Skip to main content
Fig. 2 | BMC Bioinformatics

Fig. 2

From: Machine-learning algorithms based on personalized pathways for a novel predictive model for the diagnosis of hepatocellular carcinoma

Fig. 2

Construction of diagnostic model. AConstruction of a PDS matrix for the training cohort. B Principal component analysis of selected pathways. C Elastic-net penalized regression models with EPSGO were performed to obtain the optimal hyperparameters α and λ (α = 0.85372, λ = 0.004230762, deviance = 0.03503). Among them, α represents the balance between lasso and ridge penalties. A closer α to the arrow direction indicates that the model was more like lasso regression, otherwise it was more like ridge regression. And the total amount of penalization was controlled by λ. As for CVM (cross-study validation method), the deviance of cross-study validation, was used to measure the effectiveness of modeling. Thus the lowest point of the curve with the minimum deviance was the final EPSGO solution. D The selection of non-zero coefficients regard to hyperparameter λ. Each curve corresponds to a predictor. The numbers above the box mean the numbers of non-zero coefficients with their corresponding log(λ). And the Y-axis was each predictor’s coefficient, gradually approaching 0 as λ increases. E Heatmap of 24 non-zero coefficient pathways. F Cross-study validation for estimated probabilities of each sample. PC, principal component; EPSGO, efficient parameter selection via global optimization; NT, non-tumoral

Back to article page