Table 4. Characteristics and validation performance of ML^a prediction models in included studies (n=23).

Author	Predictors categories	ML algorithms	Best model	Internal validation (test set performance)	External validation performance	Validation methods
Akazawa and Hashimoto [10]	MRIb, laboratory parameters, demographic characteristics	Multimodal DLc, XGBoostd, VGG16e	Multimodal DL	AUCf=0.73 (95% CI 0.66‐0.80), Accuracy=0.68	Not reported	Random split (8:2), cross-validation
Akazawa and Hashimoto [18]	Radiomics features, clinical variables	LRg	LR	AUC=0.69 (95% CI 0.62‐0.75)	AUC=0.70 (95% CI 0.66‐0.73)	Internal: random split (7:3), external: another institution
Chen et al [12]	Clinical variables	Bayesh, MLPi, DTj, KNNk, LR, RFl, SVMm, XGBoost	Bayes	AUC=0.82 (95% CI 0.80‐0.84), Sensitivity=0.93, Specificity=0.42, F score=0.92	AUC=0.85 (95% CI 0.83‐0.87), Sensitivity=0.95, Specificity=0.50, F score=0.96	Internal validation: 10-fold cross-validation, (8:2 split), multicenter external validation
de Reus DC et al [19]	Tumor type, ECOGn score, surgical procedure, preoperative platelet count	Not reported	Not reported	Not reported	AUC=0.63 (95% CI 0.58‐0.68), Sensitivity=0.74, Specificity=0.41, F score=0.33	Multicenter external validation
Li et al [20]	Demographic characteristics, laboratory parameters, imaging characteristics, pathological characteristics	LR	LR	AUC=0.80	Not reported	Random split (training set:test set=7:3)
Liu et al [21]	MRI	DL	VGG16	Accuracy=0.75, Sensitivity=0.73, Specificity=0.77	Not reported	5-fold cross-validation
Mo et al [22]	Clinical variables	DNNo	DNN	Accuracy=0.91, Sensitivity=0.89, Specificity=0.92, Precision=0.92	Not reported	Training:test=6:4
Park et al [23]	Laboratory parameters, surgical parameters, MELDp score, demographic characteristics	LR, Elastic Net, SVM, RF, XGBoost, NNq	LR	AUROCr=0.84, AUPRs=0.82	Not reported	Training:test=7:3, feature selection via nested cross-validation
Shi et al [13]	Tumor type, ECOG score, surgical procedure, preoperative platelet count	LR, KNN, DT, XGBoost, RF, SVM	XGBoost	AUC=0.85 (95% CI 0.82‐0.87), Accuracy=0.77, Recall=0.85, F score=0.78, Precision=0.72	AUC=0.80(95% CI 0.77‐0.86), Accuracy=0.73, Recall=0.73, F score=0.73, Precision=0.73	Internal validation: random split (7:3 ratio), external validation: independent cohort
Shi et al [24]	Surgical parameters, laboratory parameters, demographic characteristics	LGBt, XGBoost, CatBu, AdaBv, LR, LSTMw, MLP	LGB	AUC=0.93, Accuracy=0.87, Sensitivity=0.8, Specificity=0.85	Not reported	Training:test =2:1, ADASYNx was used to address data imbalance
Stehrer et al [25]	Surgical parameters, laboratory parameters, demographic characteristics	RF	RF	Regression performance: significant correlation between predicted and actual values; mean error 7.4 (SD 172.3) mL	Not reported	Random split (training:test=8:2), performance evaluation: correlation and mean error between predicted and actual values
Sun et al [26]	Surgical parameters, laboratory parameters, demographic characteristics	LR	LR	AUC=0.73 (95% CI 0.67‐0.79), Accuracy=0.88	Not reported	Random split (training set:test set=7:3), 5-fold cross-validation
Wang [27]	Radiomics features, clinical variables	LR, SVM, RF, SGDy, KNN	LR	AUC=0.83, Accuracy=0.80, Sensitivity=0.75, Specificity=0.83	Not reported	Random split (training set:test set=7:3), 5-fold cross-validation
Wakiya et al [28]	Surgical parameters, laboratory parameters, tumor markers	DT	DT	Accuracy=0.80, Sensitivity=1, Specificity=0.66	Not reported	Random split (training set:test set=3:1)
Xu [29]	Radiomics features, clinical features	SVM	SVM	AUC=0.87, Accuracy=0.85, Sensitivity=0.72, Specificity=0.89	Not reported	Random split: training:validation:test=6:2:2
Xue et al [30]	Laboratory parameters	LR, DT, XGBoost, CNNz, LSTM	XGBoost	AUC=0.72, Accuracy=0.87, Precision=1, Recall=0.18, F score=0.31	Not reported	Random split (training set:test set=7:3), 5-fold cross-validation
Yang et al [31]	Demographic characteristics, surgical parameters, laboratory parameters	XGBoost, LR, LGBM, RF, SVM	RF	AUC=0.86, Accuracy=0.78, Sensitivity=0.86, Specificity=0.81	Not reported	Random split into training and internal validation sets; 15-fold cross-validation conducted on the training set
Yang et al [32]	MRI-anatomical-clinical features, morphological features	LR, SVM, RF, XGBoost	XGBoost	AUROC=0.88 (95% CI 0.74‐1.00), Accuracy=0.85, Sensitivity=0.90, Specificity=0.81	AUROC=0.82 (95% CI 0.68‐0.96), Accuracy=0.78, Sensitivity=0.81, Specificity=0.75	Data from 2 medical centers
Yin et al [33]	CTab-based radiomics features, clinical factors	DNN, LR	DNN	AUC=0.92, Accuracy=0.75, Sensitivity=0.30, Specificity=0.83	Not reported	Random split (training set:test set=7:3), temporal split, class imbalance handling: SMOTEaa
Zheng et al [34]	Radiomics features, clinical factors, laboratory parameters	SVM	SVM	AUC=0.87 (95% CI 0.76‐0.94), Accuracy=0.76, Sensitivity=1, Specificity=0.65	AUC=0.81 (95% CI 0.72‐0.87), Accuracy=0.79, Sensitivity=0.87, Specificity=0.65	Center 1: partitioned into training and internal test sets. Center 2: designated as the external test set.
Zheng et al [35]	Tumor characteristics, surgical parameters, laboratory parameters	RF, MDNac	RF	AUC=0.79 (95% CI 0.65‐0.93), Accuracy=0.82	Not reported	Random split (training set:test set=8:2), bootstrap
Zong et al [36]	Multiparametric MRI	DL	MS-3D-ResNetad	AUC=0.87 (95% CI 0.86‐0.89), Accuracy=0.85, Sensitivity=0.86, Specificity=0.85	Not reported	Random split (training set:test set=7:3)
Li [37]	Clinical risk factors in obstetrics	LR, DT, KNN, BPNNae, XGBoost, LGBM	LR	AUC=0.88 (95% CI 0.83‐0.92), Accuracy=0.77, Sensitivity=0.84, Specificity=0.67, PPVaf=0.78, NPVag=0.75	Not reported	5-fold cross-validation

^aML: machine learning.

^bMRI: magnetic resonance imaging.

^cDL: deep learning.

^dXGBoost: extreme gradient boosting.

^eVGG-16: visual geometry group - 16 layers.

^fAUC: area under the curve.

^gLR: logistic regression.

^hBayes: naïve Bayes.

ⁱMLP: multilayer perceptron.

^jDT: decision tree.

^kKNN: k-nearest neighbors.

^lRF: random forest.

^mSVM: support vector machine.

ⁿECOG: eastern cooperative oncology group.

^oDNN: deep neural network.

^pMELD: model for end-stage liver disease.

^qNN: neural network.

^rAUROC: area under receiver operating characteristic curve.

^sAUPR: area under the precision versus recall curve.

^tLGB: light gradient boosting machine (LightGBM).

^uCatB: categorical boosting (CatBoost).

^vAdaB: adaptive boosting (AdaBoost).

^wLSTM: long short-term memory.

^xADASYN: adaptive synthetic sampling.

^ySGD: stochastic gradient descent.

^zCNN: convolutional neural networks.

^aaSMOTE: synthetic minority over-sampling technique.

^abCT: computed tomography.

^acMDN: mixture density network.

^adMS-3D-ResNet: multi-stream 3D residual network.

^aeBPNN: back propagation neural network.

^afPPV: positive predictive value.

^agNPV: negative predictive value.