Background

Prostate cancer patients with pelvic lymph node metastasis (PLNM) have poor prognosis. Based on EAU guidelines, patients with >5% risk of PLNM by nomograms often receive pelvic lymph node dissection (PLND) during prostatectomy. However, nomograms have limited accuracy, so large numbers of false positive patients receive unnecessary surgery with potentially serious side effects. It is important to accurately identify PLNM, yet current tests, including imaging tools are inaccurate. Therefore, we intended to develop a gene expression-based algorithm for detecting PLNM.

Methods

An advanced random forest machine learning algorithm screening was conducted to develop a classifier for identifying PLNM using urine samples collected from a multi-center retrospective cohort (n = 413) as training set and validated in an independent multi-center prospective cohort (n = 243). Univariate and multivariate discriminant analyses were performed to measure the ability of the algorithm classifier to detect PLNM and compare it with the Memorial Sloan Kettering Cancer Center (MSKCC) nomogram score.

Results

An algorithm named 25 G PLNM-Score was developed and found to accurately distinguish PLNM and non-PLNM with AUC of 0.93 (95% CI: 0.85–1.01) and 0.93 (95% CI: 0.87–0.99) in the retrospective and prospective urine cohorts respectively. Kaplan–Meier plots showed large and significant difference in biochemical recurrence-free survival and distant metastasis-free survival in the patients stratified by the 25 G PLNM-Score (log rank P < 0.001 and P < 0.0001, respectively). It spared 96% and 80% of unnecessary PLND with only 0.51% and 1% of PLNM missing in the retrospective and prospective cohorts respectively. In contrast, the MSKCC score only spared 15% of PLND with 0% of PLNM missing.

Conclusions

The novel 25 G PLNM-Score is the first highly accurate and non-invasive machine learning algorithm-based urine test to identify PLNM before PLND, with potential clinical benefits of avoiding unnecessary PLND and improving treatment decision-making.

中文翻译：

---

用于检测前列腺癌盆腔淋巴结转移的非侵入性 25 基因 PLNM-Score 尿液检测

背景

前列腺癌盆腔淋巴结转移（PLNM）患者预后较差。根据 EAU 指南，列线图显示 PLNM 风险 >5% 的患者通常在前列腺切除术期间接受盆腔淋巴结清扫术 (PLND)。然而，列线图的准确性有限，因此大量假阳性患者接受了不必要的手术，并可能产生严重的副作用。准确识别 PLNM 非常重要，但目前的测试（包括成像工具）并不准确。因此，我们打算开发一种基于基因表达的算法来检测 PLNM。

方法

采用先进的随机森林机器学习算法筛选来开发识别 PLNM 的分类器，使用从多中心回顾性队列 ( n  = 413) 收集的尿液样本作为训练集，并在独立的多中心前瞻性队列 ( n  = 243)中进行验证）。进行单变量和多变量判别分析来测量算法分类器检测 PLNM 的能力，并将其与纪念斯隆凯特琳癌症中心 (MSKCC) 列线图评分进行比较。

结果

开发了一种名为 25 G PLNM-Score 的算法，发现可以准确地区分 PLNM 和非 PLNM，回顾性和前瞻性尿液中的 AUC 分别为 0.93 (95% CI: 0.85–1.01) 和 0.93 (95% CI: 0.87–0.99)分别为队列。 Kaplan-Meier 图显示，按 25 G PLNM 评分分层的患者的生化无复发生存率和无远处转移生存率存在巨大且显着的差异（对数等级分别为P  < 0.001 和P  < 0.0001）。它避免了 96% 和 80% 不必要的 PLND，在回顾性和前瞻性队列中分别仅缺失 0.51% 和 1% 的 PLNM。相比之下，MSKCC 评分仅保留 15% 的 PLND，缺失 0% 的 PLNM。

结论

新型 25 G PLNM-Score 是第一个基于机器学习算法的高精度、非侵入性尿液检测，可在 PLND 之前识别 PLNM，具有避免不必要的 PLND 和改善治疗决策的潜在临床益处。