JMIR Publications

ABSTRACT

Background:

Identifying and managing serious spinal pathology (SSP), such as cauda equina syndrome or spinal infection, is challenging. Traditional red flag questioning is increasingly criticised, and improving decision-making is being actively researched.

Objective:

We aimed to improve serious pathology identification by constructing and validating a decision support tool using Artificial Intelligence (AI) that combines current evidence and expert knowledge.

Methods:

A modified RAND appropriateness procedure, including variable, structure and probability elicitation was deployed to build a Bayesian AI model of reasoning elicited from 16 experts over 3 rounds. The causal model was designed to predict the likelihood of a patient with a particular presentation having an SSP. An established framework directed a 4-part validation that included comparison of the model with consensus statements, practice guidelines and recent research. Clinical cases were entered into the model and the results compared to clinical judgement from spinal experts.

Results:

The model included 38 variables in three domains of risk factors (10 variables), signs & symptoms (17 variables) and judgement factors (11). Comparison with the evidence showed the model is typically consistent but needs changes to e.g., 2 of 11 judgement factors. Case analysis showed cauda-equina-syndrome, space-occupying-lesion, cancer and inflammatory condition identification performed well across validation domains. Fracture performed less well, but with well-defined reasons for the erroneous results.

Conclusions:

A knowledge-based AI system for decision support for SSP was constructed. The tool can be completed in a time period compatible with a patient contact and shows encouraging validity. Further work to improve the existing model and include treatment decision making is needed alongside prospective validation. The prototype tool is ready to be taken forward for refinement and clinical testing.