Fig. 5: ML models for discriminating liver fibrosis regressors following long-term AVT.

a AUC values (95% CI) for ML models developed using five feature selection methods in conjunction with four ML algorithms. These models were constructed from the training set and subsequently validated in the testing set derived from long-term patients in the 4D-DIA-MS discovery cohort. The AUC values (95% CI) for the optimal long-term panels are highlighted in dark red. Training set: n = 15 for R; n = 7 for NR. Testing set: n = 10 for R; n = 4 for NR. b ROC curve analysis of the selected long-term panel, used to monitor fibrosis regression status in long-term patients from the PRM-MS validation cohort (CD163 was not detected, which was imputed using the mean abundance of IDH1 in the training set). Data were presented as AUC values with 95% CI. PRM-MS validation cohort: n = 42 for R; n = 12 for NR. c AUC values (95% CI) of the ML models across clinically relevant populations, including patients with baseline Ishak scores ≥ 3, non-steatosis patients at baseline, and those exhibiting a ΔIshak ≥ 1 post-long-term AVT. These values are presented for both the 4D-DIA-MS discovery cohort (comprising the training set, testing set, and the aggregate of both sets) and the PRM-MS validation cohort (CD163 was not detected, which was imputed using the mean abundance of IDH1 in the training set). Detailed subgroup sample sizes are provided in Supplementary Table 1. Abbreviations: 4D-DIA-MS four-dimensional data-independent acquisition mass spectrometry, AUC area under the curve, PLS partial least squares, PRM-MS parallel reaction monitoring mass spectrometry, SVM support vector machine.