Development of a machine-learning based model for predicting multidimensional outcome after surgery for degenerative disorders of the spine.

BACKGROUND: It is clear that individual outcomes of spine surgery can be quite heterogeneous. When consenting a patient for surgery, it is important to be able to offer an individualized prediction regarding the likely outcome. This study used a comprehensive set of data collected over 12 years in an in-house registry to develop a parsimonious model to predict the multidimensional outcome of patients undergoing surgery for degenerative pathologies of the thoracic, lumbar or cervical spine. METHODS: Data from 8374 patients (mean age 63.9 (14.9-96.3) y, 53.4% female) were used to develop a model to predict the 12-month scores for the Core Outcome Measures Index (COMI) and its subdomain scores. The data were split 80:20 into a training and test set. The top predictors were selected by applying recursive feature elimination based on LASSO cross validation models. Based on the 111 top predictors (contained within 20 variables), Ridge cross validation models were trained, validated, and tested for each of 9 outcome domains, for patients with either "Back" (thoracic/lumbar spine) or "Neck" (cervical spine) problems (total 18 models). RESULTS: Among the strongest outcome predictors in most models were: preoperative scores for almost all COMI items (especially axial pain (back or neck) and peripheral pain (leg/buttock or arm/shoulder)), catastrophizing, fear avoidance beliefs, comorbidity, age, BMI, nationality, previous spine surgery, type and spinal level of intervention, number of affected levels, and surgeon seniority. The R2 of the models on the validation/test sets averaged 0.16/0.13. A preliminary online tool was programmed to present the predicted outcomes for individual patients, based on their presenting characteristics. https://linkup.kws.ch/prognostictool . CONCLUSION: The models provided estimates to enable a bespoke prediction of the outcome of surgery for individual patients with varying degenerative pathologies and baseline characteristics. The models form the basis of a simple, freely-available online prognostic tool developed to improve access to and usability of prognostic information in clinical practice. It is hoped that, following confirmation of its validity and practical utility, the tool will ultimately serve to facilitate decision-making and the management of patients' expectations.

Can the Charlson Comorbidity Index be used to predict the ASA grade in patients undergoing spine surgery?

BACKGROUND: The American Society of Anaesthesiologists' Physical Status Score (ASA) is a key variable in predictor models of surgical outcome and "appropriate use criteria". However, at the time when such tools are being used in decision-making, the ASA rating is typically unknown. We evaluated whether the ASA class could be predicted statistically from Charlson Comorbidy Index (CCI) scores and simple demographic variables. METHODS: Using established algorithms, the CCI was calculated from the ICD-10 comorbidity codes of 11'523 spine surgery patients (62.3 ± 14.6y) who also had anaesthetist-assigned ASA scores. These were randomly split into training (N = 8078) and test (N = 3445) samples. A logistic regression model was built based on the training sample and used to predict ASA scores for the test sample and for temporal (N = 341) and external validation (N = 171) samples. RESULTS: In a simple model with just CCI predicting ASA, receiver operating characteristics (ROC) analysis revealed a cut-off of CCI ≥ 1 discriminated best between being ASA ≥ 3 versus < 3 (area under the curve (AUC), 0.70 ± 0.01, 95%CI,0.82-0.84). Multiple logistic regression analyses including age, sex, smoking, and BMI in addition to CCI gave better predictions of ASA (Nagelkerke's pseudo-R2 for predicting ASA class 1 to 4, 46.6%; for predicting ASA ≥ 3 vs. < 3, 37.5%). AUCs for discriminating ASA ≥ 3 versus < 3 from multiple logistic regression were 0.83 ± 0.01 (95%CI, 0.82-0.84) for the training sample and 0.82 ± 0.01 (95%CI, 0.81-0.84), 0.85 ± 0.02 (95%CI, 0.80-0.89), and 0.77 ± 0.04 (95%CI,0.69-0.84) for the test, temporal and external validation samples, respectively. Calibration was adequate in all validation samples. CONCLUSIONS: It was possible to predict ASA from CCI. In a simple model, CCI ≥ 1 best distinguished between ASA ≥ 3 and < 3. For a more precise prediction, regression algorithms were created based on CCI and simple demographic variables obtainable from patient interview. The availability of such algorithms may widen the utility of decision aids that rely on the ASA, where the latter is not readily available.