How should we judge AI tools that claim to predict which cancer patients will respond to immunotherapy, and what evidenc

Judging AI tools intended for clinical decision-making requires a high bar for evidence, as the stakes involve patient safety and treatment efficacy. To evaluate these tools, one should look for the following criteria:

1. Robust Validation

• External Validation: A tool must be tested on data from institutions, patient populations, and imaging/sequencing equipment entirely different from those used to train the original algorithm. This ensures the model isn't just "memorizing" the specific characteristics of a single hospital's data.
• Prospective Studies: Retrospective studies (looking back at historical data) are prone to biases. The gold standard is a prospective clinical trial where the AI’s prediction is tracked against actual patient outcomes in real-time.

2. Clinical Utility vs. Statistical Accuracy

• Actionability: It is not enough for an AI to be "accurate." It must demonstrate that using the tool improves patient outcomes compared to standard clinical practice. Does it help clinicians choose a more effective treatment, or does it merely confirm what doctors already know?
• Comparison to Standard of Care: The AI should be benchmarked against current predictive biomarkers, such as PD-L1 expression levels or tumor mutational burden. If the AI provides no additional value over these established tests, its clinical utility is limited.

3. Transparency and Interpretability

• Avoid "Black Boxes": In a medical context, it is vital to understand why an AI makes a prediction. If the model can point to specific biological features or imaging characteristics that lead to its conclusion, clinicians can verify if those features align with known cancer biology.
• Feature Bias: Ensure the model isn't relying on "shortcut" data—such as patient age, socioeconomic status, or hospital site—rather than biological signals linked to immunotherapy response.

4. Generalizability and Diversity

• Population Representation: Immunotherapy response can vary across different genetic ancestries and demographic groups. An AI trained predominantly on data from one specific population may fail or perform poorly when applied to a more diverse patient base. The training data must be transparently reported.

5. Regulatory and Ethical Oversight

• Clinical Authorization: In the U.S., look for FDA clearance or approval (such as De Novo or PMA pathways). Regulatory bodies evaluate these tools for safety and effectiveness specifically for their intended clinical use.
• Ongoing Monitoring: Trustworthy tools are treated as dynamic. Because medical treatments and data change, these tools require "post-market surveillance" to ensure their accuracy remains stable as they are deployed in different clinical settings.

The Bottom Line: To be trustworthy enough to guide treatment, an AI tool should function as a decision-support system, not a decision-maker. It should be used to augment a multi-disciplinary team’s expertise, supported by peer-reviewed evidence that shows it consistently identifies patients who will benefit from immunotherapy while minimizing unnecessary exposure to toxic side effects for those who will not.