douglas and Johnson’s counties have outreach programs to provide help to individuals in need, who are outreached based on different sources of referrals. We have developed a machine learning system that identifies individuals in need to support and at risk of having a potentially fatal outcome from a mental or behavioral health crisis to prioritize proactive outreach efforts.

Background
In the last 10 years, the number of suicides and overdose deaths in the US has been alarmingly increasing. Sadly, in most cases, this fatal outcome is the culmination of a series of events that are related to previous hospitalizations, emergency department visits, and suicidal ideations. Douglas Kansas and Johnson Kansas counties are no exceptions. Both counties have a special crisis response team for dealing with behavioral crises to help people in need. Their actual process is reactive, people will be outreach because they are in a crisis, having an interaction with the crisis line, having been referred by a family member, having an interaction with law enforcement, etc. None of these referrals look at the different interactions individuals have had with different systems in the past, not as isolated events, but as a history of each individual that can be relevant to identify if they could have a behavioral crisis with a potentially fatal outcome.

Scope
In partnership with the Douglas County Criminal Justice Coordinating Council and Johnson County Mental Health Center, we developed a machine learning model that assesses the risk of an individual living or commuting in both counties of having a potentially fatal outcome from a behavioral health crisis in the following months. Also, the model prioritizes who should be outreached from the crisis response teams of each county taking into consideration their own constraints in resources. This solution enables the crisis team in each county to outreach people in need in a preventive way, shifting away from reacting, and alleviating the burden on other healthcare system resources that are shared with other types of medical emergencies.

Data
The model used data from both counties -when is available- related to different interactions that their citizens have with different systems.

  • Ambulance runs in each county with information from paramedics that attended the run.
  • Emergency room visits with information about the medical examiner.
  • Criminal justice records
  • Medical examiner records
  • Police department records
  • Crisis line records

Analysis
The data-driven early warning system (EWS) takes all available data to identify events that each individual has had in the past interacting with different health and law enforcement systems in both counties, and uses machine learning methods to detect patterns that precede individuals at risk of having a behavioral health crisis. The model took into consideration different acuities and outcomes of behavioral health crises and focuses on those that could have a potentially fatal outcome in the next 6 months which includes deaths by suicide or overdose, suicide attempts, and overdose attempts. While the model was trained on data from both counties, it generates output tailored to each county’s resource constraints, prioritizing interventions for those at the highest risk.

  • Expanding the labels to have different levels of acuities:  “Very high Acuity”, “High Acuity”, “Medium Acuity”.
  • Expanding the time horizon to 1, 3, 6, and 12 months
  • Generate a risk profile that can summarize the risk of an individual with the different levels of acuities and time horizons.
  • Augment relevant information on the journey of the individuals at risk by adding to the risk profile the recent events a person has had.