Table 2. The characteristics of the included studies for diagnosing NSCLC through the gene profiles analyzed by AI models.
| Authors | Publication year | Number of datasets | Number of cases | Number of genes (total) | Subtypes (cases) | Training set (cases) | Validation set (cases) | Test set (cases) | Independent test datasets (cases) | Classifier | Results | Conclusion | ||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| ACC | SN | SP | AUC | Precision | ||||||||||||
| Xiao et al. (31) | 2017 | 1 | 162 | 1,385 | TCGA: ADC (n=162) | NR | NR | NR | 0 | DL-based multi-model (KNN, SVM, DT, RF, GBDT) | KNN: 88.00%; SVM: 97.20%; DT: 96.80%; RF: 93.20%; GBDT: 96.80%; majority voting: 97.20%; DL-based method: 99.20% | DT: 97.37% | NR | NR | DT: 98.46% | The DL-based multi-model algorithm could obtain more information to achieve the accuracy of 99.20% for distinguishing ADCs from normal |
| Yuan et al. (32) | 2020 | 1 | 150 | 1,100, 260, 43 (n=20,502) | GEO: ADC (n=77), SCC (n=73) | NR | NR | NR | 0 | SVM, RF, RIPPER | SVM: 0.867; RF: 0.880; RIPPER: 0.867 | SVM: 0.987; RF: 0.974; RIPPER: 0.867 | SVM: 0.740; RF: 0.781; RIPPER: 0.872 | NR | SVM: 0.800; RF: 0.772; RIPPER: 0.877 | Analyzing the gene expression dataset of NSCLC subtypes, the RIPPER algorithm yielded the almost equal performance of subtyping NSCLCs compared with the SVM/RF classifier |
| Podolsky et al. (33) | 2016 | 3 | 480 | NR | DFCI: ADC (n=139), SCC (n=21), other (n=26), normal (n=17); UMD: ADC (n=86), normal (n=10); BWHD: ADC (n=150), other (n=31) | 235 | 96 | 149 | 0 | KNN, NB, SVM, DT | NR | NR | NR | KNN, k=1: 0.87; KNN, k=5: 0.96; KNN, k=10: 0.97; NB_normal: 0.85; NB_histogram: 0.84; SVM: 0.91; C4.5 DT: 0.92 | NR | Compared with other machine learning algorithms, SVM was the optimal tool in NSCLC morphology classification based on gene expression level evaluation |
| Cai et al. (34) | 2015 | 2 | 1,099 | 16 (n=45) | TCGC: ADC (n=126), SCC (n=134); GEO: SCLC (n=28); TCGA: ADC (n=452), SCC (n=359) | 288 | 0 | 811 | 0 | RF and multi-SVMs | Training datasets: 86.54%; Independent datasets: 84.60% | Training datasets: 84.37%; Independent datasets: 85.52% | NR | NR | Training datasets: 66.79%; Independent datasets: 85.94% | The accuracies of multi-SVM model with such 16 top features for diagnosing NSCLC subtypes were 86.54% and 84.6% in the training and test set, respectively |
| Li et al. (35) | 2018 | 2 | 853 | 20 (n=107) | TCGA: ADC (n=286), normal (n=59); GEO: ADC (n=387), normal (n=121) | 2/3 of each dataset | 0 | 1/3 of each dataset | 0 | RF, SVM, and ANN | TCGA: 98.68%; GSE68465: 99.51%; GSE10072: 97.91%. | TCGA: 99.28%; GSE68465: 99.95%; GSE10072: 98.05% | TCGA: 95.68%; GSE68465: 92.83%; GSE10072: 97.75% | NR | NR | Machine learning models with twenty ADC signature genes were robust for early ADC diagnosis |
| Dong et al. (36) | 2019 | 1 | 369 | 699 | TCGA: ADC (n=369) | NR | NR | NR | 0 | SVM, KNN, LR, RF, gcForest and the ensemble MLW-gcForest | Methylation: 0.751; RNA: 0.689; CNV: 0.645; multi-modal: 0.908 | Methylation: 0.763; RNA: 0.679; CNV: 0.677; Multi-modal: 0.882 | NR | Multi-model: 0.96 | Methylation: 0.771; RNA: 0.659; CNV: 0.675; Multi-modal: 0.896 | MLW-gcForest algorithm had an AUC of 0.96 and an accuracy of 0.908 for ADC staging, better than those achieved by traditional machine learning algorithms |
| Yang et al. (37) | 2020 | 2 | 600 | 42, 26, 16 (n=528) | TCGA: ADC (n=470); GSE62182: ADC (n=94); GSE83527: ADC (n=36) | 376 | 94 | 0 | 130 | SVM | NR | NR | NR | TCGA: 0.62; GSE62182: 0.66; GSE83527: 0.63 | NR | The 16‑miRNA signature analyzed by LIBSVM algorithm showed a similar ability to classify ADC pathological stages to that of the combinations of 42 or 26 miRNAs |
NR, not reported; AI, artificial intelligence; DL, deep learning; SVM, Support Vector Machine; KNN, K-nearest neighbors; GBDT, gradient boosting decision trees; LR, logistic regression; RF, Random Forest; DT, Decision Tree; ANN, artificial neural networks; NB, Naive Bayes; RIPPER, Repeated Incremental Pruning to Produce Error Reduction algorithm; ADC, adenocarcinoma; SCC, squamous-cell carcinoma; NSCLC, non-small cell lung cancer; SCLC, small cell lung cancer; TCGA, The Cancer Genome Atlas; GEO, Gene Expression Omnibus; DFCI, Dana-Farber Cancer Institute; UMD, University of Michigan Dataset; BWHD, Brigham and Women’s Hospital Dataset; CNV, copy number variation; AUC, Receiver-operating characteristic (ROC) curve; ACC, accuracy; SN, sensitivity; SP, specificity.