In an era of rapid technology development in surgery, urologists remain at the forefront of surgical innovation. Incorporating new technology into routine clinical practice requires analysis of both safety and efficacy. Although healthcare research often relies heavily on randomised trials, limitations of randomised trials in surgical research are well known [1]. In the early stages of surgical innovation, outcomes are influenced by the surgeon’s learning curve (LC). Critical LC evaluation provides a nuanced understanding of early outcomes, as well as providing valuable information of what is required to achieve proficiency.

In this issue of the BJUI, Lenfant et al. [2] describe the LC of single-port (SP) robot-assisted extraperitoneal prostatectomy (SP-EPP) using the cumulative sum (CUSUM) method. Over a 19-month study period 150 consecutive cases were performed by a single surgeon (Kaouk). Anticipated outcomes were defined from a reference multiport (MP) cohort. Acceptable levels of complications were reached after ˜30 cases. Expected operative time based on the MP cohort was 168 min, whereas the mean observed SP-EPP operative time was 190 min with improvement occurring in four phases of 30–40 cases each. Beyond ˜120 cases, the surgeon reached a mastery level, where the study-defined trifecta outcome (surgical margin status, postoperative complications, and biochemical recurrence) threshold was reached.

Learning curves can be considered to have three components [3]. First is the ‘starting point’ and the second is the ‘slope’, which reflects how fast a person learns a new task. In the Lenfant et al. [2] paper, the study surgeon Kaouk, began his SP learning curve with a baseline experience of >2000 MP robot-assisted radical prostatectomies (RARPs). In addition, he undertook cadaveric and dry laboratory training. Together, these factors would undoubtably have affected his starting point, as well as the slope of the LC, contributing to a relatively short journey in transitioning to safe and proficient SP surgery. Consequently, the authors sensibly acknowledge the potential limitations of generalising their findings to novice surgeons. The third component to the LC is the ‘plateau’. This is where incremental change in the measured outcome is no longer significant. However, a key point here is that reaching a plateau does not necessarily equate to competence or safe performance [4].

Analysis of the surgical LC is complex. The surgical learning environment is prone to bias. Risk-adjustment models that could remove the noise from confounding factors such as patient-to-patient variation and operating room team composition would be immensely useful [4, 5]. It may be possible to remove the effect of variation by using some type of regression analysis, but this is uncommonly done. Ideal LC assessment would include multivariate analysis accounting for patient and surgical team factors [3].

Outcomes or endpoints also need to be selected carefully as few variables will assess true competency [4]. Generally, it is recommended that parallel analyses of several variables should be undertaken [5]. Operative time is often used as a surrogate for performance; however, faster procedures do not always imply better outcomes and may reflect the proficiency of the surgical team rather than individual ability. Lovegrove et al. [6] in their paper on a structured and modular training pathway for RARP, took a unique approach to LC assessment, assessing technical skill as an independent variable. Using healthcare failure mode effect analysis (HFMEA), to breakdown the surgical procedure into a series of critical steps, they were able to analyse the LC for each sub-step separately. As surgeons, we know that for each procedure there will be aspects that are more challenging than others. In keeping with this, the LCs for each of the sub-steps in the Lovegrove et al. [6] study varied. Manual assessment of performance of so many discrete steps is time consuming. However, the large volume of data generated in the sub-steps of robotic surgery lends itself to machine learning and we are beginning to see its application in robotic surgical education.

Surgeons are regularly presented with new learning opportunities. We must acquaint ourselves with methods for LC analysis to inform quality assurance projects and training. The LC literature is much more extensive in the industrial space and there is little overlap with surgical LC [5]. The Woodall et al. [5] recent critique of the use of CUSUM methods with surgical LC data suggests that industrial LC models have not been applied to full advantage in surgical applications. Ongoing work with LC analysis and integration of machine learning will be an important part of future innovation in robotic surgery.

在外科技术快速发展的时代，泌尿科医师始终处于外科创新的前沿。将新技术纳入常规临床实践需要对安全性和有效性进行分析。尽管医疗保健研究通常严重依赖随机试验，但外科研究中随机试验的局限性是众所周知的 [ 1 ]。在外科创新的早期阶段，结果受外科医生学习曲线 (LC) 的影响。关键 LC 评估提供了对早期结果的细致入微的理解，并提供了达到熟练程度所需的宝贵信息。

在本期BJUI 中，Lenfant等人。[ 2] 使用累积总和 (CUSUM) 方法描述单端口 (SP) 机器人辅助腹膜外前列腺切除术 (SP-EPP) 的 LC。在为期 19 个月的研究期间，一名外科医生 (Kaouk) 连续实施了 150 例病例。预期结果是从参考多端口 (MP) 队列中定义的。在～30个病例后达到可接受的并发症水平。基于 MP 队列的预期手术时间为 168 分钟，而观察到的平均 SP-EPP 手术时间为 190 分钟，改善发生在四个阶段，每个阶段 30-40 例。超过 120 个病例后，外科医生达到了精通水平，达到了研究定义的三重结果（手术切缘状态、术后并发症和生化复发）阈值。

可以认为学习曲线具有三个组成部分 [ 3 ]。第一个是“起点”，第二个是“斜率”，它反映了一个人学习新任务的速度。在 Lenfant等人中。[ 2] 论文中，研究外科医生 Kaouk 以 >2000 MP 机器人辅助根治性前列腺切除术 (RARP) 的基线经验开始了他的 SP 学习曲线。此外，他还接受了尸体和干实验室培训。总之，这些因素无疑会影响他的起点，以及 LC 的斜率，导致过渡到安全和熟练的 SP 手术的旅程相对较短。因此，作者明智地承认将他们的发现推广到新手外科医生的潜在局限性。LC 的第三个组成部分是“平台期”。这是测量结果的增量变化不再显着的地方。然而，这里的一个关键点是达到平台期并不一定等同于能力或安全表现 [ 4 ]。

手术 LC 的分析是复杂的。外科学习环境容易产生偏见。可以消除混杂因素（例如患者与患者的差异和手术室团队组成）中的噪音的风险调整模型将非常有用 [ 4, 5 ]。可以通过使用某种类型的回归分析来消除变异的影响，但这种做法并不常见。理想的 LC 评估将包括考虑患者和手术团队因素的多变量分析 [ 3 ]。

还需要仔细选择结果或终点，因为很少有变量会评估真正的能力 [ 4 ]。通常，建议对多个变量进行平行分析 [ 5 ]。手术时间常被用作绩效的替代指标；然而，更快的手术并不总是意味着更好的结果，并且可能反映了手术团队的熟练程度而不是个人能力。洛夫格罗夫等人。[ 6] 在他们关于 RARP 的结构化和模块化培训途径的论文中，采用了一种独特的 LC 评估方法，将技术技能作为一个独立变量进行评估。使用医疗保健故障模式效应分析 (HFMEA)，将外科手术分解为一系列关键步骤，他们能够分别分析每个子步骤的 LC。作为外科医生，我们知道对于每个手术，都会有比其他手术更具挑战性的方面。为了与此保持一致，Lovegrove等人 中每个子步骤的 LC 。[ 6] 研究各不相同。手动评估如此多离散步骤的性能非常耗时。然而，机器人手术的子步骤中产生的大量数据有助于机器学习，我们开始看到它在机器人手术教育中的应用。

外科医生会定期获得新的学习机会。我们必须熟悉 LC 分析方法，以便为质量保证项目和培训提供信息。LC 文献在工业领域要广泛得多，并且与手术 LC [ 5 ]几乎没有重叠。伍德尔等人。[ 5 ] 最近对 CUSUM 方法与手术 LC 数据使用的批评表明，工业 LC 模型尚未在手术应用中充分发挥优势。正在进行的 LC 分析和机器学习集成工作将成为机器人手术未来创新的重要组成部分。