CaresAI predicts TNM staging from TCGA pathology reports using BERT embeddings

CaresAI participated in the SMM4H-HeaRD 2026 shared task by predicting Tumor, Node, and Metastasis (TNM) stage labels independently from Cancer Genome Atlas (TCGA) pathology reports. The study framed the problem as three multi-label classification tasks, exploring classical and deep learning approaches using TF-IDF features and embeddings from ClinicalBERT, BioBERT, and PubMedBERT.

• WRN achieved AUROC scores of 0.839 (T), 0.8502 (N), and 0.803 (M) with F1-scores of 0.622, 0.702, and 0.9337 respectively during training.
• LightGBM with TF-IDF performed best on the training phase with AUROC scores of 0.9368 (T), 0.9524 (N), and 0.8311 (M).
• Test set 1 results showed Macro-F1 scores of 0.978, 0.957, and 0.879 for T, N, and M categories.
• Performance declined on test set 2, with a drop in Macro-F1 score from 0.938 to 0.858 across all stages.

The authors note that the performance decline suggests limitations in model generalizability, sensitivity to class imbalance, and challenges in processing lengthy clinical documents, indicating that further optimization is required before clinical use.