Hierarchical Temporal Transformer for Cancer Grade Prediction and Cross-Cancer Transfer Learning from Pathology Reports

Hierarchical Temporal Transformer (HTT) predicts cancer grade from TCGA pathology reports across 14 tumour types, reaching AUROC 0.925 ± 0.007 on unseen types.

• Why it matters: Prior language models typically read reports one at a time and are tested only on tumour types seen during training, leaving uncertainty about using visit timing and cross-type transfer.
• What they did: HTT uses an adapted BiomedBERT encoder plus a cancer-type-conditioned temporal transformer, with an interval-aware time code (HTT-IA) tested on synthetic AUROC benchmarks and on 4,786 TCGA reports in 14 types.
• The result: On the growth benchmark, HTT-IA achieved AUROC 0.843 (+0.053, 0.026 to 0.082) versus 0.790 for the initial time code, and on unseen tumour types it matched in-distribution performance and a single-report baseline.

AI summary of the abstract · may contain errors · not medical advice · read the paper.