Read the result.
Understand the limits.

Define the outcome precisely

Objective response, disease control, pathological complete response, and survival are different outcomes. A model evaluated for one cannot automatically be assumed to predict another. The research pages retain each study’s reported endpoint and explain material ambiguities.

Separate development and evaluation

Independent evaluation matters because a model may learn patterns specific to its development cohort. Patient-level separation is especially important when several cells, images, or samples come from one person. Preprocessing and feature selection must avoid information from held-out data.

Reporting reference: TRIPOD+AI, BMJ 2024

Look beyond a single accuracy figure

Sensitivity and specificity describe performance at a particular threshold. Predictive values also depend on the frequency of the outcome in the evaluated population.

Research findings and clinical use are different stages

Biomarkers can inform cancer treatment, but a matching marker does not guarantee benefit. The projects presented here are exploratory analyses and do not establish a clinically validated treatment-selection test.

Good machine learning practice addresses representative datasets, independent testing, and performance in the intended context. A development framework is not evidence that a particular product has regulatory authorization.