Evolutionary Multi-objective Optimization of Hyperparameters for Decision Support in Healthcare

Abstract

Hyperparameter optimization is a complex task in machine learning that involves finding the optimal values for the hyperparameters that govern the behavior of a learning algorithm. However, the space of possible hyperparameters is usually vast, which is very common in practical application domains like prediction or classification in medical applications. Therefore, manually exploring all possible combinations is practically impossible. Then again, traditional single-objective optimization techniques often focus on maximizing a specific metric, such as accuracy, but may overlook the trade-offs between different performance measures. Balancing several performance metrics, like accuracy and recall, is paramount in the medical domain, where decisions directly impact patient outcomes. This work introduces a multi-objective evolutionary approach for hyperparameter tuning for machine learning models in deep vein thrombosis prediction, an area that has received limited exploration. The contribution lies in utilizing a multi-objective approach to achieve a balanced accuracy-recall trade-offs aligned with established thresholds or standards in the medical domain. By simultaneously optimizing accuracy and recall, the proposed approach provides a set of Pareto-optimal solutions consisting of optimized hyperparameters, allowing decision-makers to select the best compromise solution based on their preferences or current demands. Overall, this work advances the understanding of the benefits of multi-objective hyperparameter optimization in the medical domain and in the case of improved decision-making. It paves the way for future research in utilizing multi-objective hyperparameter optimization for machine learning models in healthcare applications, contributing to advancing machine learning techniques to improve patient outcomes.