Classification of breast cancer recurrence based on imputed data: a simulation study

Area of Science:

• Medical Statistics
• Machine Learning in Healthcare
• Biostatistics

Background:

• Limited research exists on statistical classification accuracy for breast cancer recurrence, especially with imputed missing data.
• Understanding classifier performance with missing data is crucial for reliable medical event prediction.

Purpose of the Study:

• To compare the performance of binary classifiers (logistic regression, linear discriminant analysis, quadratic discriminant analysis) on breast cancer recurrence.
• To evaluate how imputed missing data affects classifier accuracy and discriminative ability.

Main Methods:

• Simulated incomplete datasets with 15-60% missingness under Missing At Random (MAR) and Missing Completely At Random (MCAR) mechanisms.
• Employed various imputation techniques: mean, hot deck, k-nearest neighbor, multiple imputations via chained equation, expectation-maximization, and predictive mean matching.
• Compared classification accuracy and Area Under the Receiver Operating Characteristic (ROC) curves for each classifier.

Main Results:

• Linear discriminant analysis achieved the highest classification accuracy (73.9%) using mean imputation with 45% missing data under MCAR.
• Logistic regression with predictive mean matching imputation yielded the highest Area Under ROC curves (0.6418) at 30% missingness (MCAR).
• K-nearest neighbor imputation achieved the highest Area Under ROC curves (0.6428) at 60% missing data under MCAR.

Conclusions:

• The choice of imputation method significantly impacts the performance of statistical classifiers for breast cancer recurrence.
• Different classifiers exhibit varying robustness to missing data, highlighting the need for careful selection in medical applications.
• Findings provide insights into optimizing statistical models for predicting breast cancer recurrence in the presence of missing data.