Abstract
Healthcare must deliver high quality, high value, patient-centric care while improving access and costs even as aging and active populations increase demand for services like knee arthroplasty. Machine learning and artificial intelligence (ML/AI) using past clinical data primarily replicates existing cause-to-effect actions. This is insufficient to forecast outcomes, costs, resource utilization and complications when radical process re-engineering like COVID- inspired telemedicine occurs. To predict episodes of care for innovative arthroplasty patient journeys, a sophisticated integrated knowledge network must model optimal novel care pathways. We focus on the first step of the patient journey: shared surgical decision making. Patient engagement is critical to successful outcomes, yet existing methods cannot model impact of specific decision variables like interactive clinician/caregiver/patient participation in pre- and post-operative rehabilitation, and other factors like comorbidities. We demonstrate coupling of simulation and AI/ML for augmented intelligence musculoskeletal virtual care decisions for knee arthroplasty. This novel coupled-solution integrates critical data and information with tacit clinician knowledge.
Introduction
When patients experience knee pain due to degenerative or inflammatory changes, there are several potential treatment and risk mitigation decision pathways that are addressed prior to undergoing knee arthroplasty. In a value-based payment model, there is an inherent bias to focus on conservative interventions prior to recommending knee arthroplasty. However, patient knowledge and understanding is often limited or misguided because of inconsistent information sources and widely varying local clinical practice. A conservative anti-inflammatory approach may treat knee pain with oral nonsteroid anti-inflammatory drugs (NSAIDs) and/or injectable corticosteroids as the outcome. If/as such conservative treatments are ineffective, partial or full joint repair or replacement may occur. There are many decision points in the knee joint resurfacing process that are left to the surgeon without shared decision making. These include technical aspects such as cemented or uncemented implantation or resurfacing the patella. Other considerations include pain management, rehabilitation, discharge planning, equipment needs at home, and comorbidity impact on healing and well-being.
Each of these factors can be addressed through a statistical model, using tools like artificial intelligence and machine learning (AI/ML) to identify optimal surgical or non-surgical clinical solutions. Application of AI/ML has been inconsistently adopted by the arthroplasty community, however. Part of the challenge for such AI/ML applications is that such systems rely primarily on data that has been collected and curated from existing and prior clinical cases, and much of that data is derived from the clinicians’ point of view. In the emerging value-based reimbursement paradigm, patients’ assessment and feedback becomes a significant weighted portion of the overall “optimal” process evaluation. Because AI/ML tools rely on actual data generated from each experience, there is no easy way to generate sufficient statistically valid data for a future “to be” process that has never existed. Even if a pilot study were undertaken for a specific new knee arthroplasty process intervention, exploration of multiple potential interventions of interest could consume considerable time and, could introduce unsupportable costs and risks.
One opportunity that emerged out of the COVID-19 pandemic is the introduction and use of telemedicine for many parts of the pre- and post-surgical patient journey. As documented in the literature, virtual musculoskeletal patient triage became a feasible extension to the orthopedic surgeon’s office. These successes have inspired discussion about using telemedicine to enhance the patient’s pre- and post-surgical rehabilitation and to also engage the patient much more directly in many key decisions. For example, it is perceived that improving patient education and dialog through an integrated telemedicine and knowledge network could serve to improve clinical outcomes, perceived patient experience, and better value metrics.
This paper presents a new foundational method and approach by creating an ensemble modeling tool that extends and integrates AI/ML, based on as-is data, with simulation modeling, which can represent many potential “to be” future care paradigms. In addition, this paper is based on the belief that optimal clinical and patient care may best come from an ensemble modeling tool that integrates clinician and patient perspectives, priorities, and values, creating an augmented intelligence resources. The augmented intelligence ensemble modeling does not replace human decision-making; instead it provides the human stakeholders with the best available decision information for their use.
Take, as an example, a 58-year old patient who used to be physically active, but can no longer stand or walk and is now obese and pre-diabetic. The obesity and diabetes pose risks for successful knee arthroplasty, and the younger age of the patient might suggest that a full knee replacement may not last their full active lifetime. The clinician and patient decision processes regarding full or partial replacement or other interventions involves quite a few unstructured facts and decisions related to quality of life, risk tolerance, patient commitment to pre- and post-surgical medical care, strengthening, and rehabilitation. One might posit – and model -- a novel intervention whereby the patient agrees to a very carefully designed prehabilitation and medical treatment process overseen by a telemedical care management team, and an equally carefully designed telemedical post-surgical recovery process. Such an innovative care paradigm, which might even include diet, mental coaching, life style changes, and constant clinical caregiver intervention could be modeled with a simulation tool like the one described in this paper. That tool would, however, benefit by inclusion and integration of AI/ML modules into an ensemble tool that could analyze and recommend optimal decisions at each step of the patient’s journey. By including the patient and clinicians into each decision point, this ensemble tool could augment the intelligence of all stakeholders, allowing them the ultimate choice of the next process steps; hence supporting precision and personalisation.
Through published literature, it is known that unstructured decision making in dynamic and complex environments is not only challenging but in almost every situation the decision maker is undoubtedly faced with incomplete information1. Actionable data and tools of knowledge management are essential to ensuring accuracy in the decision-making process2. Given today’s healthcare contexts, where tremendous data is generated from multiple sources, there is great demand for short cycle data when making critical decisions, many of which may have far-reaching, life-altering consequences. More than ever, an integrated knowledge network is a strategic necessity. However, while we recognize the importance it has been difficult to model and demonstrate the benefits of maximizing essential knowledge assets to enhance clinical decision making and enhance process flows. In the following, we demonstrate this by defining and using coupled simulation models and AI/ML models and answer the research question “How can we use simulation and AI/ML coupled modelling to derive better value in specific healthcare contexts?”.
Background
Before designing and developing an appropriate integrated knowledge network that supports the episode of care, it is necessary to first understand key, and often unique, aspects of the healthcare industry, the significant challenges and the underlying healthcare value proposition. In sharp contrast to other industries, healthcare has a unique structure where the receiver of the services (the patient) is often not the predominant payer for those services (the insurance company or government)3. Healthcare interventions are typically complex and involve a variety of players such as providers, payers, patients and regulators4. The complexity of the system leads to economic dilemmas such as moral hazards, considerations pertaining to cost versus quality and information asymmetry. These issues have the potential to create obstacles when trying to deliver efficient and effective healthcare5,6. High quality data and germane knowledge play a vital role in overcoming these obstacles and can only be obtained through the design of a technology enabled system, defined as the intelligence continuum7,8,9.
In addition, in today’s challenging COVID-19 context there is even more pressure on healthcare organizations to provide high value, high quality care to patients and subscribe to a value-based care paradigm. Value is yet to be clearly defined; some view it as quality/cost but the construct of quality itself is rather nebulous. By using coupled simulation and AI/ML modelling, it is possible to demonstrate the anticipated benefits of different scenarios and thereby get a better understanding of the value to be derived in a specific context. We illustrate this in the context of musculoskeletal virtual care (MSKVC).
A Taxonomy for Coupled Simulation and AI/ML Models
The strength of simulation models is their ability to capture complex clinical workflows with resource constraints. However, as part of episode of care and care continuum, such clinical workflows also involve many complex decision points where the use of an AI/ML model would be more appropriate. Thus, we propose an ensemble simulation modeling tool that can tie the two approaches (AI/ML and simulation) together, presenting the clinician and patient with best options as care proceeds.
This ensemble modeling tool could be viewed as coupling the simulation tool with the AI/ML tool in a client-server mode. Conceptually, one type of model (client) may utilize services of another type of model (server) on an as-needed basis. Further, in utilizing these services, the communication channels may be synchronous or asynchronous. A synchronous coupling might be needed for a real-time decision like recommending terminating an excessive exercise routine that is exceeding risk thresholds. An asynchronous coupling might be appropriate to extend or alter a prehabilitation period or regimen to assure a safer and more satisfactory post-surgical outcome.
In the context of our proposed simulation and AI/ML ensemble, we get four different couplings as shown in figure below (Figure 1).
Figure 1.
Synchronous and asynchronous couplings of Simulation and AI/ML models and their associated software architectures.
The two diagrams a. and b. in the figure above illustrate the software realization of these couplings. Note that each direction is independent, that is, if a simulation model as a client is invoking the services of an AI/ML model as a server, then the arrow is from the simulation model to the AI/ML model. Of course, at the communication level, the requests and responses will be bidirectional. The coupling a. requires synchronous communication where one entity can invoke the other entity during active run-time to obtain results for a set of inputs sent over the network via blocking send/receive communication primitives. If the communication is asynchronous, then it may not require a real-time communication between the entities as shown in b. In fact, it could be implemented via a shared resource (file, database, cloud-storage, etc.). Colloquially, we may refer to the synchronous coupling as tight coupling and the asynchronous coupling as loose coupling. In summary, therefore, here are the four different ways of combining simulation models with AI/ML models:
Simulation model client to AI/ML model server with synchronous communication. We denote this as SAMsync, where SAM stands for Simulation And ML/AI.
AI/ML model client to simulation model server with synchronous communication. We denote this as MASsync, where MAS stands for ML/AI And Simulation.
Simulation model client to AI/ML model server with asynchronous communication. We denote this as SAMasync, where SAM stands for Simulation And ML/AI.
AI/ML model client to simulation model with asynchronous communication. We denote this as MASasync, where MAS stands for ML/AI And Simulation.
As a concrete implementation, we have created a Python-based library (PyCPN) that can be used in conjunction with Colored Petri Nets (CPN) Tools simulation software to provide a realization of a tight (synchronous) coupling between a CPN simulation model and a Python-based AI/ML model10,11. The core of these libraries are primitives to open a connection to a remote process (client mode), accept a connection from a remote process (sever mode), send a string of arbitrary length to a remote process, and receive a string of arbitrary length from a remote process. These libraries allow dynamic creation of connections during the simulation of a CPN model and multiple simultaneous connections are supported. These libraries implement a protocol with blocking send/receive communications primitives on top of the TCP/IP for passing messages synchronously between a CPN model and the external Python application containing AI/ML model as shown below (Figure 2):
Figure 2.
Synchronous coupling and communication realization between a CPN simulation model and a Python-based AI/ML model.
SAMsync Synchronous (Tight) Coupling Example-Musculoskeletal Virtual Care (MSKVC)
To illustrate our approach and use, we created a CPN simulation model of a Musculoskeletal Virtual Care (MSKVC) process12. The overall patient flow and triaging described in the aforementioned paper is as follows:
“Patients seeking musculoskeletal care are introduced by contacting the centralized appointment desk originating most from a provider referral, emergency department (ED), or urgent care consults. Patients are diverted to a virtual short questionnaire that can be administered electronically or via telephone. The questionnaire would assess the acuity, severity or urgency and a secondary questionnaire that includes basic patient demographics, pertinent comorbidities, and key elements would provide insight into the nature of the patients’ complaints, including type, location, character, acuity, recent surgical history, and newly obtained imaging. An acuity and urgency-based stratification of patients would be conducted based on outcomes of the virtual musculoskeletal triage questionnaire, where patients with chronic conditions and established follow-up patients are offered routine virtual visits according to availability and geographical location. Conversely, derangements that are deemed to be acute, especially among patients with recent surgical history, are directed to a virtual musculoskeletal triage channel where a live interview with a musculoskeletal provider would guide down-stream disposition. Based on the results of the triaging questionnaire, coupled with providers’ assessment, patients would be referred to an urgency-appropriate disposition ranging from direct surgical admission, referral to the emergency department (ED) or orthopaedic acute care centre after direct surgical admission, referral to the emergency department (ED) or orthopaedic acute care centre after direct notification, and provision of a plan-of-care, or requesting appropriate imaging. Alternatively, patients could be scheduled for an on-demand (immediate), urgent (same-day), expedited (within 72 hours), or routine virtual visit with the appropriate member of an orthopaedic clinical care team.”
Central to this MSKVS is a questionnaire that can be evaluated by an AI/ML model. In the simulation model of the underlying triaging process, at the decision-making state, the collected patient profile can be communicated vi a synchronous channel to the AI/ML model and the result can then be communicated to the simulation model to proceed with the rest of the simulation steps. Of course, the AI/ML model is not capable of accounting of various resources, or patient arrival rates, etc. So, we do need a simulation model to optimize the process side.
Utilizing the synchronous communication library and primitives described above, we open a connection to the Python process running the AI/ML model as highlighted in the simulation model snapshot below (Figure 3).
Figure 3.
CPN simulation model snapshot showing connection establishment to an external Python process running an AI/ML model.
Our hierarchical CPN simulation model is divided into separate sub-modules13,14. The figure below (Figure 4) shows two snapshots of the sub-module responsible for handing the questionnaire. Once the patient is done answering the questionnaire, the highlighted code segment in the figure on the left gets executed when the associated transition highlighted in green is fired. As a result, the collected patient profile is converted into a string and sent to the Python process. At this point, the simulation is stopped until a result is received form the AI/ML model. The result is then evaluated, and patient is routed to the appropriate next stage based on the outcome from the AI/ML model. This is depicted in the figure on the right. For case depicted, the patient had the following profile: age 41 with a recent injury to foot and a recent surgery but has no swelling and can move. Thus, per result from AI/ML model, the patient is sent for an on-demand virtual visit.
Figure 4.
Snapshot of the CPN simulation model sending patient profile to the AI/ML model (left) and routing of the patient based on received result (right).
SAMasync Asynchronous (Loose) Coupling Example-Blood Sample Matching Lab Process
As part of our ongoing efforts focused on improving patient safety and creating a value proposition for the healthcare, especially in the context of episode and continuum of care, we are creating several other models of various hospital and clinical flows and operations. We use one such model to illustrate the concept of asynchronous coupling described above. In particular, we give details of this coupling using the model of a lab process for matching blood. In this case, we have a data analytics model that can give the frequencies of various markers of interest in blood samples based on some given profile. The simulation model then reads this data asynchronously to initialize its blood sample generator function. In this case, the there is no need for the data model to be coupled synchronously with the simulation model.
In the simulation model, the representation of blood sample consists of a list of segments. Each segment contains a list of marker position that indicate a marker of interest at that position. The frequencies of these markers vary. For example, the notation [(1,[]),(2,[]),( (3,[8]),(4,[11,12])] represents a blood sample where there are no markers of interest in segments 1 and 2 but it has a marker of interest at position 8 in segment 3 and positions 11 and 12 in segment 4. The frequency data obtained from the data analytics model is shown below (Table 1):
Table 1.
Marker frequency data for blood samples obtained from the data analytics model.
| Segment | Position:Frequency List |
|---|---|
| 1 | 1:0.0000001, 2:0.0000001, 3:0.0000001, 4:0.0005, 5:0.0000001 |
| 2 | 6:0.006, 7:0.0045 |
| 3 | 8:0.14, 9:0.135 |
| 4 | 10: 0.0485,11:0.4055, 12:0.468 |
The simulation model reads this data during the initialization step and generates a pool of blood samples to be used in the matching process. This is depicted by the two model snapshots in figure below (Figure 5). It is worth noting the savings in terms of both cost and time with the simulation model if one were to optimize or re-engineer the underlying process using a wet-lab with real blood samples!
Figure 5.
Snapshot of the lab process simulation model showing generation of blood samples based on data supplied by a data analytics model.
Conclusion
In today’s healthcare environment, there is increased pressure to deliver high quality, high value, and patient-centered care while simultaneously improving access and controlling cost. Focusing on a cause-to-effect chain of actions is insufficient to provide statistical modelling that can forecast patient clinical outcomes, resource utilization and complication risks. We propose an ensemble simulation modeling tool that can tie the two approaches (AI/ML and simulation) together, presenting the clinician and patient with best options as care proceeds.
We demonstrated the benefits coupled simulation and AI/ML modeling and illustrated with a use case of musculoskeletal virtual care (MSKVC) which is integrated with knee arthroplasty. In this way we show how it is possible to yield real-time integration of critical data and information, while simultaneously capturing tacit knowledge that can enhance optimal clinical outcomes and patient experience. Our second example dealt with the model of a lab process for matching blood. In this case, we used a data analytics model off-line for the frequencies of various markers of interest in blood samples based on some given profile. This input was then used by the simulation model to drive the process.
We also proposed a taxonomy and an approach for coupling simulation models with AI/ML and data analytics models for a better clinical informatics. However, it should be noted that such couplings are a part of a continuum rather than isolated instances. The conceptual diagram in figure (Figure 6) illustrates the various facets of this couplings that can co-exist. In fact, when dealing with episode of care and continuum of care, all possible combinations of these coupling may need to be realized. We also contribute to the area of expandable AI.
Figure 6.
A conceptual diagram representing a continuum of synchronous (tight) and asynchronous (loose) coupling between simulation models and AI/ML models.
References
- 1.Award E, Ghaziri H. Knowledge Management. Upper Saddle River, New Jersey: Prentice Hall; 2004. [Google Scholar]
- 2.Wickramasinghe N, Bali R, Lehany B, Gibbons C, Schaffer J. Healthcare Knowledge Management Primer. New Jersey: Routledge; 2009. [Google Scholar]
- 3.Wickramasinghe N, Schaffer J. Realizing Value Driven e-Health Solutions. IBM Center for the Business of Government. 2010. [online] http://www.businessofgovernment.org/report/realizing-value-driven-e-health-solutions (accessed 1 March 2022)
- 4.Wickramasinghe N, Schaffer J, Bali R. The Health Care Intelligence continuum: key model for enabling KM initiatives and realising the full potential of SMT in healthcare delivery. International Journal of Biomedical Engineering and Technology. 2008;1(4):415–27. doi: 10.1504/IJBET.2008.020070. [Google Scholar]
- 5.Boyd JR. Coram, R. Boyd: The Fighter Pilot Who Changed the Art of War. Vol. 2002. New York: Little, Brown & Co; 1976. Destruction and Creation. [Google Scholar]
- 6.von Lubitz D, Wickramasinghe N. Healthcare and technology: the doctrine of network-centric healthcare. Intl. J. Electronic Healthcare. 2006;4:322–44. doi: 10.1504/IJEH.2006.010440. [DOI] [PubMed] [Google Scholar]
- 7.Alberts DS, Garstka JJ, Stein FP. CCRP Publication Series, Dept. of Defense. Washington, DC: 2000. Network Centric Warfare: Developing and Leveraging Information Superiority; pp. 1–284. [online] http://www.dodccrp.org/publications/pdf/Alberts_NCW.pdf (accessed 1 March 2022) [Google Scholar]
- 8.Cebrowski AK, Garstka JJ. Network-centric warfare: its origin and future. US Nav. Inst. Proceedings. 1998;124(1):28–35. [online] https://www.usni.org/magazines/proceedings/1998/january/network-centric-warfare-its-origin-and-future (accessed 1 March 2022) [Google Scholar]
- 9.Boyd JR. Patterns of Conflict, Unpublished Briefing. 1987. [online] https://danford.net/boyd/patterns.pdf (accessed 1 March 2022)
- 10.Gehlot V. From Petri Nets to Colored Petri Nets: A Tutorial Introduction to Nets Based Formalism for Modeling and Simulation. Winter Simulation Conference (WSC) 2019. pp. 1519–1533. doi: 10.1109/WSC40007.2019.9004691.
- 11.Jensen K, Kristensen LM, Wells L. Coloured Petri Nets and CPN Tools for modelling and validation of concurrent systems. International Journal on Software Tools for Technology Transfer. 2007;9(3):213–54. [Google Scholar]
- 12.King D, Emara AK, Ng MK, Evans PJ, Estes K, Spindler KP, Mroz T, Patterson BM, Krebs VE, Pinney S, Piuzzi NS, Schaffer JL. Transformation from a traditional model to a virtual model of care in orthopaedic surgery: COVID-19 experience and beyond. Bone & Joint Open. 2020;1(6):272–80. doi: 10.1302/2046-3758.16.BJO-2020-0063.R1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Gehlot V, Sloane EB, King D, Schaffer JL, Kirk M, Miller E, Wickramasinghe N. COVID-19 and Beyond: Taming the Virtual Healthcare Tiger with Colored Petri Nets based Modeling and Simulation. Proceedings of the Hawaii International Conference on System Sciences (HICSS) 2022. pp. 3773–3782. doi: 10.24251/HICSS.2022.460.
- 14.Jensen K, Kristensen LM. Modelling and Validation of Concurrent Systems. Berlin-Heidelberg: Springer-Verlag; 2009. Coloured Petri Nets. [Google Scholar]