Editorial: Predicting surgical satisfaction using artificial neural networks

Abstract

In this issue of the Journal of Neurosurgery: Spine, the clinical article titled “Use of artificial neural networks to predict surgical satisfaction in patients with lumbar spinal canal stenosis” by Azimi and colleagues1 represents the first application of an artificial neural network (ANN) in predicting outcomes from spinal surgery. Although ANNs have been used in the biomechanical assessments of spine disease, to date the literature lacks any examples of the clinical applications of ANNs in patients with spine disease. The authors undertook this initiative by identifying 168 patients who had undergone surgery for lumbar stenosis and who had completed a series of preoperative grading instruments. Along with age, sex, and symptom duration, scores on these instruments were collectively taken as the inputs for the ANN. The network was trained on half of the patients by pairing the inputs with known outputs. When a novel cohort of patients was then used to test the model, it statistically outperformed a standard logistic regression (LR) model in terms of accuracy in predicting surgical satisfaction. Notably, the authors incorporated only preoperative data points in an effort to predict long-term outcome. One could reasonably expect that the most important factors in predicting this include biomechanical and surgical factors such as sagittal balance, the number of levels decompressed, and whether or not the interspinous ligaments are preserved during decompression, just to name a few. Accurate prediction of longer-term outcomes would probably need such variables incorporated. It would not only be simple to model an end point such as whether or not a patient receives a spinal fusion, but this would have practical implications if such a model were ever to be used in clinical practice. One of the popular criticisms of ANN modeling is that the variables exist as a “black box,” with the exact relationships and interaction between them remaining somewhat abstract. Although this is true, what can still be gleaned easily from the model is which of the variables is most salient in predicting outcome. A simple visual representation of the model in the authors’ Fig. 3 readily reveals that the stenosis ratio has the greatest impact in predicting patient satisfaction. We remain quite a distance from seeing the use of ANNs in practice. To date there are well under a dozen examples of ANN models applied to neurosurgical diseases. The few examples that exist, however, illustrate the power of this form of modeling over traditional linear models, and also over clinical intuition alone. To date ANNs have been shown to outperform LR models at predicting 1-year survival in patients with brain metastases,4 have been shown to be more accurate than LR models at predicting in-hospital survival in patients with traumatic brain injury,6 more accurate than either LR models or clinicians in predicting in-hospital survival from traumatic brain injury,5 more accurate than LR models in predicting survival from intracerebral hemorrhage,3 and more accurate at predicting which patients with subarachnoid hemorrhage are likely to suffer from cerebral vasospasm.2 That the demonstrated accomplishments of ANNs in neurosurgery can be listed so briefly hints at their unexplored potential. It is worth emphasizing that the authors used a commercially available software package to generate their ANN model. This nicely highlights the ease with which such models can be developed. As clinical decision making becomes increasingly complex, clinical decision support tools will perhaps play an increasingly useful role in clinical practice. At the present time, however, such clinical informatics remain widely underdeveloped and underutilized, especially in the realm of neurosurgical diseases. To this end, the authors have contributed a simple proof of concept of the application of ANNs to neurosurgical decision making. (http://thejns.org/doi/abs/10.3171/2013.10.SPINE13851)