Table 3.
Publications relevant toMLon treatment response and survival prediction
| Publication | Feature extraction method | Prediction model | Sample size | Data type | Performance | Validation method | Feature selection/input | Highlight/advantage | Shortcoming |
|---|---|---|---|---|---|---|---|---|---|
| Jiang et al. [104] | MRRN-based model | MRRN-based model | 1210 | CT Images | DSC (0.68–0.75) | 5-fold cross-validation | 3D image features | The model can accurately track the tumor volume changes from CT images across multiple image resolutions | The model does not predict accurately enough when the tumor size is small |
| Qureshi [106] | NA | RF; SVM; KNN; LDA; CART | 201 | Molecular structure and somatic mutations of EGFR | Accuracy (0.975) | 10-fold cross-validation | 4 clinical features + 4 protein drug interaction features + 5 geometrical features | The model integrates multiple features for data training, and achieves better performance than other benchmarked models | Among the possible 594 EGFR mutations available in the COSMIC database, the model only considers the most common 33 EGFR mutations for model training |
| Kapil et al. [107] | AC-GAN | AC-GAN | 270 | Digital pathology images | Lcc (0.94); Pcc (0.95); MAE (8.03) | Hold-out | PD-L1-stained tumor section histological slides | The model achieves better performance than other benchmarked, fully supervised models | In the experiments, the use of PD-L1 staining for TPS evaluation may not be accurate enough |
| Geeleher et al. [109] | NA | Ridge regression model | 62 | RNA-seq | Accuracy (0.89) | Leave-one-out cross-validation | Removed low variable genes | The model can accurately predict the drug response using RNA-seq profiles only | The training sample size is small |
| Chen et al. [123] | Chi-square test + NN | NN | 440 | RNA-seq | Accuracy (0.83) | Hold-out | RNA-seq of 5 genes | The model uses multiple laboratory datasets for training to improve its robustness | The model doesn’t consider demographic and clinical features, which may affect the prediction |
| LUADpp [125] | Top genes with most significant mutation frequency difference | SVM | 371 | Somatic mutations | Accuracy (0.81) | 5-fold cross-validation | Somatic mutation features in 85 genes | The model can predict with high accuracy with only seven gene mutation features | Mutation frequency may be impacted by the sampling bias across datasets; LD may also affect the feature selection |
| Cho et al. [126] | Information gain; Chi-squared test; minimum redundancy maximum relevance; correlation algorithm | NB; KNN; SVM; DT | 471 | Somatic mutations | Accuracy (0.68–0.88) | 5-fold cross-validation | Somatic mutation features composed of 19 genes | To improve performance, the model uses four different methods for feature selection | The training cohort consists of only one dataset |
| Yu et al. [128] | Information gain ratio; hierarchical clustering | RF | 538 | Multi-omics (histology, pathology reports, RNA, proteomics) | AUC (> 0.8) | leave-one-out cross-validation | 15 gene set features | The study uses an integrative omics-pathology model to improve the accuracy in predicting patients’ prognosis | Cox models may be overfitted in multiple-dimension data |
| Asada et al. [130] | Autoencoder + Cox-PH + K-means + ANOVA |
SVM | 364 | Multi-omics (miRNA, mRNA) | Accuracy (0.81) | Hold-out | 20 miRNAs + 25 mRNAs | The study uses ML algorithms to systematically model feature extraction from multi-omics datasets | The model does not consider the impact of clinical and demographic variances in data training |
| Takahashi et al. [131] | Autoencoder + Cox-PH + K-means + XGBoost/LightGBM |
LR | 483 | Multi-omics (mRNA, somatic mutation, CNV, mythelation, RPPA) | AUC (0.43–0.99 under different omics data) | Hold-out | 12 mRNAs, 3 miRNAs, 3 methylations, 5 CNVs, 3 somatic mutations, and 3 RPPA |
The study uses ML algorithms to systematically model feature extraction from multi-omics datasets | The datasets collected in this study contain uncommon samples between different omics datasets, which may cause bias in model evaluation |
| Wiesweg et al. [136] | Lasso regression | SVM | 122 | RNA-seq | Significant hazard ratio differences | Hold-out | 7 genes from feature selection model + 25 cell type-specific genes |
The ML-based feature extraction model performs better than using any single immune marker for immunotherapy response prediction | The metrics used in this study does not perceptual intuition. Using accuracy or AUC may be better |
| Trebeschi et al. [137] | LR; RF | LR; RF | 262 | CT imaging | AUC (0.76–0.83) | Hold-out | 10 radiographic features | The model can extract potential predictive CT-derived radiomic biomarkers to improve immunotherapy response prediction | The predictive performance between different cancer types is not robust |
| Saltz et al. [142] | CAE [143] | VGG16 [144] + DeconvNet [145] | 4612 (13 cancer types) |
Histological images | AUC (0.9544) | Hold-out | Image features of H&E-stained tumor section histological slides | The model outperforms pathologists and other benchmarked models | The predictive performance between different cancer types is not robust |
Note: MRRN, resolution residually connected network; CART, classification and regression trees; AC-GAN, auxiliary classifier generative adversarial networks; Lcc, Lin’s concordance coefficient; Pcc, Pearson correlation coefficient; MAE, mean absolute error; TPS, tumor proportional scoring; LD, linkage disequilibrium; Cox-PH, Cox proportional-hazards; ANOVA, analysis of variance; miRNA, microRNA; RPPA, reverse phase protein array; CAE, convolutional autoencoder; mRNA, messenger RNA; PD-L1, programmed cell death 1 ligand 1; COSMIC, the Catalogue Of Somatic Mutations In Cancer; EGFR, epidermal growth factor receptor. Compared with hold-out, cross-validation is usually more robust, and accounts for more variance between possible splits in training, validation, and test data. However, cross-validation is more time consuming than using the simple holdout method.