Low back pain a leading cause of disability worldwide and is associated with large costs for people with back pain and for society as a whole.¹ In up to 95% of people with low back pain a specific cause cannot be found and limits clinicians ability to implement the most appropriate treatment.² It is now recognised that low back pain is a heterogenous condition, with the potential to have many underlying features. For example, changes to spine health may increase nociceptive activity (a neural input to the brain). The strength of this nociceptive input may also be increased by changes in the spinal cord and/or the brain, and by psychological factors (e.g. anxiety), prior to being perceived as pain of varying intensity (Figure 1).³ Attempts have been made to sub-group people with low back pain, but these have been largely unsuccessful.⁴ Part of the reason for this is that some approaches focus on pain symptoms and/or duration of pain, which is simplistic.⁵ Other approaches have focused solely on changes in the spine^6,7 or specific pain mechanisms.⁸ Using artificial intelligence and machine learning approaches presents a unique opportunity to detect clinically important patterns in individuals with low back pain who may benefit from targeted treatments.

Figure 1. Various factors that can be associated with the nervous system output of pain in low back pain.

In our work presented in npj Digital medicine here,⁹ we conducted a systematic review to determine how artificial intelligence has been utilised in low back pain research to date. To give insight into the path ahead, we also completed two additional systematic reviews of commonly used low back pain classification schemes (McKenzie method¹⁰ and STarT Back tool¹¹).

We identified 48 studies which used artificial intelligence tools for low back pain. The main results of the study showed that artificial intelligence has primarily been used for the binary classification of low back pain (low back pain versus no back pain). Of these, 25 had assessed classification of low back pain, with 20 reporting an accuracy of over 80% (Figure 2). Most models were applied on small data sets with limited features.

The McKenzie method and STarT Back tool had all been assessed in key domains, however had mixed results for their efficacy. For example, only one of four STarT Back and four of 11 McKenzie studies had significant improvements in pain intensity in trials when compared to non‑classified interventions or no treatment.

Figure 2. Stages of development of artificial intelligence in low back pain research.

Our work shows there is scope to utilise artificial intelligence in low back pain research, but the key steps in using such tools are yet to be performed. A key potential advantage of this approach over existing approaches is that previous classification schemes have not considered all key systems associated with low back pain and tend to have mixed effectiveness.

Artificial intelligence could bridge this gap, by utilising a data driven approach to recognise patterns in clinical presentations. Future research of artificial intelligence should examine a large range of clinical features (e.g. spinal tissue, psychosocial and central nervous system factors) to attempt the subclassification of low back pain. Furthermore, based on our framework here, the classification tool needs to consider reliability, validity, prognostic efficacy and how well it improves pain intensity, disability, and healthcare costs in clinical trials, if it is to have an impact how treating the burden of low back pain.

Acknowledgments: We would like to acknowledge all our co-authors for their contributions to the original manuscript.

References

1  Vos, T. et al. Global, regional, and national incidence, prevalence, and years lived with disability for 328 diseases and injuries for 195 countries, 1990-2016: a systematic analysis for the Global Burden of Disease Study 2016. The Lancet 390, 1211-1259, doi:10.1016/s0140-6736(17)32154-2 (2017).

2   Maher, C., Underwood, M. & Buchbinder, R. Non-specific low back pain. The Lancet 389, 736-747 (2017).

3   Tagliaferri, S. D. et al. Domains of chronic low back pain and assessing treatment effectiveness: A clinical perspective. Pain Practice 1, doi:10.1111/papr.12846 (2019).

4   Fairbank, J. et al. The role of classification of chronic low back pain. Spine 36, S19-S42 (2011).

5   Gifford, L. Aches and pains.  (CNS Press Ltd, 2014).

6  Jensen, R. K., Kent, P., Jensen, T. S. & Kjaer, P. The association between subgroups of MRI findings identified with latent class analysis and low back pain in 40-year-old Danes. BMC Musculoskeletal Disorders 19, 1-12 (2018).

7   Jensen, R. K., Jensen, T. S., Kjaer, P. & Kent, P. Can pathoanatomical pathways of degeneration in lumbar motion segments be identified by clustering MRI findings. BMC Musculoskeletal Disorders 14, 198 (2013).

8  Smart, K. M., Blake, C., Staines, A. & Doody, C. Clinical indicators of ‘nociceptive’,‘peripheral neuropathic’and ‘central’mechanisms of musculoskeletal pain. A Delphi survey of expert clinicians. Manual Therapy 15, 80-87 (2010).

9  Tagliaferri, S. et al. Artificial Intelligence to Improve Back Pain Outcomes and Lessons Learnt from Clinical Classification Approaches: Three Systematic Reviews. npj Digital Medicine 3, doi:https://doi.org/10.1038/s41746-020-0303-x (2020).

10 McKenzie, R. & May, S. The lumbar spine: mechanical diagnosis & therapy. Vol. 1.  (Spinal Publications New Zealand, 2003).

11 Hill, J. C. et al. A primary care back pain screening tool: Identifying patient subgroups for initial treatment. Arthritis Care & Research 59, 632-641 (2008).