Assessment of the learning curve of robotic liver resection using complexity-adjusted CUSUM analysis Authors: Jiwon Han, Hyokee Kim, Wan Joon Kim Published: 2026 ABSTRACT Robotic liver resection (RLR) has progressively been integrated into minimally invasive hepatobiliary surgery. However, comparative evidence evaluating the learning curves of RLR and laparoscopic liver resection (LLR) while accounting for surgical complexity remains limited. This study analyzed the learning curve of RLR through complexity-adjusted cumulative sum (CA-CUSUM) analysis based on operative time and the IWATE difficulty score. This retrospective study was conducted using data from minimally invasive liver surgeries performed by a surgeon at the single institution between 2017 and 2025. Donor hepatectomies and combined organ resections were excluded. A CA-CUSUM analysis adjusted for the IWATE score was utilized to evaluate the learning curves. A total of 55 RLR and 149 LLR cases were analyzed. The CA-CUSUM analysis revealed a breakpoint at case 11 for RLR and case 50 for LLR ( p<0.05). The initial 11 RLR cases had a median IWATE score of 9, showing that high-complexity cases were performed from the beginning of the robotic experience. Comparing the competency phase (cases 1–11) and efficiency phase (cases 12–55) of RLR, the median operative time decreased from 215 to 145 min ( p<0.05) and median estimated blood loss decreased from 1,000 to 450 ㎖ ( p<0.05). Other safety-related outcomes were comparable between the two phases, with zero conversions to open surgery. Adjusting for procedural difficulty, the complexity-adjusted analysis identified an operative-time breakpoint after 11 cases in RLR, compared to 50 cases in LLR. For a surgeon with extensive laparoscopic background, this rapid transition reflects technical adaptation and platform-specific skill transfer. These findings highlight the efficiency of transitioning to robotic liver surgery when leveraging prior minimally invasive experience.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s11701-026-03909-1. Introduction Minimally invasive liver surgery (MILS) is widely accepted as a standard approach for the treatment of both benign and malignant liver tumors [1]. Compared with traditional open hepatectomy, laparoscopic liver resection (LLR) has demonstrated substantial clinical benefits, including decreased postoperative pain, shorter hospital stays, and diminished overall complication rates compared to conventional open hepatectomy [2, 3]. However, notwithstanding its widespread adoption, LLR involves inherent technical limitations that challenge surgical precision. Even with advanced visualization, the primary constraints of laparoscopy stem from rigid, non-articulating instrumentation. Manipulating straight instruments through a fixed abdominal pivot point restricts the angles of approach, making parenchymal transection along the curved diaphragmatic surface technically demanding. Consequently, performing anatomical resections in difficult anatomical regions, such as the posterosuperior segments, remains a challenging task under conventional laparoscopy [4, 5]. To broaden the clinical application of MILS, robotic liver resection (RLR) has been increasingly adopted alongside ongoing advancements in minimally invasive techniques. The robotic platform complements traditional approaches by providing features such as a stable three-dimensional stereoscopic view, articulating instruments with seven degrees of freedom that effectively restore intuitive hand-eye coordination, and tremor filtering algorithms. These features reduce physical and cognitive demands, facilitating precise tissue dissection and intracorporeal suturing within narrow surgical fields [6, 7]. Existing meta-analyses suggest that the robotic approach is associated with a shorter learning curve compared to laparoscopy, requiring fewer cases to achieve proficiency [8, 9]. Furthermore, the enhanced dexterity of the robotic system may facilitate an earlier adoption of complex cases, allowing experienced laparoscopic surgeons to perform high-complexity liver resections without a prolonged stepwise adaptation period. Therefore, this study aimed to compare the learning curves of RLR and LLR performed by a single surgeon at a single center. By utilizing a complexity-adjusted cumulative sum (CA-CUSUM) model adjusted for the IWATE difficulty score, this study evaluates the learning curve of RLR derived from operative time as the primary outcome, while comparing phase-specific perioperative outcomes to evaluate clinical safety during technical adaptation. Methods Study design and patients Conducted as a single-center, retrospective investigation, this study analyzed cases carried out by a single surgeon from 2017 to 2025. The patients undergoing MILS, specifically comparing LLR and RLR were included. RLR was performed utilizing the da Vinci Xi surgical system (Intuitive Surgical Inc., Sunnyvale, CA, USA). Donor hepatectomies and combined organ resections were excluded from the analysis. Demographics, tumor characteristics, and perioperative parameters were comprehensively reviewed. R software (version 4.5.2; R Foundation for Statistical Computing, Vienna, Austria) was utilized to perform all statistical analyses. The annual breakdown of total liver resections performed by the operating surgeon, including open, laparoscopic, and robotic procedures, is summarized in Online Resource 1. Extended cholecystectomy was excluded because it was not considered a primary liver resection for the purpose of this institutional volume analysis. Surgical technique for robotic liver resection Under general anesthesia with endotracheal intubation, central venous pressure (CVP) was strictly maintained below 5 cmH₂O during the parenchymal transection phase to minimize bleeding from the hepatic veins. All robotic liver resections were supported by a dedicated hepatobiliary surgical team, including experienced scrub nurses and surgical assistants. The composition of the operating room staff remained largely stable, which facilitated stabilized patient positioning, efficient robotic docking, and minimized turnover-related procedural delays. To establish a pneumoperitoneum of 12 mmHg, a 12-mm camera port was inserted via an open technique at the umbilicus. Three additional robotic ports (8-mm or 12-mm) were strategically positioned to ensure adequate triangulation and avoid instrument collision. Additionally, one assistant port (5-mm or 12-mm) was placed on the patient’s left side for suction and retraction. The Pringle maneuver was prepared prophylactically using an extracorporeal method, to be tightened in an intermittent fashion (15 min of ischemia followed by 5 min of reperfusion) in the event of significant bleeding during parenchymal transection. Depending on the resection type, inflow control was achieved via either the Glissonian approach or individual hilar dissection. Indocyanine green (ICG) (1.25 mg) was injected intravenously to achieve negative or positive staining under near-infrared fluorescence (Firefly imaging) to precisely delineate the segmental boundaries. Parenchymal transection was performed by employing both a Robotic Vessel Sealer and a Robotic Maryland Bipolar Forceps. Small vessels and bile ducts were coagulated or clipped, whereas larger structures or major hepatic veins were divided using robotic staplers or secured with Hem-o-lok clips. Once transection was finalized, specimen extraction was conducted using a retrieval bag through a Pfannenstiel incision or an extended umbilical incision. Hemostasis was achieved using bipolar coagulation, argon plasma coagulation, or surgical sutures. Routine placement of a closed-suction drain near the operative bed was carried out via one of the lateral port sites. Assessment of surgical complexity Stratification of the surgical complexity of each procedure was accomplished using the validated IWATE difficulty scoring system [10, 11]. This system quantitatively evaluates individual case difficulty and categorizes it into four distinct levels (low, intermediate, advanced, and expert) by assigning points based on six critical preoperative parameters: tumor size, tumor location, proximity to major vessels, the extent of hepatic resection, liver function, and background liver status [11]. Complexity-adjusted CUSUM (CA-CUSUM) modeling To control temporal variations in case difficulty, a complexity-adjusted cumulative sum (CA-CUSUM) analysis was performed applying operative time residuals. The estimation of the structural breakpoint was performed via segmented linear regression analysis using R software. First, expected operative time (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i}$$\end{document}) was predicted through a baseline linear regression model grounded in the validated IWATE difficulty score: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i}$$\end{document} = \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} + \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:\cdot\:{IWATE}_{i}$$\end{document}. The CA-CUSUM value for each chronological procedure was subsequently calculated as the cumulative sum of the differences between the observed operative time (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{O}_{i}$$\end{document}) and the predicted expected operative time (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i}$$\end{document}):\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathrm{CA-CUSUM}}_{n}=\mathop{\sum}\limits_{i=1}^{n}\left({O}_{i}-{E}_{i}\right)$$\end{document} The precise statistical breakpoint representing the transition to technical proficiency was determined using segmented linear regression analysis. A single structural change point was estimated from the linear model of the CA-CUSUM values plotted against the chronological case sequence. To validate the statistical significance of the identified breakpoint, Davies’ test was performed to evaluate the presence of non-constant regression parameters across the change point. Following breakpoint identification, cases were categorized into a competency phase and an efficiency phase. Results Annual trends and baseline patient characteristics The annual case volumes and relative proportions of laparoscopic liver resection (LLR) and robotic liver resection (RLR) from 2017 to 2025 are presented in Fig. 1. Prior to the initiation of the robotic program (2017–2020), a baseline of 35 LLR cases was established. Following the introduction of RLR in 2021, the annual robotic volume expanded from 1 case (4.5% of the total minimally invasive volume) to 33 cases (63.5%) in 2025, systematically surpassing the annual LLR volume (19 cases, 36.5%) (Fig. 1). A total of 204 minimally invasive liver resections were analyzed, comprising 149 LLRs and 55 RLRs. Baseline patient and perioperative characteristics were highly comparable between the two cohorts (Table 1), demonstrating no statistically significant differences in terms of sex, age, Body Mass Index (BMI), Child–Pugh score, or tumor characteristics, including size and number. Major liver resection, defined as the resection of three or more segments, was also similarly distributed between the groups (23% in LLR vs. 10% in RLR, p=0.454). However, the proportion of malignant cases was significantly higher in the LLR group than in the RLR group (99% vs. 89%, p<0.05). The overall distribution of IWATE difficulty classifications (low, intermediate, advanced, and expert) was similar between both groups ( p=0.497). Fig. 1Annual trends of minimally invasive liver resection from 2017 to 2025, (a) represents annual case volume of laparoscopic liver resection (LLR, gray bars) and robotic liver resection (RLR, red bars), and (b) represents the relative proportion of each surgical approach per year Annual trends of minimally invasive liver resection from 2017 to 2025, (a) represents annual case volume of laparoscopic liver resection (LLR, gray bars) and robotic liver resection (RLR, red bars), and (b) represents the relative proportion of each surgical approach per year Table 1Baseline and perioperative characteristics of patientsCharacteristicLLR(N = 149)RLR(N = 55)p-value*Sex, n (%)0.586F35 (23%)15 (27%)M114 (77%)40 (73%)Age (year), median [IQR]64 [57, 70]63 [55, 69]0.386BMI (kg/m²), median [IQR]24.1 [22.5, 26.9]24.2 [22.2, 26.1]0.533Malignancy, n (%) < 0.05 Benign1 (0.7%)6 (11%)Malignant148 (99%)49 (89%)Child-Pugh, n (%)1.000A148 (99%)55 (100%)B1 (0.7%)0 (0%)IWATE score, median [IQR]7.00 [5.00, 9.00]8.00 [5.00, 10.00]0.223IWATE difficulty, n (%)0.497Low15 (10%)6 (11%)Intermediate53 (36%)14 (25%)Advanced49 (33%)19 (35%)Expert32 (21%)16 (29%)Tumor number, n (%)0.5551144 (97%)52 (95%)24 (2.7%)3 (5.5%)31 (0.7%)0 (0%)Tumor size (cm), median [IQR]2.50 [1.60, 3.50]2.60 [1.80, 4.30]0.315Tumor location, n (%)0.272Anterior-Inferior73 (49%)22 (40%)Posterior-Superior76 (51%)33 (60%)Resection type, n (%)0.454Minor114 (77%)45 (82%)Major35 (23%)10 (18%)Proximity to major vessel, n (%)41 (28%)20 (36%)0.232Operative time (min), median [IQR]156 [105, 219]149 [115, 212]0.964*p-values were calculated using the Wilcoxon rank-sum test for continuous variables and Fisher’s exact test for categorical variables Table 1Baseline and perioperative characteristics of patientsCharacteristicLLR(N = 149)RLR(N = 55)p-value*Sex, n (%)0.586F35 (23%)15 (27%)M114 (77%)40 (73%)Age (year), median [IQR]64 [57, 70]63 [55, 69]0.386BMI (kg/m²), median [IQR]24.1 [22.5, 26.9]24.2 [22.2, 26.1]0.533Malignancy, n (%) < 0.05 Benign1 (0.7%)6 (11%)Malignant148 (99%)49 (89%)Child-Pugh, n (%)1.000A148 (99%)55 (100%)B1 (0.7%)0 (0%)IWATE score, median [IQR]7.00 [5.00, 9.00]8.00 [5.00, 10.00]0.223IWATE difficulty, n (%)0.497Low15 (10%)6 (11%)Intermediate53 (36%)14 (25%)Advanced49 (33%)19 (35%)Expert32 (21%)16 (29%)Tumor number, n (%)0.5551144 (97%)52 (95%)24 (2.7%)3 (5.5%)31 (0.7%)0 (0%)Tumor size (cm), median [IQR]2.50 [1.60, 3.50]2.60 [1.80, 4.30]0.315Tumor location, n (%)0.272Anterior-Inferior73 (49%)22 (40%)Posterior-Superior76 (51%)33 (60%)Resection type, n (%)0.454Minor114 (77%)45 (82%)Major35 (23%)10 (18%)Proximity to major vessel, n (%)41 (28%)20 (36%)0.232Operative time (min), median [IQR]156 [105, 219]149 [115, 212]0.964*p-values were calculated using the Wilcoxon rank-sum test for continuous variables and Fisher’s exact test for categorical variables Baseline and perioperative characteristics of patients *p-values were calculated using the Wilcoxon rank-sum test for continuous variables and Fisher’s exact test for categorical variables Complexity-adjusted learning curve analysis To validate the clinical applicability of the difficulty scoring system, a linear regression analysis was performed within the robotic surgery cohort. For RLR cases, the IWATE difficulty score positively correlated with total operative time, yielding an estimated slope of 17.1 min (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{R}^{2}=0.29$$\end{document}, p<0.05). This confirms that the scoring system accurately reflects technical complexity and provides meaningful predictability for surgical duration in robotic liver resection (Fig. 2). The baseline expected operative-time model was \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i}=54.72\:+16.6\times\:{IWATE}_{i}\:$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left({R}^{2}=0.282,\:p<0.001\right)$$\end{document} for LLR and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}_{i}=43.09\:$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$+17.10\times\:{IWATE}_{i}\:$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\left({R}^{2}=0.291,\:p<0.001\right)$$\end{document} for RLR. Continuous segmented regression on CA-CUSUM identified significant breakpoints at case 50 for LLR (95% CI; 48.70-50.74, Davies’ test p<0.001) (Fig. 3a), and case 11 for RLR (95% CI; 9.83–13.12, Davies’ test p<0.001) (Fig. 3b). Residual diagnostics and sensitivity analyses confirmed model stability without influential outliers altering the primary transition breakpoints (Online Resource 2). For the initial 11 cases, the median IWATE score in the robotic group was higher than in the laparoscopic group (9 vs. 5), though the difference lacked statistical significance ( p=0.141; Fig. 4). Fig. 2Linear regression model illustrating the correlation between the IWATE difficulty score and total operative time with the red line indicating the trend Linear regression model illustrating the correlation between the IWATE difficulty score and total operative time with the red line indicating the trend Fig. 3Complexity-adjusted cumulative sum (CA-CUSUM) analysis for learning curve evaluation, (a) the learning curve of laparoscopic liver resection (LLR) and (b) the learning curve of robotic liver resection (RLR) with the gray shaded areas defining the competency phase (thin gray regression lines) and the unshaded areas defining the efficiency phase (thick red regression lines) Complexity-adjusted cumulative sum (CA-CUSUM) analysis for learning curve evaluation, (a) the learning curve of laparoscopic liver resection (LLR) and (b) the learning curve of robotic liver resection (RLR) with the gray shaded areas defining the competency phase (thin gray regression lines) and the unshaded areas defining the efficiency phase (thick red regression lines) Fig. 4Comparison of the IWATE difficulty score within the initial 11 operations between the laparoscopic liver resection (LLR, gray box) and robotic liver resection (RLR, red box) with the individual data points presented as black dots Comparison of the IWATE difficulty score within the initial 11 operations between the laparoscopic liver resection (LLR, gray box) and robotic liver resection (RLR, red box) with the individual data points presented as black dots Competency vs. Efficiency in robotic liver surgery Determined by the CA-CUSUM breakpoint, the RLR cases were split into two defined clinical phases, consisting of a competency phase (cases 1–11) and an efficiency phase (cases 12–55). Between these phases, the median operative time significantly decreased from 215 min to 145 min ( p<0.05), indicating improved operative efficiency, and median estimated blood loss significantly decreased from 1,000 to 450 mL ( p<0.05), indicating improved technical efficiency and procedural control (Table 2). Intraoperative blood transfusion rates (36% vs. 11%, p=0.067) showed a downward trend without statistical significance, and no conversions to open surgery occurred across either phase (Table 2). All resections achieved macroscopic clearance (R2, 0%). Microscopic negative margins (R0) was achieved in 73% (8/11) during the competency phase and 75% (33/44) during the efficiency phase (Table 2). Microscopic margin involvement (R1) occurred in 2 cases (18%) and 6 cases (14%), respectively, primarily representing parenchymal-sparing detachment along major vascular structures (e.g., middle or right hepatic veins). In the robotic cohort, minor complications (Clavien–Dindo Grades I–II) comprised transient paralytic ileus ( n=1), Clostridioides difficile-associated diarrhea ( n=1), acute prostatitis ( n=1), and isolated postoperative fever requiring antibiotic escalation ( n=3). Grade III complications were predominantly intra-abdominal fluid collections or hematomas; six cases required ultrasound-guided percutaneous catheter drainage, and one bile leak was managed via endoscopic retrograde cholangiopancreatography (ERCP) with concurrent ICU admission for pneumonia. Grade IV morbidity involved one case of contrast-induced anaphylactic shock managed with immediate pharmacological stabilization. One Grade V mortality occurred due to progressive postoperative liver failure after a major hepatectomy. Postoperative morbidity rates were comparable between the competency and efficiency phases (30% vs. 24%, p=0.703). One 90-day mortality occurred during the early RLR experience (the competency phase). A 58-year-old male with hepatitis C cirrhosis (Child-Pugh class A) and a 3-cm segment 7 HCC underwent a robotic right hemihepatectomy (IWATE score: 8). Operative time was 299 min with an estimated blood loss of 1,000 mL. Hemoglobin levels were stable without perioperative red blood cell transfusions (14.3 g/dL preoperatively to 12.5 g/dL postoperatively). Postoperatively, the patient developed portal vein thrombosis complicated by progressive ISGLS(International Study Group of Liver Surgery) Grade C liver failure and hepatorenal syndrome. Despite listing for salvage liver transplantation with a MELD score of 40, the patient died of liver failure on postoperative day 25. Table 2Comparison between competency and efficiency phases in robotic liver surgeryCharacteristicCompetency(N = 11)Efficiency(N = 44)p-value*Sex, n (%)1.000F3 (27%)12 (27%)M8 (73%)32 (73%)Age (years), median [IQR]60 [56, 63]63 [55, 71]0.417BMI (kg/m²), median [IQR]24.42 [22.04, 26.12]24.06 [22.23, 26.03]0.599Malignancy, n (%)1.000Benign1 (9.1%)5 (11%)Malignant10 (91%)39 (89%)Child-Pugh, n (%)A11 (100%)44 (100%)IWATE score, median [IQR]9.00 [4.00, 10.00]8.00 [5.00, 10.00]0.581IWATE difficulty, n (%)0.389Low2 (18%)4 (9.1%)Intermediate1 (9.1%)13 (30%)Advanced5 (45%)14 (32%)Expert3 (27%)13 (30%)Tumor number, n (%)0.495110 (91%)42 (95%)21 (9.1%)2 (4.5%)Tumor size (cm), median [IQR]2.80 [1.90, 3.60]2.35 [1.80, 4.40]0.681Tumor location0.495Anterior-Inferior3 (27%)19 (43%)Posterior-Superior8 (73%)25 (57%)Resection type, n (%)1.000Minor9 (82%)36 (82%)Major2 (18%)8 (18%)Proximity to major vessels, n (%)3 (27%)17 (39%)0.728Operative time (min), median [IQR]215 [145, 280]145 [109, 191] < 0.05 Estimated blood loss (mL), median [IQR]1,000 [650, 2,150]450 [200, 1,000] < 0.05 Intraoperative blood transfusion, n (%)4 (36%)5 (11%)0.067Conversion, n (%)0 (0%)0 (0%)Margin status, n (%)0.859R08 (73%)33 (75%)R12 (18%)6 (14%)R20 (0%)0 (0%)Not applicablea1 (9.1%)5 (11%)Length of stay (days), median [IQR]10.0 [8.0, 19.0]9.0 [8.0, 13.0]0.468Postoperative morbidity, n (%)4 (36%)10 (23%)0.443Clavien-Dindo ≥ III, n (%)3 (27%)6 (14%)0.36230-day mortality, n (%)1 (9.1%)0 (0%)0.200*p-values were calculated using the Wilcoxon rank-sum test for continuous variables and Fisher’s exact test for categorical variables. Analyses were performed using available cases, and missing values were not imputed. Variables with fewer than two levels were not testeda Margin status evaluated as ‘Not applicable’ (n = 1, 9.1% in the competency phase and n = 5, 11% in the efficiency phase) represents benign lesions where formal pathological margin evaluation was not clinically indicated Table 2Comparison between competency and efficiency phases in robotic liver surgeryCharacteristicCompetency(N = 11)Efficiency(N = 44)p-value*Sex, n (%)1.000F3 (27%)12 (27%)M8 (73%)32 (73%)Age (years), median [IQR]60 [56, 63]63 [55, 71]0.417BMI (kg/m²), median [IQR]24.42 [22.04, 26.12]24.06 [22.23, 26.03]0.599Malignancy, n (%)1.000Benign1 (9.1%)5 (11%)Malignant10 (91%)39 (89%)Child-Pugh, n (%)A11 (100%)44 (100%)IWATE score, median [IQR]9.00 [4.00, 10.00]8.00 [5.00, 10.00]0.581IWATE difficulty, n (%)0.389Low2 (18%)4 (9.1%)Intermediate1 (9.1%)13 (30%)Advanced5 (45%)14 (32%)Expert3 (27%)13 (30%)Tumor number, n (%)0.495110 (91%)42 (95%)21 (9.1%)2 (4.5%)Tumor size (cm), median [IQR]2.80 [1.90, 3.60]2.35 [1.80, 4.40]0.681Tumor location0.495Anterior-Inferior3 (27%)19 (43%)Posterior-Superior8 (73%)25 (57%)Resection type, n (%)1.000Minor9 (82%)36 (82%)Major2 (18%)8 (18%)Proximity to major vessels, n (%)3 (27%)17 (39%)0.728Operative time (min), median [IQR]215 [145, 280]145 [109, 191] < 0.05 Estimated blood loss (mL), median [IQR]1,000 [650, 2,150]450 [200, 1,000] < 0.05 Intraoperative blood transfusion, n (%)4 (36%)5 (11%)0.067Conversion, n (%)0 (0%)0 (0%)Margin status, n (%)0.859R08 (73%)33 (75%)R12 (18%)6 (14%)R20 (0%)0 (0%)Not applicablea1 (9.1%)5 (11%)Length of stay (days), median [IQR]10.0 [8.0, 19.0]9.0 [8.0, 13.0]0.468Postoperative morbidity, n (%)4 (36%)10 (23%)0.443Clavien-Dindo ≥ III, n (%)3 (27%)6 (14%)0.36230-day mortality, n (%)1 (9.1%)0 (0%)0.200*p-values were calculated using the Wilcoxon rank-sum test for continuous variables and Fisher’s exact test for categorical variables. Analyses were performed using available cases, and missing values were not imputed. Variables with fewer than two levels were not testeda Margin status evaluated as ‘Not applicable’ (n = 1, 9.1% in the competency phase and n = 5, 11% in the efficiency phase) represents benign lesions where formal pathological margin evaluation was not clinically indicated Comparison between competency and efficiency phases in robotic liver surgery *p-values were calculated using the Wilcoxon rank-sum test for continuous variables and Fisher’s exact test for categorical variables. Analyses were performed using available cases, and missing values were not imputed. Variables with fewer than two levels were not tested a Margin status evaluated as ‘Not applicable’ (n = 1, 9.1% in the competency phase and n = 5, 11% in the efficiency phase) represents benign lesions where formal pathological margin evaluation was not clinically indicated Discussion This study demonstrated a compressed, complexity-adjusted learning curve for RLR compared to LLR (11 vs. 50 cases). Unadjusted learning curve models are inherently prone to selection bias because case difficulty typically increases as a surgeon’s experience accumulates [12]. By adjusting for case complexity via the IWATE score, this CA-CUSUM analysis confirms that the accelerated achievement of RLR efficiency was independent of early case selection. To contextualize these findings within broader surgical research, technical proficiency is traditionally assessed using chronological case volume or raw operative times [12, 13]. However, these methods inherently assume uniform case complexity over time [13]. In minimally invasive hepatobiliary surgery, this assumption is often invalid because surgeons dynamically select more complex cases as they gain experience [12, 14]. Evaluating learning curves solely based on unadjusted metrics can therefore introduce a significant confounding bias [12]. To address this, Cumulative Sum (CUSUM) analysis has been utilized to pinpoint where operative parameters stabilize [15]. While standard CUSUM plots cumulative variance against the cohort mean, it does not account for case-by-case difficulty, potentially misinterpreting longer operative times in complex cases as a lack of technical progress [12, 15]. Complexity-adjusted CUSUM (CA-CUSUM) models resolve this limitation by incorporating validated scoring systems, such as the IWATE criteria, into a regression analysis [12, 16]. The IWATE system effectively stratifies procedural difficulty (low, intermediate, advanced, and expert) based on tumor characteristics and the extent of resection, a utility confirmed in a recent systematic review by Linn et al. [17]. By utilizing the residuals between actual and expected operative times, CA-CUSUM successfully isolates true technical efficiency from the confounding effects of case selection [12, 15]. Despite these methodological advantages, comparative data utilizing CA-CUSUM to evaluate the learning curves of RLR and LLR remain limited [12]. Our complexity-adjusted approach addresses this gap, allowing for a more rigorous interpretation of the specific factors driving our compressed 11 case learning curve. Although there was a baseline imbalance in malignancy proportions (99% malignant in LLR vs. 89% in RLR, p < 0.05), this difference did not influence intraoperative decision-making or parenchymal transection techniques. The benign tumors in the RLR group, consisting of hepatic hemangioma, multilocular simple cyst, mucinous cystic neoplasm, hepatocellular adenoma, bile duct adenoma, and regenerating nodule combined with bile duct hamartoma and hemangioma required formal parenchymal resections. Furthermore, because the IWATE difficulty score stratifies surgical complexity based on anatomical and functional metrics, procedural difficulty was effectively adjusted within the CA-CUSUM model regardless of underlying pathology. Anatomical location dictates the technical difficulty of minimally invasive liver resections. In this cohort, 60% of the RLR procedures involved the posterosuperior segments. Laparoscopic resection of these segments is mechanically restricted by instrument angles along the diaphragmatic surface [4, 5]. The robotic platform overcomes these constraints through articulating instruments and stable, multi-axis retraction [6]. Importantly, the high initial difficulty (median IWATE score of 9) demonstrates that the observed 11-case efficiency breakpoint occurred without early case-selection bias. Postoperative morbidity analysis demonstrated that rapid technical adoption did not compromise patient safety. Interventions for major complications, such as percutaneous drainage or ERCP, reflected the high baseline procedural complexity involving major vascular proximity and posterosuperior segmentectomies [18]. The incidence of these interventions remained stable between the competency and efficiency phases. The significant reduction in median estimated blood loss (1,000 to 450 mL, p<0.05) alongside operative time optimization reflects improved parenchymal transection stability and confident vascular control in the efficiency phase (Table 2). Furthermore, the absence of open conversion across both phases stands in contrast to the biliary or hemorrhagic spikes frequently reported in early laparoscopic series, reinforcing the inherent stability of the robotic platform [2, 18]. Beyond individual surgeon experience, institutional infrastructure and team proficiency dictate the learning curve trajectory [19]. Notwithstanding the requirement for docking procedures, the RLR and LLR groups showed equivalent median operative time without statistical significance(149 vs. 156 min, p=0.964). This operative efficiency is directly attributable to a dedicated surgical nursing team whose procedural proficiency mitigated setup overhead and docking sequences, thereby optimizing the 11-case transition phase. The role of prior laparoscopic volume in shaping the robotic learning curve warrants analysis. Ratti et al. [20] compared RLR learning curves between surgeons with disparate laparoscopic backgrounds (1,200 vs. 100 cases) and reported comparable learning phases for low-to-intermediate complexity cases, citing the platform’s ergonomic advantages. However, their protocol mandated a stepwise progression for high complexity resections [20]. Conversely, the surgeon in our cohort immediately initiated RLR with high-complexity cases (initial median IWATE score of 9) and achieved efficiency by case 11. This indicates that extensive laparoscopic experience, synergized with robotic technology, minimizes the need for a prolonged, stepwise adaptation period, even when managing expert-level anatomical challenges from the outset. The capacity of the robotic platform to compress the learning curve remains evident when controlling for training sequence. Efanov et al. [21] demonstrated that even when skills were extrapolated from RLR to LLR—giving the laparoscopic arm a theoretical advantage—the learning curve for posterosuperior resections remained significantly shorter in RLR than in LLR (16 vs. 29 cases) [21]. This confirms the inherent technical advantages of the robotic platform, independent of historical training sequences. Several limitations of this study should be acknowledged. First, while constrained by potential selection bias inherent to its retrospective, single-center design, this study mitigated the risk by adjusting for case complexity via the IWATE score. Second, because all operations were performed by a single surgeon with substantial prior laparoscopic experience (149 LLR cases overall, with 35 cases prior to the RLR program), this compressed 11-case learning curve reflects the transferability of laparoscopic skills and platform adaptation in an experienced hepatobiliary surgeon. Therefore, these findings cannot be directly generalized to low-volume centers or trainees lacking a solid laparoscopic foundation. Third, although surgical difficulty was adjusted using the IWATE scoring system, this index primarily elaborates anatomical and resection-specific complexity and does not fully account for other surgical difficulties such as repeat hepatectomy, or extensive adhesions. Future prospective studies applying comprehensive complexity models may help further delineate learning curves. Finally, long-term oncological outcomes were not evaluated, necessitating future multi-center prospective validation [22, 23]. Furthermore, future dedicated analyses are warranted to specifically evaluate the unique learning curve and perioperative feasibility of pure single-port robotic liver resections [24, 25, 26]. Conclusion In conclusion, complexity-adjusted analysis demonstrated a shorter learning curve for RLR (11 cases) than LLR (50 cases), despite a high initial case complexity (median IWATE score of 9). This rapid technical adaptation indicates that prior laparoscopic experience combined with the technical features of the robotic platform facilitates procedural efficiency when transitioning to robotic liver resection. This transition was achieved with optimized operative efficiency and while maintaining comparable perioperative outcome. Institutional support, particularly from a specialized surgical team, remains an important factor in mitigating setup overhead and stabilizing early robotic procedures. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2