A federated approach to identify women with early-stage cervical cancer at low risk of lymph node metastases.

Objective: Lymph node metastases (pN+) in presumed early-stage cervical cancer negatively impact prognosis. Using federated learning, we aimed to develop a tool to identify a group of women at low risk of pN+, to guide the shared decision-making process concerning the extent of lymph node dissection.

Methods: Women with cervical cancer between 2005 and 2020 were identified retrospectively from population-based registries: the Danish Gynaecological Cancer Database, Swedish Quality Registry for Gynaecologic Cancer and Netherlands Cancer Registry. Inclusion criteria were: squamous cell carcinoma, adenocarcinoma or adenosquamous carcinoma; The International Federation of Gynecology and Obstetrics 2009 IA2, IB1 and IIA1; treatment with radical hysterectomy and pelvic lymph node assessment. We applied privacy-preserving federated logistic regression to identify risk factors of pN+. Significant factors were used to stratify the risk of pN+.

Results: We included 3606 women (pN+ 11%). The most important risk factors of pN+ were lymphovascular space invasion (LVSI) (odds ratio [OR] 5.16, 95% confidence interval [CI], 4.59-5.79), tumour size 21-40 mm (OR 2.14, 95% CI, 1.89-2.43) and depth of invasion>10 mm (OR 1.81, 95% CI, 1.59-2.08). A group of 1469 women (41%)-with tumours without LVSI, tumour size ≤20 mm, and depth of invasion ≤10 mm-had a very low risk of pN+ (2.4%, 95% CI, 1.7-3.3%).

Conclusion: Early-stage cervical cancer without LVSI, a tumour size ≤20 mm and depth of invasion ≤10 mm, confers a low risk of pN+. Based on an international privacy-preserving analysis, we developed a useful tool to guide the shared decision-making process regarding lymph node dissection.