Cancer Data, AI, and Hype Detection
Educational orientation
Expanded educational edition: 2026-09-10. This page is for critical thinking about data and AI, not clinical deployment guidance.
By the End of This Chapter
Identify leakage, population shift and missing evidence of utility.
Learning matrix and gaps in the short edition
Core Idea
Cancer data can come from pathology, imaging, sequencing, registries, trials, wearables, claims, and notes. Each source has missingness, bias, measurement limits, and governance constraints. AI systems can help organize patterns, but a model is not clinically useful just because it scores well on a dataset.
Be suspicious of claims that skip cohort design, external validation, calibration, subgroup performance, and intended use.
Step-by-Step Explanation
1. Start with intended use and unit
Before model architecture, define population, prediction time, available information, and desired outcome. An image, specimen, and patient are different units. Leakage occurs when training receives information unavailable in intended use, including the same patient's data in testing or variables produced after the event. Randomly splitting rows does not necessarily solve this.
2. A score does not finish evaluation
Discrimination describes separation between groups; calibration compares predicted probabilities with observed frequencies. External validation tests transport to genuinely distinct data with a relevant population, site, or period. Clinical utility, workflow impact, errors, harms, and subgroup performance still require assessment. A model may rank well but give wrong probabilities; correct ranking may also improve no relevant outcome.
3. Calibration with simple numbers
Imagine 100 hypothetical predictions of 10% risk, all for the same event and period. With a well-calibrated model, such groups would tend to have close to 10 events, allowing for sampling variation. This does not mean we can identify which ten people will experience the event. If the frequency is systematically 30%, probabilities are underestimated even if the model ranks people reasonably well. This separates calibration from discrimination. Neither result removes the need for external validation, bias and leakage assessment, or clinical-utility evaluation in intended use.
Supporting reading: Collins et al. (2024): TRIPOD+AI.
The Concept Illustrated
Worked Example
A fictional classifier predicts diagnosis using a “treatment started” column filled only after confirmation. High accuracy may come from future information. Removing the column is necessary but insufficient: duplicates, patient-level splitting, selection bias, and external validation still need checking. Only then can use be investigated for improving a clinical workflow or outcome.
Essential Vocabulary
AUC-ROC: area under the curve relating sensitivity to the false-positive rate across thresholds. It summarizes ranking/discrimination, not the accuracy of predicted probabilities or clinical benefit.
Leakage: training signal that would not exist in real use.
External validation: testing in an independent setting.
Calibration: whether probabilities match reality.
Intended use: the exact task and context.
What This Does Not Mean
An AI demo, preprint, or leaderboard score is not evidence of patient benefit.
Builder/Data Lens
Before modeling, document provenance, labels, endpoints, splits, governance, and failure modes.
Check Your Understanding
Does a high AUC at one hospital establish benefit for patients at another?
Answer and reasoning
No. AUC measures discrimination in that dataset. Transport, calibration, thresholds, subgroups, and impact evaluation are additional questions. External validation also does not automatically establish clinical utility.
Optional Deeper Exploration
To audit a headline, write down: which task, which data, which comparator, which uncertainty, and which evidence stage? A reporting checklist improves transparency; it does not certify freedom from bias, regulatory approval, or clinical benefit.
Go Deeper
References
- National Cancer Institute. Cancer Statistics
- U.S. Food and Drug Administration. AI/ML-Enabled Medical Devices