Impact of the Power of Adaptive Weight on Penalized Logistic Regression: Application to Adipocytic Tumors Classification

Abstract

Introduction: MRI-based texture features in adipocytic tumors to serve as non-invasive predictive biomarkers that can provide precise outcomes for decision-making. Power of adaptive weight and the initial weight for the adaptive Lasso is one of the important parameters. This study aimed to compare the impact of the initial weight together with the power of adaptive weight for this adaptive Lasso under high-dimensional sparse data with multicollinearity. Methods: All independent variables in the Monte Carlo simulation were generated using the Toeplitz correlation structure. Performance of the initial weight together with the power of adaptive weight on penalized approaches was evaluated using the mean of the predicted mean squared error (MPMSE) for simulation study and the area under the receiver operator characteristic curve (AUC), precision, recall, F1-score, and the classification accuracy of models for real-data applications. Results: The simulation study showed that the smallest MPMSE value was obtained from the square root of the adaptive Lasso together with the initial weight using Lasso. Additionally, the results of this approach on the real-data application achieved high performance to distinguish the intramuscular lipomas from well-differentiated liposarcomas: the values of AUC, accuracy, precision, recall, and F1-score for the model based on penalized logistic regression classifier were 0.935, 0.928, 0.919, 0.921, and 0.925 respectively, and 0.946, 0.935, 0.932, 0.934, and 0.930 respectively for the model based on support vector machine classifier. Both the simulation study and the real-data application presented that the square root of the adaptive Lasso together with the initial weight using Lasso was the best option under high-dimensional sparse data with multicollinearity. Conclusion: Our finding showed that the power of adaptive weight on penalty function and the initial weight can affect certain the classification accuracy of machine-learning model. In practice, if choosing these parameters are appropriate, it produces models that have good performance.