Building and also experimenting v an agent-based design to study the population-level impact of CommunityRx, a clinic-based community resource referral intervention
Building and also experimenting through an agent-based design to examine the population-level impact of CommunityRx, a clinic-based community source referral treatment Stacy Tessler Lindau, Jennifer A. Makelarski, Chaitanya Kaligotla, Emily M. Abramsohn, David G. Beiser, Chiahung Chou, Nicholson Collier, Elbert S. Huang, Charles M. Macal, Jonathan Ozik
*

*

*

*

*

CommunityRx (CRx), one information technology intervention, gives patients v a personalized list of healthful community resources (HealtheRx). In repeated clinical studies, nearly half of those who obtained clinical “doses” that the HealtheRx shared their information with others (“social doses”). Clinical trial architecture cannot totally capture the affect of information diffusion, which deserve to act together a force multiplier because that the intervention. Furthermore, trial and error is required to understand how intervention shipment can optimize social spread under differing circumstances. Come study info diffusion from CRx under varying conditions, we constructed an agent-based model (ABM). This study defines the design building procedure and illustrates exactly how an ABM offers insight around information diffusion through in silico experimentation. To build the ABM, we built a synthetic populace (“agents”) using publicly-available data sources. Utilizing clinical psychological data, we emerged empirically-informed procedures simulating agent activities, source knowledge evolution and also information sharing. Utilizing RepastHPC and also chiSIM software, we replicated the treatment in silico, simulated info diffusion processes, and generated emergent info diffusion networks. The CRx ABM was calibrated using empirical data come replicate the CRx intervention in silico. We offered the ABM come quantify info spread via society versus clinical dosing then carried out information diffusion experiments, comparing the society dosing impact of the treatment when ceded by physicians, registered nurses or clinical clerks. The synthetic population (N = 802,191) exhibited diverse behavioral characteristics, including activity and knowledge development patterns. In silico delivery of the intervention was replicated v high fidelity. Large-scale info diffusion networks emerged among agents exchanging source information. Varying the propensity for details exchange resulted in networks with different topological characteristics. Community source information spread out via social dosing was virtually 4 fold that from clinical dosing alone and did not vary by shipment mode. This study, using CRx as an example, displayed the procedure of building and also experimenting through an ABM to study information diffusion from, and the population-level influence of, a clinical information-based intervention. If the focus of the CRx ABM is come recreate the CRx intervention in silico, the general process of model building, and also computational testing presented is generalizable come other massive ABMs of info diffusion.

You are watching: The lowest resource utilization is usually found in


CommunityRx (CRx) is a clinic-based intervention that gives patients v information about community sources for health-maintenance and also promotion. Prior work uncovered that nearly half of people exposed to CRx re-superstructure their source information with others. This examine describes construction of and also experimentation with an agent-based model (ABM) to research the potential affect of CRx and also other health information interventions on the broader community via social spread out or “dosing” from world directly exposed to the intervention. We display how we incorporated clinical trial, demographic and epidemiologic data and also expert informant insights to develop and assign actions to a synthetic study populace (agents) in the population. Using CRx clinical trial data, us then ceded the treatment to these agents and also simulated info spread. We describe in silico testing to show insights about information spread produced by the ABM that complement clinical attempt findings. This examine shows how data indigenous individual-level clinical and populace studies have the right to be offered to develop a computational activities to evaluate the wider impact the a health information intervention. In addition to inspiring integration of individual-level and also systems science approaches to the research of wellness information interventions, this study permits peer review to educate model iteration and experimentation.


Citation: Lindau ST, Makelarski JA, Kaligotla C, Abramsohn EM, Beiser DG, Chou C, et al. (2021) Building and also experimenting with an agent-based model to research the population-level impact of CommunityRx, a clinic-based community source referral intervention. ivorycrimestory.com Comput Biol 17(10): e1009471. Https://doi.org/10.1371/journal.pcbi.1009471

Editor: Benjamin Muir Althouse, institute for disease Modeling, united STATES

Received: November 20, 2020; Accepted: September 23, 2021; Published: October 25, 2021

Data Availability: The design code and also the workflow code offered to carry out the parameter an are characterization experiments room publicly available at https://github.com/jozik/community-rx).

Funding: Research report in this publication was sustained by the national Institute top top Aging that the nationwide Institutes of wellness R01AG047869 (S.T.L, J.A.M., E.M.A., D.G.B., C.C., N.C., E.S.H., C.K., C.M.M., J.O.), R01AG064949 (S.T.L., J.A.M., E.M.A., E.S.H.), and also K24AG069080 (E.S.H.), the nationwide Institute of minority Health and Health Disparities R01MD012630 (S.T.L., J.A.M., E.M.A.), the national Institute of Diabetes and Digestive and Kidney illness P30DK092949 (E.S.H.) and also the national Heart, Lung and Blood academy K23HL145090-01 (E.T.). The complete amount the the project costs were financed v federal money. The funders had actually no function in study design, data collection and analysis, decision to publish, or preparation of the manuscript. The content is exclusively the duty of the authors and also does not necessarily stand for the main views the the nationwide Institutes that Health.

Competing interests: I have read the journal’s policy and also the authors of this manuscript have the following contending interests: Under the regards to prior room of Health and Human Services, Centers for Medicare & medical allowance Services capital (1C1CMS330997), innovators were expected to construct a sustainable organization model to continue and also support the design that to be tested after ~ award funding ended. Dr. Stacy Lindau is the founder and owner that a recently acquired social affect company NowPow, LLC. No the college of Chicago nor college of Chicago medicine is endorsing or promoting any type of NowPow/MAPSCorps reality or its business, products, or services. No other authors the this manuscript have competing interests.


Introduction

Over the last five years, through the shift to value-based care, community resource referral technologies have actually been embraced by many U.S. Health systems.<1–3> these platforms, which aim to attach patients to community-based sources for health-related socioeconomic requirements (e.g. Food and housing support, transportation) serve to advancement health systems’ populace health treatment delivery models and also their initiatives to minimize inequities as result of social and also structural determinants of health and disease.<4,5> mainstream data sources about healthful community resources are frequently outdated, idiosyncratic, analog, absence eligibility information and also must be accessed outside the usual clinical informatics workflow.<3,6> together a result, clinician and patient efforts to exploit these resources to promote wellness or control with illness are regularly frustrating and inefficient.<3> Sustainability of these interventions in healthcare practice relies not only on their seamless integration v usual clinical workflows, but also on your demonstrated influence on separation, personal, instance and populace health-related outcomes and also costs.

In comparison to medical drug and an equipment interventions, info interventions delivered by a clinician come a patient can easily spread to other neighborhood members with social networks. Evidence suggests that society networks can potentiate the influence or credibility of health-related info <7,8> and that patients typically view information acquired from a clinician as very trustworthy.<9> Therefore, info spread v social networks from an intervention yielded in the clinical setting has the intriguing potential to meaningfully affect people in the community beyond those who directly received the intervention. Alternatively, an unfavorable information or misinformation spread might be harmful come individual and also public health and wellness efforts. However, classic clinical attempt designs, which commonly assess the affect of an intervention on the separation, personal, instance patient, execute not capture the dynamic, multi-level affect of information-based health and wellness interventions.

CommunityRx is a scalable, evidence-based information technology intervention designed together with a broad diversity of stakeholders, including residents, on Chicago’s south Side come promote population health by connecting patients to health-promoting ar resources (“resources”).<10,11> CommunityRx was created, in part, as result of systematic deficiencies in tendency data resources which under-represent health-promoting businesses and organizations in greater poverty communities.<6> these deficiencies impede the initiatives of healthcare experts to meaningfully and also equitably execute ~ above clinical best practice accuse that suggest a wide variety of community resources for various health conditions.

The primary mode whereby CommunityRx disseminates source information is via the “HealtheRx,” a published list of regional resources personalized to the patient’s age, gender, house address, health and wellness conditions, and also preferred language (Fig 1). A HealtheRx is auto-generated at the allude of medical treatment using software program algorithms incorporated with the digital medical record (EMR) to enhance individuals come an indicated set of neighborhood resources (Fig 2). With the press of a digital button in the EMR by a clinician (e.g. Physician, nurse or clerk), extant data in the patient’s chart (e.g., age, gender, residence address, wanted language, ICD-9/10 codes) space consumed right into the CommunityRx algorithm to create a HealtheRx.


Indicated resources for each condition and also status are informed by best clinical practice guidelines, skilled opinion and community member input around self-care activities for an ext than 30 usual health problems (e.g., diabetes, hypertension, obesity), social conditions (e.g. Food insecurity, real estate instability, residential violence) and statuses (e.g., newborn, adolescent, pregnant). Every HealtheRx contains resources to address an individual’s simple needs (e.g., food and also housing), physical and also mental well-being (e.g. Fitness, nutrition, yoga), condition self-management (e.g. Cigarette smoking cessation, load loss) and also caregiving (e.g. Cancer support, respite care, residence hospice) needs. Clinicians deliver the HealtheRx to patients at discharge from a clinical visit, including emergency and also hospital care. An independent quasi-experimental examine of CRx comparing world who received one or an ext HealtheRx to matched controls discovered a far-reaching decrease in hospitalization amongst Medicare beneficiaries and also a decrease in emergency room utilization amongst Medicaid beneficiaries.<10,12,13> expense savings end the short term were modest come neutral. Delivery of the intervention by a salesperson is much less costly than shipment by a nurse or physician. Experimentation is required to notify the lowest price delivery technique while maintaining details spread come the wider population.

Inspired by the Kilbridge et al. Prescription management process for drugs, community resource prescribing calls for several steps.<11,14> The community source prescribing “decision-making” process is initiated throughout the clinical encounter, sustained by the software algorithms described above. The personalized prescription is auto-generated native extant data in the patient’s digital medical record (EMR) (e.g. Age, gender, home address, preferred language, trouble list) plus any new data (e.g. A new diagnosis or society risk) entered throughout the visit. When the HealtheRx is generated and printed, that is “fulfilled.” Fulfillment starts with “evaluation” as soon as a clinician or clinical staff member evaluate the HealtheRx v the patient and also is “dispensed” when the patience contacts or visits the community source provider. The community resource is “administered” once the patience uses, obtain or spend the recommended resources. Kilbridge’s medicine e-prescribing design does not contemplate the possibility of sharing or diffiusion (although prescription medicines are sometimes shared).<15> However, in the CRx iteration of the e-prescribing procedure model, we did anticipate the prescribed information can be shared. We thus designed our main research come assess because that sharing.

Systems science approaches are progressively being used to complement classic epidemiologic and also ecological studies,<18> although few clinical trials have actually incorporated a systems scientific research approach. Observational and also experimental studies can take place in tandem with, and inform, the building of computational models.<19–21> Agent-based modeling (ABM) is a systems science simulation methodology that captures emergent behavior among individuals (“agents”) together a duty of interactions in between agents and their environment. ABM is an especially useful for researching phenomena influenced by second-order effects resulting native stochastic and also network interaction (e.g. The population-level impact of one individual-level intervention where individuals connect with and impact each other’s behavior).<22> Additionally, ABM permits linkage the individual an option behavior come emergent population-level outcomes (like total clinic access time or other health maintenance activities). In contrast to other approaches (e.g. Differential equations modeling) these micro-to-macro level dynamics make ABMs an especially well suitable to study an informational intervention. ABM can additionally serve as an reliable adjunct for experimentation that is exterior the minimal scope the a clinical trial or as well time-consuming or otherwise costly to conduct in vivo.

ABM has been provided to research the spread of transmittable disease<23–25>, impacts of environmental exposures <26,27> and also health-related interventions and policies.<28–36> Barbrook-Johnson et al. Offered ABM to examine the impacts of the tell ME intervention, an information-based treatment that deployed publicly health communication strategies to reduce the spread of influenza.<36> The tell ME ABM was designed come predict the effect of communication strategies top top influenza spread, but did no run simulations particularly to model info flow and dynamics. A summary of the validated phone call ME ABM to be presented together a to teach tool, to show how such models deserve to be offered to educate policy, even in the absence of finish data. Best practice dictates that the an initial step in any systems modeling advance is to produce the suitable model design and also then subject the design to evaluation including peer review.<37> We have actually previously explained in the systems science literature the technical style details that the ABM supplied in this study as well as the connected model validation.<38,39>

Informed by and extending this work, the purposes of this study are to: (1) describe the interdisciplinary an approach of design building and also data sources used to construct an ABM to study an information-based wellness intervention, and also (2) explain an in silico experiment to illustrate just how an ABM can generate insights to enhance a clinical trial of one information-based intervention. First, making use of the advance of the CRx ABM together a instance study, we define interdisciplinary model structure methods that can generalize to other information-based health care interventions. Next, to illustrate exactly how ABM experimentation can enhance findings created using timeless clinical psychological methods, we describe an in silico experiment come deepen our understanding of the info sharing findings from connected clinical trials. Come our knowledge, this is the an initial study to use ABM to study the population-level details diffusion dynamics result from a clinical information intervention.


1.0 reason for structure the CommunityRx ABM

The CRx clinical trials provide very first order evidence of information spread native the civilization exposed come the treatment (patients and also clinicians) to others in your social networks. The CRx ABM was designed come simulate and also experimentally probe, in silico, the flow and also spread that this info at scale, across the larger population under varying conditions. Spread out of community source information from patients to others in the ar is a preferable dynamic that could serve together a force multiplier because that CRx and also other information-based wellness interventions. Return we observed no empirical proof of negative information or misinformation spread in former studies, expertise the flow of negative or misinformation is likewise of interest in future instances the the model. Here, we describe the intersiciplinary model-building procedure and one intervention distribution experiment conducted in silico to notify future iterations and implementation that the intervention.


2.0 Interdisciplinary model-building process

The interdisciplinary procedure for building the CRx ABM associated several steps, educated by extant literature<21> and also illustrated in Fig 3. Investigators, including physician scientists, epidemiologists, biostatisticians and systems scientists from the college of Chicago organic Sciences division and the Decision and Infrastructure Sciences department and Consortium for progressed Science and Engineering at Argonne nationwide Laboratory, met on regular basis over the food of numerous years to: (1) come to a mutual understanding around and vocabulary to permit interdisciplinary examine of the CRx intervention, (2) identify and also integrate publicly use, clinical trial and also other data sources with the ABM drawing on complementary expertise (e.g. The equipment scientsts were knowledgeable with time usage data and also the rather were knowledgeable with publicly health and also trial data), (3) fill data gaps by desiging and also implementing new data collection strategies illustration on qualitative research study expertise of both groups and also survey research skills of the biomedical scientists, (4) describe and validate the model,<39> and (5) design and also implement in silico testing (a focus of this paper).


In addition to observational data native publicly available demographic, economic and also epidemiologic datasets<10,11,40,41> and also primary data collection from clinical trials, the CRx ABM likewise used data sourced from professional informants. Experienced informants—referred come as subject matter professionals in the systems science literature—are generally used in agent-based modeling to fill gaps in extant data.<42> among the skilled informants connected in developing the CRx ABM were human being who live in, mutual demographic attributes with and also had decades of experience serving and providing medical treatment for people in the target geography. Primary data collection, including two clinical trials (CRx-1 and CRx-2), occurred in parallel v the CRx ABM building, v each technique informing the various other (Fig 3).


3.0 overview of the CRx Agent-Based Model

Here, we administer an overview of the CRx ABM with focus on exactly how the ABM to be built, using the Grimm et. Al. Overview, architecture concepts and Details (ODD) an overview protocol.<43> comprehensive description that the model design, calibration and also validation, including all of the contents indicated because that a complete ODD, have been previously published.<38,39> Language offered verbatim from the Grimm et. Al. Summary ODD template is italicized and also bold in the sections below and also many section headers derive straight from the Grimm protocol. The model and workflow code supplied to implement the parameter an are characterization experiments room publicly obtainable at https://github.com/jozik/community-rx. A description of version processes that occur in every timestep is detailed in appendix A in S1 Text. The CRx ABM was applied in C++ making use of the Repast because that High-Performance computing ABM toolkit and also the Chicago Social communication Model (chiSIM) framework.<44–47> The CRx study was approved by the university of Chicago Institutional testimonial Board.


3.1 CRx ABM Purpose and also Questions

The in its entirety purpose the the CRx ABM is come (a) show the flow and spread of resource info from primary agents (those who got one or an ext clinical doses or HealtheRxs) to rather in the community and also (b) conduct experiments ~ above how distribution of the HealtheRx and also other variable problems can affect the flow and also spread of info to the community. Specifically, we space addressing the adhering to questions: (1) can we recreate, in silico, distribution of the CRx intervention, including generation of personalized HealtheRxs (“clinical dosing”)? (2) have the right to we recreate the basic dynamics of source information diffusion (“social dosing”) following in silico delivery of one intervention yielded to main agents at the suggest of clinical care? (3) how does variation in clinical parameters (e.g. That delivers the HealtheRx come the primary agent) influence information diffusion come the broader population?


3.2 CommunityRx ABM Patterns

To create that us have developed a working computational laboratory for experimentation, the CRx ABM building procedure involved a methodical examination the patterns at the individual and also system level that room salient come the function of the model. To think about our model realistic enough for its purpose, we provided the complying with critieria: (a) the demographic attributes of the synthetic populace would reflect those of the yes, really population, (b) the synthetic populace would exhibit diverse behavior characteristics, consisting of agent task and knowledge advancement patterns, (c) in silico distribution of the treatment would it is in replicated through high fidelity, (d) large-scale info diffusion networks would emerge amongst agents exchanging source information, (e) differing the propensity for info exchange would an outcome in networks with different topological characteristics. Check of these trends is gift in the results section that this paper.


3.3 CommunityRx ABM Entities

In general, the main materials of a population-based ABM encompass model entities and also behaviors. Entities include a synthetic populace of agents who statistically correspond to the population in the examine area and also an environment, which frequently refers come a physical environment. Actions include agent behaviors and also a collection of interactions for each agent, such as receiving or share information.<42> The CRx ABM includes the complying with entitiesa synthetic populace of agents and a geographical setting that together room statistically representative that the neighborhood under study (16 ZIP codes on Chicago’s southern Side, a mostly African American/Black demographic wherein the CRx intervention was created and studied). The state variables characterizing this entities are listed in Table A in S1 Text.

Using the publicly obtainable Synthetic Populations and also Ecosystems of the world (SPEW) 2016 dataset (http://stat.cmu.edu/~spew/resources/),<40> we built a synthetic population of world (agents) ages 16 years and also older (the HealtheRx for human being younger 보다 16 was typically given come an accompanying adult). The individual characteristics of the synthetic population and their distribution are presented in the outcomes section and Table B in S1 Text. Population characteristics were static.

Building the environment connected imbuing the version with the physical locations (“places”) that agents can occupy within the 16 ZIP password (106 mi2) CRx study. This an ar comprises a small less than half of Chicago’s geography. The physical setting included: (1) households, workplaces, and schools (data because that these areas were likewise obtained from the SPEW dataset); (2) health care sites (clinics) wherein agents might receive a HealtheRx, and also (3) health-promoting community-based organizations and business (places) wherein agents might go come use solutions (“resources” are particular places providing certain services ) prescribed top top the HealtheRx.

Primary model habits include (a) procedures of details sharing (how details flows and also spreads or diffuses from an individual agent to rather in the community) and also (b) decision habits (how agents choose to utilize resources as a role of information received). This processes and also behaviors room described listed below under “Important Processes.”


3.4 CommunityRx ABM spatial and also temporal resolution

The CRx ABM is spatially explicit. Each location in the synthetic environment, including health-promoting resources, are provided at an actual physical location with a solved latitude and also longitude in the 16 ZIP code study region. Each time step in the model represents one hour the simulated time and also simulations space run for a period of at least 4 weeks. We usually observed stable habits by mainly 3 of the simulation. Agents space assigned particular activity schedules because that every 24 hours, representing a simulated day. The level of the CRx ABM is in recreating the CRx treatment within the characterized spatial and also temporal resolution.

Structurally, the CRx ABM is a time-stepped task simulation in i beg your pardon agent states are update at each simulation time term (hourly), based on agent behaviors and interactions. Using a demographically matched, randomly selected daily task schedule, the simulation ingredient of the version determines one agent’s task for the current time step and moves every agent to the activity’s location. Co-located agents share info about details community resources based on the task being performed. Based on information received and retained through this process, each agent decides even if it is or not to perform certain health-promoting tasks (a subset the all possible activities) utilizing community-based sources that were prescribed to the agent by the CRx intervention. Thus, we deserve to derive population-level metrics of source information diffusion. Further, together a duty of information diffusion and also agent decisions about health-promoting activities, we can likewise derive population-level metrics of resource use. A summary of version processes that happen in every time action is noted in the complying with subsection and summarized in paragraph 3 of attachment A in S1 Text.


3.5.1 Agent activity behavior.

In addition to sociodemographic characteristics, agents were imbued with task schedules because that each simulated day. Agent task schedules were developed from the publicly accessible 2016 American Time use Survey (ATUS) dataset (https://www.bls.gov/tus/data.htm).<40> This dataset consisted of 10,493 activity schedules covering weekdays and also weekends, each linked with the demographic features of survey respondents. By equivalent the SPEW and ATUS datasets ~ above age, gender, race and also ethnicity, each agent in the synthetic populace was assigned a daily task schedule, in 1-hour increments, from a distinct collection of accessible ATUS schedules because that weekdays and also for weekends.

This equivalent resulted in a collection of available schedules for each certified dealer ranging between 2 to 1496. For each simulated day, an agent was randomly assigned a unique task schedule indigenous the collection of obtainable matched schedules. These daily schedule assignments resulted in agents being co-located in distinctive places. Details sharing emerged when co-located people interacted. The big range of obtainable schedules for the daily random assignment was a modeling selection to introduce randomness of accessible activity schedules for each demographic subgroup in our agent population (alternatively, agents would follow the same schedules every simulated day). When a larger sample of obtainable schedules does undoubtedly introduce stochasticity, it was deemed far better than the instance of having teams of agents with very few schedules, which could result in unrealistically repetitive task patterns. Future occupational will focus on the implication of this randomization on the resulting co-location networks. We additionally do no account because that the seasonality the schedules as result of the lack of accessible granular data. The ramifications of this modeling choices are a limitation of ours model and the emphasis of future work.

Using the ATUS data, us classified all activities into one of two subsets: (1) activities of day-to-day living (e.g., sleeping, washing and grooming, talking on the phone) or (2) physical and also mental wellness maintenance or promo (“health maintenance”) activities at both the individual (e.g., health-related self-care, law yoga) and household level (e.g., obtaining medical treatment for a child or an adult in the household). Using an iterative data collection and discussion process with skilled informants, us then contrasted the subset of wellness maintenance ATUS tasks to the finish list of company types listed on a HealtheRx (“service types” refers to kinds of solutions that a place might offer in contrast to a “resource” which refers to a particular place giving a particular service). Each ATUS activity from this subset to be mapped come one or much more relevant HealtheRx locations where an agent might reasonably carry out that ATUS activity. For example, the ATUS activity “providing medical care to family members children” was mapped come the HealtheRx solutions “fill prescriptions,” “medical supplies,” and “home care.” expert informants to be presented v both the subset of ATUS tasks and HealtheRx service types and discussed the likelihood of an agent making use of a service while performing an task for every mapping until agreement was achieved. Using the CRx source inventory, each service kind was associated with all places in the physical atmosphere that noted that service.

We computed the average variety of minutes per week invested doing each activity for every agent across all task schedules the were demographically suitable to that agent. Next, the moment per week during which a offered HealtheRx organization could be offered was identified by summing each agent’s mean time act every ATUS task mapped to that HealtheRx service and also dividing by the total number of agents for each matched schedule. This mean was computed within every of four period groups (Fig 4). These groups were characterized using demographic, life course and also healthcare plan considerations and also are consistent with commonly applied strata for age in the clinical and populace health literature.


3.5.2 Agent understanding evolution.

Agents kept dynamic β scores for resources they knew about. In ~ initialization, agents to be seeded with knowledge around resources using a distance-based arbitrarily assignment algorithm. Agents to be seeded with expertise of 10 to 100 resources located within a low distance radius (parameterized together 1 mile), 1 come 5 sources within a tool distance (1–3 miles) and 1 come 5 sources for long ranges (3+ miles). The first two main were taken into consideration a burn-in period for the whole simulation. We usually observed stabilized actions by main 3 of the simulation. Results reported in this document use calculation from week 3 that the simulation.

The β scores were raised by info dosing, a factor based on exposure to resource information and also the resource of the information. Agents might be exposed to resource information through doctors, nurses, clinical clerks, social contacts and also by first-hand use of the resource. The clinician sources listed information come agents by transferring a HealtheRx during a health care visit in ~ a clinic (“clinical dose”). Society dosing arisen when agents co-located at places (e.g., a barber shop) and also exchanged source information (e.g. Around the gym). Agents supplied a source by going come its location (subsequently described as simply “use”). The functional type of β score evolution and also its parameterization were created in consultation with experienced informants and through sensitivity analyses as formerly described.<39> The β score for agent i about source j at time t evolved according to the adhering to functional kind and is explained below:

*
(1)

The decay parameter λ (0.991 for the simulation run developing Fig 5, through an allowable variety of values between 0.991 and also 0.9994) accounted for receding expertise of a resource. The source of source information was linked with a value, εx, where x ∈ (Doctor, Nurse, Clinical Staff, Use, Peer, None}. “Use” indicated an agent had been to or spend a given resource following the decision function in Eq 1.


Note: Each pillar (n = 3) to represent a distinctive agent. Each row represents a unique source (n = 8). Each black dot shows the β scores (left y-axis) at in allude in time in hrs (bottom x-axis). Details dosing events (receipt the information around a offered resource) that developed during a provided hour are indicated by vertical lines as: receipt that a HealtheRx (blue), receipt of information about resources from a social call (green) and also use the a resource (red).

https://doi.org/10.1371/journal.pcbi.1009471.g005


For the simulation run used to produce Fig 5, the adhering to values were offered εx = 0.05, 0.15, 0.25, 0.2, 0.9, 1. Resources expected to have actually a greater influence top top the certified dealer had greater dosing worths (corresponding to smaller εx values). Together modeled, εx only accounts because that the source from which the info was received and not the type of source information received. This modeling an option was based on one the our certain questions: exactly how does the information source influence details diffusion dynamics? This question is important to real-world deliberations around how ideal to supply community source referrals in the clinical context. The ABM is designed so it might easily be iterated to encompass a resource-specific info dynamic. Furthermore, info dynamics may vary by factors like the stigma linked with information around a offered resource type (e.g. HIV/AIDS support group, food pantry or substance usage counseling). Future model advancement could it is in iterated to account for this kind of dynamic. Together referenced previously, the knowledge evolution process does no account because that a an unfavorable effect (where an unfavorable or incorrect information could lead come an agent avoiding a details resource). Although associated studies have surfaced no proof for an adverse information spread, that is definitely possible. This variable is a limitation the the design left for future design development.

Agents kept knowledge around a maximum of 200 resources; sources with reduced β scores were changed by sources with greater β scores over time. The presumption of a memory pack of 200 sources was make to stand for the bounded expertise and ability of the agents come maintain information for their decision-making (described below). The CRx ABM permits for the size and also heterogeneity of agents’ cognitive capacity to be parameterized. V the dynamics that agents’ β scores, the CRx ABM models internal agent states with respect to resource knowledge.


3.5.3 Agent details sharing behavior.

Information sharing was likewise modeled to count on the nature of individuals’ tasks (e.g., over there is no information exchange when an agent was sleeping, vice versa, information have the right to be exchanged as soon as co-located agents were working out in a gym). The propensity for an agent to receive details (“p-score”) to be dependent ~ above the activity in which the agent was engaged. Every ATUS task in ours subsets to be assigned a propensity for details sharing. Details sharing to be assumed come occur in between two or more co-located world by means of face-to-face conversation. (Therefore, details shared digitally, because that example, was not captured in the model.)

Because we discovered no extant data to notify assumptions about propensity for info sharing throughout various activities, us surveyed professional informants in iterative rounds to create p-scores because that each appropriate activity. The survey elicited the likelihood (p-score = none, low, medium, high) that the respondent would receive resource information from an additional person if doing a offered activity. Because that example, “sleeping” uniformly produced a p-score that “none.” activities like “helping family members adults” and “grocery shopping” produced a p-score that “medium” and “socializing, relaxing, and recreation as part of job” and also “obtaining medical and also care solutions for household adult” developed scores the “high.” Agents in ~ a ar share details with various other co-located agents based upon a threshold identified by the propensity for source information sharing during that activity and individual random draws against that threshold. The attributes of agents receiving details were no a factor in the info sharing dynamic. The design of this device was chosen to reflect the an essential information sharing dynamic commensurate with the intended propensity of source information sharing identified with skilled opinion, and also can be considered as a limitation enforced by the data ~ above the model. However, this info sharing dynamic does allow us to compare social dosing effect to the straight clinical dosing effect. The parameter values for these effects were an important to version calibration and also was previously described in detail.<39>


3.5.4 certified dealer decision-making.

We then occurred processes because that agents’ decisions about using a resource (recall, characterized as a certain service at a particular place). These decisions to be dependent ~ above agents’ information-sharing behavior and also their knowledge evolution. Every of the parameters involved in the decision processes was educated by empirical data and also expert opinion (Table A in S1 Text). Sensitivity analyses performed on selected parameters have actually been previously described.<39>

The agent’s decision to usage a source was modeled together a binary choice A/B Decision Model. Each activity was defined by a decision type–A/B decision or not. Tasks classified as health maintenance or promotion activities (the activities that the CRx intervention targets) faced an A/B decision choice. Other tasks not related to health-maintenance behaviors did not require a decision choice–agents would continue with such activity at the designated location. Agents presented v an ATUS activity in their daily schedule that was subject to an A/B decision confronted a choice to use a resource or not. For example, when presented v the ATUS activity “fitness,” the agent chose whether to usage a fitness basic (Decision A) or not (Decision B). If the agent decided to use the resource, lock proceeded v the activity in your schedule mapped to that service. If they determined not to use the resource, they ongoing doing the previous activity in your schedule.

The adhering to agent attributes dictated the decision procedure for each agent i∈P¸ where ns is the set of every agents:

a role of their expertise (βi,j) about a specific resource j∈ R, where R is collection of all resources (e.g., a community center), wherein they might use a particular service (e.g. A gym),the distance to the resource for the certified dealer (δi,j), andthe inherent inertia the they necessary to overcome for using that resource (γj).

An agent’s decision to usage the resource, Decision A, was made as soon as the agent’s activation level (threshold) was gone beyond by the agent’s mix of knowledge and also effort forced (distance and inertia). The connection at the time (t) is offered by:

*
(2)

To provide an illustrative example–consider one agent adhering to their daily hourly schedule till they pertained to an activity requiring an A/B decision: walk to a gym or not. The agent uses the decision calculus described above to choose in between the gym (A) or their existing (B) activity. If the agent knows of no gym, the β will certainly be low. If the gym is far away, the δ will be high. The certified dealer will proceed their previous task (a B decision in Eq 2 above). The dynamic defined in Eqs 1 & 2 permit us come isolate and also measure the result of information dosing ~ above knowledge about and use of selected sources for health and wellness maintenance or promotion tasks that to be the focus of the CRx intervention. The version does not encompass other wellness promotion and also maintenance tasks occurring at other places, for example going for a walk external or casual support team at a home.


3.6 CRx ABM submodel: transporting the HealtheRx in silico

To apply the in silico treatment to the fabricated population, agents to be assigned health conditions and also a wanted language utilizing data native the very same 26,558 HealtheRxs. Agent qualities were statistically matched by gender, age, race and ethnicity to the patients because that whom the in vivo HealtheRxs had been generated. HealtheRxs were distributed to agents at clinic areas where CRx to be deployed in vivo throughout the CRx-1 and also CRx-2 clinical trials.<10,11>


3.7 version implementation and also validation

A comprehensive description of the version calibration and validation procedure has been previously published.<39> The community Rx ABM determined each agent’s activity at every hourly time step and moved agents come their equivalent locations. Agent state and also location changes were selectively logged, together were the different species of details dosing events. The total variety of agents who determined to use a specific source following the ab decision process, was also logged. To calibrate and also validate the model, we selected source use in our agent population for 10 certain resources (clinics) that forced an A/B decision and also calibrated our parameter set where the design output (average source use) was within a defined selection of empirically it was observed clinic visits. A random forest model to be iteratively fitted to model evaluations come characterize the model parameter room against observed empirical data.<39> Table A in S1 Text explains the parameters offered in the CRx ABM.


4.0 In silico experimentation

Experimentation to be done using the Extreme-scale Model exploration with Swift<51> structure on the Midway2 computer cluster at the college of Chicago and the Bebop swarm at Argonne nationwide Laboratory.


4.1 Generating endogenous information diffusion networks

We simulated the info diffusion processes within the CRx ABM and generated endogenous details diffusion networks based upon hourly snapshots the co-located agents exchanging information based upon εx worths (described in ar 3.5.2) and also the propensity to share information. In these networks, separation, personal, instance agents were network “nodes,” and information exchange events between agents to be the network “edges.” The web links in the networks signified the pathways v which information around resources was exchanged. We created an exemplar age-stratified image of these networks utilizing Gephi (https://gephi.org) (Fig 6).


4.2 experiments on propensity for info sharing

We ran computational experiments, varying input parameter values that administer the CRx ABM’s details diffusion processes. This process helped united state analyze how propensity to share information (p-score) affects populace level details networks. We parameterized the low, medium, and high p-scores into adjustable rates of details exchanges to change a basic generator that the information diffusion network. We show distinction in network level distributions throughout different levels of p-scores (Fig 7).


4.3 Experiment to determine optimal setting of CommunityRx intervention delivery

To estimate the size of the spread of HealtheRx details from society dosing alone, us calculated the proportion of agents who preserved HealtheRx info from society dosing solely to those who preserved HealtheRx information from clinical dosing (using logged simulation data). Utilizing a calibrated parameter set (technical description of version calibration and validation space summarized in section 3.7 and described in information in <39>), we supplied the ABM to operation 4 week long simulations, with 15 runs because that each the 3 HealtheRx shipment modes (delivered through a physician, nurse, or clinical clerk). We typically observed stable habits by main 3 that the simulation. We used week 3 outputs come report results (the very first two weeks were used as the burn-in duration for the simulation). Parameter values for the 45 speculative runs are shown in Table C and D in S1 Text. Because that qualitative comparison of geographical spread the information throughout the 16 ZIP codes in the CRx ABM, we plotted the residence location the agents who retained information from straight clinical versus social dosing throughout week 3 that the simulation (Fig 8 and also Appendix B in S1 Text).


Establishing a functioning model

The mean age of the synthetic populace (N = 802,191) to be 44 (range: 16–94). The majority of the populace was female (56%), non-Hispanic (84%), and also African American or black (59%) (Table B in S1 Text).

Agents, on average, spent the most time engaging in tasks during i beg your pardon they could use the complying with HealtheRx company types: spirituality services, group exercise classes, wade groups, assist paying mortgage and also rent, and getting fresh fruits and also vegetables. The average number of minutes per week agents invested doing tasks mapped to these five self-care service species varied by period (Fig 4). For example, contrasted to agents younger than 30 year old, agents 65 years and also older spent, top top average, 4.6 times much more minutes every week doing tasks where they can use spiritual care services.

Agent knowledge around resources progressed over time, dependent on one agent’s information dosing events. Fig 5 mirrors the β score dynamics for three agents’ knowledge around eight different resources over time. Together expected, source knowledge boosted over time v successive info dosing events (i.e., receiving a HealtheRx, utilizing a resource, or getting information around a source from another agent), particularly those from sources with greater influence (like receiving a HealtheRx and resource use) and decreased when information dosing occasions did no occur.

By tracking info exchange occasions that emerged as agents relocated through their task schedules, we generated emerging information diffusion networks. The links in networks signified the pathways v which information around community sources was exchanged. An example diffusion network, generated over the course of a single hour for two age groups—agents 16 come 25 year old and agents 65 or an ext years old—showed differential clustering by age groups (Fig 6). Information exchange “hubs” varying from predominantly younger to mainly older agents emerged.

As we readjusted the rate of propensity for information exchange from low to medium to high, the network degree distributions adjusted (Fig 7); the variety of links (pathways v which details was exchanged) in between agents increased. At the greatest rate of details exchange, the tail of the circulation extended throughout three order of magnitude contrasted to the lowest rate of information exchanged. Idealized network models have been displayed to exhibit exponential,<52> extended exponential,<53> or scale-free<54> level distributions. In contrast, these emergent networks, endogenously created from the local interactions the co-located agents adhering to their individualized schedules, displayed level distributions with facility structures, which are not as easily fit to simplified sensible forms.


In silico clinical versus social dosing experiment

The geographic spread of HealtheRx information via society dosing much exceeded direct clinical dosing alone (Fig 8). Compared to clinically dosed agents (mean = 29,028; variety 27,271–30,701), an typical of 4.2 times an ext agents (range 3.9–4.6) retained HealtheRx information from social dosing alone (mean = 123,562; range 118,175–126,658). Table E in S1 message details the results from every of the 45 simulation runs. The sample of information spread by agents receiving a HealtheRx at a clinic visit was similar when the HealtheRx was delivered by physicians, nurses or clinical clerks (Appendix B in S1 Text).


Discussion

CRx is an information-based treatment that connects patient to healthful ar resources. In former studies,<10,11> we found that nearly half of patients and also clinicians who received the intervention shared information about resources v others, providing proof of a possibly potent however previously unexplored pathway through which a clinic-based intervention have the right to impact population health. Based on these observations, this study explains how we built a computational laboratory to facilitate the quantification that the population-level impact of CRx. Prior researches of similar information-based interventions have actually relied on typical observational and experimental designs.<5> this designs catch the interventions’ individual-level impact, but not the population-level effects resulting native the ease v which health-promoting details from the intervention have the right to spread. Information technology is enabling the rapid expansion of information-based interventions in windy health and health care.<55> come our knowledge, ABM has actually not yet been used to research the population-level influence of a health and wellness information intervention delivered during a clinical encounter.

In comparison to a clinical trial, a computational activities is designed to efficiently explore “what-if” concerns at large-scale. In the case of CRx, a clinical trial inquiry “What is the impact of a community resource information intervention on center age and also older adult patients’ self-efficacy for self-care?” We carried out an experiment to resolve the question: “What is the result on info diffusion if the treatment is ceded by a doctor versus a nurse versus a clerk?” Overcoming speculative capacity of a solitary clinical trial, the ABM will certainly also permit us come ask other crucial questions like: “What if the patients who get the treatment are young adults?” Or, “What if the treatment is delivered based on screening versus making use of a universal approach? What will certainly be the impact on health outcomes and also how will the population effects of the treatment be different?”

This examine shows how an interdisciplinary team of biomedical and also systems scientists collaborated to develop a computational laboratory. We have demonstrated that it is realistic enough for the purpose, through observing trends that replicated intuitive or previously-known results. Together expected, we demonstrated that synthetic agents displayed diverse behavior characteristics, reflective that the known geographical population, and also that the man-made intervention might be delivered to this agents in silico. We additionally showed that agents’ source knowledge advanced as they were exposed to source information and evolved differently depending upon the details source. We depicted the introduction of information diffusion networks and also how the level distributions of these networks adjusted as the worths of crucial parameters prefer the propensity of details sharing were varied. Lastly, us demonstrated the computational feasibility of using extremely granular activity data, an important feature of a examine seeking ultimately to i found it when, where and also how info spreads in between people in a population. Combined with former validation work, these results assist to create the model’s readiness because that hypothesis-testing experimentation.<38,39>

To illustrate the potential that the ABM as a enhance to clinical methods, we likewise conducted one in silico experiment come examine how specific source information disseminates from an individual exposed to the CRx intervention in a clinical setting (“clinical dose”) come other individuals in the population through society dynamics (“social dose”). With 45 speculative runs simulating mixing behaviors in a geographical population, this examine finds that personalized community source information delivered to a patient during a healthcare visit diffusion to others in the community. Complementing the three-month trial’s observation that the total number of people reached by HealtheRx info was in ~ least dual the variety of middle-aged and also older patients in the treatment group, the ABM in silico experiment approximates this ratio to be about 4-fold, when including all people ages 16 and older in the geographic population. We additionally find that the mode of clinical distribution (physician, nurse, clerk) likely has tiny relative impact on the size or dynamics of spread, saying that clinical implementation of community resource referral solutions can focus much less on that delivers the treatment and more on optimizing patient and also clinician access to it. This empirical and simulated observations offer life to the principle of patient society dosing, a previously overlooked mechanism, and also potential force multiplier v which community resource referral and other health-promoting info interventions may supply value beyond the patient who receives direct intervention during a medical care visit. To our knowledge, this is the an initial study to research the population-level diffusion dynamics that a clinical info intervention.

Most ABMs are developed using data native extant literature, an additional datasets and also expert informants.<42> In 2010, Morell and colleagues propose an integrated technique where data and also insights produced through observational evaluations are supplied to journey building and parameterization that ABMs.<21> Public health researchers consequently voiced support for this safety approach.<19,20> structure on and also extending this ideas, the CRx ABM was constructed in tandem with both observational and also experimental studies and also was parameterized making use of data from this studies. Building the ABM come simulate agent-level information sharing, a dynamic observed in CRx trials, revealed a must assign a propensity for details sharing during agent activities. Because we had actually not guess this data need at the outset of the trial, we created this parameter set by surveying professional informants. The enhancement of major data arsenal to estimate the propensity for information-sharing throughout daily activities would boost the worth of population-based time use surveys, favor ATUS, to in silico testing with information-based interventions.

In parallel through ABM building, iteration and also experimentation, we proceed to run clinical observational and experimental trials come deepen our knowledge of how information about resources spreads and also the impact of the source information on a theory-driven collection of health and healthcare outcomes.<56,57> this empirical data native clinical trials will certainly be offered to additional parameterize and improve the ABM end time, which, in turn, will certainly enable much more advanced in silico experimentation. For instance, the CRx ABM have the right to be provided to to compare community source use in relationship to varying prices of details spread and also network size and other attributes to inform assessments and also decision-making regarding resource allocation. These, and also other experiments, room a emphasis of future work.

Although ABM-based experimentation enables rapid iteration the community-wide trials under varying problems that would certainly be infeasible to conduct in the real-world, there are limitations. The CRx ABM is thrust by a decision-making version that is a simplified depiction of determinants influencing how people decide to usage health-promoting community resources. For example, we only account for face-to-face information sharing, not various other modalities like social media. The validity that ABM-based testing is dependency on the important factors captured by the underlying behavior models. Part agent characteristics that are likely dynamic in the real-world are static in this first iteration of the design (e.g., certified dealer activation). Modeling behavior readjust is an active area the systems science research;<58> we intend that brand-new discoveries will educate iterations of our version over time. Also, the CommunityRx ABM just accounts for knowledge about and usage of regional health-related resources that could be listed on the HealtheRx. Wellness maintenance tasks could occur exterior of these sources (e.g., walking in the neighborhood, an informal support team at someone’s home). We likewise have a limited understanding that which and how countless resources a human being needs to preserve health, just how this info is maintained in memory and how these determinants vary among individuals. Lastly, the propensity for details sharing during details activities is critical variable in our ABM, however our approximates may be minimal by usage of experienced informant data. This limitation reveals a data need that will be crucial to future computational research studies of information-based interventions and also presents an interesting opportunity for brand-new empirical research. Although these are crucial limitations, the in silico laboratory is flexible; we can iterate the decision model in tandem with emergent empirical data and developments in the field much more generally. Thus, if the emphasis of the CRx ABM is ~ above simulating the CRx intervention, the process of version building, and also computational testing presented is generalizable to other massive ABMs, for example those modeling details diffusion processes. However, generalizability may be restricted because the version was validated only versus the data the were used to notify the model building.

See more: Businesses Of All Sizes Are Increasingly Looking For Global Opportunities Because

We have demonstrated the procedure and feasibility of complete clinical trials and also systems science methods to construct a flexible laboratory for studying the population health affect of one information-based intervention delivered to individuals at the suggest of clinical care. By using open resource tools and sharing ours methods,<38,39> we aim to build trust in our approach, prompt feedback native peers, and enable others to use, iterate and learn from our model. In enhancement to advancing knowledge specific to understanding and also valuing the influence of CRx, this work serves to advancement knowledge and testing the a fast-growing range of information-based health interventions being occurred for distribution to promote patience and population health.