Fig. 1: Overview of real-world cohorts and study design.

a, Overview of training data and the real-world Harvard-NSCLC and Harvard-PAN (melanoma, renal, breast, bladder, oesophagus, others) evaluation cohorts (total, n = 3,476). Harvard-PAN tumours with fewer than 100 patients per entity were pooled as ‘Others’. All patients were treated with immunotherapy at the Dana-Farber Harvard Cancer Center (DFHCC). External and biological validation was done in the prospectively collected TRACERx NSCLC cohort (n = 464) b, Overview of model development. A deep-learning system able to automatically predict thymic health, as a proxy for thymic functionality, based on standard-of-care chest CT scans was developed using 5,674 independent CT scans of the training data. We applied the model to the standard-of-care CT scans from the Harvard-NSCLC and Harvard-PAN cohorts as well as the external prospectively collected TRACERx cohort for statistical analysis. SSL, self-supervised learning. c, Overview of cohort descriptions. d, Representative images of high, average and low thymic health. The anatomical overview of cancer types in a was created in BioRender; Birkbak, N. https://BioRender.com/aa6hkul (2026).