机器学习预测挽救性睾丸取精成功率:方法优势与临床转化局限
Comment on: predicting sperm retrieval success in salvage testicular sperm extraction: a machine learning perspective
这篇评论讨论用机器学习预测非梗阻性无精子症患者挽救性睾丸取精成功率。Hazir等基于503例多中心队列,XGBoost模型AUC为0.96,SHAP显示BMI、睾丸探查范围、LH和精液量为最重要变量。作者指出单双侧探查是术中变量,无法用于术前咨询;常规与显微镜下取精失败后的临床情境不同;模型缺乏外部验证和校准评估,需进一步研究其真正临床效用。
Non-obstructive azoospermia (NOA) remains one of the most challenging conditions in male infertility [1]. Although microdissection testicular sperm extraction (mTESE) has improved the chances of sperm retrieval, a substantial proportion of men experience an unsuccessful first procedure. For these men, the decision to undergo another surgical attempt requires balancing an uncertain probability of sperm retrieval against the additional surgical, financial, and psychological burden of salvage surgery. Reliable prediction of sperm retrieval could therefore substantially improve patient counseling and shared decision-making. In this context, Hazir et al. provide a timely contribution by applying machine-learning (ML) algorithms to a multicenter cohort of 503 men undergoing secondary testicular sperm extraction after a previous unsuccessful testicular sperm extraction (TESE) [2]. Sperm retrieval was achieved in 41.9% of patients, and among the ten evaluated algorithms, XGBoost achieved an area under the receiver operating characteristic curve (AUC) of 0.96. SHapley Additive exPlanations (SHAP) identified body mass index (BMI), unilateral versus bilateral testicular exploration, luteinizing hormone (LH), and semen volume as the most influential variables, with testicular volume, infertility duration, and other hormonal parameters providing additional predictive information [2]. These findings support the growing interest in ML as a means of integrating multiple clinical variables whose relationship with residual spermatogenesis is unlikely to be purely linear [3, 4]. The study has several important strengths. Its multicenter design, large cohort, and use of explainable ML methods represent relevant methodological advantages. In particular, SHAP analysis improves model interpretability by illustrating the relative contribution of individual predictors, an important step toward the potential clinical translation of ML-based prediction tools. However, translating this predictive performance into clinical practice requires careful consideration of whether the variables included in the model are available at the time of preoperative counseling. This issue is particularly relevant to the extent of testicular exploration. In the study, TESE was initially performed on one testis, with exploration of the contralateral testis only when sperm retrieval from the first side was unsuccessful [2]. Therefore, whether exploration was unilateral or bilateral was determined during surgery and was unknown at the time of preoperative counseling. Including this intraoperative variable may have contributed to the high predictive performance of the model. Repeating the analysis without this variable could provide a more clinically applicable estimate of preoperative predictive performance. Another important consideration is that not all secondary TESE procedures represent the same clinical scenario. Approximately one-third of men in the cohort underwent conventional rather than microscopic secondary TESE [2]. More importantly, the prognostic meaning of a previous unsuccessful conventional TESE may differ substantially from that of a failed mTESE. Salvage mTESE after conventional TESE can identify focal areas of spermatogenesis missed during the initial procedure, with clinically meaningful retrieval rates reported in selected cohorts [5]. Conversely, repeating mTESE after an unsuccessful procedure performed with extensive microscopic exploration represents a different and probably more challenging scenario [6, 7]. A subgroup analysis comparing men undergoing salvage mTESE after failed conventional TESE with those undergoing repeat mTESE after failed mTESE would therefore be particularly informative. Previous and salvage surgical techniques, the extent of the initial exploration, and potentially surgeon and embryology laboratory experience should therefore be considered when developing future prediction tools. The availability of histopathological information from the previous TESE is another particularly valuable feature of the study. Hypospermatogenesis, maturation arrest, Sertoli cell-only syndrome, and preserved spermatogenesis were significantly associated with secondary retrieval outcomes [2]. This is consistent with previous evidence. In a multicenter study, Boeri et al. reported successful sperm retrieval in 49.2% of men undergoing salvage mTESE after failed conventional TESE, with hypospermatogenesis emerging as the only independent predictor of successful retrieval [5]. Similarly, a systematic review and meta-analysis identified histological pattern, together with age, testicular volume, FSH, and LH, as potentially informative predictors of salvage mTESE outcomes [7]. Interestingly, histology was not among the highest-ranking SHAP features in the model by Hazir et al. [2]. This apparent discrepancy illustrates an important distinction between statistical importance and clinical relevance. ML algorithms identify the combination of variables that best optimizes prediction within a given dataset; feature ranking does not necessarily reproduce the biological or prognostic hierarchy established in clinical practice. SHAP analysis is valuable for interpreting ML models, but the importance assigned to individual predictors should always be considered in the context of established clinical evidence. Given that men with Klinefelter syndrome (KS) accounted for 12.7% of the cohort and KS was significantly associated with retrieval outcome, a subgroup analysis in this population would have been particularly informative [2]. In a recent large Italian multicenter cohort of 383 men with nonmosaic KS undergoing TESE, sperm retrieval was achieved in 29.8% of patients, with higher preoperative total testosterone levels and favorable testicular histology independently associated with successful retrieval [8]. Assessing the performance and feature importance of the proposed ML model specifically in men with KS could therefore provide valuable information on its applicability to this clinically relevant subgroup. A further consideration concerns the transition from discrimination to clinical utility. Previous ML approaches to sperm retrieval prediction have shown promising but generally more modest performance. Zeadna et al. reported an AUC of approximately 0.81 using gradient-boosted trees [3], while Bachelot et al. subsequently reported an AUC of 0.90 in a prospectively tested cohort [4]. Although an AUC of 0.96 indicates excellent overall discrimination, its clinical relevance across the heterogeneous surgical and etiological subgroups included in the study remains to be established. Despite its multicenter nature, the study by Hazir et al. remains retrospective and lacks independent external validation [2]. This is particularly relevant in NOA, where patient characteristics, surgical technique, tissue processing, and embryology expertise may vary substantially across institutions. Moreover, AUC alone does not establish whether a model improves clinical decision-making. Calibration is essential if predicted probabilities are to be communicated meaningfully to couples. Decision-curve analysis could further determine whether using the model provides a net clinical benefit across relevant probability thresholds. Ultimately, the question is not simply whether an algorithm can distinguish successful from unsuccessful procedures, but whether its prediction would change counseling or the decision to proceed with another surgical attempt. Hazir et al. should be commended for addressing an important and difficult clinical problem using a large multicenter dataset. Notwithstanding these promising results, external validation remains the critical next step before clinical implementation. Confirmation of model performance across independent populations and healthcare settings will be essential to establish its generalizability and clinical utility. The next step will be to determine whether this excellent discrimination can be maintained in independent populations using variables that are genuinely available at the time of preoperative counseling, while accounting for the technique and findings of the previous sperm retrieval attempt and the underlying etiology of NOA. Ultimately, ML models may complement established prognostic factors and clinical expertise. Such an approach could contribute to increasingly individualized counseling for men considering salvage sperm retrieval and represents an important direction for future research.
This is a preview of subscription content, access via your institution
Access options
Access through your institution
Subscribe to this journal
Receive 12 print issues and online access
251,40 € per year
only 20,95 € per issue
Buy this article
- Purchase on SpringerLink
- Instant access to the full article PDF.
39,95 €
Prices may be subject to local taxes which are calculated during checkout
Additional access options:
References
Minhas S, Boeri L, Capogrosso P, Cocci A, Corona G, Dinkelman-Smit M, et al. European Association of Urology Guidelines on male sexual and reproductive health: 2025 update on male infertility. Eur Urol. 2025;87:601–16. https://doi.org/10.1016/j.eururo.2025.02.026.
Hazir B, Sahin A, Gultekin MH, Ozer C, Kayra MV, Hasirci E, et al. Predicting sperm retrieval success in salvage testicular sperm extraction: a machine learning perspective. Int J Impot Res. 2026. https://doi.org/10.1038/s41443-026-01291-y
Zeadna A, Khateeb N, Rokach L, Lior Y, Har-Vardi I, Harlev A, et al. Prediction of sperm extraction in non-obstructive azoospermia patients: a machine-learning perspective. Hum Reprod. 2020;35:1505–14. https://doi.org/10.1093/humrep/deaa109.
Bachelot G, Dhombres F, Sermondade N, Haj Hamid R, Berthaut I, Frydman V, et al. A machine learning approach for the prediction of testicular sperm extraction in nonobstructive azoospermia: algorithm development and validation study. J Med Internet Res. 2023;25:e44047 https://doi.org/10.2196/44047.
Boeri L, Bebi C, Dente D, Greco E, Turetti M, Capece M, et al. Outcomes and predictive factors of successful salvage microdissection testicular sperm extraction (mTESE) after failed classic TESE: results from a multicenter cross-sectional study. Int J Impot Res. 2022;34:795–9. https://doi.org/10.1038/s41443-021-00487-8.
Ramasamy R, Ricci JA, Leung RA, Schlegel PN. Successful repeat microdissection testicular sperm extraction in men with nonobstructive azoospermia. J Urol. 2011;185:1027–31. https://doi.org/10.1016/j.juro.2010.10.066.
Zhang F, Dai M, Yang X, Cheng Y, Ye L, Huang W, et al. Predictors of successful salvage microdissection testicular sperm extraction (mTESE) after failed initial TESE in patients with non-obstructive azoospermia: a systematic review and meta-analysis. Andrology. 2024;12:30–44. https://doi.org/10.1111/andr.13448.
Raffo M, Boeri L, Iafrate M, Negri F, Birolini G, Ramadani R, et al. Serum total testosterone and testicular histology help predict positive sperm retrieval in nonmosaic Klinefelter patients undergoing testicular sperm extraction: real-life data from a large multicenter cohort. Fertil Steril. 2026;125:45–53. https://doi.org/10.1016/j.fertnstert.2025.07.040.
Funding
This work received no external funding.
Author information
Authors and Affiliations
Division of Experimental Oncology/Unit of Urology; URI; IRCCS Ospedale San Raffaele, Milan, Italy
Massimiliano Raffo, Edoardo Pozzi, Fausto Negri, Francesco Montorsi & Andrea Salonia
University Vita-Salute San Raffaele, Milan, Italy
Massimiliano Raffo, Fausto Negri, Francesco Montorsi & Andrea Salonia
Department of Urology, University of Turin, Città della Salute e della Scienza, Turin, Italy
Marco Falcone
Department of Urology, Biruni University, School of Medicine, Istanbul, Turkey
Marco Falcone
Authors
- Massimiliano Raffo
- Edoardo Pozzi
- Fausto Negri
- Marco Falcone
- Francesco Montorsi
- Andrea Salonia
Contributions
MR and EP conceived the manuscript and drafted the initial version. FN critically revised the manuscript and contributed to the intellectual content. MF, FM, and AS critically reviewed and revised the manuscript. All the authors approved the final version of the manuscript.
Corresponding author
Correspondence to Massimiliano Raffo.
Ethics declarations
Competing interests
Marco Falcone is a Deputy Editor of the International Journal of Impotence Research and will not be involved in the peer-review process or in any editorial decisions regarding this manuscript. All other authors declare no competing interests.
Declarations of AI generative
ChatGPT-5.5 (OpenAI) was used exclusively to assist with English language editing and grammatical refinement. All scientific content, interpretations, and conclusions were developed by the authors, who take full responsibility for the manuscript.
Additional information
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
About this article
Cite this article
Raffo, M., Pozzi, E., Negri, F. et al. Comment on: predicting sperm retrieval success in salvage testicular sperm extraction: a machine learning perspective. Int J Impot Res (2026). https://doi.org/10.1038/s41443-026-01378-6
Received: 26 August 2026
Revised: 28 September 2026
Accepted: 06 October 2026
Published: 16 October 2026
Version of record: 16 October 2026
DOI: https://doi.org/10.1038/s41443-026-01378-6
来源:Nature 健康护理主题 · nature.com