LAGOS: learning health systems and how they can integrate with patient care

CA and TA

CA is a systematic coding and categorisation method for investigating texts and resolving quantitative description of features.12 13 TA provides the systematic element characteristic of CA, while also combining analysis of frequency with analysis of in context meaning, providing a more truly qualitative analysis.12 CA and TA are established methodologies regularly used in clinical, nursing and other healthcare research contexts.12–15 In this paper, CA of the definitions for LHS provided in the literature identified a large number of related concepts. TA of these concepts refined the key themes that define the LHS and identified their relationships within the context of LHS.

The pathway and layer-based perspective

Development of a pathway theory and analytical framework can highlight underlying values and fundamental relationships between otherwise fragmented concepts.16 Different from the clinical or treatment pathways common to medicine and nursing, pathway theory is designed around consistent values and beliefs,16 in this case, the established levels of medical practice application.11 17 18 Starting with this application pathway and using it as the basis for ordering all other pathways, we sought to bring together the domains of LHS and CCPS documentation and arrange them on the basis of how each applies to the provision of population, evidence-based or precision medicine.

Case studies

CS are a method for conducting and presenting comparative research into subject areas that include qualitative and mixed-mode information science inquiry.19 CS are frequently used within information science and are considered to be as well-developed as any other scientific method.20 A range of CS types exist, including: exploratory, explanatory, descriptive, intrinsic, instrumental and collective.21 22 CS allow researchers to capture the knowledge of practitioners using a broad variety of data sources to ensure the knowledge is considered through multiple lenses.23 24 CS in computing and information sciences tend to be more open methodologically, accepting of methods from both natural and social sciences. In this way they tend towards being more hybrid in nature when contrasted to those in the medical sciences that rely more on interviews and data collection techniques involving individuals. The taxonomy and LAGOS framework are evaluated in the context of three subprojects of the PamBayesian project.

Characterisation and conceptualisation of LHS

Most LHS papers define LHS using definitions proposed in two seminal Institute of Medicine (IoM) workshop publications. Prior to the first contemporary LHS, the IoM described LHS as a system in which knowledge generation is so embedded into the core of the practice of medicine that it is a natural outgrowth and product of the healthcare delivery process and leads to continual improvement in care.25 Their latter and more highly cited definition describes an LHS as a system in which progress in science, informatics and care culture align to generate new knowledge as an ongoing natural by-product of the care experience, and seamlessly refine and deliver best practices for continuous improvement in health and healthcare.26 This definition fails to describe attributes that would contribute to aspects of patient care that LHS systems should target: quality, safety, efficiency and efficacy. As a consequence, the LHS domain has seen little development of these aspects.5 6 Similarly, the IoM definition does little to develop an understanding of the attributes and concepts that underpin implementation and usage of LHS in clinical practice.5

The unified conceptualisation and framework for systematically characterising LHS derives from our following strands of previous research:

• Concept: The lack of understanding of the concept of LHS was described as the research community awareness challenge.6

• Taxonomy: Developed a taxonomy describing the entire scope of LHS as observed in the current literature showing how each type positions within the larger learning health organisation.5

• Framework: Developed a unifying framework showing the role and context for LHS and its integration into the learning healthcare organisation.5

• Implementation: Identified the benefits, barriers and potential facilitators for LHS, and comparative analysis of how these may have arisen from, or be related to, those that have impacted EHR implementation during the preceding 30 years.7

• Realisation: In this paper we demonstrate how the taxonomy, framework and factors for successful LHS implementation can be applied to a contemporary project to identify those aspects that constitute LHS, the type of LHS, the barriers to which facilitators may be applied and the benefits that may result.

From the concepts identified in LHS definitions found across literature where authors self-identified their works as LHS, five themes converged which self-align into a learning life-cycle shown in figure 1. Similarly, EBM is best presented as a learning life-cycle which comprised five themes: assess, ask, acquire, appraise and apply.27 We propose that the themes defining LHS and precision medicine are also a form of a life-cycle consisting of the following five phases:

Request permission

• Patientdata: Collections of patient data in the form of EHR are seen as a near-inexhaustible source from which to learn new knowledge.28 29

• Computingsystem: Not limited to the computers used to access or store patient data, but also includes those which contain the programmes and perform the machine learning, prediction and other computing necessary to learn and apply knowledge.29 30

• Learning: The concept of learning as it relates to LHS are the processes that analyse data to derive or generate new knowledge.28–31

• New knowledge: The new knowledge learnt from patient data can advance our understanding of the underlying mechanisms of disease and patients’ response to treatment.29 32

• Near realtime: The current driving ambition for LHS is to expedite the process, often described in terms of a 17-year lag, of getting knowledge to inform clinical decisions from scientific discovery to clinical use.28 29 32 33

The LAGOS framework

LAGOS is an acronym for the five pathways shown in figure 2. These pathways are Learning health systems, Applications, Guidance, Operational and Systems. They focus and converge on the patient, and broadly define the areas covered by the Framework. The intended focus of LAGOS is the individual patient presenting before the clinician. Each pathway radiates from the most general or broadest application of that pathway’s scope, towards the centre, which represents the most specific application, and which is directed towards the individual patient. Thus, within LAGOS, as the viewer moves along each pathway towards the patient in the centre, the focus of elements at each layer shifts from a population-based focus to a precision medicine focus. Precision medicine itself forms part of a life-cycle whenever the knowledge learnt from engaging LHS influences or impacts future decisions on health policy, population medicine, clinical practice guideline (CPG) or the development of new health-based computing and learning systems.

Request permission

The LHS pathway

The complete LHS taxonomy shown in figure 3 is represented in the LHS pathway of LAGOS. At the outer edge is surveillance, which operates as an automated alert process within information technology systems that (A) monitors the entire population’s EHR for diagnosis or clinical coding of a range of communicable diseases, and (B) can also be programmed to monitor for adverse treatment outcomes. Those LHS types at the second layer primarily work with or on the basis of EBM, or are used by clinicians in review of their, or other clinicians, treatment outcomes. This includes cohort identification which is most often used within learning health organisations to identify groups of patients based on one or more similar characteristics. Below are those LHS most proximal to the patient and which can be engaged by the clinician in direct patient care. These models support personalised clinical decision-making and predict risks and outcomes that may result for an individual patient from receiving the selected treatment. It is these LHS that directly meet the definition of being precision medicine.

Request permission

Figure 3

Learning health systems (LHS) taxonomy.5

The guidance and operational pathways

The full details of our process of developing the taxonomy, definitions and characterisations for CCPS are described in ref 9. Here we provide a brief summary necessary for the reader to understand the pathways. CCPS define healthcare policy and procedure and are arranged in a hierarchy describing both their operational nature and distance from the patient, as shown in figure 4. For example, policy is both the furthest from the individual patient in that it is set by governments to guide health services for entire populations, and least operational in that it is the most general document and least likely to be based on evidence-based science. CCPS can also be described based on whether their primary intention is guidance or operationalisation. Policy and the levels of CPG operate primarily in the guidance space and are described on the guidance pathway based on their proximity to the patient; policy is primarily population based while local CPGs are closest to the individual patient. There is overlap between the guidance and operational pathways in that local CPGs can also be seen with operational content. On the operational pathway, local CPGs lie at the population end as they would be applied to the general population of a health district diagnosed with a particular condition, while care plans, care maps and treatment protocols are found along the pathway converging towards individual patient treatment.

Request permission

Figure 4

Taxonomy and hierarchy for clinical care process specifications.9

LHS in the context of generating the Realistic Synthetic Electronic Health Record

Accessing EHR for secondary use purposes such as data research, modelling and artificial intelligence training presents with challenges, notably:

1. Attaining ethics approval for access to collections of EHR.

2. Difficulty when consent is required from each individual patient.

3. Over-reliance on anonymisation that can reduce or remove important contextual detail.

The CoMSER Realistic Synthetic Electronic Health Record (RS-EHR) and ATEN Realism in Synthetic Data projects operate following the approach described in figure 5 and focus on satisfying the need for access to EHR for secondary uses relying on a privacy-preserving knowledge-intensive method to generate locally realistic, but not real, synthetic EHR without needing access to the real EHR.

Request permission

The relationship between LHS and RS-EHR can be two-way. LHS can help provide the aggregated statistical data and knowledge described as rules and relationships that exist in EHR data sets. In this way, RS-EHR generation need never be exposed to real EHR during the definition or generation of synthetic EHR. Conversely, LHS can be built, trained and validated by projects like PamBayesian using collections of RS-EHR, prior to being productionised to work for real patients and clinicians.

The LHS paradigm allowed us to fully exploit the routinely collected data from the healthcare system. This made development of knowledge-intensive methods for generating synthetic EHR successful, making it easy to create collections of realistic synthetic EHR for use in secondary uses where privacy concerns prevent release of real data. Further, development of knowledge-intensive models enables prediction of patient risk for particular negative outcomes and recommending appropriate and more effective treatments based on patient characteristics, history and current symptomatology possible.

To fulfil RS-EHR’s aims the following LHS types from the LHS pathway, which apply to the levels of medical practice from the application pathway (in brackets), are needed:

1. Cohort identification—learning evidence (EBM) and operating within the context of the learning healthcare organisation level of the system pathway to identify a prescribed cohort of patients with similar health conditions or characteristics such as demographics and symptomatology consistent with the disease to be modelled and generated.

2. Positive deviance and negative deviance—learning evidence (EBM) and operating within the context of the learning healthcare organisation level of the system pathway; of commonly used treatments, both effective and ineffective to ensure synthetic patients receive realistic treatments and outcomes.

3. Predictive patient risk modelling—specific to patient (precision medicine) and operating within the LHS level of the system pathway to identify patterns and model risk factors consistent with adverse events.

4. Clinical decision support system—specific to patient (precision medicine) and operating within the LHS level of the system pathway to identify characteristics of synthetic patients that make them compatible for generation of specific disease or treatment outcomes.

LHS in the context of patient risk and decision modelling

There are numerous approaches for developing intelligent systems supporting clinical decision-making for diagnosis, prognosis or treatment selection. Bayesian networks (BNs) are one such approach. BNs model uncertainty and allow the user to update prior belief, such as when assessing the probability for presence of a medical condition in light of new evidence (additional symptoms, risk factors and test results). However, the process of building these intelligent systems for chronic conditions is not yet fully explored and understood. Chronic conditions are particularly challenging in this context as the patient’s condition must be monitored for extended periods during which many decisions may be undertaken. Ideally, doctors and nurses should be able to monitor patients without the resource-intensive, expense and inconvenience of clinic visits, except when such visits are necessary. Current clinical records and care processes do not easily receive, integrate or enable patients in the home to collect and transmit self-monitoring data from inexpensive sensor-based devices like the Apple Watch and continuous glucose monitors.

PamBayesian is developing a new framework for distributed probabilistic decision-support systems. As shown in figure 6, PamBayesian combines patient data with clinical expertise and patient input, for use in developing intelligent systems. The novelty of this framework is the use of ‘conventional’ EHR (eg, blood tests, imaging results) combined with near real-time continuous data from local sensors for learning and providing new knowledge. This allows for autonomy in a collaborative decision-making environment that includes clinicians and patients, to avoid unnecessary visits to a clinic or hospital. Once the patient’s condition crosses the diagnostic threshold (in green), the clinician prescribes the treatment (in yellow) and treatment review (in red) thresholds. The patient self-monitors the parameters of their condition and enters these into the LHS application. If assessment and prediction of their condition rises above the treatment threshold, the patient receives treatment, be it medication or otherwise. If it rises above the treatment review threshold, the clinician is alerted that the patient requires review so that an appointment can be offered.

Request permission

To fulfil PamBayesian’s aims the following LHS types from the LHS pathway, which apply to the levels of medical practice from the application pathway (in brackets), are needed:

1. Cohort identification—learning evidence (EBM) and operating within the context of the learning healthcare organisation of the system pathway to identify patients with similar demographic and clinical characteristics.

2. Clinical decision support system—specific to patient (precision medicine) to collect and analyse daily data and operating within the LHS level of the system pathway to provide relevant patient feedback.

3. Predictive patient risk modelling—specific to patient (precision medicine) and operating within the LHS level of the system pathway to predict and identify potential future adverse events.

LHS in the context of empowering patient participation in healthcare

Despite advances in modern medicine, many chronic conditions such as diabetes and rheumatoid arthritis have generally proven incurable. The daily life of patients with chronic conditions is highly affected by disease progression; over time disease symptoms exacerbate until they overwhelm the patient. Patients must constantly evaluate their condition, making day-to-day decisions regarding care and relying on advice from their treating clinicians to guide those decisions. Again, despite medical advances, access to healthcare remains a significant issue for all patients. Regular appointments with doctors or nurses are time consuming, expensive, inconvenient and, in many cases, cannot be scheduled to coincide with times when the worst symptomatology may present.

PamBayesian aims to empower patients to undertake day-to-day self-care within boundaries; diagnostic, treatment and treatment review thresholds that are defined by the patient’s clinician. As shown on the right side of figure 7, home health monitoring devices and applications will be used to gather patient symptoms, measurements and reports about their condition, and with BN intelligence will tailor clinical knowledge and generate patient advices. In this way, PamBayesian promotes continuous monitoring of the patient’s condition while supporting patient self-management and engagement of timely interventions. PamBayesian also promotes a more effective and efficient interaction model between patients and clinicians whereby expensive and time-consuming clinic visits need only occur when a patient’s monitoring shows that their symptomatology has escalated and surpassed the treatment review threshold as discussed in the previous section.

Request permission

To fulfil PamBayesian’s aims the following LHS types from the LHS pathway, which apply to the levels of medical practice from the application pathway (in brackets), are needed:

1. Cohort identification—learning evidence (EBM) and operating within the context of the learning healthcare organisation of the system pathway to identify patients with similar demographic and clinical characteristics.

2. Clinical decision support system—specific to patient (precision medicine) to collect and analyse daily data and operating within the LHS level of the system pathway to provide relevant patient feedback.

3. Predictive patient risk modelling—specific to patient (precision medicine) and operating within the LHS level of the system pathway to predict and identify potential future adverse events.