P-021 Predicting testicular sperm extraction outcomes in non-obstructive azoospermic patients using machine learning and non-invasive biomarker detection

Abstract

Abstract Study question Can artificial intelligence predict the absence of sperm in testicular biopsies of infertile men diagnosed with non-obstructive azoospermia (NOA) using non-invasive biomarker detection? Summary answer A machine learning model based on blood biomarkers and genetic determinants was highly efficient in predicting testicular sperm extraction (TESE) outcome in NOA patients (AUC=0.65). What is known already Male infertility due to severe spermatogenic failure (SPGF) can lead to NOA, characterized by absence of sperm in the ejaculate. While some azoospermic men may father biological children via TESE, this procedure is unsuccessful in those exhibiting the Sertoli cell-only (SCO) phenotype. An immune-mediated component has been revealed for SCO, based on the association of the MHC class II region. This genetic region is crucial in many autoimmune diseases, including celiac disease, which has been associated with reproductive disorders. This study aims to develop a non-invasive diagnostic method to predict TESE outcome in NOA patients. Study design, size, duration A study including a total of 293 infertile men diagnosed with SPGF was conducted, comprising 143 SCO patients and 150 infertile men due to SPGF but without SCO. Different machine learning models were evaluated to predict the phenotype (SCO or non-SCO) and, consequently, the TESE failure, in patients suffering from idiopathic male infertility due to SPGF. Participants/materials, setting, methods Three machine learning models (artificial neural networks, random forest, and logistic regression) were evaluated and compared based on standard metrics. Four parameters (genetic and hormonal) obtained through non-invasive procedures were included in the models: 1) polygenic risk scores (PRS) based on association data of celiac disease, 2) presence of specific amino acid residues in the MHC class II protein HLA-DRβ1, 3) values for follicle-stimulating hormone (FSH), and 4) FSH to luteinizing hormone ratio (FSH/LH). Main results and the role of chance The best performing machine learning model, an artificial neural network, achieved the following results on the testing dataset: AUC=0.65, accuracy=0.61, sensitivity=0.62 and specificity=0.61. The presence of specific HLA-DRβ1 residues was the most discriminative variable for distinguishing between SCO and non-SCO phenotypes, supporting the hypothesis of an immune-mediated component in most SCO patients. This finding aligns with the HLA-DRB1 locus being the main genome-wide association with SCO. Additionally, the PRSs calculated for each infertile man based on genome-wide association data for celiac disease were significantly higher in SCO patients compared to non-SCO patients (PRS model P-value=2.48E-02), highlighting its potential as a predictive parameter. Moreover, FSH levels and their ratio to LH contributed to SCO prediction, consistent with previous associations between elevated levels of FSH and this phenotype. Limitations, reasons for caution While the model demonstrated a promising AUC in diagnosing SCO patients and predicting unsuccessful TESE, further refinement is needed through incorporating of additional informative biomarkers. The performance of the model was limited by small sample size, which constrained the training and testing datasets. Therefore, replication in larger cohorts is warranted. Wider implications of the findings Machine learning models predicting TESE outcomes could provide infertile men due to SPGF a very valuable probability estimate of success based on a pre-biopsy histological phenotype diagnosis. While further improvements to the reported model are needed, it has the potential to help SCO patients avoid unnecessary invasive procedures. Trial registration number No

Ähnliche Arbeiten