Introduction A key decision for assessment of Low back pain (LBP) is identifying serious underlying conditions such as Cauda Equina Syndrome, infection, fracture or space-occupying lesions. Previous decision support tools for LBP deployed rule-based recommendations, yet Artificial Intelligence has enabled ‘intelligent’ decision support tools, with Bayesian Networks particularly suitable for complex conditions such as LBP. This study aimed to test whether clinical knowledge could be elicited to construct a Bayesian Network to support clinicians’ detection of serious pathology masquerading as LBP.

Methods A modified-RAND appropriateness procedure elicited knowledge from 16 domain experts from General Practice, Rheumatology and Musculoskeletal specialties. This comprised a four-stage process using bespoke online tools interleaved with face-to-face meetings; 1) Variable elicitation, 2) Structure elicitation, 3) Probability elicitation 4) Validation. Independent experts in spinal pathology reviewed the initial tool and its outputs.

Results The tool includes background risk factors (e.g. trauma, age), signs and symptoms (e.g. bladder disturbance, inflammatory symptoms) and derived judgement factors (e.g. cord compression, fracture). The tool has an interactive online interface, requiring real-time patient inputs from the subjective assessment, then gives a judgement comparing baseline to the current patient. Content validation suggested no missing elements to the model, but may require more detail for clinical understanding of terms. Face validation exposed some inconsistency in clinical reasoning, particularly for spinal infections and fractures.

Conclusion The structured elicitation method yielded a reasoning model using expert clinician knowledge, establishing consensus amongst participants about its content. Further iterations to expand this to common LBP presentations should follow.