Sepsis Prediction by Using a Hybrid Metaheuristic Algorithm: A Novel Approach for Optimizing Deep Neural Networks

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Computational Biology

Background:

• Early sepsis diagnosis significantly reduces mortality risk.
• Gradient-based algorithms in neural networks for sepsis estimation can get stuck in local minima.
• Swarm intelligence and evolutionary approaches offer promising alternatives for optimization problems.

Purpose of the Study:

• To propose a novel hybrid metaheuristic algorithm (HMS-PSO) for optimizing deep neural network weights.
• To apply the optimized deep neural network for the early diagnosis of sepsis.
• To enhance the reliability, durability, and adjustability of sepsis prediction models.

Main Methods:

• Developed a hybrid metaheuristic algorithm combining Human Mental Search (HMS) and Particle Swarm Optimization (PSO) to overcome local minima.
• Integrated the HMS-PSO algorithm with a deep neural network (DNN), creating the HMS-PSO-DNN model.
• Validated the HMS-PSO algorithm against PSO and HMS using benchmark functions.
• Applied the HMS-PSO-DNN model to a dataset of 640 patients (aged 18-60) for sepsis prediction.

Main Results:

• The HMS-PSO algorithm demonstrated superior performance, reliability, and adjustability compared to PSO and HMS on benchmark functions.
• The HMS-PSO-DNN model achieved a lower Mean Squared Error (MSE) of 0.22 over 30 independent runs compared to other algorithms.
• The model exhibited enhanced accuracy, robustness, and overall performance in predicting sepsis.

Conclusions:

• The proposed HMS-PSO hybrid algorithm effectively optimizes deep neural networks for sepsis prediction.
• HMS-PSO-DNN offers a reliable and accurate approach for the early diagnosis of sepsis.
• This method improves upon traditional gradient-based algorithms by avoiding local minima and enhancing predictive performance.