Applying machine learning on patient-reported data to model the selection of appropriate treatments for low back pain: A Pilot Study

The objective of this pilot study was to determine whether machine learning can be applied on patient-reported data to model decision-making on treatments for low back pain (LBP). We used a database of a university spine centre containing patient-reported data from 1546 patients with LBP. From this dataset, a training dataset with 354 features (input data) was labelled on treatments (output data) received by these patients. For this pilot study, we focused on two treatments: pain rehabilitation and surgery. Classification algorithms in WEKA were trained, and the resulting models were validated during 10-fold cross validation. Next to this, a test dataset was constructed - containing 50 cases judged on treatments by 4 master physician assistants (MPAs) - to test the models with data not used for training. We used prediction accuracy and average area under curve (AUC) as performance measures. The interrater agreement among the 4 MPAs was substantial (Fleiss Kappa 0.67). The AUC values indicated small to medium (machine) learning effects, meaning that machine learning on patient-reported data to model decision-making processes on treatments for LBP seems possible. However, model performances must be improved before these models can be used in real practice.