TABLE 5.
Modeling and model training strategy.
| Article | Binary classifier | Swallow sample (aspirated/unsafe vs. normal) | Reference test | Training strategy |
|---|---|---|---|---|
| Frakking et al. (2022) | SVM | 18 vs. 106 | VFSS | 50:50 training-to-testing ratio, 5-fold CV for hyperparameter tuning |
| Lee et al. (2006) | RBF | 94 v. 100 | VFSS | 10-fold CV |
| Lee et al. (2011) | 3 channels (airway invasion, valleculae clearance and pyriform sinuses bolus clearance) on 9 classifiers (LDA Euclidean, LDA Mahalanobis, NN (10, 20, 30 HUs), PNN, and KNN (K = 11, 21, 31) | Airway invasion: 39 vs. 265 Valleculae BC: 64 vs. 61 Pyriform sinuses BC: 25 vs. 129 | VFSS | 10-fold CV |
| Merey et al. (2012) | LDA w/Euclidean, LDA w/Mahalanobis, SVM linear, SVM RBF, SVM RBF + B2 optimizer | 94 vs. 544 | VFSS | 8-fold CV, bootstrapping to balance class |
| Park et al. (2022) | Logistic Regression, Decision Tree, Random Forest, SVM, GMM, XGBoost | N/A (per-patient) | VFSS and spirometry | - |
| Sarraf Shirazi et al. (2014) | SVM | N/A (per-patient) | VFSS or FEES | Leave-one-out |
| Sarraf Shirazi et al. (2012) | Minimum distance classifier | N/A (per-patient) | VFSS or FEES | Leave-one-out |
| Fuzzy k-means clustering | 32 vs. 128 | |||
| Sejdic et al. (2013) | Bayes | - | VFSS | Leave-one-out |
| Shu et al. (2022) | SVM, k-means, Naive Bayes, ANN | 378 vs. 1701 | VFSS | 10-fold CV |
ANN, artificial neural network; BC, bolus clearance; CV, cross-validation; FEES, fiberoptic endoscopic evaluation of swallowing; GMM, gaussian mixture model; HU, hidden units; KNN, k-nearest-neighbor; LDA, linear discriminant analysis; N/A, not applicable; NN, feed-forward non-linear classifier; PNN, probabilistic neural network; RBF, radial basis function; SVM, support vector machine; XGBoost, Extreme gradient boosting; VFSS, videofluoroscopic swallowing study; w/: with.