Artificial intelligence in the detection of low back pain

Accurate clinical discrimination of subjects with back pain is difficult. As an aid to discrimination a probabilistic neural network (PNN) was constructed to differentiate categories of paraspinal muscular fitness. The electromyogram (EMG) power spectra from 65 subjects with and without chronic back pain were used to train the PNN. The PNN was tested by comparing the clinical historical diagnosis of back pain in 33 subjects to the PNN classification. Subjects were placed on a test frame in 30 degrees of lumbar forward flexion. An isometric load of 2/3 maximum voluntary contraction (MVC) was held constant for 30 s whilst surface EMGs were recorded at the level of the left 4th/5th interspace. The raw EMG was filtered, digitized and power spectra were calculated using the Fast Fourier Transform. The power spectrum was loaded into the input layer of a three layer PNN and propagated to the output layer that classified the spectrum as normal, abnormal, or unclassifiable. Ten of eleven normal subjects were correctly classified (specificity 91%). Nine of eleven chronic back pain subjects were correctly classified (sensitivity 82%). One trained athlete and one acute back pain were classified correctly. The system was unable to classify subjects with a past history of back pain that was not chronic. Diagnosis of low back dysfunction using a PNN has been shown to be an accurate method of categorizing normal and chronic back pain subjects. The results in subjects with a past history of back pain at any time of their life illustrate the difficulty of classification of these subjects. Spectral shape and PNN techniques may be useful in identifying subjects with back pain who may be at a high risk in the workplace.