Fig. 8: Performance of the 10-M model for distinguishing thyroid cancer differentiation states. | npj Digital Medicine

Fig. 8: Performance of the 10-M model for distinguishing thyroid cancer differentiation states.

From: Developing a thyroid cancer differentiation state classification system using deep residual networks and metabolic signature profiling

Fig. 8

10-M ResNet classifier cleanly separated PDTC and ATC from other thyroid states in metabolomic space. a Workflow for constructing the 10-M classifier neural network model to stratify thyroid cancer differentiation status, with feature selection based on SHAP value rankings. The model was developed using the FUSCC cohort, comprising normal (n = 57), FTC (n = 23), PTC (n = 103), PDTC (n = 17) and ATC (n = 10) samples. b Bar chart showing the average SHAP values for the top 10 metabolites contributing to the ResNet model’s predictions. The most influential metabolites include PC (17:0_24:0), PC (36:4), Hex1Cer (d18:1_22:0), N-acetylaspartic acid, ADP-ribose, ribose-5-phosphate, 1-methyl-nicotinamide, PC (32:1), UDP-glucose and ChE (20:3). c ROC curve for the validation cohort, showing an average AUROC of 0.98. d The dimensionality reduction embedding plot illustrates the model’s accuracy in predicting PDTC. e The dimensionality reduction embedding plot illustrates the model’s accuracy in predicting ATC.

Back to article page