Table 1 AI methods on osteoporotic fracture prediction
Author | Time | Country | Mortality | Subjects | AI algorithm | Train/validation/test set | Best result |
|---|---|---|---|---|---|---|---|
Whittier et al. [36] | 2022 | Canada | HR-pQCT | 5873 patients | Fuzzy c-means clustering | 446 train and 5873 test | HR = 2.96 |
Shimizu et al. [37] | 2022 | Japan | Database | 7033 patients | DT, Feature Selection and Relative Importance, ANN, and SVM | 75% train and 25% test | AUC = 0.74 |
Kong et al. [38] | 2022 | Korea | Database | 1595 participants | DeepSurv | 1416/1595 train (fivefold CV), and 179/1595 test | C-index = 0.614 |
Dong et al. [39] | 2022 | USA | Database | 4461 subjects and 15,524 spine radiographs | GoogLeNet | 76.5% train, 8.5% validation, and 15% test | AUC-ROC = 0.99, AUC-ROC = 0.82, sensitivity = 59.8%, PPV = 91.2%, and F1 score = 0.72 |
Chen et al. [40] | 2022 | China | Database | 14,419 patients | XGBoost combining MLP | 80% train and 20% test | AUC = 0.9 (approximately), accuracy = 90.38%, and F1 score = 0.9037 |
Ulivieri et al. [41] | 2021 | Italy | Database | 172 women | 2 derivative algorithms of ANN | 90/172 train and 82/172 test, then reverse to 82/172 train and 90/172 test | AUC = 0.896, accuracy = 82.93%, sensitivity = 82.14%, specificity = 83.72% |
Nissinen et al. [42] | 2021 | Finland | DXA | 2949 + 459 women and 115 men | Convolutional neural network (CNN) | 2949/3523 train (tenfold CV), and 574/3523 test | AUC = 0.64, accuracy = 52.0%, sensitivity = 67.8%, specificity = 51.4% |
de Vries et al. [43] | 2021 | Netherland | Database | 7578 patients | CR, RSF and ANN-DeepSurv model | 100% train and 100% test | C-index = 0.625 |
Wu et al. [44] | 2020 | USA | Database | 5130 men | RF, NN, LR, and gradient boosting, | 80% train (tenfold CV), and 20% test | AUC = 0.71, Accuracy = 0.88 |
Villamor et al. [45] | 2020 | Spain | Database | 137 patients | SVM, RBF, LR, SNN, and RF | 101/137 train (tenfold CV), and 36/137 test | Accuracy = 0.86 |
Galassi et al. [46] | 2020 | Spain | Database | 137 patients | SVM, LR, DT, and RF | 70% train (twofold CV), and 30% test | Accuracy over 87%, Specificity over 92%, and Sensitivity over 83% |
Engels et al. [47] | 2020 | Germany | Database | 288,086 individuals | SL, XGBoost, LR, RF, SVM and RUS | 80% train (tenfold CV), 20% test | AUC = 0.72 |
Almog et al. [48] | 2020 | USA | Database | 6,329,986 patients | Crystal Bone | 50% train (threefold CV), 50% test | AUC = 0.81 |
Su et al. [49] | 2019 | USA | Database | 5994 men | CARTs | tenfold CV | AUC = 0.726 |
Muehlematter et al. [50] | 2019 | Switzerland | CT | 60 stable and 60 unstable vertebrae of 58 patients | MLP, ANN, RF, SVM, and naïve Bayesian classifier | 2/3 train (tenfold CV), 1/3 test | AUC = 0.97 |
Ferizi et al. [51] | 2019 | USA | Database | 92 women | linear models, SVM, DT, KNN, and EL | 22/23 train (23-fold CV), 1/23 test | Specificity = 0.83(adjusted), Accuracy = 0.71(adjusted), Precision = 0.68, F1-score = 0.67(adjusted) |
Kruse et al. [52] | 2017 | Denmark | Database | 10,775 women | Standardized variable means, Euclidean distances, and Ward's D2 method of HAC | Not required | Nine (k = 9) clusters were identified |
Kruse et al. [53] | 2017 | Denmark | Database | 4722 women and 717 men | Classification Tree, BAT, BGLM, PLS, KNN, LogitBoost, BGAM, HDDA, RF, C5.0, CIT, LMT, SGB, QDA, LDA, BFDA, BMARS, NSC, SVMRW, NN, NNFE, XGB, CIRF, and AB | 75% train (fivefold CV), 25% test | AUC = 0.92 |
Schuler et al. [54] | 2010 | Australia | CT | 100 | InShape model | Not presented | R = 0.91 |