An Agent-Based Explanation for SPMI Living Situation Changes

Problem:




A Severely and Persistently Mentally Ill (SPMI) patient is generally defined as someone with a diagnosis of Schizophrenia, Bipolar Disorder, or Major Depressive Disorder, and this group constitutes about 1.7% of the US population. The housing environment of the SPMI population in the United States has changed drastically over the past 60 years, most notably in the percentage of the population living in prison versus the percentage living in community based care and private residences.


Many of the changes have been successful. The number of SPMI patients living in the community increased, and although SPMI patients can still be institutionalized against their will, the likelihood of this happening is much lower than in the past. Yet a growing minority of people with severe illness are worse off because they are homeless or incarcerated.


In this case study, IBM Global Research and Otsuka Pharmaceuticals used an agent-based approach to model these remarkable swings in SPMI living situations over the second half of the 20th century. They wanted to better understand the situation and possible improvements that could be made.


Solution:




The important thing to note about SPMI housing options, is that some types of housing situations (jails, prisons and long-term hospitals) keep SPMI patients longer when they have mental health relapses, while the majority of housing situations seek to evict patients who misbehave in this same way. This fundamental difference was the key rule that the model sought to demonstrate as a major influencing force on the dramatic housing situation changes undergone by SPMIs over the second half of the 20th century.


SPMI Housing Situation Changes

SPMI Housing Situation Changes


The SPMI housing agent-based model contained components based upon the SPMI population’s real world housing environment.


First, the model included the “housing cycle”. When SPMIs left one housing situation, they were faced with a decision about where to live next. SPMI housing choices were thus a continuous cycle. Upon a patient’s arrival at the housing decision point, the patient was randomly placed into one of the seven housing types with certain probabilities. The housing types included:

  • Jail or prison 
  • Long term hospital 
  • Community hospital 
  • Assisted living facility
  • Shelter or transitional housing 
  • Private or subsidized residence 
  • Homeless

The housing types had baseline lengths of stay for SPMI patients. Long term hospitals tended to keep SPMI patients for longer than community hospital emergency rooms, etc.


The SPMI patient agents cycled between the decision point and actually residing in the location decided upon. Each patient had a “time to mental health crisis”, representing the various severity levels of mental illness. Time to crisis was uniformly distributed between 5 and 250 days for each SPMI agent in the model and was static for each patient for the entire model run. In this model, a mental health crisis was defined as a symptom flare-up that would make it clear to people interacting with the agent that the agent was not mentally stable. Examples included a psychotic episode, hallucinations, deep depression, etc. In the model animation, agents were closer to the color pure red when their time to crisis was closer to the minimum, and closer to the color pure green when it was closer to the maximum. The model was populated with 1,000 such SPMI patient agents.


The seven housing types responded to a SPMI patient’s mental health crisis differently. While a long-term hospital or a prison typically extended length of stay for a patient having a crisis, most other types of housing sought to evict or discharge such patients.

Healthcare Agent Based Simulation Model Structure

The housing cycle of SPMI agents

The model also included a ceiling placed by the United States legislation on the percentage of patients who can possibly live in long-term hospitals. The ceiling was gradually lowered each decade, which reflected what happened in real life. With each subsequent lower ceiling on the percentage of agents capable of being housed in long term hospitals, private housing added the majority of the formerly hospitalized agents with a small increase to the jail and prison populations.


The model metrics included the baseline lengths of stay and likelihood of going to each housing type upon reaching the housing decision point. Included were the date, percentage of the total population currently living in each housing situation, and the average time to crisis for the group currently living in each of the seven types of housing.


The model closely matched real world experience about the health of SPMI people living in each of the housing categories. The least healthy SPMI patients tended to pool in long-term hospitals and prisons. As long-term hospitals lost capacity, the worst SPMI agents increasingly moved into jails and prisons.


The reason for this phenomenon of the lowest time to crisis agents pooling in long-term hospitals and prisons had to do with the unique way in which these two housing types treated patients who have a crisis while housed in their custody. The less healthy the person, the stronger the reinforcing effect to stay in a long term hospital or prison and to leave all the other housing options. An ethical problem arises when legislation places a limit on the number of SPMI patients who can live in long term hospital, resulting in prison being the only housing environment which has both the capacity and a response of increased length of stay upon a mental health crisis event.


Outcome:




This agent-based model improved the understanding of Severely and Persistently Mentally Ill housing dynamics in multiple important ways.


First, the model showed that even a weak length of stay alteration due to SPMI crises produced a very strong effect. SPMI patients were in a state of flux until arriving in a place that will not let them leave.


Second, increasing patient time to crisis has a significant positive impact on the population’s housing makeup. In a second model run, the researchers increased patient time to crisis by 45 days in the year 2000 and examined what would happen by 2030. As expected, a healthier population (a population with higher time to crisis values) had weaker length of stay reinforcing effects and thus lowered the prison population rate. This insight had implications for policy makers looking to evaluate the return on investment of projects designed to improve the lives of the mentally ill. This agent-based simulation allowed them to understand that proactive governments can indeed lower prison rates by improving mental health in a geographic area.

Project presentation by Kyle Johnson from IBM Global Business Services:


Finally, the model raised ethical concerns for future mental health policies in the United States. The suggested ways to improve the situation were either increase long term hospital capacity or find a way to increase patient time to crisis. One effort to improve SPMI patient health in the information technology realm is currently underway with a partnership between Otsuka Pharmaceuticals and IBM. This project seeks to use care coordination information technology to help a geographic area’s many health care providers work together to efficiently treat SPMI patients.


More details on the project can be found in the accompanying article.

More Case Studies

  • Modeling Operations at Pharmaceutical Distribution Warehouses
    Cardinal Health, a billion dollar pharmaceutical distribution and logistics firm, manages multiple products from brand name pharmaceuticals and generic drugs to over the counter drugs, health & beauty items and their own private label. They face a multitude of typical distribution warehouse challenges that are further complicated by the nature of pharmaceutical products. Brian Heath, Director of Advanced Analytics at Cardinal Health, and an experienced user of AnyLogic software, employed agent based modeling to solve various business problems, saving Cardinal Health over $3 Million annually.
  • Utilizing Real-World Data and Multimethod Modeling to Battle Food Insecurity
    Issues such as whether households get enough food, how it is distributed within the household and whether that food fulfils the nutrition needs of all members of the household is an ongoing problem in developing countries. An International Civil Agency employed a Major Consulting Firm to analyze food insecurity in developing countries and cultivate a system that enables an economy to prepare for and halt possible food insecurity.
  • Simulation of Maternity Ward Operations
    This model simulates the maternity ward in a hospital currently under construction. The purpose of the model is to support discussions related to which resources, capacity, and work methods are required on the new ward. The project was carried out for Karolinska University Hospital in the Stockholm County, Sweden.
  • Evaluating Hospital Inpatient Care Capacity
    Stockholm County, Sweden was in the process of building a new, highly specialized hospital. The Health Administration of the county questioned whether they would get an acceptable level of care production with the current investments and reasoning concerning various operational and strategic issues. To find the answers, they used simulation modeling in AnyLogic.
  • Handling Total Care Need for Dialysis Patients
    The County of Stockholm (Sweden), like any country or region, experiences a continuous need to handle the healthcare necessities of various patient groups. Each group can be seen as a subpopulation, with its own distinctive traits, characteristics, and challenges. The discussed simulation project focused on the dialysis patients, a group who needs to visit caregiving facilities frequently.
  • Disaster Response Applications Using Agent-Based Modeling
    In an effort to find practical operational solutions for response to an unexpected crisis or natural disaster, Battelle, world’s largest, non-profit, independent R&D organization, needed to test the effectiveness of a 48 hour shelter-in-place order for an Improvised Nuclear Device scenario. The goal was to reduce radiation dosages received during an uncoordinated mass evacuation, by comparing immediate evacuation and shelter-in-place order.
  • Evaluating Healthcare Policies to Reduce Rates of Cesarean Delivery
    The challenge of reducing the cesarean delivery rate has been recognized by numerous researchers for years. For the first time, in research conducted for the Washington State, Alan Mills, FSA MAAA ND, a research actuary, and his colleagues reproduced this part of the United States healthcare system in a simulation model to allow the stakeholders, including health agencies, insurers, clinicians, and legislators, to test their assumptions on the model to find the right solutions.
  • Shaping Healthcare Policy Using Simulation
    An initiative by the Department of Mechanical and Industrial Engineering at the University of Toronto, the Centre for Research in Healthcare Engineering (CRHE), was in response to the immediate and compelling desire for efficiency and quality improvements in the Canadian healthcare system.
  • Simulating Bicycle Sharing System with the Help of AnyLogic System Dynamic Modeling
    Sustainable bicycle transportation systems have generated broad expectations and have recently attracted focus in the scope of urban planning by alleviating traffic congestion and reducing noise and air pollution. Such bike sharing systems are easy-to-use and generally promoted by the public administration, operating by advertisement companies, transportation operators and, and public organizations.
  • Outpatient Appointment Scheduling Using Discrete Event Simulation Modeling
    Indiana University Health Arnett Hospital, consisting of a full-service acute care hospital and a multispecialty clinic, faced poor statistics because the number of no-show patients (those who don’t show up for their scheduled appointments) rose dramatically to 30%. This was primarily connected to the fact that clinic schedules were driven by individual preferences of the medical staff, which led to increased variations in scheduling rules. To eliminate the problem, the client wanted to develop a scheduling methodology that would benefit the clinic, doctors, and patients.