Serious spinal pathologies (SSPs) are rare [1], difficult to identify [2], and have potentially life-changing implications for patients with low back pain (LBP). Examples include, but are not limited to, cauda equina syndrome (CES), spinal malignancies, spinal infections, and vertebral fractures. In clinical practice, red flag questions [3] are routinely used by health care professionals during their assessment to attempt to assess whether a suspicion of an SSP might be raised; however, many patients will report ≥1 of these symptoms without having a serious underlying pathology [4]. A number of studies have questioned the validity of red flags as indicators [2,4]. The red flag questions represent a complex reasoning algorithm that takes place in the decision-making process on the diagnosis or treatment pathway. A recent study exploring the use of clinical decision support to manage LBP in community pharmacies revealed clinicians’ uncertainty when screening for SSP [5]. Therefore, developing effective decision support to aid clinicians in the reasoning process is desirable; for example, a clinical decision support system (CDSS) for predicting the need for surgical intervention in patients with LBP has been implemented with some success in the Netherlands [6], but the authors of the study [6] have identified that the results could be improved with a Bayesian network (BN) technique [7]. BNs are capable of using expert opinion combined with evidence where it is available to give probabilistic outcomes for a given scenario [8]. BNs are graphical models that represent the probability distribution and conditional independencies among variables by a directed acyclic graph. BNs are ideally suited to reasoning under uncertainty because the conditional independencies encoded in their structure are used to compute complex probabilistic queries in an efficient way. Their graphical structure offers a suitable medium to encode domain knowledge. They can be combined with machine learning algorithms to learn from data.

A BN has been used to predict mental health conditions; however, this study relied upon a large data set (>17,000 data points) over multiple time points [9]. Indeed, the availability of large data sets containing outcome measures and a relatively low rate of errors is usually needed in the majority of development approaches used. Musculoskeletal medicine does not typically have the luxury of such data. As such, evidence, patient data, and expert opinion will need to be combined to replicate the reasoning process in musculoskeletal and spinal conditions, as demonstrated by Sheng et al [10].

The aim of our study was 2-fold: first, to develop and build a BN using a 3-stage elicitation, and second, to establish the accuracy of the BN following a robust validation framework. The remainder of this paper will discuss the development of these 2 aims, referring to the “BN.” We use this term intentionally to focus attention on the knowledge-based artificial intelligence (AI) sitting behind a potential CDSS.We consider that a CDSS comprises: a user interface, a software package to integrate with an electronic health record and an underlying knowledge-based AI model (the BN). As such the elements of the CDSS require separate development studies (Figure 1). The design and usability of a CDSS are beyond the scope of this study.

To be able to build the BN and represent relationships between the reported symptoms, risk factors, and potential outcomes, a formal elicitation process was needed to determine the variables to be included within the model. The considered approaches for the elicitation included a Delphi study [11], a RAND appropriateness method [12], and a focus group method [13]. After running an internal pilot with clinical colleagues [14], we deemed the RAND method to be the most appropriate. An outline of the elicitation procedure is provided in the Elicitation Process subsection. This paper focuses on the health care application of this method to SSP in LBP. For a full report of the mathematics behind the elicitation procedure and construction of the BN, the interested reader is directed to our protocol [14] and Multimedia Appendix 1 [8,12,14-20].

Ethics approval was granted by the Queen Mary University of London Ethics of Research Committee (QMREC2018/48 027). Informed consent was obtained from all participants in the study. All participant contributions were anonymized in the data collection. Participants were offered reimbursement of expenses and a shopping voucher in exchange for participation.

The elicitation consisted of 3 distinct stages. In the first stage (variable elicitation), the 16 clinical experts considered which variables should be included in the model. In the second stage (structure elicitation), we asked the experts to link the variables together via the relationships between the variables to form the structure of the model. The third stage involved a probability elicitation to populate the model with appropriate values.

A web-based tool was developed to conduct the elicitation (Figures 2-4). The elicitation was based on an imposed structure of medical reasoning: risk factors, signs and symptoms, and judgment factors. The term judgment factors was used instead of diagnosis because it would be unrealistic to claim that this tool could reach a diagnosis. Rather, we aimed to produce a risk assessment or judgment as to what may be influencing the patient’s presentation.

Participants were asked to consider a patient scenario and decide what factors they would consider when assessing the patient and deciding on the next course of action. They were asked to consider risk factors, signs and symptoms, and judgment factors with respect to clinical reasoning, with explanations of each provided as part of the tool. A small number of factors were provided for them as an illustration (Figure 2).

Participants were able to use the tool (Figure 2) to drag and drop the provided variables into any category they deemed appropriate as well as add their own recommendations for variables. They could then rank the variables within each category to signify the importance of their inclusion in the resulting BN. During the web-based phase, this work was conducted individually. During the workshop, participants had the opportunity to discuss the variables in each category before revisiting the elicitation tool and completing the task again. In stage 2, participants used grids (Figure 3) to express how they believed the variables should be related to each other (ie, establishing the edges in the BN), which we used to form the structure of the BN. Participants entered a number ranging from 0 to 3 to signify the strength of the relationship, with 0 representing no relationship and 3 representing a strong relationship. Again, participants worked individually on the web, then came together in a workshop to form consensus. Consensus was calculated by taking the median scores and calculating the interpercentile range [12]. Variables with a high median score (ie, most of the participants thought that there was a strong relationship) and a low interpercentile range (ie, there was consensus) were included and formed an edge in the BN.

Stage 3 comprised a web-based elicitation only. The participants were guided through a series of questions to elicit the probabilities of events in the BN. An illustrative example can be found in Figure 4. Participants were given example scenarios with a clinical context and asked to estimate the probability of the scenarios. This facilitated the elicitation of partial node probability tables, with a plan to refine these with machine learning at a later stage.

A basic validation of the resultant model was planned. Typically, in a machine learning algorithm, a training set and a test set of data would be used to validate the model. However, this is not possible owing to the constraints on the available data in musculoskeletal conditions. Therefore, to test the model, the framework for validating an expert-elicited BN proposed by Pitchforth and Mengersen [21] was used to conduct some preliminary model testing. The results of validating the model will help us plan for the next stages of work, that is, identifying what further development may be needed on the model.

The validation framework used is summarized in Table 1. This framework recommends 7 domains of validation. In this study, it was possible to use 4 (57%) of the 7 domains; the domains concurrent, convergent, and discriminant require a comparator model; however, no appropriate alternative SSP model could be found. Further work is therefore required to identify a suitable comparator model in an alternative field of medical diagnosis. The methods used for the other 4 domains are detailed in Table 1.

The judgment factors that had been clearly defined by the participants were CES, space-occupying lesion, fracture, infective condition, inflammatory condition, and nerve root condition (encompassing radiculopathy and radicular pain from L1 to S1). As such, it was possible to find published guidance on the signs and symptoms as well as risk factors for nomological validation. Cord compression was not defined well enough and could have been construed as a sequela of the other pathologies listed. Irritability of pain, depression, stress, and anxiety were only examined in relation to a patient with LBP, rather than in terms of the medical diagnosis of these conditions. Therefore, it was not appropriate to compare these disorders with published diagnostic criteria because the participants did not intend to diagnose them; they only wanted to identify the disorders’ impact on LBP or symptoms.

The research literature was explored for guidelines and consensus statements on cauda equina, space-occupying lesion, fracture, infective condition, inflammatory condition, and nerve root condition. Comparison was made with the recent consensus study by Finucane et al [22] on the variables cauda equina, space-occupying lesion/cancer (malignancy), fracture, and infective condition (spinal infection). For inflammatory condition, comparison was made with the study by McCrum [23]. For nerve root condition, comparison was made with the National Institute for Health and Care Excellence (NICE) guidelines for lumbar radiculopathy [24]. Each paper identified a series of clinical signs and symptoms as well as risk factors for the conditions.

Five spinal experts were recruited via convenience sampling. The experts did not participate in the initial elicitation or in any other aspect of the study. The model structure was presented to the experts using a web-based tool [15] and an explanation of what a BN is. The experts were given the opportunity to explore the structure of the BN independently and comment as they felt necessary.

The spinal experts were asked to review the model and comment on the appropriateness of individual variables and the appropriateness of connections among the variables as well as suggest any new variables or relationships that they thought were missing from the model.

The Gwet agreement coefficient 2 test [25] for agreement was used to assess the responses of the spinal experts compared with those of the original study participants. The results were interpreted using the Landis and Koch criteria [26].

A series of patient case studies were taken from the literature [22,27] and clinical practice, with known outcomes. The cases, which are summarized in Table 2, were used to make predictions using the resulting model and compared with the actual outcomes.

Sixteen participants were recruited to the study (Figure 5). In stage 2, those who failed to complete the web-based elicitation before the study deadline (8/16, 50%) were excluded from further participation. The remaining participants (8/16, 50%) attended the first face-to-face workshop. Of these 8 participants, 7 (88%) attended the second face-to-face workshop, whereas 1 (12%) dropped out. Participant numbers between 7 and 15 are recommended [12].

A BN was formulated from the results of the 3 stages of the study. The risk factor variables (pink) and signs and symptoms variables (blue) in Figure 6 represent the specific questions that require answers to produce a probability of a condition (green). For variables with no user input, the BN will accept unknown and use a prior baseline assumption.

Participants were asked to list the variables pertinent to patients presenting with LBP and rank them in order of importance. During the 2 face-to-face workshops, participants were clear on the need to explore the role of spinal red flags and exclude SSP before any other aspect of LBP. They produced a list of categories and expressed how they believed the elicited variables should sit within these categories. In the second workshop, these categories were used to produce 2 grids. In the grids, participants gave a value ranging from 0 to 3 to denote the strength of the relationship between the variables in the rows and the variables in the corresponding columns. This yielded a clear delineation in the results: SSPs were consistently scored very highly, with a low interpercentile range, and all other factors received a low score. As such, the resulting BN primarily contains variables and edges associated with SSP.

The 3-stage RAND process has allowed us to build a BN model that provides a prediction of SSP. The expertise of the participants in this study centered around spinal masquerading conditions, and the perception that these were the not-to-be-missed pathologies were judged to be key to the success of the model.

Our study has followed a well-trodden path of manually building a BN using expert-elicited data [28,29]. The rise in advanced automated machine learning techniques has led to a decline in the more time-consuming manual building approaches. However, this manual method is suited to the problem of SSP owing to the lack of structured data [8]. Moreover, the deep integration of experts within the construction process helps engender trust in the model, which is critical to its success [30]. We completed the elicitation procedure with relatively low time burden on the experts by using a bespoke web-based elicitation tool, thereby negating one of the main disadvantages of the manual method. This also had the advantage of eliciting a range of views and establishing a consensus, whereas other studies often rely on the expertise of a single clinician [31,32].

Many studies have been published on building and using BNs as decision support tools; however, very few of them report the validation of such tools. In a review of prognostic models in obstetrics, only 8.7% of the tools were validated [30]. In using a comprehensive published framework, we have taken this a step further in establishing clinical confidence in the model (and areas that require further refinement), which will be required if the BN is to be used in clinical practice [33,34].

The results presented are promising for improving the identification and management of SSP. It is currently considered that clinicians in primary care make diagnoses with approximately 90% accuracy [35]. The use of a CDSS is intended to enhance, not replace, the expertise of clinicians [8]. Even if this BN could help to improve the accuracy rate by only 5 percentage points, the number of people who experience the long-term consequences of missed pathology would be halved. The last estimate of National Health Service expenditure on missed CES alone was £185 million (US $233 million) over 10 years for litigation costs [36]. Reducing the cases of missed SSP could significantly reduce these costs. If the BN were successfully deployed in a CDSS, it could help to improve the recognition of these rare conditions and reduce missed diagnoses.

Reducing variation in care is a key facet of the National Health Service long-term plan [37]. LBP is a complex condition to manage, particularly because of the many available pathways, including, but not limited, to orthopedics, rheumatology, musculoskeletal triage services, physiotherapy, self-management, pain management, and neurosurgery. This BN, as part of a CDSS, can systematically help clinicians to identify which patients with LBP are at risk of SSP and require escalation to specialist services, reducing unnecessary waits for inappropriate services. This would, in turn, leave conservative management options open for those who are at low risk [6].

The validity and reliability of red flag questions, designed to identify SSP, have been questioned in the recent literature [1,2,4,38]. Finucane et al [22] have opted to use a consensus method to provide a framework for the use of red flags in clinical practice. Our study used a similar methodology, recognizing the difficulty in bridging the gap between the lack of evidence for red flags and their regular use in clinical practice. By focusing on the SSP aspects of managing LBP in this study, we have prepared the model to consider safety in clinical use. In deploying this tool into a clinic, there is scope for machine learning from real-life data, thereby adding to the research knowledge about these difficult-to-manage cases.

Reducing unnecessary care using a CDSS has been proven in a wide range of health care settings. Compliance with guidelines represents reduction in unnecessary care such as overprescribing [39] and helps to manage waiting lists by ensuring that patients are referred to the correct service. The use of a CDSS has been shown to increase compliance with imaging guidelines in LBP [40], reduce prescription errors [41], improve the risk management of venous thromboembolism [42], and standardize cancer care in accordance with the evidence base [43]. These studies encompass a wide range of health care settings from community care to an intensive care unit. This study is the first step in developing a similar capability in SSP in LBP.

The BN’s good performance, in particular on CES, inflammatory condition, and space-occupying lesion, is encouraging. In this regard, the elicitation process has been a success.

The erroneous results may be due to flaws in the elicitation process or the biases of the recruited clinicians. Certain risk factors identified in the literature were either not identified in this study or ranked too low to be included in the final model. A known issue with BNs is the need to constrain the number of variables and relationships to make the BN computationally viable [8]. This runs the risk of leaving out variables that a clinician may normally consider as part of the clinical reasoning process. However, the benefit of the Bayesian method is the combination of expert knowledge with research knowledge, as well as the ability to interrogate the model to explore these cases. This study underlines the need for transparency in decision-making and a combination of knowledge sources, both of which are possible with the Bayesian method.

Furthermore, the spinal experts involved in the validation were asked to recommend additional factors to be added to the model if they were judged to be missing. There was no consensus on any additional factors from the experts, perhaps reflecting the difficulty of assessing these pathologies, even for experienced clinicians.

This study provides a blueprint for the development of expert-elicited BNs and their validation. The elicitation procedure described here would be applicable for any musculoskeletal condition and, indeed, a wide range of medical conditions. In addition, this study goes a step further than previous development studies in demonstrating the utility and practicality of using a validation framework to assess the output of the elicitation and identify the areas of future development. This is a crucial next step toward building trustworthy BNs that have more chance of success in terms of adoption and implementation in health care.

The next stage of development will involve building a web-based interface as a prototype CDSS knowledge-based AI system. It will need to be modified not only with regard to the clinical information collected here but also with regard to how risks are presented to the clinician, what action to take, and how urgently. An example design pipeline is presented in Figure 1. Further work on the usability and acceptability of the tool will be required to develop an interface that could put this information to best use in clinical practice. A mechanism to report outcomes to enable the AI learning component will be required. Further work to develop its capability to assess the optimum treatment for LBP that is not SSP should also follow.

This knowledge-based AI CDSS is a further step forward in providing clinicians in practice with access to expert reasoning. This method presents the combined reasoning of clinical experts, validated by the evidence base. The use of a BN as part of a CDSS enables expert recommendations to be made to patients from their first contact in the care pathway.

This study has demonstrated that knowledge elicitation from multiple clinical experts was an effective way of developing a BN for complex decision-making within the constraint of a lack of structured data. The key positive features are that this was a systematic elicitation of complex information, it involved multiple expert participants, it led to consensus, and its results are largely consistent with independent expert judgment.

The resultant BN represents a model of clinical reasoning regarding SSP, with the potential capacity to incorporate machine learning and improve during use. The use of a validation framework has yielded important insights into areas where the BN has high levels of accuracy and areas that require further development. The structure and content of the BN have been independently reviewed with high levels of agreement between the study participants and the independent reviewers. The validation of the BN’s predictive ability highlighted issues with particular pathologies such as nerve root condition and fracture. The concurrent validation using research literature has provided well-defined avenues for improving the predictions that currently do not provide adequate accuracy.

Having the ability to identify SSP in LBP is a promising first step on the way to developing a BN for LBP as a whole. With this essential element in place, the elicitation process should be rerun to build models designed to guide the active management decisions for the mechanical and psychosocial aspects of LBP.

A prototype CDSS will be built using the BN to be used in a time frame appropriate for clinical consultations and be ready for testing in a clinical setting to guide further development.