Table 2.

Examples of n-grams that were relevant in identifying acute low back pain notes when using convolutional neural network-based architecture trained with manual annotations. The n-grams’ relevance was determined by analyzing the neurons of the convolutional neural networks activating the max-pooling layers and their log-odds to contribute to the final output. Log-odds were filtered per notes and then averaged over all the notes and evaluation folds.

Type	Acute LBPa-related predictive n-grams
Diagnosis	Muscle spasm lower back Acute LBP flare Been having acute back pain Acute midline LBP Sports acute bilateral LBP Acute low back pain Acute LBP
Related conditions	Gait abnormality Showed significant disk herniation Intermittent sciatica Spinal stenosis
Medications	Back pain flare prescribed flexeril Cyclobenzaprine Flexeril Naproxen for acute low back Prescribed muscle relaxant
Recommendations	Back brace for back pain Obtain lumbar spine MRIb Recommendation RTWc visit RTW full duty quick

^aLBP: low back pain.

^bMRI: magnetic resonance imaging.

^cRTW: return-to-work.