Deep learning and cancer biomarkers: recognising lead-time bias

We are concerned that Ole-Johan Skrede and colleagues1 did not take into account lead-time bias in their novel deep learning scanning algorithm for histopathology slides that predicts survival of patients with colorectal cancer. Survival is the time from cancer diagnosis to death. Before cancer screening was introduced, survival was a valid estimate of prognosis for patients with cancer. However, following the introduction of early detection through screening, survival estimates are notoriously affected by lead-time and overdiagnosis bias.