Mayo Clinic Researchers Are Developing Hypothesis-Driven AI for Oncology

Researchers at the Mayo Clinic have announced a research initiative focused on developing what they are calling "hypothesis-driven AI" for oncology — an approach designed to address growing concerns about the interpretability and reliability of black-box machine learning models in clinical settings.

The initiative reflects a broader tension in medical AI: the models that perform best on benchmark metrics often function in ways that are opaque even to the researchers who built them, making clinical validation and regulatory approval significantly more difficult.

The Problem With Standard Medical AI

Current approaches to medical AI predominantly use deep learning architectures — convolutional neural networks for imaging, transformer models for clinical text — that are trained end-to-end to optimize predictive performance. These models can achieve impressive accuracy on held-out test sets, but they often do so by learning correlations in training data that may not reflect the underlying biology and may not generalize to new patient populations or clinical settings.

This creates two distinct problems for clinical deployment. First, it makes it difficult for clinicians to understand why a model is making a particular prediction — limiting the ability to catch errors before they affect patient care. Second, it makes regulatory approval more difficult: the FDA has been cautious about approving black-box AI systems for high-stakes clinical decisions, and the EU AI Act classifies AI systems used in healthcare as high-risk, requiring transparency and human oversight.

The Hypothesis-Driven Alternative

The Mayo Clinic's approach starts from clinical hypotheses grounded in established biological mechanisms rather than from raw data. Rather than asking "what patterns in this data predict patient outcomes?", hypothesis-driven AI asks "given what we know about the biology of this cancer type, what features should be predictive, and can we train a model that uses those features?"

This approach constrains the model's learning in ways that are biologically motivated, making the model's reasoning more interpretable and its predictions more likely to generalize. It also creates a tighter feedback loop between AI development and basic science: when the model underperforms, researchers can investigate whether the underlying biological hypothesis is wrong or whether the model failed to implement it correctly.

Regulatory and Legal Implications

From a regulatory standpoint, hypothesis-driven AI offers significant advantages. The FDA's AI/ML-based Software as a Medical Device framework emphasizes the importance of transparency and the ability to explain model decisions — requirements that are much easier to satisfy when the model is built around explicit biological hypotheses. Similarly, the EU AI Act's conformity assessment requirements for high-risk AI in healthcare are substantially easier to satisfy when the AI system's logic can be traced to established scientific knowledge rather than learned correlations in training data.