A study by Jun S Kim (New York City, USA) and colleagues has demonstrated that an artificially intelligent (AI) system can better predict risk factors for anterior lumbar surgery than logistic regression.
Statistical modelling is commonly used in clinical research to isolate independent risk factors for postoperative complications. However, logistic regression is prone to inaccuracies resulting from an inability to capture the interactions between such factors.
In commonly-used statistical models, “complex interplay between risk factors is rarely accounted for, which can lead to inaccurate patient morbidity and mortality,” the authors write. “Neural networking is a machine-learning classification system inspired by the human brain.”
“Say you have a patient undergoing posterior cervical fusion with diabetes, cardiovascular disease, kyphotic alignment on X-ray and myelopathy,” Kim told Spinal News International. “A patient is not the simple sum of these features but a complex nonlinear interplay between all these features. A deep learning set up can model this to make an accurate prediction, given that it is fed enough data.”
Neural networks can “learn”
This type of network holds a large cluster of “neurons” which “collectively but uniquely weigh the importance of input variables.” By honing the particular neurons themselves, a system can “‘learn’ through repetitive epochs.”
Kim and colleagues hypothesised that such AI systems can “better predict postoperative complications than logistic regression.”
To test their claim, the researchers performed a retrospective cohort analysis on patients undergoing anterior lumbar fusion surgery. Data was obtained from 2011–14 US national surgical records. Splitting the cohort randomly into AI and logistic regression groups, the first two years of data (2011–13) were used to “train” the AI system. The trained model was tested on cases from 2014 “to simulate real world performance.”
To account for class imbalance throughout use of the 2011–2014 data the team used a random under-sampling algorithm, in addition to Bayesian regularisation to prevent overfitting.
“Models were trained with 17 key demographic and operative variables as predictors,” Kim et al write. “We defined postoperative complication as venous thromboembolism, surgical site infection, cardiac complications or mortality.” Thirty-eight patients out of the final study population (n=78) experienced complications. The team evaluated model efficacy using area under the receiver-operator curve (AUROC) as well as accuracy itself.
AI’s accuracy outstripped that of logistic regression
The accuracy of the trained neural network model in predicting complications in this patient cohort (95%) was shown to far outstrip that of standard logistic regression (62%).
AUROC results, too were much greater for the AI model (97%) than logistic regression (61%). In addition, the AI model appeared to be far more fine-tuned to both sensitivity and specificity (92% and 90%, respectively) than logistic regression (62% and 64%).
In this study, which Kim and colleagues explain was the “first case of using AI in spine literature with AUROC and accuracy values”, the model’s demonstrated prediction abilities “far exceeds those of logistic regression.”
Speculating as to the reasons for this dramatic difference, the authors write, “The power of this network lies in its simplicity, with only one hidden layer comprised of neurons.” Machine learning algorithms, he said, are often successful classifiers. However, they can be difficult to interpret because of their complexity.
Kim and colleagues concluded, “The combination of interpretability and accuracy suggests these algorithms can be applied to real time clinical workflow.”
The study was presented as an E-poster at the Scoliosis Research Society Annual Meeting (SRS; 7–9 September, Philadelphia, USA).
