Impact of missing data in training artificial neural networks for computer-aided diagnosis

Artificial neural networks (ANN) are frequently used in the development of Computer-Aided Diagnosis systems for breast cancer detection and diagnosis. One class of models uses descriptions of mammographic lesions encoded following the BI-RADS lexicon. Data sets that have been carefully curated to ensure completeness are generally used; however, in routine practice, some information is typically missing in clinical databases. The impact of missing data on the performance of a feedforward, back-propagation ANN, as measured by the area under the Receiver Operating Characteristic curve, was found to be much higher when data were missing from the testing set than when data were missing from the training set. This empirical study highlights the need for additional research on developing robust clinical decision support systems for realistic environments in which key information may be unknown or inaccessible.