Table 3

Checklist of key elements that should be described in machine learning papers, structured by manuscript section

DataMinimise risk of overfitting by the use of multiple, heterogeneous and independent datasets.
Provide a complete description of the data acquisition process.
Describe the basic technical information of the imagery and the use of any preprocessing methods.
Define a reliable gold standard for all data used to train, validate and test the model.
Algorithms that localise lesions on images and videos, reliable gold standard input for the model should incorporate annotations by multiple experts.
Provide detailed information on ethics approval concerning the use of patient data.
Algorithm architectureProvide a basic description of the algorithm architecture and clear-cut motivation for the most relevant technical choices.
Describe extensive technical details in separate technical publications, or in supplementary materials.
Experimental set-upDescribe the experimental set-up of the algorithm and choose the appropriate performance metrics.
Primary outcome parameters should be based on the envisioned clinical application of the model.
Do not optimise hyperparameters on test set.
Ensure that training, validation and test sets are always split on a patient-basis.
Report a complete overview of all evaluated models to prevent ‘cherry-picking’ of the best performing algorithms.
Results should be presented with caution and in a structured approach.
Include a section where data selection bias, overfitting and generalisability are explicitly discussed.
Describe the necessary steps towards clinical implementation.