Table 1 Overview of included studies on machine learning and LBP.
Study | Year | N | N LBP | N CON | Type LBP | AI/ML techniques | Utilised for | Summary | Inputs | Train/Test | Sen | Sp | Acc | AUC | Conclusions |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Case−control | |||||||||||||||
Abdullah et al.49 | 2018 | 310 | 210 | 100 | Unclear | K-Nearest Neighbour, Principal Component Analysis, Random Forest | Classification | To predict spinal abnormalities using machine-learning techniques | Pelvic incidence, pelvic tilt, lumbar lordosis angle, sacral slope, pelvic radius, degree spondylolisthesis, pelvic slope, direct tilt, thoracic slope, cervical tilt, sacrum angle and scoliosis slope | Yes | — | — | 0.85 | — | Authors concluded that the KNN classifier outperformed the RF classifier. |
Al Imran et al.50 | 2020 | 310 | 210 | 100 | Unclear | Random Forest, K-Nearest Neighbour, Support Vector Machine | Classification | Enhancing classification performance in low-back pain symptoms | Pelvic incidence, pelvic tilt, lumbar lordosis angle, sacral slope, pelvic radius, degree spondylolisthesis, pelvic slope, direct tilt, thoracic slope, cervical tilt, sacrum angle and scoliosis slope | Yes | — | — | 0.92 | — | Authors concluded that the application of the genetic algorithm-based feature selection approach can improve classification accuracy. |
Ashouri et al.20 | 2017 | 52 | 52 | 28 | Chronic | Support Vector Machine | Classification | Spinal 3D kinematic assessment to classify individuals with chronic low-back pain using machine learning | Five trunk flexion and extension parameters | Yes | 1.00 | 1.00 | 1.00 | — | Authors concluded that quantitative techniques provide clinicians and practitioners with improved discriminating means for predicting and diagnosing low-back disorders. |
Bishop et al.23 | 1997 | 183 | 183 | 80 | Unclear | Artificial Neural Network | Classification | Classifying low-back pain from dynamic motion characteristics | Trunk range of motion and movement velocity | Yes | — | — | 0.86 | — | Authors concluded a neural network based on kinematic data is an excellent predictive model for the classification of low-back pain. |
Bounds et al.53 | 1990 | 200 | 200 | 0 | Other | Multi-Layer Perception, K-Nearest Neighbor | Classification | A comparison of neural networks to other pattern recognition approaches for low-back pain | NR | Yes | — | — | 0.95 | — | Authors concluded that MLP and RBF networks outperform clinicians. |
Caza-Szoka et al.54 | 2015 | 65 | 43 | 22 | Chronic | Naïve Bayes | Classification | Bayesian learning for electromyography in chronic low-back pain | Electromyography data | Yes | — | — | 0.70 | — | Authors concluded this paper outlined the advantage of Naïve Bayesian classification models. |
Caza-Szoka et al.24 | 2016 | 24 | 24 | 12 | Chronic | Artificial Neural Network | Classification | Electromyography array for predicting chronic low-back pain | Electromyography of the paraspinal muscles | Yes | — | — | 0.80 | — | Authors concluded that a nonlinear analysis can be used for CLBP detection. |
Chan et al.55 | 2013 | 40 | 20 | 20 | Chronic | Artificial Neural Network, Artificial Neural Network, Multi-Layer Perception, Decision Tree | Classification | A smart phone-based gait assessment to identify people with low-back pain | Gait features | No | — | — | 0.88 | — | Authors concluded it is feasible to develop a mobile-based tele-care system for monitoring gait. |
Darvishi et al.25 | 2017 | 160 | 160 | 92 | Unclear | Artificial Neural Network, Logistic Regression, K-Nearest Neighbor | Classification | Prediction of low-back pain severity in industrial workers based on personal, psychological, and occupational factors | Age, gender, body mass index, smoking status, alcohol status, family history, SMWL, job stress, job satisfaction, job security, social relations, force, repetition, posture, and career length | Yes | — | — | 0.92 | — | Authors concluded that a neural network prediction model was more accurate than regression methods. |
Du et al.57 | 2018 | 171 | 88 | 83 | Chronic | Support Vector Machine | Classification | Using surface electromyography to detect chronic low-back pain | Electromyography data | Yes | — | — | 0.98 | — | Authors concluded the models recognised chronic low-back pain with high accuracy. |
Hu et al.28 | 2018 | 44 | 44 | 22 | Chronic | Artificial Neural Network | Classification | Deep learning to identify low-back pain during static standing | Angular rotation, linear translation and centre of pressure measures | Yes | — | — | 0.97 | 0.99 | Authors concluded that the deep learning neural networks could be used to accurately differentiate LBP populations from healthy controls using static balance performance. |
Hung et al.29 | 2014 | 52 | 52 | 26 | Chronic | Artificial Neural Network, Principal Component Analysis | Classification | Electromyography to classify low-back pain from lifting capacity evaluation | Erector spinae muscle activity (including 30 and 50% loading) during lifting tasks | No | 0.90 | 0.88 | 0.89 | 0.93 | Authors concluded that features with different loadings (including 30 and 50% loading) during lifting can distinguish healthy and back pain subjects. |
Jin-Heeku et al.32 | 2018 | 1510 | 1510 | 883 | Unclear | Support Vector Machine | Classification | Analysis of sitting posture predicting low-back pain | Data from pressure sensors to assess sitting posture | Yes | 1.00 | 1.00 | 1.00 | — | Authors concluded that a support vector machine can classify individuals with CLBP. |
LeDuff et al.34 | 2001 | 59 | 59 | NR | Unclear | Artificial Neural Network | Classification | Data mining medical records to understand low-back pain treatment pathways | Number of contacts with the different kinds of health professionals, medicines and total costs | Yes | — | — | 0.91 | — | No specific conclusions. |
Melo Riveros et al.40 | 2019 | 310 | 310 | 210 | Other | Artificial Neural Network, K-Means Clustering, Self-Organising Map | Classification | Diagnosing spinal pathology from low-back positional characteristics | Pelvic incidence, pelvic inclination, angle of lordosis, sacral slope, pelvic radius and degree of spondylolisthesis | Yes | 0.79 | 0.92 | 0.83 | — | Authors concluded the solution obtained with self-organising maps provides better results with respect to the solution obtained with K-means. |
Oliver et al.41 | 1995 | 98 | 98 | 62 | Other | Artificial Neural Network | Classification | Electromyography to predict low-back pain. | Electromyography data (power spectra) | Yes | 0.82 | 0.91 | 0.92 | — | Authors concluded that the electromyography signals and ML techniques may be useful for identifying back pain patients. |
Oliver et al.42 | 1996 | 60 | 60 | 27 | Chronic | Artificial Neural Network | Classification | Electromyography to predict low-back pain | Electromyography data (power spectra) | Yes | 0.80 | 0.79 | — | — | Authors stated that artificial intelligence neural networks appear to be a useful method of differentiating paraspinal power spectra in back pain sufferers. |
Olugbade et al.62 | 2015 | 53 | 23 | 30 | Chronic | Support Vector Machine | Classification | Pain level prediction and classification using kinematics and muscle activity | Trunk flexion kinematics and EMG, sit-to-stand kinematics and EMG and depression | Yes | — | — | 0.94 | — | Authors concluded the model had very good performance due to thorough analyses. |
Parsaeian et al.44 | 2012 | 34,589 | 34,589 | 7286 | Unclear | Artificial Neural Network | Classification | Predicting low-back pain based on lifestyle and psychosocial characteristics | Age, sex, education level, urban versus rural, smoker versus non-smoker, strenuous versus non-strenuous working conditions, BMI, mental health disorders and marital status | Yes | — | — | — | 0.75 | Authors concluded that an artificial neural network approach yielded better performance than logistic regression but that the difference would not be clinically significant. |
Sandag et al.63 | 2018 | 310 | 210 | 100 | Unclear | K-Nearest Neighbour, Logistic Regression, Naïve Bayes, Random Forest, Decision Tree | Classification | Classification of low-back pain using K-Nearest Neighbour algorithm | Pelvic incidence, pelvic tilt, lumbar lordosis angle, sacral slope, pelvic radius, degree spondylolisthesis, pelvic slope, direct tilt, thoracic slope, cervical tilt, sacrum angle and scoliosis slope | Yes | — | — | 0.92 | — | Authors concluded K-Nearest Neighbour approaches could be used to help further classify low-back pain individuals. |
Silva et al.47 | 2015 | 12 | 12 | 5 | Unclear | Support Vector Machine | Classification | Identifying low-back pain in golfers off muscle activity and swing kinematics | Electromyography during golf swing and kinematic variables of golf swing | Yes | — | — | 1.00 | — | Authors concluded that low-back pain golfers showed different neuromuscular coordination strategies when compared with asymptomatic golfers. |
Ung et al.64 | 2014 | 94 | 47 | 47 | Chronic | Support Vector Machine | Classification | Multivariate classification of chronic low-back pain on structural MRI data | Structural brain MRI data | Yes | — | — | 0.76 | — | Authors concluded support vector machines could classify chronic low-back pain based on grey matter changes. |
Karabulut et al.58 | 2014 | 310 | 210 | 100 | Unclear | Synthetic Minority Technique, Logistic Model Tree | Diagnosis | Automated predictions of vertebral pathologies with a logistic model tree | Pelvic incidence, pelvic tilt, lumbar lordosis angle, sacral slope, pelvic radius, degree spondylolisthesis, pelvic slope, direct tilt, thoracic slope, cervical tilt, sacrum angle and scoliosis slope | Yes | — | — | 0.90 | — | Authors concluded that the machine-learning techniques reasonably accurate classification. |
Mathew et al.38 | 1988 | 200 | 200 | 200 | Other | Fuzzy Logic | Diagnosis | Classifying nerve root compression, simple low-back pain, spinal pathology and abnormal illness behaviour. | Age, sex, site of pain, duration of pain, type of onset, relationship to physical activity and movement, neurological symptoms, inappropriate symptoms, red- and yellow-flags in history and spinal deformity | Yes | — | — | 0.90 | — | Authors stated that the AI techniques can be used for the differential diagnosis of low-back disorders and can outperform clinicians. |
Mathew et al.61 | 1989 | 150 | 150 | 0 | Surgery | Computer Diagnostic System | Diagnosis | Prediction of operative findings in low-back surgery | Age, sex, site of pain, duration of pain, type of onset, relationship to physical activity and movement, neurological symptoms, inappropriate symptoms, red- and yellow-flags in history and spinal deformity | Yes | — | — | 0.92 | — | Authors concluded that this computer system has the potential to facilitate assessment on a large number of patients. |
Vaughn et al.65 | 1998 | 198 | 198 | 0 | Other | Multi-Layer Perception | Diagnosis | Knowledge extraction from a multilayer network for low-back classification | Demographic data, present and past symptoms, pain description/behaviour, finding from physical examination (lumbar spinal movements, tension tests, neurological tests), Oswestry Disability Index, Zung depression index, modified somatic perception questionnaire, the distress and risk assessment method | Yes | — | — | 0.96 | — | Authors concluded that future work should seek to automatically endure a valid rule for each input case to enhance the network. |
Vaughn et al.66 | 2001 | 196 | 196 | 0 | Other | Multi-Layer Perception | Diagnosis | MLP network for the classification of low-back pain | Demographic data, present and past symptoms, pain description/behaviour, finding from physical examination (lumbar spinal movements, tension tests, neurological tests), Oswestry Disability Index, Zung depression index, modified somatic perception questionnaire, the distress and risk assessment method | Yes | — | — | 0.77 | — | Authors concluded a full explanation facility interprets the output on a case-by-case basis. |
Vaughn et al.48 | 2001 | 198 | 198 | 198 | Other | Artificial Neural Network | Diagnosis | Classifying nerve root compression, simple low-back pain, spinal pathology and abnormal illness behaviour | Demographic data, present and past symptoms, pain description/behaviour, finding from physical examination (lumbar spinal movements, tension tests, neurological tests), Oswestry Disability Index, Zung depression index, modified somatic perception questionnaire and the distress and risk assessment method | Yes | — | — | 0.82 | — | Authors stated that application of the method leads to the discovery of a number of mis-diagnosed training and test cases and to the development of a more optimal low-back-pain MLP network. |
Sari et al.45 | 2012 | 169 | 169 | 110 | Unclear | Artificial Neural Network, Fuzzy Inference System | Other | Predicting low-back pain intensity based on pain intensity and skin resistance | Skin resistance and pain intensity | Yes | — | — | — | — | Authors stated that their designed systems are effective to predict the pain intensity level objectively. |
Cohort | |||||||||||||||
Magnusson et al.37 | 1998 | 27 | 27 | 0 | Chronic | Artificial Neural Network | Classification | Range of motion and motion patterns following rehabilitation in low-back pain | Trunk motion data from eight motion tests | Yes | — | — | 0.78 | — | Authors stated that a neural network based on kinematic variables is an excellent model for classification of low-back-pain dysfunction. |
Azimi et al.21 | 2014 | 168 | 168 | 0 | Spinal Stenosis | Artificial Neural Network | Prognosis | Predicting surgical satisfaction for lumbar spinal canal stenosis with artificial neural networks | Age, pain intensity, stenosis ratio, walking distance, Japanese Orthopaedic Association score for assessing LBP, and Neurogenic Claudication Outcome Score | Yes | — | 0.41 | 0.97 | 0.81 | Authors concluded that artificial neural network approach more accurate in predicting 2-year post-surgical satisfaction than a logistic regression model. |
Azimi et al.22 | 2015 | 402 | 402 | 0 | Recurrent | Artificial Neural Network | Prognosis | Predicting recurrent lumbar disc herniation with artificial neural networks | Age, sex, duration of symptoms, smoking status, recurrent LDH, level of herniation, type of herniation, sports activity, occupational lifting, occupational driving, duration of symptoms, visual analogue scale, the Zung Depression Scale, and the Japanese Orthopaedic Association Score | Yes | — | 0.46 | 0.94 | 0.84 | Authors concluded that artificial neural networks can be used to predict recurrence of lumbar disc herniation. |
Barons et al.52 | 2013 | 701 | 701 | 0 | Unclear | Artificial Neural Network, Latent Class Analysis, Logistic Regression | Prognosis | Determining who benefits from cognitive behavioural therapy | RMDQ, FABQ, PSE, SF-12, HADS | No | — | — | 0.61 | — | Authors concluded that artificial neural networks would be the best candidate to support treatment allocation. |
Hallner et al.27 | 2004 | 71 | 71 | 0 | Acute | Artificial Neural Network | Prognosis | Identifying individuals at risk of chronic low-back pain based on yellow-flags | Pain intensity at the beginning of hospitalisation, Beck Depression Inventory and Kiel Pain Inventory | Yes | 0.73 | 0.97 | 0.83 | — | Authors concluded that this model could contribute to the early detection of risk factors for patients with acute low-back pain, and could assist with avoiding chronicity. |
Jarvik et al.30 | 2018 | 4665 | 4665 | 0 | Acute | LASSO Model | Prognosis | Predicting recovery from acute low-back pain in older adults | Age, gender, race, ethnicity, education, employment status, marital status, smoking status, the duration of current episode of back or leg pain, back-related claim or lawsuit, patient confidence that their back or leg pain would be completely gone or much better in 3 months, baseline pain-related characteristics, baseline psychological distress, baseline falls, BMI, comorbidity score, baseline diagnosis, spine-related interventions and opioid prescriptions | Yes | — | — | — | 0.75 | Authors concluded that baseline patient factors were more important than early interventions in explaining disability and pain after 2 years. |
Jiang et al.31 | 2017 | 78 | 30 | 48 | Chronic | Support Vector Machine | Prognosis | Electromyography for prediction of recovery following functional restoration | Electromyography during left lateral bending, right lateral bending, left turning, right turning | Yes | 1.00 | 0.94 | 0.97 | 0.89 | Authors stated that the tools can be used to identify patients who will respond to functional restoration rehabilitation. |
Shamim et al.46 | 2009 | 501 | 501 | 0 | Surgery | Fuzzy Inference System | Prognosis | Prediction of poor outcomes following lumbar disc surgery | Sex, BMI, occupation, marital status, use of oral corticosteroids, multilevel disease, epidural steroid injection, duration of symptoms, duration of non-operative treatment, extent of changes on MRI, previous spine surgery, emergency versus elective surgery, operative time, intraoperative complications, operating surgeon and post-op complications | No | 0.88 | 0.86 | — | — | Authors concluded a fuzzy inference system is a sensitive method of predicting patients who will fail to improve with surgical intervention. |
Other | |||||||||||||||
Kadhim et al.33 | 2018 | 10 | 10 | 0 | Unclear | Fuzzy Inference System | Classification | A decision support system for back pain diagnosis | Sex, height, weight, age and a series of clinical symptoms | No | — | — | 0.84 | — | Author stated that the proposed system can be used by domain experts (physicians) to help enhance decision-making. |
Lee et al.19 | 2019 | 53 | 53 | 0 | Chronic | Support Vector Machine | Classification | Prediction of clinical pain intensity from functional connectivity and autonomic states | Functional connectivity and heart rate variability | Yes | — | — | 0.92 | 0.97 | Authors concluded that a machine-learning approach model identifies putative biomarkers for clinical pain intensity. |
Lin et al.60 | 2006 | 180 | 180 | 0 | Unclear | Naïve Bayes | Diagnosis | A decision support system for low-back pain diagnosis | Gender, age, current pain symptoms, clinical pain history, pregnancy history, number and tingling | No | — | — | 0.73 | — | Authors concluded the system provides an easy-to-follow framework for low-back pain. |
Andrei et al.51 | 2015 | 260 | 260 | 0 | Other | Fuzzy Inference System | Prognosis | Computer-aided patient evaluation of low-back pathology | Pain, calories, flexion, extension, rotation and lateral flexion range of motion | No | — | — | 0.98 | — | Authors concluded a complex fuzzy system is essential for lumbar spine pathology. |
Li et al.59 | 2017 | 100 | 100 | 0 | Unclear | Artificial Neural Network, K-Nearest Neighbor, Fuzzy Inference System | Prognosis | Probabilistic Fuzzy classification for Stochastic data | Pain area, height and width of pain area and ratio | No | — | — | NR | — | Authors concluded more information can be extracted from limited samples using a PFC approach. |
Dickey et al.56 | 2000 | 9 | 9 | 0 | Chronic | Artificial Neural Network | Other | Relationship between pain and spinal motion characteristics in low-back pain | 32 spinal motion parameters | No | — | — | 0.99 | — | Authors concluded they observed clear patterns of segmental spinal motion in low-back pain. |
Liszka-Hackzell et al.35 | 2002 | 40 | 40 | 0 | Other | Artificial Neural Network | Other | Categorising individuals with low-back pain based on self-report and activity data | Unclear | Yes | — | — | — | — | Authors stated that that neural network techniques can be applied effectively to categorising patients with acute and chronic low-back pain. |
Liszka-Hackzell et al.36 | 2005 | 18 | 18 | 0 | Chronic | Artificial Neural Network | Other | Analysis of night-time activity and daytime pain in chronic low-back pain | Measures of sleep quality through actigraphy | Yes | — | — | — | — | Authors concluded that daytime pain levels are not correlated with sleep the night before, nor with the night following. |
Meier et al.39 | 2018 | 20 | 20 | 0 | Chronic | Multivariate Patten Analysis | Other | Predicting neural adaptions based on psychosocial constructs | Bilateral fear-related brain regions including the amygdala, hippocampus, thalamus, anterior cingulate, insula, and medial prefrontal, and orbitofrontal cortices | Yes | — | — | — | — | Authors stated the approach might ultimately help to further understand and dissect psychological pain-related fear. |
Gal et al.26 | 2015 | 15 | 15 | 0 | Unclear | Fuzzy Inference System | Treatment allocation | Computer-assisted prediction of low-back pain treatment | Sex, age, disability level, daily activity expressed in calories and trunk mobility measures | No | — | — | — | — | Authors concluded the system has the ability to identify the correct treatment and can ensure the quality of the treatment. |
Oude et al.43 | 2018 | 45 | 45 | 0 | Unclear | Boosted Tree, Decision Tree, Random Forest | Treatment allocation | To determine if self-referral is possible in individuals with low-back pain | Age, well-being index, duration of pain, use of analgesics, history of trauma, use of corticosteroids, presence of specific serious disease, weight loss in past month, constant pain, night-time pain, pain with lifting/sneezing/coughing, radiating pain, reduced muscle strength, cauda equina symptoms, referral preference | Yes | — | — | 0.72 | — | Authors stated that the study showed possibilities of using ML to support patients with LBP in their self-referral process to primary care. |
