Evolutionary-Based Deep Learning Network Model using Adaptive Mixing Differential Evolution and Application in Acute Pulmonary Embolism

Abstract

Introduction:

Acute pulmonary embolism (APE) is characterized by high incidence and mortality, along with non-specific clinical manifestations. Its common symptoms such as dyspnea, chest pain, cough, and hemoptysis can also appear in other diseases, frequently resulting in the oversight of APE patients and raising the risk of misdiagnosis and mortality. Current clinical risk stratification for pulmonary embolism usually depends on hemodynamic evaluation, the pulmonary embolism severity index, echocardiography, and myocardial injury markers. However, these assessment methods tend to be complex, time-consuming, invasive, and lack repeatability. Therefore, developing a more efficient and accurate tool for APE prediction and analysis is crucial.

Objectives:

To achieve precise prediction and analysis of APE patients using accessible clinical data, we developed an evolutionary-based deep learning network AlexNet model (EDLAlexNet) that leverages blood biochemical indices, vital signs, clinical parameters, and clinical characteristics. The goal is to provide a reliable clinical tool for the assessment and management of APE with high accuracy, specificity, sensitivity, and a favorable AUC.

Methods:

We developed the EDLAlexNet model, which incorporates a novel evolutionary computation method called adaptive mixing differential evolution (MIXDE) integrating Q-learning and opposition-based learning. The performance of the MIXDE algorithm was statistically validated on standard test datasets. Subsequently, the MIXDE-based EDLAlexNet was used to analyze data from intermediate-low-risk and high-risk pulmonary embolism patients.

Result:

The results for APE using EDLAlexNet showed promising performance, achieving an accuracy of 93.76%, specificity of 89.46%, sensitivity of 95.74%, and an AUC of 0.9527. These outcomes demonstrate the model’s effectiveness in precisely predicting and analyzing APE patients.

Conclusion:

Overall, EDLAlexNet, which integrates the MIXDE algorithm, exhibits excellent performance in APE prediction and analysis. It shows potential as a valuable clinical tool for the assessment and management of APE, addressing the limitations of current assessment methods.