P-016 Can we predict sperm DNA fragmentation values using patients’ semen parameters and clinical characteristics combined with machine learning?

Abstract

Abstract Study question Which clinical factors and semen parameters best predict sperm DNA fragmentation percentage (DF%) using machine learning, and which model performs best for the prediction? Summary answer Progressive sperm motility, sperm morphology, primary or secondary infertility, varicocelectomy, and smoking predicted DF% across models, with Gamma Generalized Linear Model providing the best fit. What is known already Numerous studies have shown that sperm DNA fragmentation has a potential clinical significance in assisted reproduction outcomes. However, current testing methods are non-standardised, time-consuming and expensive. Predicting DF% during clinical consultations could inform clinical decision-making. No reliable pipelines exist to predict sperm DNA fragmentation from clinical and seminal fluid data. Therefore, we investigated the best predictors of DF% using machine learning, which can handle complex datasets and detect non-linear relationships. By integrating clinical and semen parameters, a prediction tool could be built to enable clinicians to quickly and objectively predict sperm DNA fragmentation. Study design, size, duration This cross-sectional study analyzed data from 236 men undergoing seminal fluid analysis recruited from two private IVF centres. In-person interviews collected demographic (age, body mass index), clinical (primary/secondary infertility, varicocelectomy), and smoking status between July 2023 and April 2024. All participants provided informed consent following ethical approval. Participants/materials, setting, methods Among the male participants (mean age: 34.72 years), 138 (58.5%) had primary infertility, 36 (15.3%) had secondary infertility, and 62 (26.3%) were fertile. Semen analysis followed World Health Organization (WHO) 2010 guidelines, assessing volume, motility, concentration, and morphology. DF% was assessed using the Sperm Chromatin Dispersion (SCD) test. Six statistical models, including linear and generalized linear models (GLM) with and without bootstrapping, were applied to identify DF% predictors. Main results and the role of chance Linear regression identified progressive sperm motility (B = -0.171, p = 0.031), normal sperm morphology (B = -1.195, p = 0.004), infertility status (B = -3.925, p = 0.005), and varicocelectomy (B = -6.686, p = 0.001) as significant predictors, explaining ∼20% of DF% variability (R² = 0.199). Lower DF% was observed in primary infertility cases compared to secondary infertility across all models. Internal validation through bootstrapping (1,000 resamples) reinforced the stability of identified predictors and uncovered smoking as a significant factor (B = -3.343, p = 0.046), minimizing potential sampling bias. The GLM validated these predictors, showing slightly improved explanatory power (R² = 0.209) and consistency for progressive sperm motility (B = -0.199, p = 0.014), normal sperm morphology (B = -1.162, p = 0.005), infertility status (B = -7.131, p = 0.012), and varicocelectomy (B = -6.642, p = 0.001). The Gamma GLM further refined predictions, addressing DF’s skewed distribution and delivering superior fit indices (AIC = 1791.53, BIC = 1829.49, deviance = 83.495, χ²(9) = 103.19, p < 0.001). Predictor consistency across models highlights robust relationships, while the Gamma GLM’s improved fit and tailored assumptions make it the most reliable approach for predicting DF% in clinical contexts. Limitations, reasons for caution The study is limited by the small sample size and restricted demographics, reducing generalizability and diagnostic power. Wider implications of the findings These preliminary findings contribute to developing machine learning models for predicting sperm DF%, utilising clinical factors and semen analysis parameters to support clinical decision-making in fertility assessments. Trial registration number No

Ähnliche Arbeiten