Skip to main content

Table 2 Characteristics of ML-based predictive model of HCC patients after hepatectomy

From: Current updates in machine learning in the prediction of therapeutic outcome of hepatocellular carcinoma: what should we know?

Author Study type No. of patients Model Outcomes AUC/C-index Conclusion
Hamamoto [43], 1995 Retrospective Single center 65 ANN Death In the study for predicting the died of hepatic dysfunction, ANN predicted the outcome of 11 patients in the validation group and achieved the accuracy of 100%
Ho [44], 2012 Retrospective Multicenter 427 ANN and DT 1,3,5-year DFS D: 0.977 and 0.734 (1-year)
0.989 and 0.825 (3-year)
0.963 and 0.675 (5-year)
V: 0.777 and 0.718 (1-year)
0.774 and 0.561 (3-year)
0.864 and 0.627 (5-year)
The ANN outperforms DT in predicting DFS in post-surgical HCC patients
Xu [48], 2012 Retrospective Multicenter 336 SVM RR The SVM based on IHC features could identify HCC patients who are easily recurrence after surgery, and the predictive accuracy of SVM was 66.5%
Chiu [45], 2013 Retrospective Multicenter 434 ANN 1,3,5-year survival D: 0.980 (1-year)
0.989 (3-year)
0.993 (5-year)
V: 0.875 (1-year)
0.798 (3-year)
0.810 (5-year)
The ANN model can process a greater number of predictors and had better accuracy than the traditional LR model
Qiao [46], 2014 Prospective Multicenter 725 ANN 5-year survival D: 0.855
V: 0.829
The ANN model outperforms both Cox and other staging systems in predicting survival in HCC patients who have received surgical resection
Cai [36], 2015 Retrospective Single center 299 BN 10-month survival The BN model had 67.2% of accuracy to classify the survival time of post-surgical HCC patients
Akai [49], 2018 Retrospective Single center 127 RSF DFS, OS 0.611
RSF can predict the individual risk for each patient on DFS and OS
Wang [23], 2019 Retrospective Single center 167 DCNN RR 0.825 Combined clinical information and radiomics features can effectively predict early recurrence of HCC patients
Kim [50], 2019 Retrospective Single center 167 RSF1*
Early recurrence
Lately recurrence
Early recurrence: 0.671(RSF1)
Early recurrence: 0.737(RSF1)
Compared to another two RSF models, combined clinicopathologic-radiomic RSF model achieved the highest predictive power for the recurrence within 2 years after surgery of HCC, and has fair predictive performance for lately recurrence
Xu [19], 2019 Retrospective Multicenter 1139 SVM
RR The accuracy of SVM, RF and BN model was 0.46, 0.48 and 0.56, respectively, in validation group form another independent institution. The BN model could contribute to HCC recurrence research
Mai [47], 2020 Retrospective Single center 353 ANN PHLF 0.880(D)
The risk of severe PHIF in HCC patients after surgery based on ANN model, can be accurately divided into 3 groups
Saillard [55], 2020 Retrospective Multicenter 522 CNN1#
OS D: 0.75(CNN1)
V: 0.68(CNN1)
Two CNN models based on histological features form WSIs performed well for predicting OS of HCC patients after surgery, and both CNN models outperformed the CS that the score included the relevant clinical, biological and pathological features
Schoenberg [51], 2020 Retrospective Single center 180 RF DFS D: 0.766(0.627–0.904)
V: 0.788(0.658–0.919)
RF model based on clinical and laboratory variables, can accurately predict DFS after surgery of HCC
Wang [52], 2020 Retrospective Multicenter 201 RF 5-year survival D: 0.980
V: 0.758
RAD model integrated with RF in a valid method to predict 5-year survival of post-operative HCC patients
Liao [53], 2020 Retrospective Multicenter 645 RF 1,3,5-Y survival V1: 0.626(1-year)
V2: 0.600(1-year)
RF model based on 46 histopathplogical features, was able to stratify post-surgical patients of HCC into long and short-term groups. And the RF model showed similar accuracy with TNM staging systems
Saito [54], 2020 Retrospective Multicenter 158 SVM RR The SVM model based on digital pathological images has the accuracy of 89.9% for prediction of HCC recurrence after surgery
  1. *RSF1: clinicopathologic model using RSF; **RSF2: radiomic model using RSF; ***RSF3: combined clinicopathologic-radiomic model using RSF
  2. #CNN1: convolutional neural network model with automatically method to processing WSI imaging; ##CNN2: convolutional neural network model with an attention mechanism for WSI annotation by pathologist
  3. ML machine learning, HCC hepatocellular carcinoma, AUC area under the curve, ANN artificial neural network, DT decision tree, DFSdisease free survival, D development cohort, V validation cohort, SVM support vector machine, RR recurrence rate, IHC immunohistochemistry, LR logistic regression, BN bayesian network, RSF recurrence-free survival, OS overall survival, DCNN deep convolutional neural network, RF random forest, PHIF posthepatectomy liver failure, WSI whole-slide imaging, RAD radiomic, TNM tumor node metastasis