Model Frameworks for Health Care

From SystemsWiki

Revision as of 07:42, 22 September 2017 by Phoenix (Talk | contribs)
(diff) ← Older revision | Current revision (diff) | Newer revision → (diff)
Jump to: navigation, search
  • Constructing and Extending Integrated Dynamic Health Care Models: A Draft Framework
    • Geoff McDonnell, University of New South Wales
    • Nathaniel Osgood, University of Saskatchewan



  • Limitations in popular dynamic modelling methods continue to restrict their effective use in shaping policy change, regional service development and individual care in health systems.
  • Models can be considered as a mix of theory and data which elegantly or descriptively represent interactions among components that directly or indirectly contribute to health care.
  • Different modelling methods resonate with different stakeholders and disciplines, and a judicious mix of individual based models and aggregate pattern oriented models may offer a way to more effective model use at multiple levels in health care.
  • A middle level of services has evolved in Technical IT Frameworks and may be usefully extended to health and health care services. This middle services level interacts with policy and individual care levels.
  • The challenge is to place abstract policy models with the concerns of the everyday lives of real people, linking the sweep of history with individual biographies.

Key Ideas

Overall Aims

  • Facilitate construction of useful models that are actually used to take effective action in redirecting change at the policy, services and individual care level.
  • Develop useful, reusable, and flexible abstract representations of interacting structures/agents to computationally produce the behaviors of interest over time and run virtual what-if intervention experiments.
  • Coherently combine Theory, Data, Inferences, Decision Choices, Learning and Persuasion aspects of model building and use, including compelling and engaging visualizations.
  • Support the entire modeling process, including initial problem framing, conceptualization, qualitative reflection, participation and Critical Systems Thinking (IM-1204), consisting of selecting the appropriate context, clarifying values and checking models against reality.
  • Improve tools, training and ongoing education.
    • Models informed with mobile device data
    • Housekeeping including change awareness
    • Evaluation of impacts and learning effects

Support of the entire modeling process

Traditionally, modelling tools have supported more of the quantitative aspects of model building, calibration, sensitivity analysis and scenario testing. The qualitative aspects of conceptualizing problem situations and eliciting stakeholder concerns also need to be formally captured. This will help ensure that these concerns are translated into requirements which can then be traced to features of the completed model and scenarios. Linking stakeholder perspectives and values to the decision making processes can be performed using a qualitative approach like Critical Systems Heuristics. Another useful construct for value networks is the services-oriented architecture, which is widely used in IT services planning. A service is represented as an interaction between a stakeholder need and a resource capable of being used to meet that need. Stakeholder beliefs and values as well as derivative considerations involving the structure of the physical system drive the scope and level of detail of a model and set the scene for using the model and the group model building process for persuasion and emancipation. The model represents shared understanding of the opportunity for improvement and in addition to rigor and relevance may require data at the individual level in order to show who benefits and who pays for proposed changes. Similar constructs can be used to help individuals perceive their own better futures to motivate them to engaged self-care of chronic illness. In addition to their own data, simple descriptive visceral models that resonate with personal desires and experiences may require explicit individual models (with agent-based models as an example).


Why include an Agent-Based Modelling Framework?

  • Within the discussion below, we place an emphasis on agent-based modelling for several reasons. Two broad classes of those reasons are discussed below

General Motivations for Use of Agent-Based Modelling

One general class of motivations for building a reusable and general framework for agent-based models is the general attractiveness of such models. We discuss some specific motivations for the use of such models here.

  • We wish to capture aspects of heterogeneity that are cumbersome or infeasible to capture in a traditional aggregate model. There are several important reasons for wishing to do so: To give the option of representing detailed interventions that differ in response to these factors (e.g. which treat people differently based on previous history of treatment, or based on network position), to allow for calibrating against data related to these factors (e.g. longitudinal data), or to allow for understanding investigating the differential or equity effects of interventions in light such factors (e.g. how an intervention affects those in rural vs. urban areas, affects those of different ethnicities or different age groups, etc.)
  • Examples of heterogeneous attributes that we may wish to capture include – but are not limited to – the following:
    • Different geographic, spatial or topological contexts For example, we may wish to prioritize contact tracing based on a contact’s network position, focus vaccination on “bridging” individuals within a network, or which are located beyond a certain distance from a network, or understanding equity concerns with respect to the rural/urban divide, or across neighborhoods of the city)
    • Aspects of individual history ( “biography”) For example outcomes of past encounters of this individual with respect to the health care system, or records of past bouts of illness or treatment that may affect the outcome of the current bout.
    • States simultaneous with respect to several evolving processes (e.g. co-morbid conditions). Representing such heterogeneity in an aggregate model is typically cumbersome and error prone, requiring the stratification of the population by different attributes – both static and dynamic – and involving messy (and, too often, error-prone) specification of transition rules.
    • Continuous attributes. While aggregate models can do a good job of capturing continuous evolution of global attributes, when characterizing a population, they impose an “attribute based disaggregation” strategy that either summarizes an attribute across an entire subpopulation or forces discrete categories upon attributes – whether static or dynamic – that are often far more convenient to treat as continuous. Examples would include birth weight, degree of glycemic control or other physiological categories, location of birth, current weight, age, degree of belief or risk perception, skill level, etc. Within an aggregate model, each additional category added to a stratified model directly increases the size of the model. By contrast, within an agent-based model, there is no special privilege or performance advantage conferred upon to discrete attributes – a continuous attribute can be maintained just as easily as a discrete attribute, and there is no significant performance disadvantage to maintaining properties of an agent in a more detailed (even continuous) form. A notable side benefit – enough to be an attraction itself – of maintaining attributes in a continuous form is the flexibility of the aggregation measure employed. With a continuous attribute (e.g. age), one has the option of opportunistically employing different categories in different contexts – for example, summarizing by 5 year age categories in one case, according to 2-year age categories for children in another, in other cases, summarizing data for neonates according to month or weeks of age. Such flexibility in the aggregation achieved atop agent-based models can prove of considerable value when calibrating and parameterizing a model.
    • Linkages to other individuals. For example, we may wish to maintain records of the persistent connections between or relationships between an individual and others (e.g. between a mother and her children, or between co-workers).
  • We seek to explore the impact of aggregation/disaggregation on model behavior. While aggregate models allow limited representation of heterogeneity via subscript-based stratification, such stratification tends to not only be cumbersome and of limited scalability, but also quite inflexible. To add a new dimension of heterogeneity into a model is often a very time-consuming – and frequently delicate and error prone – task. As a result, the use of aggregate models not only discourages representation of heterogeneity in the first place, but also tends to raise significant barriers to investigating the impact on model behavior and intervention outcomes of representing heterogeneity (i.e. of disaggregation with respect to some dimension of aggregation). By contrast, it is frequently quite simple and straightforward to add a new dimension of heterogeneity into an agent-based model, thereby allowing ready investigation as to the value of capturing a particular type of heterogeneity within a model.
  • We need to capture distinct localized perception among agents to characterize agent behavior or dynamics. Situations where an individual’s behavior or evolution depends on their local context can be viewed as one more motivation for representing heterogeneity, but it is a particularly central one. For example, we may seek to represent how an individual’s degree of risk perception – and, by extension, risk behavior – with regards to an infectious pathogen or a behavioral risk factor such as drinking or smoking depends on the health status of people around that individual, or how the local perception of highway traffic (from the point of view a given vehicle) dictates braking behavior of that vehicle. Representing such dependence can be of central importance to understanding e.g. how changing the individual’s information access or mental models would change system behavior. By contrast, all too often aggregate models are “behaviorally impoverished”, with fixed or globally dictated transition rates defined, and decision making defined only at an “average” level across a whole population or subpopulation. Fortunately, agent-based models provide ready ways of situating agents in contexts, and a platform in which one can characterize decision-making at that individual level, such as might be based on heuristics or formalisms such as discrete choice theory, and informed by e.g. individual-level sensor data.
  • To communicate with and secure buy-in from stakeholders who conceptualize or communicate most effectively about the model at an individual model.
  • When there is much greater familiarity with or capacity to characterize dynamics at an individual level
  • We wish to characterize behavior at multiple scales. In modeling, we have a frequent need to characterize contacts which exhibit multiple levels of context – for example, individuals, neighborhoods, schools, city, and province. In some cases, each of these levels is associated with distinctive processes, distinctive patterns of emergent behavior (to which we may wish to calibrate the model), decision making stakeholders, information and material feedbacks, etc. In contrast to aggregate models (where variables relating to different scales of a situation – e.g. are placed in the model with no significant reflection of the hierarchy involved), the nesting structure of agent-based models naturally mirrors that (i.e. exhibits an isomorphism to) nesting structures in the external world. This both facilitates the collection of information within the model (e.g. in a way that allows leveraging and comparison against hierarchical linear models), and allows for natural and intuitive navigation of the model by stakeholders.
  • Capacity to use synthetic data sets and compare against “synthetic ground truth”. Because agent-based models are articulated at an individual level, we can readily apply widely-used techniques that operate using individual-level data. For several reasons – including the fact that aggregate models do not follow the trajectories of specific individuals over time – there are far more limited options for applying such techniques in conjunction with even a highly stratified aggregate model. Example applications include the following:
    • Evaluating robustness of study design. Using an agent-based model, we can apply a chosen study design on a synthetic population (e.g. a randomized control trial), simulate the evolution of that study on that synthetic population or cohort, and evaluate the consistency of simulated study outcomes compared to the underlying “synthetic ground truth” represented in the model.
    • Evaluating robustness of statistical inference and measures. Like study design criteria, statistical inferences commonly used in epidemiology typically operate on individual-level data. In a similar manner to the above, we can apply statistical measures – for example, administrative data algorithms to infer the presence of some underlying condition – to synthetic populations, and evaluate the robustness of the classifications, and the impact of the algorithm on the perceived outcome measures (e.g. on the perceived prevalence of diabetes).

Relative Advantages of Applying the Proposed Framework to Agent-Based Models

Beyond the general motivations for using agent-based modelling, there are some particular reasons why a framework of the sort discussed here will offer particular value within the agent-based modelling context.

  • Capacity to create reusable objects. The object-oriented approach that underlies most agent-based modeling packages permits ready reusable, modular pieces that can be provided in binary form by third parties. While various approaches – from the cut-and-paste reuse of “molecules” to Sandia’s textual model composition system to Powersim’s built-in objects – have attempted to provide similar modularity for System Dynamics models, such approaches fall far short of routine reuse in object-oriented software engineering. As a result, it is far easier to create a “plug and play” system making use of third-party binary components within Agent-Based Models than in System Dynamics toolsets.
  • Extensibility of Agent-Based modelling packages. The fact that agent-based modelling frameworks are commonly built upon widely-used general-purpose programming languages facilitates extensibility. In addition, the open-source character of Repast and the construction of AnyLogic atop the extremely popular Eclipse platform opens additional opportunities for ready adaptation.
  • Availability of enabling third-party interfacing technologies. While System Dynamics packages have traditionally exhibited a far more limited capacity for interfacing with 3rd party tools and technologies, the platforms used for Agent-Based modelling can readily articulate with such external resources. Examples would be logging libraries (e.g. log4j), Aspect-oriented programming technologies (e.g. AspectJ), visualization libraries (e.g. network workbench), database interfaces and persistence technologies (e.g. hibernate), statistical and mathematical packages.

Advantages for Working with AnyLogic

AnyLogic is most notably distinguished as an Agent-Based modelling platform in two regards:

  • Its support for “declarative modelling” (allowing a modeller to focus more on what is being described, rather than how it is to be realized), and its capacity to combine System Dynamics, Agent Based Networks and Process Centric Discrete Event Methods. Within this work, we hope to strongly leverage the declarative character of the modelling in two ways.
    • Firstly, we hope to build atop and extend the declarative features within our own, with many of the tools that we provide being realized using declarative specifications native to Anylogic (e.g. predefined varieties of parameterized statecharts).
    • Secondly, we anticipate that just as AnyLogic’s declarative features (and the reduction in software engineering they support) have served to attract a large user population to AnyLogic, the declarative elements and code-saving elements of our framework will attract many of the same users.
  • AnyLogic also offers significant advantages of being built atop of Java technology (opening up the opportunity to use a vast number of 3rd party libraries and frameworks) and Eclipse. This will provide considerable benefit in the extension of the framework to support integration with other computational methods, including Bayesian Inference, connectionist (e.g. neural network) computations, Discrete Choice Modelling, or to represent simple heuristic-based decision making, and advanced analytical approaches such as Markov Chain Monte Carlo.

Explicit Linkage of Policy, Service and Individual Care Levels

  • We envision a framework for a collection of interoperable models linking policy, structural and behavioral change, with individualised practice change through services change. Based on the benefits involved, we anticipate the application of a Local Services Orientation (viewing a service as an interaction between a need and a resource in a value network), applying the technical services framework to the less well defined and designed health and human care services. Aligned with this goal, we further seek to bridge the gap between abstract aggregate policy models and concrete clinical practice models using a middle services layer.

Use of the agent-based approach will aid in representing several levels of context, in making equity tradeoffs (combining bads and goods among individuals and groups, largely submerged in aggregation methods) explicit in order to do sensitivity analysis, and make additional assumptions (e.g. about aggregation measures) explicit.

  • With this frame, we seek to take advantage of the capacity of agent-based models to easily represent multiple scales and levels of context and greater detail to create hybrid “zoomable” models for integrating networked health and social care services at a regional district or community level. By supporting – among other things – far richer intervention selection, understanding of interventional effects, and behavioural models, this should result in a more grounded and descriptive approach than higher level aggregation.
  • As a theoretical basis for this framework, we will seek to expand on the Context, Mechanism and Outcome approach described by Pawson and Tilley for Realistic Evaluation of Social Programs.By capturing both localized perception and decision-making, spatial and topological context, and multiple level/scale models, we believe that the framework will help enable application of the “Nested adaptive cycles” concept in Socio-ecological systems, which addresses improvement and decline over Life Cycles, Spatial distributions and Levels of Organization.

We seek to focus on connecting 3 levels of decision interactions:

  • Policy Process Change, according to:
    • Time frame and windows
    • Spatial context
    • Level of organization
  • Services Process Life Cycle
    • Access
    • Capability (resources and composition)
    • Delivery
    • Control
  • Clinical Care Process/Task Interaction
    • Exchange of value
    • Moments of truth: Rather than focusing on day-to-day steady state, some of the defining dynamics that dictate organizational evolution and patient experience are decided in a “punctuated” fashion, based on specific contexts of strong perceived importance.

Linking Dynamic Models and Data

Model in Loop Planning and Action

Model in Loop" Planning & Execution: Adaptive sampling, information filtering and routing to specialist decisional services (implemented via software defined networks like those implemented via OpenFlow. Sensing and models (notes from Nate Osgood)

  • Self-management including informal carers

Helps manage more effectively Improved Sense of self-efficacy Permits learning of how body responds to insulin, exercise, eating This learning is not contingent upon a patient-specific model, but could be enhanced by this Allows sharing of learning with others of self-management expertise

Particularly key during pregnancy intergenerational effects Later risk of Type 2 diabetes for mother High degree of motivation

  • Service delivery

Lower acute cases (e.g. ESRD, blindness, amputations) The more effect the self-management, the lower and more deferred the burden on the health care system

Key issue for rural China: Lower complexity of cases seen, so that more complex cases are manageable by the small number of highly qualified doctors present in rural regions The majority of cases be handled by the far less deeply trained medical professionals in tertiary centers (e.g. village clinics)

Information collected by devices (with some modeling) allows more efficient management via understanding of risk factors facing a particular individual Forgetting to measure glycemic levels Neglecting to take insulin after measuring Sedentary behavior Depression Smoking (a known risk factor for diabetes) Overwhelmed by caregiving role Indigent (e.g. migratory workers)

This could work towards risk-factor specific Reminders (e.g. reminders to check insulin levels) Surveys (to better understand individual risk factor based on responses to contextually triggered on-device survey instruments)

Ability to probabilistically detect some early stage complications (e.g. polyurea), to allow for more focused & efficient (less scattershot or delayed) testing Earlier detection means lower costs & healthcare burden down the road

Capacity to better leverage of chart records (eHealth or not) with (electronic) risk factor data Possible for leveraging for telehealth (to give better patient profile)

  • Public Health (using data not only from sensors, but also from contextually triggered surveys)

Alerting to epidemiology of risk factors Identifying at-risk individuals for clinical or public health intervention (critically, per Pentland's exploratory work, this could include probabilistic inference to identify potential latent diabetics -- who are very numerous in China [an estimated 70% of Chinese rural diabetics in Anhui are undiagnosed] Earlier diagnosis of such diabetics could greatly reduce the future burden on the healthcare system, particularly for acute cases

Understanding risk factors in different regions City vs. rural

Designing more effective interventions e.g. data can lend understanding of how social networks diffuse new norms regarding Food consumption Physical activity Interactions between behavior change across multiple risk factors (e.g. those who are motivated to invest in managing their diabetes may make further attempts on quitting smoking [one of the biggest risk factors for diabetics] More effective peer-help programs (forging new social networks for exchange of understanding & support) Clustering groups according to barriers that face Group for those who are both trying to quit smoking & manage diabetes at same time Groups who circulate in common areas

Context specific intervention design (e.g. design for urban or rural regions that takes into account the different character of the built environment and social networks present)

Lower overall costs Lower pressure

  • Where do models come in? At all levels

Processing the data (machine learning models & filtering) Clinical & public health operational insight: inferential models Identify undiagnosed diabetics Identify those suffering important risk factors e.g. depression sedentary behavior Early stage renal complications (e.g. polyuria)

Understanding of individual physiological dynamics: physiological models publically available diabetes models, jim rogers & ed gallaher's model of anemia management, etc.) Could further be used for actual training on how to most effectively respond to conditions (although probably not anytime soon in China) Intervention-oriented Public health models leveraging deep understanding of individual-level heterogeneity in risk factors, degrees of self-management, diagnosis


Questions & Comments to Geoff McDonnell
Personal tools