Fig. 4: HPC performance in mesothelioma survival prediction and risk stratification.

a Forest plot for the Cox proportional hazards model used in survival prediction and values represent log hazard ratios (centre) with 95% confidence intervals (error bars), obtained from the Cox model, including Significant HPC's p-values, confidence intervals, log hazard ratios, and pathologist annotations on HPC morphology. The Cox model was trained on n = 512 independent patients, each linked to a survival outcome. All samples are biological replicates, and the unit of study is the patient. No technical replicates were used in the analysis (scale bar, 400 μm). b UMAP plot of tile vector representations, color-coded by Cox model predicted hazard ratio for each HPC. c Kaplan-Meier plots showing low-risk and high-risk patient groups for LATTICe-M and TCGA-MESO datasets, with their reported p-value performance scores. Shaded bands represent 95% confidence intervals. d Kaplan-Meier plots comparing the performance of our survival predictor, pre-trained on the LATTICe-M dataset, tested on TCGA-MESO epithelioid cases, versus predictions based on the traditional systemic overall grade of the TCGA-MESO epithelioid cases. Shaded bands represent 95% confidence intervals. e An example of a SHAP decision plot for a high-risk (red line) and low-risk (blue line) patient, displaying the percentage of HPCs within their samples that contribute to their outcomes. Source data are provided as a Source Data file.