Item 12 — How missing data were handled.

Clearly describe how missing data were handled. For example, describe processes to replace them with approximate, predicted, or proxy values. Discuss biases that imputed data may introduce.

Explanation and Elaboration: Missing data is an inherent challenge in real-world clinical and research datasets, often arising from incomplete patient records, technical limitations, or variations in data collection procedures. Mishandling missing values can introduce significant bias, skew algorithmic training, and ultimately lead to misguided inferences. There is no single solution to missing data. Replaced missing values can introduce bias, but excluding records with missing values may also create bias. Whatever approach is taken should be carefully considered, explained and justified.

Examples
Where scans were missing, empty (zero-filled) scans with the same resolution and orientation as the other scans were inserted. Sparsified training was implemented as an augmentation layer by randomly setting secondary sequences to zero with independent probabilities of 20%. This percentage approximated the frequency of missing sequences in the clinical dataset. The NiftyNet pipeline includes histogram normalization and whitening…. These layers were adjusted to ensure that zero-filled volumes representing missing sequences were unaffected. Consequently, missing sequences in the original data and from the sparsified training augmentation layer were fed to the network as all zero matrices, which, as a result of the whitening layer, corresponds to the mean intensity of the available data. [Materials and methods (31), CC BY 4.0]
The validation data of HMH and MDACC were complete datasets without missing values. UTHSCSA, however, had 1.34% missing data. Thus, of 20 features each for all 500 cases (ie, 10 000 observations from this site), 134 observations were missing. We used one-hot encoding … to vectorize the input parameters, and after the parameters were vectorized, the binary values in the missing data became all zeroes (0,0), compared with (0,1) for “yes” and (1,0) for “no”. Interestingly, continuous input variables like age, height, and weight were and would always be available in real-time workflow. [Materials and methods (32), CC BY 4.0]