Predicting Male Fertility Using Machine Learning: A Semen Parameters Based Analysis with the VISEM Dataset

Technical Summary: Predicting Male Fertility Using Machine Learning with the VISEM Dataset

Problem Statement
Male infertility affects approximately 17.5% of the global adult population, with male factors contributing to nearly half of all infertility cases. The clinical cornerstone for evaluating male reproductive health is semen analysis, which assesses sperm concentration, motility, and morphology according to World Health Organization (WHO) guidelines. However, traditional semen evaluation relies on manual microscopy and expert interpretation, rendering the process labor-intensive and susceptible to observer bias, inter-laboratory variability, and inconsistent diagnoses. While artificial intelligence (AI) and machine learning (ML) have shown promise in reproductive medicine, previous studies have largely focused on predicting isolated parameters (e.g., motility as a continuous variable) or performing image-based tracking. There remains a gap in the literature regarding systematic frameworks that utilize standardized semen attributes to categorize fertility status into clinically relevant groups (Fertile, Sub-Fertile, and Infertile) using supervised learning.

Methodology
This study addresses the aforementioned gap by applying supervised machine learning models to the publicly available VISEM dataset, which contains semen analysis records from 85 male participants. The methodology followed a structured workflow:

1. Data Preprocessing and Feature Engineering: The dataset was cleaned to remove missing values and unnecessary spaces. Key features were extracted and normalized, including sperm concentration, total sperm count, normal spermatozoa percentage (derived using the Teratozoospermia Index), progressive motility, non-progressive motility, and immotile sperm percentage.
2. Classification Criteria: Samples were labeled into three distinct classes based on WHO thresholds for progressive motility, morphology, and sperm concentration:
   • Fertile (Fast): Progressive motility ≥40%, morphology ≥4%, concentration ≥15 million/mL.
   • Sub-Fertile (Average): Progressive motility 20–39%, morphology 2–3%, concentration 10–14 million/mL.
   • Infertile (Slow): Progressive motility <20%, morphology <2%, concentration <10 million/mL.
3. Model Benchmarking: The study utilized the LazyPredict framework to automate the training and evaluation of over 40 machine learning algorithms. This approach allowed for a rapid comparison of models without extensive hyperparameter tuning.
4. Evaluation Metrics: Model performance was assessed using 5-fold cross-validation to ensure robustness. Key metrics included accuracy, F1-score, balanced accuracy, and Receiver Operating Characteristic - Area Under Curve (ROC-AUC) analysis using a one-vs-rest approach for the multiclass problem. Statistical significance was tested using paired t-tests.

Key Contributions

• Comprehensive Classification Framework: Unlike prior works focusing on regression of single parameters, this study implements a multi-class classification system that categorizes male fertility into Fertile, Sub-Fertile, and Infertile groups based on a combination of WHO-defined thresholds.
• Systematic Algorithm Comparison: The research provides a benchmark evaluation of over 40 ML classifiers on the VISEM dataset, identifying the most effective algorithms for this specific clinical task.
• Interpretability and Correlation Analysis: The study includes detailed correlation analyses and feature visualizations (density plots, pair plots, and heatmaps) to demonstrate the interdependencies between sperm concentration, morphology, and motility, reinforcing the biological basis for the classification.

Results
The experimental results demonstrated that machine learning models can effectively distinguish between fertility categories based on seminal parameters:

• Top-Performing Models: Among the 40+ algorithms tested, the Nearest Centroid classifier achieved the highest performance with an accuracy of 94.2%. It was closely followed by Support Vector Machines (SVM) and Quadratic Discriminant Analysis (QDA), both achieving accuracies around 94%. Gaussian Naive Bayes and Random Forest also showed strong performance, with Gaussian Naive Bayes reaching 91%.
• Robustness Validation: The Nearest Centroid model was validated through 5-fold cross-validation and ROC-AUC analysis. The model achieved the following AUC scores:
  • Fertile Class: 0.95
  • Sub-Fertile Class: 1.00
  • Infertile Class: 0.97
• Confusion Matrix Analysis: The model achieved perfect classification for the Fertile and Infertile groups. Misclassification occurred in only one sample within the Sub-Fertile group (misclassified as Infertile), suggesting high reliability in distinguishing extreme categories while highlighting minor overlap in borderline cases.
• Feature Correlation: Correlation analysis revealed a moderate positive correlation (0.50) between sperm concentration and progressive motility, and a weaker positive correlation (0.35) between morphology and motility. Visualizations confirmed that the "Fast" (Fertile) class consistently exhibited higher median values across all three key parameters compared to the "Average" and "Slow" classes.

Significance and Claims
The paper posits that machine learning models, particularly the Nearest Centroid classifier, can provide fast, accurate, and objective assessments of semen quality. The study claims that these findings illustrate the potential of ML to support clinical decision-making in andrology and assisted reproductive technologies (ART) by reducing human error and variability associated with manual analysis.

The authors emphasize that this work shifts the focus from isolated parameter prediction to a comprehensive classification of fertility-relevant parameters suitable for clinical settings. While acknowledging limitations such as the small sample size (85 participants) and the potential for overlap in borderline cases, the study concludes that data-driven tools can enhance the consistency of fertility diagnostics. The authors suggest that future work could integrate additional biological markers (e.g., DNA fragmentation, hormonal profiles) and explore ensemble methods or neural networks to model complex nonlinear interactions, ultimately contributing to more personalized fertility treatment strategies.