Improving performance of extreme learning machine for classification challenges by modified firefly algorithm and validation on medical benchmark datasets

The extreme learning machine (ELM) stands out as a contemporary neural network learning model designed for neural networks, specifically emphasizing those with a single hidden layer. This model has gained significant importance in recent years and is frequently employed in research projects due to its reputation as one of the swiftest and most robust methods. ELM is distinguished by its ability to obtain accurate results without the need for prolonged training, setting it apart from other classifiers. Additionally, its reduced reliance on human intervention significantly diminishes the likelihood of errors. Despite their considerable potential, ELMs are not extensively employed. One contributing factor could be the ongoing challenges that ELM is yet to overcome, requiring successful resolution. A prevalent issue is the model’s performance being notably dependent on the weights, biases within the hidden layer, and the quantity of neurons in that layer. Optimizing the number of neurons, referred to as hyperparameter optimization, falls under the category of NP-hard optimization problems. The second challenge lies in training the ELM, which involves establishing the weights and biases tailored for a specific task, presenting another NP-hard challenge. The research presented in this manuscript concentrates on addressing both aspects: optimizing hyperparameters, specifically the number of neurons in the hidden layer, and training the network to fine-tune the weights and biases. The main goal of this research is to effectively resolve both optimization and training by utilizing an improved swarm intelligence algorithm. As a result, both issues were addressed using an adapted version of the firefly algorithm. The proposed approach was tested and validated on twelve authentic datasets and four synthetic datasets designed for classification purposes. One of the forefront tasks among them involves the fetal nonstress test, commonly known as the cardiotocography problem, requiring the interpretation of data from two wearable sensors to discriminate between 3 and 10 imbalanced classes. The obtained outcomes are compared with the results reached by similar state of the art approaches, and the simulations show that the firefly algorithm improved by the group search operator can lead to superior performance. Additionally, enhancements of proposed method are confirmed by rigid statistical tests and results of best generated model for significant heart disease dataset are interpreted by valuable Shapley Additive Explanations (SHAP) tool.