# Cancer: Can AI Models Truly Predict Tumor Stages? > Source: > Published: 2026-07-10 22:41:01+00:00 # Cancer: Can AI Models Truly Predict Tumor Stages? AI's potential for predicting cancer stages using deep learning is promising, yet challenges in model generalization remain. Are these models ready for clinical use? The ambition to harness [artificial intelligence](/glossary/artificial-intelligence) in predicting cancer stages is undeniably high. The recent exploration into using AI to independently predict Tumor, Node, and Metastasis (TNM) stage labels, anchored on the Cancer Genome Atlas pathology reports, marks another chapter in this venture. Framed as multi-label [classification](/glossary/classification) tasks, this study stands as the sixth shared task of the SMM4H-HeaRD 2026 conference. ## AI Meets Cancer Pathology In dissecting the methodology, researchers deployed both classical and contemporary [deep learning](/glossary/deep-learning) approaches. The arsenal included Term Frequency-Inverse Document Frequency (TF-IDF) features and embeddings from ClinicalBERT, BioBERT, and PubMedBERT. These were combined with Logistic Regression, Light Gradient Boosting Machine (LightGBM), Feed-Forward Neural Networks (FFNN), and Wide Residual Networks (WRN). But what exactly did they find? Individual embeddings, it turns out, were on par with TNM label classification. However, when combined, they enhanced predictive power. WRN, for instance, achieved impressive AUROC scores of 0.839 for T, 0.8502 for N, and 0.803 for M, with corresponding F1-scores of 0.622, 0.702, and 0.9337. ## The Star Player: LightGBM with TF-IDF raw performance, LightGBM combined with TF-IDF took the spotlight. It outperformed others with AUROC scores of 0.9368 for T, 0.9524 for N, and 0.8311 for M, alongside F1-scores of 0.7559, 0.7384, and 0.7017, respectively, during training. Codabench test results showed Macro-F1 scores of 0.978, 0.957, and 0.879 for test set 1, while test set 2 yielded 0.807, 0.767, and a perfect 1.0. Yet, as any [machine learning](/glossary/machine-learning) enthusiast knows, the real test is in generalization. Here, the performance took a hit. The evaluation phase saw a notable drop in Macro-F1 scores from 0.938 to 0.858 between the two test sets. Clearly, the models struggle with generalizability, sensitivity to class imbalances, and processing lengthy clinical documents. ## Ready for the Clinic? Now, let’s apply some rigor here. While the study provides a solid baseline model and a reproducible pipeline, it’s far from ready for clinical application. The results are promising, but the drop in the evaluation phase highlights significant hurdles. Are we really ready to trust these models in high-stakes environments like hospitals? Color me skeptical, but the journey from promising research to practical implementation is anything but straightforward. What they're not telling you is that without further [optimization](/glossary/optimization) and validation, real-world application remains a distant goal. Before these models can transition from the lab to the clinic, they require significant refinements. So, where does this leave us? AI's potential in healthcare is tantalizing, but as always, the devil is in the details. Will these models become indispensable tools for oncologists, or will they remain academic exercises? Only time, and rigorous testing, will tell. Get AI news in your inbox Daily digest of what matters in AI. ## Key Terms Explained [Artificial Intelligence](/glossary/artificial-intelligence) The science of creating machines that can perform tasks requiring human-like intelligence — reasoning, learning, perception, language understanding, and decision-making. [Classification](/glossary/classification) A machine learning task where the model assigns input data to predefined categories. [Deep Learning](/glossary/deep-learning) A subset of machine learning that uses neural networks with many layers (hence 'deep') to learn complex patterns from large amounts of data. [Evaluation](/glossary/evaluation) The process of measuring how well an AI model performs on its intended task.