Fig. 4: Machine learning of IMC data predicts clinical outcomes.

a, Schematic of the deep-learning-based strategy involving deep residual networks (Resnet50) architecture on the ImageNet dataset for feature extraction from IMC image channels. b–d, Fivefold cross-validation across clinical outcomes: histological patterns, sex (male or female), BMI (less than 30 or 30 or higher) and age (younger than 75 years of age or 75 years of age or older) n = 416; survival (less than 3 years, 3 years or longer) n = 407; progression status (progression or no progression) n = 404; stage (I–II or III–IV) n = 415; smoking status (smoker or non-smoker) n = 414 using frequency of cell types (b); spatial distribution of lineage markers (c) and spatial distribution of all markers (d). The size of the bubble represents deviation from baseline, with blue and grey indicating an improvement or worsening in predictive performance, respectively. The line in the bar plot represents the baseline. Schematics in a–d were created with BioRender. e, Accuracy of clinical progression prediction in patients with stage I LUAD (n = 286) using clinical variables, cell frequency, lineage marker and ‘all markers’ models. Comparison between the clinical variables and the cell frequency model: *P = 0.0319. Comparison between the clinical variables and lineage marker model: ****P < 0.0001. Comparison between the clinical variables and all markers model: ***P = 0.0001. Comparison between the cell frequency and lineage marker model: *P = 0.0321. f, Accuracy of clinical progression prediction in patients with stage I LUAD; discovery cohort (n = 286) and validation cohort (n = 60; 120 cores) in the cell frequency, lineage and all markers models compared with baseline. The size of the bubble represents deviation from baseline, with blue and grey indicating an improvement or worsening in predictive performance, respectively. g, Accuracy of clinical progression prediction in patients with stage I LUAD (validation cohort n = 60; 120 cores) using combinations of top-ranked (left) and neighbourhood-derived (right) lineage markers. For all combinations, see Supplementary Table 15. Data shown as mean ± s.e.m. (b–e). One-way ANOVA with Tukey multiple comparison test was used for statistical analysis (e).