Introduction: Lumbar disc herniation (LDH) is a common identifiable causes of low back pain with or without unilateral leg pain and is often treated surgically with microdiscectomy when non-operative measures fail to provide relief for at least 6 weeks. However, there is much debate about the efficacy of surgery compared to continuing conservative treatment. Discectomy has been found to be more effective in alleviating short-term symptoms, but is associated with higher rates of adverse events, in addition to re-herniation requiring re-operation in 6% of patients and recurrent low back pain and persistent disability in up to 25% of patients undergoing microdiscectomy. Therefore, developing robust prognostic tools for risk prediction of these adverse outcomes would be valuable in guiding treatment and patient expectations.
Method: We are performing a retrospective review of 120 patients who presented at Spine Service clinics and subsequently underwent microdiscectomy surgery at St. George Private Hospital between January 2010 and December 2019. Demographic data, clinical data and patient reported outcomes are being collected for each patient. Patient reported outcomes are collected Spine Service forms, based on AAOS-MODEM’s outcomes, and include pain, impact, disability and expectation profiles before and after the surgery. The data will be randomly split into training, validation and test sets in a 60%/20%/20% ratio. The training set will be subjected to train various machine learning algorithms, including gradient boosting (ensemble of weak prediction models), extreme gradient boosting, decision tree, elastic net regularisation, random forest, generalised linear modelling, and neural network.
The hyperparameters will be tuned on the validation set, before the model is assessed on the test set. Discrimination will be measured as AUROC values in the test set. Calibration will be determined using the calibration slope on the test set. Effective techniques can be applied to other surgical fields and can pave the way for external validation and impact studies that contribute to the integration of predictive modelling into clinical practice.