Impact of missing data in evaluating artificial neural networks trained on complete data

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Machine Learning for Healthcare

Background:

• Accurate breast lesion classification is crucial for patient outcomes.
• Mammographic Breast Imaging and Reporting Data System (BI-RADS) descriptors are key inputs for diagnostic models.
• Missing data in clinical datasets can significantly impair the performance of predictive models.

Purpose of the Study:

• To evaluate the effect of missing data on artificial neural network (ANN) models for breast lesion classification.
• To compare different methods for handling missing BI-RADS descriptors in ANN models.
• To assess the utility of constraint satisfaction ANNs in managing incomplete data for medical diagnosis.

Main Methods:

• Trained feed-forward, back-propagation ANNs on complete breast lesion data.
• Employed three distinct methods to estimate missing BI-RADS descriptor values.
• Evaluated a constraint satisfaction ANN capable of processing incomplete data directly.
• Compared model performance using mammographic BI-RADS descriptors.

Main Results:

• Performance of ANNs trained on complete data was affected by missing information.
• Imputation methods for missing data yielded comparable results to complete data models.
• Constraint satisfaction ANNs demonstrated robustness by accommodating missing values without prior estimation.
• ANN models showed potential in predicting benign or malignant breast lesions.

Conclusions:

• Missing data presents a challenge for ANN-based breast lesion classification.
• Constraint satisfaction ANNs offer a viable approach for handling incomplete clinical data.
• Further research is needed to develop robust clinical decision support systems for real-world scenarios with missing information.