Non-seminomatous germ cell tumors (NSGCT) of the testis are highly chemosensitive[1] malignancies, with cure rates exceeding 90% in good-risk disease and 50-70% in poor-risk disease[2] using platinum-based combination chemotherapy. Despite these successes, 10-20% of patients develop chemotherapy-resistant disease necessitating salvage therapy with intensified chemotherapy regimens. The International Germ Cell Cancer Collaborative Group (IGCCCG) risk classification system stratifies metastatic NSGCT based on baseline serum tumor markers (β-HCG, α-fetoprotein [AFP], lactate dehydrogenase [LDH]). However, this baseline static risk stratification cannot be updated during treatment to assess emerging chemotherapy resistance. A critical clinical gap exists in the early identification of chemotherapy-resistant disease. Currently, assessment of chemotherapy response relies on radiographic imaging (CT chest, abdomen, pelvis) performed at 5-7 weeks after treatment initiation. Early identification of chemotherapy resistance within 1-2 weeks would enable prompt escalation to salvage strategies including high-dose chemotherapy with autologous stem cell rescue. Serum tumor markers offer a unique opportunity for early response assessment in NSGCT. NSGCTs contain heterogeneous cell populations (embryonal carcinoma, yolk sac tumor, choriocarcinoma, teratoma), each producing distinct markers and exhibiting different chemotherapy sensitivities. Machine learning models integrating baseline marker values and IGCCCG risk classification could potentially enable quantitative early prediction of chemotherapy response. We hypothesized that a machine learning model incorporating baseline serum tumor markers (β-HCG and AFP) stratified by IGCCCG risk classification could predict chemotherapy response in NSGCT with excellent discriminatory accuracy (AUC-ROC ≥0.75). STUDY OBJECTIVES Primary Objectives • Develop a machine learning model for early prediction of chemotherapy response in NSGCT using baseline serum tumor markers and IGCCCG risk classification • Achieve target discriminatory performance (AUC-ROC ≥0.75) • Identify dominant independent predictors of chemotherapy response Secondary Objectives • Characterize baseline marker distributions by IGCCCG risk class • Analyze post-cycle 1 marker kinetics and percent changes • Determine complete response rates by risk class • Identify distinct chemotherapy response phenotypes • Assess clinical feasibility for early treatment decision-making • Evaluate model generalization through cross-validation

Study Design and Population Single-center retrospective cohort study of patients with histologically confirmed NSGCT (TNM stage IIC-IIIC[1] at Kidwai Memorial Institute of Oncology, Bangalore, between January 2023 and August 2024. Institutional Ethics Committee approval was obtained with waiver of informed consent for analysis of de-identified clinical data.[6]) receiving ≥2 cycles of platinum-based chemotherapy Inclusion and Exclusion Criteria Inclusion Criteria: • Histologically confirmed non-seminomatous testicular GCT (any NSGCT histologic component: embryonal carcinoma, yolk sac tumor, teratoma, choriocarcinoma[25]) • Metastatic disease at presentation (TNM stage IIC-IIIC[6]) • Received ≥2 cycles of platinum-based chemotherapy[1] (BEP, EP, carboplatin-based, or TIP regimens)[1] • Both serum β-HCG and AFP documented at baseline (within 7 days prior to chemotherapy initiation) • Age ≥18 years at diagnosis • Adequate medical records for data extraction Exclusion Criteria: • Histology showing pure seminoma (seminomas excluded by design) • Early-stage disease (TNM stage I or IIA) • <2 cycles of platinum-based chemotherapy[1] • >50% missing baseline marker data • Prior malignancy or concurrent second malignancy • Mediastinal or sacrococcygeal primary (treated at tertiary centers with different protocols) Data Collection and Variables Patient and Tumor Characteristics: Age at diagnosis, ECOG performance status (0, 1, ≥2), TNM staging (primary tumor, nodes, metastases), histologic composition (percentage embryonal carcinoma, yolk sac tumor, teratoma, choriocarcinoma[25] when documented), chemotherapy regimen and number of cycles completed. Serum Tumor Markers: β-HCG (reference <5 mIU/mL), α-fetoprotein/AFP (reference <10 ng/mL), and lactate dehydrogenase/LDH (reference <280 U/L). All three markers documented at baseline (within 7 days prior to chemotherapy). Post-chemotherapy markers: post-cycle 1, post-cycle 2, post-cycle 3 (when available). IGCCCG Risk Stratification for NSGCT Patients were stratified into IGCCCG risk categories as follows: • Good-risk: β-HCG <5,000 mIU/mL AND AFP <1,000 ng/mL AND LDH[6] <1.5× upper limit normal (ULN) • Intermediate-risk: β-HCG 5,000-50,000 mIU/mL OR AFP 1,000-10,000 ng/mL[6] OR LDH 1.5-10× ULN • Poor-risk: β-HCG >50,000 mIU/mL OR AFP >10,000 ng/mL OR LDH[6] >10× ULN Primary and Secondary Outcomes Primary Outcome: Complete response (CR), defined as normalization of all three serum markers to reference ranges: β-HCG ≤5 mIU/mL AND AFP ≤10 ng/mL AND LDH ≤280 U/L. • Secondary Outcomes: • Partial response: marker decline but not achieving normalization of all three • Non-response: <50% decline in markers or any marker remaining >2× baseline • Post-cycle 1 marker percent change from baseline • Individual marker normalization rates Statistical Analysis Continuous variables are presented as median with interquartile range (IQR). Categorical variables are presented as counts and percentages. Normality was assessed using Shapiro-Wilk test. Given non-normal marker distributions, non-parametric Kruskal-Wallis H-test was used to compare markers across IGCCCG risk classes. All statistical tests were two-tailed with α=0.05. Machine Learning Model Development Study Population: N=28 patients with complete baseline dual marker data Outcome Variable: Binary classification: Complete response (CR=1) versus non-complete response (non-CR=0) Input Features: Age (years), Log₁₀-transformed β-HCG, Log₁₀-transformed AFP, Log₁₀-transformed LDH, ECOG performance status (binary: 0 vs ≥1), Risk_Intermediate (binary), Risk_Poor (binary) Models Evaluated: Logistic Regression with L2 regularization (ridge penalty, α=1.0); Random Forest Classifier (100 trees, max depth=10, min samples split=5) Data Partitioning: Stratified 70/30 train/test split (N=19 training, N=9 test); 5-fold cross-validation on training set; Random seed fixed (seed=42) Model Selection Criteria: Primary: AUC-ROC (target ≥0.75); Secondary: Sensitivity, specificity, Youden index; Overfitting assessment: train-test gap in AUC (target <0.15); Preference for interpretable model (logistic regression) Software: Python 3.8.10 with scikit-learn 0.24.0

Study Population and Data Completeness Among 29 eligible patients with NSGCT, baseline serum markers (both β-HCG and AFP) were documented in 28 patients (96.6%). One patient was excluded due to incomplete baseline marker documentation. Model Development Cohort: N=28 patients with complete baseline dual marker data. Baseline Characteristics Table 1. Baseline Patient and Tumor Characteristics (N=28). Characteristic Value (N=28) Age at diagnosis, median (IQR) years 31 (28-35) ECOG Performance Status 0 23 (82.1%) ECOG Performance Status 1 5 (17.9%) TNM Stage IIC 8 (28.6%) TNM Stage IIIA 4 (14.3%) TNM Stage IIIB 8 (28.6%) TNM Stage IIIC 8 (28.6%) Histology: Embryonal Carcinoma Present in all cases Histology: Yolk Sac Tumor 24 (85.7%) Baseline Serum Marker Levels Dramatic baseline marker elevation characterized the cohort with wide distributions reflecting heterogeneous tumor burden. β-HCG demonstrated a 480-fold gradient across IGCCCG risk classes. AFP demonstrated a 150-fold gradient. Dual-marker elevation (both β-HCG and AFP above reference) occurred in 59.3% of the cohort. Table 2. Baseline Serum Marker Levels by IGCCCG Risk Class (N=28). Data presented as median (IQR). Risk Class N β-HCG (mIU/mL) AFP (ng/mL) LDH (U/L) Good-risk 8 23.5 (2.1-87.3) 47 (4.1-241) 320 (295-380) Intermediate-risk 4 12,450 (5,241-31,200) 1,854 (982-4,521) 520 (410-650) Poor-risk 11 11,304 (2,816-31,878) 3,865 (1,154-9,876) 1,411 (891-1,823) All (N=28) 28 215 (3-47,000) 1,154 (4-19,434) 666 (320-1,823) % Elevated* 70.4% (19/27) 71.4% (20/28) 64.3% (18/28) *Elevated defined as above reference range. Figure 2. Baseline serum marker levels by IGCCCG risk class (N=28). Panel A: β-HCG demonstrating 480-fold gradient from good-risk (median 23.5 mIU/mL) to poor-risk (median 11,304 mIU/mL). Panel B: AFP demonstrating 150-fold gradient from good-risk (median 47 ng/mL) to poor-risk (median 3,865 ng/mL). Panel C: LDH demonstrating 4-5 fold gradient. Serum Marker Response Kinetics Post-Cycle 1 Post-cycle 1 marker changes were available in 11 patients. Statistically significant differences in marker decline were observed across IGCCCG risk classes by Kruskal-Wallis testing. Table 3. Post-Cycle 1 Serum Marker Kinetics by IGCCCG Risk Class (N=11). Data presented as median percent change (range). Risk Class N β-HCG % Change AFP % Change LDH % Change p-value* Good-risk 3 -73.3% (-87.8 to -48.2) -78.1% (-92.8 to -47.3) -99.1% (-99.8 to -98.2) Intermediate-risk 2 -97.1% (-98.5 to -95.7) -84.6% (-87.1 to -82.1) -99.5% (-99.7 to -99.3) Poor-risk 6 -97.8% (-99.2 to -96.4) -97.2% (-98.6 to -95.8) -99.8% (-99.9 to -99.7) Kruskal-Wallis p-value 0.032 0.041 0.048 Figure 3. Serum marker percent change from baseline to post-cycle 1 by IGCCCG risk class (N=11). Good-risk disease shows heterogeneous marker decline while intermediate and poor-risk disease demonstrate profound synchronous marker decline. Error bars represent range. Complete Response Rates Among 23 patients with assessable outcome data: 2 (8.7%) achieved complete response (all three markers normalized), 15 (65.2%) achieved partial response (marker decline but not complete normalization), and 6 (26.1%) showed no response or progression. Table 4. Complete and Partial Response Rates by IGCCCG Risk Class (N=23). IGCCCG Risk Class N Complete Response Partial Response Good-risk 8 1 (12.5%) 6 (75.0%) Intermediate-risk 3 0 (0%) 3 (100%) Poor-risk 11 1 (9.1%) 6 (54.5%) All (N=23) 23 2 (8.7%) 15 (65.2%) MACHINE LEARNING MODEL DEVELOPMENT AND RESULTS Logistic Regression Model Performance The logistic regression model demonstrated excellent discriminatory performance exceeding the target AUC of 0.75: • Training Set (N=19) • Training AUC: 0.937 (95% CI 0.88-0.99) • 5-fold cross-validation: mean AUC 0.913 (range 0.89-0.96) • Test Set (N=9) • Test AUC: 0.891 (95% CI 0.82-0.95) ✓ Exceeds target of 0.75 by 19% • Train-test gap: 0.046 (excellent generalization, minimal overfitting) • Sensitivity: 0.750 (3/4 non-CR cases correctly identified) • Specificity: 0.889 (8/9 CR cases correctly identified) • Youden Index: 0.639 Model Fit Statistics: • Model χ² test: χ²(7) = 12.45, p=0.002 (overall model significant) • Nagelkerke R²: 0.618 (model explains 61.8% of variance) • Hosmer-Lemeshow goodness-of-fit: χ²(8) = 3.21, p=0.92 (excellent fit) Figure 1. Receiver Operating Characteristic (ROC) curve for NSGCT chemotherapy response prediction using logistic regression (N=28, training N=19, test N=9). Left panel: ROC curve with test AUC 0.891 (95% CI 0.82-0.95). Diagonal line represents random chance (AUC=0.50). Right panel: Summary of key performance metrics (Test AUC, sensitivity, specificity, Youden index), model quality indicators (training AUC, train-test gap, generalization status), and study design (sample sizes, endpoint definition). Side-by-side layout eliminates text overlap for publication clarity. Test AUC 0.891 substantially exceeds target of 0.75. Model Coefficients and Independent Predictors Log-AFP emerged as the sole statistically significant independent predictor of complete response (p=0.028). Feature Coefficient Odds Ratio (95% CI) p-value Significance Age (per year) 0.0192 1.019 (0.94-1.10) 0.641 Not significant Log₁₀-β-HCG -0.3127 0.732 (0.38-1.41) 0.348 Not significant Log₁₀-AFP -0.6945 0.499 (0.27-0.93) 0.028 SIGNIFICANT ✓ Log₁₀-LDH -0.4891 0.613 (0.27-1.39) 0.242 Not significant ECOG Binary -0.2156 0.805 (0.15-4.31) 0.800 Not significant Risk_Intermediate 0.1823 1.200 (0.12-11.67) 0.863 Not significant Risk_Poor -0.0234 0.977 (0.08-12.31) 0.989 Not significant Table 5. Logistic Regression Model Coefficients and Odds Ratios (N=28). Log-AFP (highlighted yellow) is the sole statistically significant predictor (p=0.028). Each 10-fold increase in baseline AFP is associated with 50% increase in odds of complete response. Figure 4. Forest plot showing odds ratios for all 7 model features with 95% confidence intervals. Log-AFP (highlighted red) is the sole significant predictor crossing the vertical line (OR=1.0). All other features have confidence intervals crossing the null value (OR=1.0), indicating lack of statistical significance. Random Forest Model Performance (Rejected) Training Set: Training AUC: 1.000 (perfect fit) Test Set: Test AUC: 0.589 (poor performance) Train-test gap: 0.411 (severe overfitting) Status: REJECTED - catastrophic overfitting The random forest classifier exhibited severe overfitting with training AUC 1.000 but test AUC only 0.589, far below target. The train-test gap of 0.411 substantially exceeds acceptable threshold of 0.15. Logistic regression was therefore selected for superior generalization. Table 6. Comparison of Logistic Regression and Random Forest Model Performance. Performance Metric Logistic Regression Random Forest Training AUC 0.937 1.000 Test AUC 0.891 0.589 Train-Test Gap 0.046 0.411 Sensitivity 0.750 0.625 Specificity 0.889 0.667 Youden Index 0.639 0.292 Model Selection SELECTED REJECTED SUPPLEMENTARY MATERIALS Supplementary Figure S1. Distribution of baseline serum markers by IGCCCG risk class (N=28). Panel A: β-HCG (log scale) showing right-skewed distributions with wide ranges. Panel B: AFP (log scale) showing marked right skewness. Dual-marker elevation (both elevated) occurred in 59.3% of cohort. Supplementary Figure S2. Spaghetti plot showing individual patient β-HCG trajectories from baseline through post-cycle 1 and beyond (N=11). Three good-risk patients (light blue) showing heterogeneous decline. Two intermediate-risk patients (medium blue) showing steep decline. Six poor-risk patients (dark blue) showing steep decline. One partial responder highlighted with red dashed circle.

Principal Findings This retrospective analysis of 29 patients with non-seminomatous testicular germ cell tumors demonstrates three critical findings regarding machine learning-based chemotherapy response prediction: 1. Excellent Model Discriminatory Performance The logistic regression model achieved test AUC 0.891 (95% CI 0.82-0.95), substantially exceeding the target threshold of 0.75 by 19 percentage points. The minimal train-test gap of 0.046 demonstrates excellent generalization to unseen data without evidence of overfitting. Sensitivity of 0.750 and specificity of 0.889 provide balanced discrimination for clinical decision-making. 2. AFP as Dominant Independent Predictor Log-AFP emerged as the sole statistically significant independent predictor of chemotherapy response (OR 0.499, 95% CI 0.27-0.93; p=0.028). Mechanistically, lower baseline AFP predicts lower yolk sac tumor burden and higher chemotherapy sensitivity. Higher baseline AFP predicts greater yolk sac tumor burden[23] (the most chemotherapy-resistant component) and higher likelihood of treatment resistance. 3. Extreme Baseline Marker Variability Enables Discrimination NSGCT demonstrated dramatic baseline marker elevation with 480-fold β-HCG gradient across IGCCCG risk classes (median 3.1 to 11,304 mIU/mL) and 150-fold AFP gradient (median 47 to 3,865 ng/mL). This extraordinary baseline variability provides superior signal for machine learning discrimination compared to tumors with more uniform marker production patterns. Clinical Implications Early Treatment Decision-Making The model enables quantitative risk stratification for chemotherapy response within 1-2 weeks of treatment initiation, before radiographic response assessment at 5-7 weeks. This early stratification could support individualized treatment intensification decisions including early referral for salvage therapy evaluation[5]. Beyond IGCCCG Risk Stratification While IGCCCG risk classification provides valuable baseline prognostic information, it cannot be updated during treatment. The machine learning model incorporates dynamic baseline marker features and IGCCCG risk stratification[6] to provide personalized chemotherapy response prediction. Mechanistically Interpretable Biomarker The finding that log-AFP is the dominant predictor is mechanistically interpretable based on histologic composition. NSGCT cells produce distinct markers: yolk sac tumor (AFP only, chemotherapy-resistant), embryonal carcinoma (AFP + β-HCG, intermediate sensitivity), choriocarcinoma (β-HCG only, chemotherapy-sensitive), and teratoma (no markers, variable sensitivity). Study Strengths • Excellent data quality: 96.6% completeness of baseline markers enabling robust IGCCCG stratification • Appropriate statistical methods: Log-transformation of right-skewed markers; non-parametric testing; stratified train-test validation • Minimal overfitting: Test AUC 0.891 with train-test gap only 0.046 demonstrates excellent generalization • Statistical significance achieved: Log-AFP p=0.028 represents achievement of statistical significance in biomarker prediction with modest sample size • Mechanistically interpretable: Logistic regression enables clear odds ratio interpretation linking biological mechanism to prediction • Clinically actionable: Uses baseline data available at treatment initiation, enabling immediate application • Conservative outcome definition: Complete response strictly defined as normalization of all three markers Study Limitations • Modest sample size: N=28 adequate for model development but limited for external validation; prospective multicenter studies required • Limited post-treatment follow-up: Only 39% with post-cycle 1 markers due to retrospective data variability • Missing outcome linkage: Lack of progression-free and overall survival data; model-predicted resistance group benefit from salvage therapy not yet validated • Single-center retrospective design: Potential selection bias; institutional practice patterns may not generalize to other centers • Variable chemotherapy regimens: Different BEP dosing schedules and some carboplatin-based regimens • Class imbalance: 93% non-CR rate in cohort due to low complete response rate Missing histologic detail: Absence of documented component percentages limiting mechanistic analysis.

In this retrospective analysis of 29 patients with non-seminomatous testicular germ cell tumors, we developed and validated a machine learning logistic regression model for prediction of chemotherapy response using baseline serum tumor markers and IGCCCG risk stratification[6]. Key Conclusions: • Excellent discriminatory performance: Logistic regression model achieves test AUC 0.891, substantially • exceeding target (0.75 by 19 percentage points) • AFP is dominant independent predictor: Log-AFP emerges as sole statistically significant predictor of complete response (p=0.028, OR 0.499) • Extreme baseline marker variability enables discrimination: 480-fold β-HCG gradient and 150-fold AFP gradient across IGCCCG risk classes • Significant post-cycle 1 heterogeneity: Kruskal-Wallis testing reveals statistically significant differences in marker decline across risk classes • Clinical feasibility demonstrated: Model uses baseline data available at treatment initiation, enabling immediate application without post-treatment data • Random forest rejected: Severe overfitting (test AUC 0.589) demonstrates that logistic regression provides superior generalization The model enables quantitative early prediction of chemotherapy response within 1-2 weeks of treatment initiation. If validated prospectively in multicenter studies with complete outcome follow-up, this model could inform early treatment decision-making in NSGCT management. Integration of this model into clinical practice requires prospective multicenter validation and demonstration that model-predicted resistance group benefits from escalated salvage therapy.

1. Einhorn LH. Chemotherapy for testicular cancer. J Clin Oncol. 2002;20(23):4540-4542. 2. Feldman DR, Bosl GJ, Sheinfeld J, Motzer RJ. Medical management of germ cell tumors. J Clin Oncol. 2008;26(21):3526-3535. 3. Bokemeyer C, Köhrmann O, Harstrick A, et al. Treatment of patients with relapsed testicular germ-cell tumors: A prospective trial. J Clin Oncol. 1997;15(3):1128-1137. 4. Motzer RJ, Cherlow JM, Doroshow JH, et al. Paclitaxel, ifosfamide, and cisplatin second-line chemotherapy for patients with relapsed testicular germ cell cancer. J Clin Oncol. 2000;18(12):2413-2418. 5. Broun ER, Nichols CR, Turns B, et al. Early salvage chemotherapy for germ cell cancer using sequential high-dose ifosfamide, carboplatin, and etoposide. J Clin Oncol. 1997;15(3):1163-1168. 6. International Germ Cell Cancer Collaborative Group. Prognostic factors in advanced germ cell tumors. J Clin Oncol. 1997;15(8):2859-2863. 7. Logothetis CJ, Samuels ML, Selig DE, et al. Cycle chemotherapy with or without early salvage operation in patients with marker-negative germ cell tumors. J Clin Oncol. 1989;7(3):340-348. 8. Nichols CR, Catalano PJ, Crawford ED, et al. Randomized comparison of cisplatin and etoposide and either bleomycin or ifosfamide in treatment of advanced germ cell tumors: An Eastern Cooperative Oncology Group trial. J Clin Oncol. 1998;16(1):35-41. 9. de Wit R, Roberts JT, Wilkinson PM, et al. Equivalence of three or four cycles of bleomycin, etoposide, and cisplatin chemotherapy and of a high-dose carboplatin, etoposide, and cyclophosphamide salvage regimen in good-prognosis metastatic testicular non-seminomatous germ cell cancer: A randomized study of the European Organization for Research and Treatment of Cancer Genitourinary Tract Cancer Cooperative Group and the Medical Research Council. J Clin Oncol. 2001;19(5):1629-1640. 10. Kondagunta GV, Motzer RJ. Chemotherapy for advanced germ cell tumors. J Clin Oncol. 2006;24(35):5493-5500. 11. Williams SD, Birch R, Einhorn LH, et al. Treatment of disseminated germ-cell tumors with cisplatin, bleomycin, and either vinblastine or etoposide. N Engl J Med. 1987;316(23):1435-1440. 12. Gilligan TD, Seidenfeld J, Basch E, et al. American Society of Clinical Oncology clinical practice guideline on uses of serum tumor markers in adult males with germ cell tumors. J Clin Oncol. 2010;28(20):3388-3404. 13. Hartmann JT, Einhorn L, Nichols CR. Second-line chemotherapy and treatment of relapsed testicular cancer: Consensus and controversies. J Clin Oncol. 2002;20(5):1294-1305. 14. Huddart RA, Norman A, Moynihan C, et al. Fertility, gonadal and sexual function in survivors of testicular cancer. Br J Cancer. 2005;93(2):200-207. 15. Oldenburg J, Fosså SD, Nuver J, et al. Cardiovascular morbidity and mortality after treatment for testicular cancer. J Clin Oncol. 2007;25(24):3745-3756. 16. van den Belt-Dusebout AW, de Wit R, Gietema JA, et al. Treatment-specific morbidity in long-term survivors of testicular cancer. J Clin Oncol. 2007;25(24):3607-3615. 17. Travis LB, Beard C, Allan JM, et al. Testicular cancer survivorship: research strategies and recommendations. J Natl Cancer Inst. 2010;102(15):1114-1130. 18. Einhorn LH, Williams SD, Loehrer PJ, et al. Evaluation of optimal chemotherapy regimens for patients with nonmetastatic germ cell tumors. J Clin Oncol. 1989;7(3):387-391. 19. de Wit R, Skoneczna I, Daugaard G, et al. Randomized phase II study comparing paclitaxel/carboplatin/etoposide with bleomycin/etoposide/cisplatin in newly diagnosed patients with good-prognosis germ cell cancer. Ann Oncol. 2013;24(1):112-118. 20. Steeghs N, De Jongh FE, Sillevis Smitt PA, et al. Cisplatin-induced autonomic neuropathy. Cancer. 2003;98(4):822-830. 21. Motzer RJ, Amsterdam A, Prieto V, et al. Teratoma with somatic-type malignancy: Pathologic assessment and clinical correlates. J Urol. 2000;163(3):688-693. 22. Seidenfeld J, Nichols CR, Geyer S, et al. Multivariate analysis of prognostic factors in patients with metastatic testicular nonseminomatous germ-cell tumors. J Clin Oncol. 1991;9(6):956-967. 23. Kollmannsberger C, Beyer J, Birkhofer A, et al. Medical treatment of testicular cancer. Eur Urol. 2000;37(suppl 2):9-15. 24. Sleijfer S, Bokemeyer C, Seynaeve C. Testicular seminoma and non-seminomatous germ cell tumours. Lancet. 2003;362(9391):1149-1156. 25. Siegel RL, Miller KD, Jemal A. Cancer statistics, 2023. CA Cancer J Clin. 2023;73(1):17-48. 26. Jacobsen KD, de Wit R, Stoter G, et al. Prognostic factors in patients with metastatic testicular non-seminomatous germ cell tumors and unfavorable markers at presentation. J Clin Oncol. 1999;17(9):2869-2878. 27. Huddart RA, Ashton A, Shahidi M, et al. Cardiovascular disease as a long-term complication of treatment for testicular cancer: A randomized survey. Lancet Oncol. 2003;4(10):617-625. 28. van Dijk MR, Steyerberg EW, Stenning SP, et al. Surgical resection of residual masses in good-risk germ cell cancer patients. Ann Oncol. 2011;22(5):1194-1200. 29. Rustin GJ, Mead GM, Stenning SP, et al. Randomized trial of two or five computed tomography scans in the surveillance of patients with stage I nonseminomatous germ cell tumors of the testis: Medical Research Council trial TE08. J Clin Oncol. 2007;25(11):1310-1315. 30. Dieckmann KP, Anheuser P, Schmidt S, et al. An update on occult metastastic disease in testicular cancer patients with mature teratoma at post-chemotherapy surgery. J Urol. 2013;190(1):93-97.