Learning curves in Versius-assisted colorectal surgery: a risk-adjusted CUSUM analysis from a single center

Abstract

Purpose

To evaluate procedure- and operator-specific learning curves in Versius-assisted colorectal cancer surgery using risk-adjusted cumulative sum (RA-CUSUM) analysis in a real-world single-center setting.

Methods

This retrospective study included consecutive adult patients with histologically confirmed colorectal cancer who underwent elective robotic surgery using the Versius system between December 2022 and August 2025. Right hemicolectomy (RHC), sigmoid resection (SR), and anterior rectal resection (AR) were analyzed. Learning curves were assessed using RA-CUSUM analysis of skin-to-skin operative time, adjusted for patient- and procedure-specific variables. Separate models were constructed for each procedure and operator.

Results

A total of 156 procedures were included: 79 AR, 51 RHC, and 26 SR. Median operative time was 215 min (IQR 185–255). Conversion to open surgery was low (3.2%), and major postoperative complications were infrequent. RA-CUSUM analysis demonstrated distinct procedure- and operator-specific learning patterns. RHC and SR reached earlier stabilization, whereas AR required a higher case volume to achieve consistent operative time reduction, with more prolonged variability in the learning phase. Oncological quality indicators were within accepted standards.

Conclusion

Learning curves in Versius-assisted colorectal surgery are procedure- and operator-dependent. RA-CUSUM provides a structured method for monitoring trends in operative performance and may support structured training pathways and quality monitoring during implementation of robotic surgery programs.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s00384-026-05189-5.

Introduction

Robotic platforms have gained increasing adoption in colorectal surgery over the past decade. However, evidence supporting a consistently shorter learning curve compared with conventional laparoscopy remains inconclusive [1, 2]. Nevertheless, robotic systems may offer technical and ergonomic advantages, particularly in anatomically confined pelvic procedures and during intracorporeal anastomosis [3].

The Versius robotic platform (CMR Surgical, Cambridge, UK) is a modular system featuring an open surgeon console and independent bedside units equipped with wristed instruments. This design aims to facilitate integration into existing operating rooms and to improve surgeon ergonomics, potentially influencing early learning dynamics during clinical adoption [4].

In 2022, the Versius system was introduced at our institution, establishing a de novo robotic colorectal surgery program in a center with extensive experience in advanced laparoscopic procedures but no prior exposure to robotic surgery. After over 2 years of clinical practice, a structured assessment of learning curves associated with Versius implementation was warranted.

Cumulative sum (CUSUM) analysis is widely used to evaluate surgical performance and learning curves; however, conventional CUSUM does not account for patient-related risk factors influencing operative outcomes. Risk-adjusted CUSUM (RA-CUSUM) incorporates case-mix variability and has been increasingly applied in minimally invasive colorectal surgery. While learning-curve analyses for the da Vinci platform are well documented, data regarding the Versius system—particularly those using risk-adjusted methodology—remain limited [5–7].

The aim of this study was to assess procedure- and operator-specific learning curves in Versius-assisted colorectal cancer surgery using RA-CUSUM analysis in a real-world single-center setting.

Materials and methods

Study design and patient selection

This retrospective, single-center study included consecutive adult patients with histologically confirmed colorectal cancer who underwent elective robotic surgery using the Versius system between December 2022 and August 2025. Robotic-assisted resection was offered whenever the platform was available and predefined inclusion criteria were met.

Exclusion criteria comprised benign disease, emergency procedures, incomplete medical records, lack of informed consent, left hemicolectomies, abdominoperineal resections, Hartmann’s procedures, multivisceral resections, and right hemicolectomies with extracorporeal anastomosis. These procedures were excluded due to limited case numbers and increased procedural heterogeneity, which would preclude reliable procedure-specific learning curve assessment. Multivisceral resections were defined as colorectal resections combined with planned en bloc removal of at least one additional organ.

According to national regulations, formal ethical approval was waived for retrospective analysis of anonymized data. All patients provided written informed consent prior to surgery, including consent for the use of anonymized clinical data for research purposes. The study was conducted in accordance with the Declaration of Helsinki.

All procedures were performed by two colorectal surgeons with extensive prior experience in laparoscopic colorectal surgery but without previous robotic experience. Before implementation of the robotic program, both surgeons had more than 10 years of experience in advanced laparoscopic colorectal surgery and each independently performed approximately 60 laparoscopic colorectal resections annually, including rectal cancer procedures. Both surgeons completed mandatory manufacturer-provided training, and initial procedures were proctored by an experienced Versius surgeon. The study was reported in accordance with the STROBE guidelines.

Surgical procedures

The analyzed procedures included right hemicolectomy (RHC), sigmoid resection (SR), and anterior rectal resection (AR), with RHC limited to intracorporeal anastomosis. In anterior rectal resections, total mesorectal excision (TME) was routinely performed for mid- and low-rectal tumors, while partial mesorectal excision was performed for upper rectal tumors according to oncological principles. Mobilization of the splenic flexure was performed selectively when required to achieve a tension-free anastomosis. For right hemicolectomy, intracorporeal isoperistaltic side-to-side stapled ileocolic anastomosis was routinely performed, with closure of the enterotomy using a continuous suture. Mesocolic dissection and vascular ligation were performed according to standard oncological principles and tumor location. Specimen extraction was standardized using a suprapubic Pfannenstiel incision whenever feasible. All operations followed standardized institutional protocols, and perioperative management adhered to Enhanced Recovery After Surgery (ERAS) principles. Cases with missing covariate data required for the regression models were excluded from the respective RA-CUSUM analyses using a complete-case approach. Procedures were ordered chronologically for each operator and procedure type.

Risk-Adjusted CUSUM Methodology

Learning curves were evaluated using RA-CUSUM analysis. Operative time (skin-to-skin time) was chosen as the primary endpoint, given the low incidence of intraoperative blood loss and major postoperative complications, precluding reliable risk-adjusted modeling.

Although console time may more directly reflect surgeon–robot interaction, skin-to-skin operative time was selected because it represents the overall efficiency of robotic procedure delivery in routine clinical practice, including docking, bedside assistance, intraoperative coordination, and workflow integration. In the setting of a newly implemented robotic program, these team-related factors constitute an integral component of the learning process. Furthermore, console time was inconsistently recorded during the earliest implementation phase, limiting its suitability for standardized longitudinal analysis.

Risk adjustment was performed using multivariable linear regression models incorporating age, sex, body mass index (BMI), and American Society of Anesthesiologists (ASA) physical status. For rectal cancer cases, preoperative radiotherapy or chemoradiotherapy and tumor height were additionally included. Separate models were constructed for each procedure type.

Separate multivariable linear regression models were developed for right hemicolectomy, sigmoid resection, and anterior rectal resection. Model construction was limited to variables considered clinically relevant determinants of operative complexity and duration. Complete-case analysis was applied, and only cases with available covariate data were included in the respective RA-CUSUM models.

The analyzed sample sizes for model construction were 51 cases for right hemicolectomy, 26 for sigmoid resection, and 79 for anterior rectal resection. Predicted operative times derived from the regression models were subsequently used to calculate cumulative observed-minus-expected deviations for each sequential case.

Inflection points and phases of curve stabilization were identified by visual assessment of sustained directional changes in the RA-CUSUM slope rather than by formal change-point statistical modeling. Accordingly, these transition points should be interpreted as descriptive indicators of operative performance trends rather than definitive thresholds of proficiency.

Statistical analysis

Continuous variables are presented as mean ± standard deviation or median with interquartile range, as appropriate. Categorical variables are reported as counts and percentages.

Group comparisons were performed using analysis of variance or the Kruskal–Wallis test for continuous variables and the chi-square or Fisher’s exact test for categorical variables. Trends across experience tertiles were assessed using the Jonckheere–Terpstra test for continuous variables and the Cochran–Armitage test for categorical outcomes.

Multivariable linear regression models were constructed for each procedure type to evaluate factors associated with operative time and learning progression. Model performance was assessed using the coefficient of determination (R²) and residual diagnostics. No transformation of variables was required, as assumptions of linearity, homoscedasticity, and normality of residuals were satisfied. A two-sided p-value < 0.05 was considered statistically significant, with Holm–Bonferroni adjustment applied for multiple comparisons. Statistical analyses were performed using R statistical software (version 4.5.2).

Results

A total of 156 robotic colorectal procedures were included in the analysis, comprising 79 anterior rectal resections (AR), 51 right hemicolectomies (RHC), and 26 sigmoid resections (SR). In the RHC subgroup, all procedures were performed with intracorporeal anastomosis. Baseline demographic and clinical characteristics are summarized in Table 1.

Table 1.

Baseline characteristics of the analyzed cohort

Variable	Value
Number of patients, n	156
Sex:
Male	80 (51.3%)
Female	76 (48.7%)
Age (years):
Median (IQR)	71 (64–77)
BMI (kg/m2):
Median (IQR)	27.8 (25.2–30.2)
ASA classification:
ASA I, n (%)	2 (1.3%)
ASA II, n (%)	59 (37.8%)
ASA III, n (%)	84 (53.8%)
ASA IV, n (%)	9 (5.8%)
Procedure type:
Anterior resection AR, n (%)	79 (50.6%)
Right hemicolectomy RHC, n (%)	51 (32.7%)
Sigmoid resection SR, n (%)	26 (16.7%)

Values are presented as median (IQR) or number (%), unless otherwise indicated

The cohort was well balanced with respect to sex (51.3% male), with a median age of 71 years (IQR 64–77) and a median BMI of 27.8 kg/m2 (IQR 25.2–30.2). Most patients were classified as ASA III (53.8%).

Perioperative outcomes are presented in Table 2.

Table 2.

Intraoperative and perioperative outcomes across all procedures

Variable	Median	IQR (Q1–Q3)	n (%)
Operative time, skin-to-skin (min)	215	185–255
Console time (min)	137	108–170
Docking time (min)	12	10–14
Intraoperative blood loss (ml)	20	0–100
Length of hospital stay (days)	6	5–6
Conversion (yes)			5 (3.2%)
Intraoperative complications (yes)			1 (0.6%)
Reoperation (yes)			11 (7.1%)
30-day readmission (yes)			4 (2.6%)

Median skin-to-skin operative time was 215 min (IQR 185–255), with a median console time of 137 min (IQR 108–170) and a median docking time of 12 min (IQR 10–14). Conversion to open surgery was infrequent (3.2%), and intraoperative complications were rare (0.6%). Most patients experienced no postoperative complications (Clavien–Dindo grade 0: 82.7%), while severe complications (Clavien–Dindo ≥ III) occurred in a small proportion of cases (Table 3). Complication grade was available for all patients. RA-CUSUM analysis of operative time demonstrated distinct procedure- and operator-specific learning patterns.

Table 3.

Postoperative complications according to Clavien-Dindo classification

Clavien–Dindo grade	n	%
0	129	82.7
1	6	3.8
2	9	5.8
3a	0	0.0
3b	10	6.4
4a	1	0.6
4b	0	0.0
5	1	0.6

Short-term oncological quality indicators were within accepted standards (Supplementary Table 1).

Procedure-specific perioperative outcomes are summarized in Supplementary Table 2. Anterior rectal resection was associated with the longest operative time and the highest rates of reoperation and anastomotic leakage, reflecting the greater technical complexity of rectal cancer surgery.

Right hemicolectomy

For Operator 1, the RA-CUSUM curve showed a typical learning pattern, with a maximum deviation point (learning peak) around the 18th case and subsequent stabilization of operative performance after approximately 25 cases (Fig. 1).

Fig. 1.

RA-CUSUM curve for operative time in right hemicolectomy (Operator 1)

The curve demonstrates an initial learning phase with a peak at approximately the 18th case, followed by a downward trend indicating progressive improvement in operative efficiency. Subsequent stabilization of the curve suggests reduced variability and increasing procedural consistency. Interpretation of inflection points should be considered exploratory and dependent on the analyzed case volume.

For Operator 2 (16 cases), an early breakpoint was observed; however, the learning curve exhibited marked oscillations, followed by a more consistent downward trend later in the series (Fig. 2). Given the limited case number, these findings should be interpreted cautiously.

Fig. 2.

RA-CUSUM curve for operative time in right hemicolectomy (Operator 2)

An early breakpoint is observed; however, the curve exhibits marked oscillations reflecting variability in operative performance. A more consistent downward trend is seen in later cases, suggesting gradual improvement. Due to the limited number of procedures, the learning pattern should be interpreted with caution.

Anterior rectal resection

RA-CUSUM models incorporated rectal cancer–specific variables.

For Operator 1 (50 cases), a prolonged learning phase was observed, with a maximum deviation point (learning peak) around the 29th case followed by a progressive downward trend in subsequent cases (Fig. 3).

Fig. 3.

RA-CUSUM curve for operative time in anterior rectal resection (Operator 1)

A prolonged learning phase is observed, with a peak around the 29th case followed by gradual reduction in variability in later cases. The extended variability reflects the technical complexity of rectal surgery.

For Operator 2 (25 cases), the peak of the learning phase occurred around the 19th case, followed by gradual improvement without a clearly defined plateau, suggesting progressive consolidation rather than a clearly defined stabilization threshold (Fig. 4).

Fig. 4.

RA-CUSUM curve for operative time in anterior rectal resection (Operator 2)

The curve shows a peak around the 19th case, followed by gradual improvement without a clearly defined plateau. This pattern suggests progressive consolidation of skills rather than abrupt stabilization of operative performance.

Sigmoid resection

For Operator 1 (22 cases), the RA-CUSUM curve demonstrated a short learning phase, with a maximum deviation point (learning peak) around the 3rd case and subsequent stabilization of operative performance after approximately 16 cases (Fig. 5).

Fig. 5.

RA-CUSUM curve for operative time in sigmoid resection (Operator 1)

A short learning phase is demonstrated, with an early peak around the 3rd case and subsequent decline. Stabilization of the curve after approximately 16 cases indicates rapid transition toward a more stable operative efficiency pattern compared with more complex procedures.

Only four sigmoid resections performed by Operator 2 were available for analysis, precluding reliable interpretation of learning patterns in this subgroup.

Overall, RA-CUSUM analysis revealed substantial inter-operator and inter-procedural variability. While right hemicolectomy and sigmoid resection were associated with earlier stabilization of operative performance, anterior rectal resection required a higher case volume to achieve consistent efficiency.

Discussion

In this study, we evaluated learning curves using a risk-adjusted approach in Versius-assisted colorectal surgery based on more than two years of routine clinical practice in a single center. By applying RA-CUSUM methodology and incorporating patient-, disease-, and procedure-specific covariates, the analysis was designed to isolate learning-related changes in operative performance from variability attributable to case mix, which represents a relevant methodological advance over conventional learning-curve analyses.

The present results demonstrate pronounced procedure- and operator-specific learning patterns. Right hemicolectomy and sigmoid resection were associated with relatively rapid stabilization of operative performance, whereas anterior rectal resection required a substantially higher case volume before consistent reductions in operative time were observed. These differences are consistent with the inherent technical complexity of pelvic dissection and reconstruction in rectal cancer surgery and underscore the importance of procedure-specific evaluation rather than extrapolation from simpler colorectal procedures.

A major strength of this analysis lies in the inclusion of rectal cancer–specific variables, such as tumor height and exposure to preoperative radiotherapy or chemoradiotherapy, within the RA-CUSUM models. Both factors are well recognized determinants of operative difficulty and duration. Their inclusion proved essential for accurate interpretation of learning curves in anterior rectal resection, as omission of these variables would likely have resulted in confounded learning patterns. These findings support the concept that learning-curve assessment in rectal surgery requires extended, procedure-tailored risk adjustment beyond generic patient characteristics.

Inter-operator variability was observed across all analyzed procedures, despite comparable backgrounds in advanced laparoscopic colorectal surgery. This observation suggests that prior laparoscopic expertise alone does not fully predict the pace or pattern of adaptation to robotic surgery. Although both surgeons shared similar laparoscopic colorectal experience, they differed by approximately 10 years in age and overall surgical seniority. These individual characteristics, while not formally assessed in this study, may have influenced individual learning trajectories. Age-related variation in cognitive flexibility, visuospatial coordination, and adaptability to new technology has been shown to affect the pace of robotic skill acquisition [8]. Future studies incorporating structured psychometric or simulator-based assessments could help quantify the role of age and general surgical maturity in shaping robotic learning curves. Consequently, individualized performance monitoring may be more informative than uniform case-number thresholds alone when guiding individualized training pathways during the early adoption phase of a robotic platform.

The present study also highlights an inherent limitation of CUSUM-based methodologies, namely the requirement for sufficient procedural volume to generate interpretable learning curves. In low-volume subgroups, such as sigmoid resection performed by Operator 2, RA-CUSUM patterns were highly sensitive to individual cases and should therefore be interpreted as exploratory. Accordingly, learning-curve conclusions should be restricted to operator–procedure combinations with adequate case numbers, while smaller series are more appropriately reported descriptively.

Interpretation of RA-CUSUM curves should also acknowledge intrinsic methodological limitations of cumulative summation analyses. Previous methodological studies have demonstrated that apparent peaks, inflection points, and stabilization phases may be influenced by total sample size, case ordering, and analytical assumptions rather than representing universally reproducible thresholds of proficiency [12, 13]. Consequently, RA-CUSUM findings should primarily be interpreted as descriptive representations of operative performance trends rather than absolute definitions of competency.

When placed in the context of existing literature on robotic colorectal surgery using the da Vinci platform, the learning-curve thresholds observed in this study appear largely comparable [5, 9, 10]. Previous reports have shown that operative efficiency is typically achieved after approximately 20–40 cases, depending on procedure complexity and surgeon experience [5, 7, 11]. The learning peaks and stabilization points identified in the present RA-CUSUM analyses for the Versius system fall within these ranges, suggesting that adoption of this modular robotic platform is not associated with a prolonged or atypical learning process compared with established robotic systems.

From a clinical perspective, these findings support the feasibility of implementing the Versius robotic system in colorectal surgery within a structured training and proctoring framework. Moreover, RA-CUSUM analysis may serve as a practical tool for continuous quality monitoring, identification of procedure-specific efficiency thresholds, and data-driven decision-making during robotic program development. Operative time reflects technical efficiency rather than oncological quality and should not be interpreted as a surrogate marker for long-term outcomes.

Although formal phase-based comparison of oncological outcomes was beyond the scope of this analysis, no evident deterioration in short-term oncological quality indicators was observed during earlier implementation phases.

Several limitations should be acknowledged. The retrospective, single-center design limits generalizability, and long-term oncologic outcomes were not assessed. While operative time was selected as the primary endpoint due to its objectivity and availability, it represents a surrogate of technical efficiency rather than surgical quality or oncological adequacy. Although short-term oncological quality indicators—including lymph node yield and R0 resection rate—were within accepted benchmarks, integration of additional performance metrics (e.g., intraoperative technical errors or long-term oncological outcomes) would strengthen the comprehensiveness of learning assessments. Learning trajectories may also have been influenced by the surgeons’ substantial experience in laparoscopic colorectal surgery, potentially shortening observed learning phases. Additionally unmeasured variables—such as psychomotor aptitude, cognitive processing, or prior exposure to simulation—were not assessed and might influence differences in inter-operator learning curves. Future research incorporating validated technical skill assessments and qualitative data (e.g., semi-structured interviews or self-assessment tools) could offer deeper insights into individual learning behaviors and their modulating factors.

Nevertheless, the use of a risk-adjusted analytical framework strengthens the validity of the findings by accounting for clinically relevant case-mix factors.

Importantly, the present findings should not be interpreted as direct measures of surgical quality or oncological proficiency. Because RA-CUSUM modeling was based on operative time, the identified learning patterns primarily reflect trends in operative efficiency and workflow adaptation rather than comprehensive technical mastery.

Conclusion

Versius-assisted colorectal surgery is characterized by distinct, procedure- and operator-dependent learning curves when evaluated using RA-CUSUM analysis. While right hemicolectomy and sigmoid resection demonstrate earlier stabilization of operative performance, anterior rectal resection requires a substantially higher case volume to achieve consistent efficiency. Risk-adjusted CUSUM provides a transparent and clinically meaningful framework for learning-curve assessment and may play a relevant role in guiding structured training pathways and monitored implementation of emerging robotic platforms in colorectal surgery.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

T.S. wrote the main manuscript. M.M prepared tables and figures. J.K. prepared statistical analysis.

Funding

This research received no external funding. The study was conducted as part of routine clinical practice and institutional quality assessment.

Data availability

All data supporting the findings of this study are available within the paper.

Declarations

Competing interests

Tomasz Sachańbiński and Maciej Miodoński are certified proctors for the Versius robotic system (CMR Surgical). Both authors have previously delivered sponsored educational lectures related to the early clinical implementation of the Versius platform. These activities were unrelated to the present study and did not influence its design, data collection, analysis, or interpretation. Jakub Kropieniewicz declares no conflicts of interest and has no affiliation with CMR Surgical.