A recent study developed models capable of “excellent discrimination” between operative and non-operative management of patients based solely on baseline preoperative values. According to the authors, patient-reported outcome measures (PROMs) were particularly instrumental in making these predictions, and they propose that operative versus non-operative decision-making of experienced surgeons “may be emulated with trained artificial intelligence (AI) algorithms”.
The authors, Alan H Daniels (Brown University, Providence, USA) and colleagues, clarify that these models predict patients who received surgery, and not necessarily those who should undergo or will do well with surgery. They suggest that “future investigations may evaluate the implementation of such models for decision support in the clinical setting”. Daniels recently presented these conclusions at the 26th International Meeting on Advanced Spine Techniques (IMAST 2019; 17–20 July, Amsterdam, The Netherlands), sponsored by the Scoliosis Research Society.
Daniels and colleagues note that adult spinal deformity patients exhibit complex and highly variable pathology and that the decision to operatively manage patients is largely subjective and varies based on surgeon training, preference, and experience. This study therefore sought to develop models capable of discriminating between patients receiving operative versus non-operative treatment.
The investigators performed a retrospective analysis of a multicentre, prospectively-defined, consecutive cohort of 1,503 adult spinal deformity patients. The cohort was divided in a 70:30 split for training and testing, and the outcome measure was operative treatment, defined as those undergoing surgery up to one year after their baseline visit.
Predictive variables included patient demographics, their past medical history, PROMs, and radiographic parameters. Statistical analysis included random forest, elastic net regression, logistic regression, and support vector machines (SVMs) with radial and linear kernels. Model performance was evaluated by AUC.
Daniels and colleagues report that predictive performance was “excellent” for SVM radial, SVM linear, and elastic net (AUC>0.9), and random forest performed “statistically inferior” to the other tested models (p>0.05).
In the training and testing sets, 69% (n=727) and 69.1% (n=311) of patients received operative management, respectively. Upon evaluation with the testing dataset, performance for SVM linear (AUC=0.91), elastic net (0.913), and SVM radial (0.914) models was “excellent”, and random forest model performed “very well” (0.83).
In the SVM radio model, the investigators found that health-related quality of life (HRQoL) metrics were particularly important for making predictions, with the top three most important variables being SRS appearance, SRS total, and ODI.
The investigators summarise: “This study utilised a variety of modern machine learning techniques to predict operative versus non-operative management of adult spinal deformity surgery patients,” and “the best models exhibited excellent discrimination”.