Table 2.
Summary of machine learning applications in this review.
| Authors | Model(s) | Cohort | Type of outcome | Results |
|---|---|---|---|---|
| Burns et al. (18) | SVM | 150 CT scans | Vertebral compression fractures | SVM achieved sensitivity of 98.7% with a false-positive rate of 0.29. |
| Hoffman et al. (26) | SVM | 27 cervical myelopathy patients | Postoperative ODI score (regression) | SVM was more accurate than multivariate linear regression for postoperative ODI. |
| Hopkins et al. (27) | DNN | 4,046 posterior spinal fusions | Surgical site infections | Neural network employed 35 input variables with a model AUC of 0.79. |
| Hopkins et al. (28) | DNN | 23,264 posterior spinal fusions | 30-day readmissions | Neural network AUC of 0.81. ACS NSQIP database study. |
| Karhade et al. (23) | ANN, BPM†, CART, SVM | 1,790 cases of spinal metastatic disease | 30-day postoperative mortality | Although the neural network had superior discrimination, the Bayes Point Machine was more calibrated and accurate overall. |
| Khan et al. (29) | CART, GAM†, MARS†, PLS†, RF, SVM | 173 cervical myelopathy patients | SF-36 | GBM and Earth models achieved AUC between 0.74 and 0.77 for predicting improvement in PCS-36 over the MCID. |
| Mehta and Sebro (30) | SVM | 370 DEXA scans | Lumbar fracture | SVM detected incidental lumbar fractures on DEXA with an AUC of 0.93 and over 94% sensitivity and specificity. |
| Ogink et al. (22) | ANN, BDT†, BPM†, SVM | 28,600 lumbar surgery patients | Non-home discharge | Neural network had the highest degree of discrimination and calibration. ACS NSQIP database study. |
| Seoud et al. (31) | SVM | 97 adolescents with scoliosis | Scoliosis classification (C1, C2 C3) | 100 surface topography measurements per patient. SVM with one-against-all strategy predicted 72% of cases. |
| Stopa et al. (21) | ANN | 144 lumbar surgery patients | Non-home discharge | External validation of ANN developed by Ogink et al. validation AUC was 0.89 with 0.50 PPV and 0.97 NPV. |
| Tee et al. (32) | CART | 806 traumatic spinal cord injury patients | Cluster analysis | Internal nodes included AIS grade, AOSpine injury morphology, anatomical region, and age. Six clusters were identified. |
| Vania et al. (33) | CNN | 32 CT scans | Spine segmentation | Outcomes included spine, background, and two masking or redundant classifications. Sensitivity and specificity of the algorithm were above 96%. |
| Varghese et al. (34) | CART | 27 pedicle screw pullout conditions | Pedicle screw pullout failure | Three input variables included foam density, screw depth, and screw angle. Correlation between observed and predicted pullout events was 0.99. |
ANN, artificial neural networks; BPM, Bayes point machines; BDT, boosted decision trees; CART, classification and regression decision trees; CNN, convolutional neural networks; DNN, deep neural networks; GAM, generalized additive models; MARS, multivariable adaptive regression splines; PLS, partial least squares; RF, random forests; SVM, support vector machines. †Indicates machine learning modalities not discussed in this review.