Award: Sanford J. Larson, MD, PhD Award
Authors: Nathan Xie; Peter Wilson; Rajesh Reddy (Sydney, Australia)
The majority of referrals for degenerative lumbar spinal conditions do not lead to surgical intervention. Stratifying referrals from primary care sources based on likeliness to proceed to surgery would not only expedite care for patients that may benefit from surgical intervention, but also allow non-operative treatments/strategies to be implemented for those who are unlikely to benefit.
By identifying clinical and imaging factors associated with progression to surgery, we aimed to develop a Machine Learning model (a branch of Artificial Intelligence) able to calculate the probability that a patient would receive surgery based on these factors.
We identified 55 factors in the literature associated with surgical progression. All patients presenting with an elective lumbar spine complaint between 2013-2018 at a single Australian Tertiary Hospital (n=326) had their medical records reviewed, with data being collected for the potential predictive factors. An Artificial Neural Network (ANN) was constructed, with the outcome being progression to spinal surgery (Yes/No). To compare it with a traditional statistical model, a Logistic Regression (LR) model was created from the same data. These were evaluated on their accuracy, discrimination (Area under ROC Curve (AUC)), and calibration (Hosmer-Lemeshow test (HLT)).
Ten clinical and imaging predictive variables were included as input in the final models. The ANN was able to predict surgical progression with 94.2% accuracy. It also exhibited excellent discriminative ability (AUC = 0.90), with good fit of the data (HLT>0.05). This was superior when compared to the LR model (Accuracy: 87.4%, AUC = 0.86, HLT >0.05).
Both the neural network and regression models predicted surgical progression with a high degree of accuracy. By demonstrating that the operating patterns of single centers can be predicted accurately, the potential for more appropriate/tailored referrals becomes possible, reducing wait-lists and increasing surgical conversion rates.