Abstract Recent scientific studies have noted that the seminal quality of males is significantly decreasing due to lifestyle and environmental factors. Clinical diagnosis of sperm quality is one important aspect of identifying the potential of semen for the occurrence of pregnancy. Due to the advances in machine learning algorithms, especially the reliable and high classification accuracy of neural network in health related problems, it is becoming possible to predict seminal quality from lifestyle data. In this respect, a few attempts were made in predicting seminal quality. These studies were conducted using imbalanced datasets, where the performance outcomes tend to be biased towards the majority class. Other studies implemented the gradient descent technique for training the neural network. The gradient descent is a local training technique that is prone to get stuck to local minima. On the contrary, meta-heuristic algorithms enable searching solutions both locally and globally. Therefore, in this study, Artificial Algae Algorithm that is improved using a Learning-Based fitness evaluation method is proposed for training Feed Forward Neural Network (FFNN). In addition, the SMOTE data balancing method was employed to balance normal and abnormal instances. Experimental analyses were carried out to evaluate the predictive accuracy of the FFNN trained using Learning-Based Artificial Algae Algorithm (FFNN-LBAAA). The results were compared with well-known machine learning algorithms, namely: Multi-layer Perceptron Neural Network (MLP), Naive Bayes (NB), Support Vector Machine (SVM), K-Nearest Neighbor (KNN) and Random Forest (RF) algorithms. The proposed approach showed superior performance in discriminating normal and abnormal semen quality instances over the other compared algorithms.

中文翻译：

---

使用基于学习的人工藻类算法训练的前馈神经网络的精液质量预测模型

摘要 最近的科学研究表明，由于生活方式和环境因素，男性的精液质量显着下降。精子质量的临床诊断是鉴定精液发生妊娠可能性的重要方面之一。由于机器学习算法的进步，特别是神经网络在健康相关问题中的可靠和高分类精度，从生活方式数据预测精液质量变得可能。在这方面，在预测精液质量方面进行了一些尝试。这些研究是使用不平衡的数据集进行的，其中性能结果往往偏向于多数类。其他研究采用梯度下降技术来训练神经网络。梯度下降是一种局部训练技术，容易陷入局部最小值。相反，元启发式算法可以在本地和全局范围内搜索解决方案。因此，在本研究中，提出了使用基于学习的适应度评估方法改进的人工藻类算法来训练前馈神经网络 (FFNN)。此外，采用SMOTE数据平衡方法来平衡正常和异常实例。进行了实验分析以评估使用基于学习的人工藻类算法 (FFNN-LBAAA) 训练的 FFNN 的预测准确性。结果与著名的机器学习算法进行了比较，即：多层感知器神经网络（MLP）、朴素贝叶斯（NB）、支持向量机（SVM）、K-最近邻 (KNN) 和随机森林 (RF) 算法。所提出的方法在区分正常和异常精液质量实例方面表现出优于其他比较算法的性能。