Table 1.
GSE96058 information.
Fig 1.
The overview of the proposed method.
Five main steps, including reading, preprocessing, feature selection, classification, and SHAP algorithm were applied to the mRNA expression data. 1) Required data was collected from the NCBI-GEO repository and organized during the reading step. 2) The pre-processing step includes two sub-steps, cross-validation and data normalization. 3) The feature-selection step contains two parts: the filter method based on ANOVA and the wrapper method based on Particle Swarm Optimization (PSO) for mRNA data, in which candidate mRNAs with more relevance to positive and negative Papillary lymph node groups were selected. 4) Multi-classifier models were utilized to evaluate the discrimination power of the selected mRNAs. 5) The SHAP algorithm was employed to discover the possible relationship between the selected mRNAs and positive and negative groups.
Table 2.
The top 109 most significant features among the 28,456 mRNAs.
Table 3.
The performance of classifiers.
Fig 2.
The confusion matrix for (a): Training set Confusion Matrix, (b): Training set ROC (c): Validation set Confusion Matrix, (d): Validation set ROC (e): Test set Confusion Matrix, (f): Test set ROC. Zero and one are negative and positive groups, respectively.
Fig 3.
The summary plot using SHAP values (a): summary plot of both classes, (b): summary plot of the negative group, (c): summary plot of the positive group.
Fig 4.
(a): Decision plot of the first sample related to negative class, (b): Decision plot of the first sample related to positive class, (c): waterfall plot of the first sample related to negative class, (d): waterfall plot of the first sample related to positive class.
Fig 5.
The force plot of the first sample in dataset (a): Force plot of negative class, (b): Force plot of positive class.
Fig 6.
The box plot of five more important mRNAs based on their SHAP values (a): GDF5, (b): BAHCC1, (c): LCN2, (d): FGF14-AS2, (e): IDH2.