Artificial Intelligence–Based Surgical Phase Analysis Enables Objective Assessment of Surgeon Skill in Robotic Distal Gastrectomy: A Multicenter Study

ABSTRACT

Aim

Artificial intelligence (AI)–based surgical video analysis can automate time‐consuming manual assessments and enable objective characterization of surgical workflows. We aimed to construct a large, multicenter, fully annotated dataset of robotic distal gastrectomy (RDG) videos and evaluate the feasibility and performance of an AI model for surgical phase recognition. We further explored whether AI‐derived phase‐specific metrics could characterize phase‐level performance differences according to surgeon experience.

Methods

We developed an image classification model to automatically identify surgical phases in RDG videos. Experienced gastric surgeons annotated nine predefined surgical phases on a frame‐by‐frame basis (1 fps). Model performance was assessed using accuracy, precision, recall, and F1‐score. In an exploratory analysis, the duration of each AI‐predicted surgical phase was quantified and compared according to each surgeon's robotic case volume.

Results

We analyzed 137 RDG videos collected from 15 institutions. The nine‐phase recognition model achieved an overall accuracy of 87.0%. Among the surgical phases, the AI‐predicted duration of right‐sided greater curvature lymphadenectomy was shorter for surgeons with higher robotic case volume than those with lower volume (3789 vs. 2547 frames, p < 0.01), indicating phase‐specific differences in operative performance.

Conclusions

The proposed AI model demonstrated robust performance in multicenter surgical phase recognition. AI‐based surgical phase analysis enabled objective characterization of phase‐specific operative profiles according to surgeon experience. These findings support the feasibility of surgical phase analysis as an exploratory framework for performance profiling and educational support in robotic gastrectomy. However, generalizability may be limited owing to institutional overlap between the training and validation datasets.

Short abstract

This multicenter study developed an AI model for automatic surgical phase recognition in robotic distal gastrectomy using a large annotated video dataset and achieved 87% overall accuracy. Exploratory analysis showed shorter AI‐derived lymphadenectomy duration among higher‐volume surgeons, supporting the feasibility of phase‐based performance profiling while highlighting the need for outcome‐based validation.

1. Introduction

In the evolution of gastric cancer treatment, laparoscopic distal gastrectomy (LDG) has shown long‐term outcomes similar to those of open distal gastrectomy [1]. Amid this progressive shift, the advent of robotic distal gastrectomy (RDG) marked a significant advancement, effectively addressing the limitations of laparoscopy, including manual tremors and surgeon fatigue [2, 3, 4, 5]. Robotic gastrectomy mitigates these challenges, facilitating a more refined surgical precision and potentially reducing short‐term complications relative to laparoscopic gastrectomy [6]. A randomized clinical trial conducted in Japan reported a complication rate (≥ Clavien–Dindo Grade III) of 5.3% for robotic gastrectomy, compared with the 16.2% reported for laparoscopic gastrectomy [7]. Despite the low complication rate for robotic gastrectomy [7], persistent problems remain: the complication rate increases with inexperienced surgeons until the learning curve is complete, and more improvement is warranted due to the inter‐surgeon variability in outcomes [8]. In addition, metrics such as operative time, blood loss, and postoperative complication rates cannot capture phase‐specific technical performance and identify which phases contribute to inefficiencies or adverse outcomes. Furthermore, clarifying the specific surgical phase to be improved is essential to develop a more efficient learning curve, and objective, phase‐specific quality metrics are required. Efficient education and surgical quality standardization across medical centers help ensure patient safety and reduce complications in a limited number of cases.

Novel technologies are being explored to support skill development and procedural standardization, thereby enhancing surgical training and optimizing outcomes. Recently, artificial intelligence (AI) has significantly advanced the field of surgical education and practice, with reports of automatic surgical phase recognition and anatomical structure identification [9, 10, 11]. AI‐based technologies can improve intraoperative decision‐making, enhance surgical precision, and facilitate real‐time guidance, ultimately improving patient outcomes [12]. Among these applications, surgical phase recognition can be applied to assess surgical skills in laparoscopic procedures [13, 14] and may facilitate efficient education and further standardization of quality in robotic gastrectomy through automatic surgical skill assessment. However, a lack of large‐scale, high‐quality annotated datasets hinders the development of robust AI‐driven models for surgical phase recognition despite these advancements.

An objective evaluation of phase‐specific performance is essential for quality control, structured training, and the safe dissemination of RDG across institutions. Variations in technical performance that are not captured by conventional global outcome measures can be identified by quantifying phase‐specific operative characteristics, including phase duration, surgical phase analysis.

Therefore, our primary aim was to establish the feasibility of AI‐based surgical phase recognition using a multicenter dataset and to characterize phase‐specific performance profiles in RDG. We constructed a fully annotated dataset of RDG and evaluated the performance of an AI‐based model for surgical phase recognition. We further investigated whether phase‐derived metrics, particularly phase duration during lymphadenectomy, could discriminate surgeon expertise, highlighting the potential role of surgical phase analysis as an objective tool for quality control and training in robotic gastrectomy.

2. Materials and Methods

2.1. Dataset for Phase Recognition

We constructed an RDG video dataset collected from 15 Japanese hospitals between June 2016 and May 2021. The number of cases from each institution were: National Cancer Center Hospital East: 12, Kobe City Medical Center General Hospital: six, Fukuoka University Hospital: four, Shiga General Hospital: two, Ehime Prefectural Central Hospital: four, Ishikawa Prefectural Central Hospital: five, Kindai University Hospital: two, Japanese Red Cross Mito Hospital: 50, Cancer Institute Hospital of Japanese Foundation for Cancer Research: 10, Osaka Medical and Pharmaceutical University: three, Saitama Prefectural Cancer Center: seven, Shizuoka Cancer Center: 20, Showa University Hospital: five, Chiba Cancer Center: five, Okayama University Hospital: two. Details of the contributing institutions, including the number of surgeons included in the data, the number of cases per institution, and the relative case volume per surgeon, are listed in Table S1.

Demographic and clinical variables, including patient age, sex, body mass index, and cancer stage, as well as operative details, such as operation time (min), volume of blood loss (mL), extent of lymph node dissection, and reconstruction method, were collected for each surgical video. The cancer staging conformed to the eighth edition of the tumor, node, and metastasis classification [15], whereas the lymph node dissection parameters and the numerical designation of lymph node stations aligned with the Third Edition of the Japanese Classification of Gastric Carcinoma and corresponding Japanese guidelines [16, 17].

To minimize the influence of case mix on the analysis of surgeon expertise, clinical and procedural variables potentially affecting surgical complexity, including clinical stage, extent of lymphadenectomy, and reconstruction method, were summarized and compared between the surgeons with > 100 cases (hereafter referred to as the higher case‐volume group) and trainee groups (Table S2). Surgeon experience was recorded as predefined ordinal categories rather than continuous cumulative case numbers; therefore, analyses treating experience as a continuous variable were not feasible.

2.2. Pre‐Processing and Annotation

Videos were recorded at 30 frames per second (fps) and converted to a standardized format (MPEG‐4 AVC/H.264), with a display resolution of 1280 × 720 pixels. RDG procedures were divided into nine defined surgical phases (Phases 1–9), and other actions were based on the previous report in principle [14]. Extracorporeal procedures, including cleaning the surgical camera and performing jejunojejunostomy during Roux‐en‐Y Loop reconstruction, were acknowledged as distinct phases, aside from phases 1–9. The definitions and descriptions of each phase are presented in Table 1. Phase 1 (preparation) involved port insertion, lateral segment retraction, and adhesiolysis. Surgical phase labels were annotated at 1 fps and independently reviewed by three experienced gastric surgeons. Any annotation differences were discussed and resolved collaboratively. The video dataset was randomly divided into training and test subsets.

TABLE 1.

Definitions of the nine surgical phases involved in robotic distal gastrectomy.

Phase	Definition	Description
Phase 1 (P1)	Preparation	The endoscopic camera is inserted into the abdominal cavity
Phase 2 (P2)	Lymphadenectomy on the left greater curvature	Approach to the omentum from the right side
Phase 3 (P3)	Lymphadenectomy on the right greater curvature	Switch of the standing position of the surgeon and the assistant
Phase 4 (P4)	Transection of the duodenum	Stapler for duodenectomy appears on screen
Phase 5 (P5)	Suprapancreatic lymphadenectomy	Approach to the suprapancreatic lymph nodes
Phase 6 (P6)	Lymphadenectomy on the lesser curvature	Approach to the lymph nodes of the lesser curvature
Phase 7 (P7)	Transection of the stomach	Stapler for gastrectomy appears on screen
Phase 8 (P8)	Reconstruction	Approach to the intestine to be reconstructed
Phase 9 (P9)	After reconstruction, to the completion of surgery	Washing instrument appears on screen
Other actions	Such as cholecystectomy	Start: previous step ends End: following step starts

2.3. Phase‐Recognition Model

Image classification was performed using a convolutional neural network (CNN) based on the EfficientNet‐B7 architecture [18]. Selected fine‐tuned hyperparameters optimized for the model's classification accuracy are summarized in Table S3. Automatic surgical phase classification was performed for phases 1–9 and other actions. All modeling procedures were performed using a script written in Python 3.6 (Python Software Foundation, Wilmington, DE, USA).

Model training and testing were performed on a workstation equipped with NVIDIA Quadro RTX 6000 GPU featuring 24 GB of VRAM (NVIDIA, Santa Clara, CA, USA) and an Intel Xeon Gold 5222 CPU@3.80 GHz with 172 GB of RAM.

2.4. Evaluation Metrics

The precision, recall, F1‐score, and overall accuracy of the metrics for evaluating the CNN model performance in the surgical phase recognition task were calculated using the following equations:

$$\text{Precision}=\frac{\mathit{TP}}{\mathit{TP}+\mathit{FP}}$$ (1)

$$\text{Recall}=\frac{\mathit{TP}}{\mathit{TP}+\mathit{FN}}$$ (2)

$$\mathrm{F}1-\text{score}=2・\frac{\mathit{\text{Precision}}・\mathit{\text{Recall}}}{\mathit{\text{Precision}}+\mathit{\text{Recall}}}$$ (3)

$$\mathit{\text{Overall accuracy}}=\frac{\mathit{TP}+\mathit{TN}}{\mathit{TP}+\mathit{FP}+\mathit{TN}+\mathit{FN}}$$ (4)

where TP, FP, TN, and FN denote true positive, false positive, true negative, and false negative counts. Model performance was evaluated on a frame‐by‐frame basis at 1 fps. For each frame, the predicted surgical phase was compared with the ground‐truth annotation provided by experienced surgeons, and TP, FP, TN, and FN were defined accordingly. Precision, recall, and F1‐score metrics were applied to assess phase‐specific performance, whereas the overall model's performance was evaluated using the global accuracy metric. A holdout validation method was adopted to assess the performance of the proposed model. The available dataset was randomly partitioned into training and validation sets to ensure that the model was evaluated using unseen data during training.

2.5. Relationships Between Each Phase Duration and the Surgeons' Robotic Case Volume

Frame counts obtained from the annotation data were used to calculate the duration of each phase. The number of frames in each phase was compared between the higher case‐volume group and trainee groups. Surgeons with an experience of > 100 robotic surgeries were classified into the higher case‐volume group, and those with an experience of < 100 surgeries were classified into the trainee group. We compared the duration between the two groups, in relation to the surgeon's skill in Phases 2–8.

2.6. Ethical approval

The Institutional Review Board (IRB) of the National Cancer Center Hospital East (Registration No. 2020‐315) and the respective IRBs of all participating institutions approved the study protocol. All participants provided written informed consent before participation. The study protocol can be freely accessed via the web pages of the participating organizations found at the URL: https://www.s‐access.ncc.go.jp/policy/#from2019‐subheadline.

2.7. Statistical Analysis

Data for surgeries performed between June 2016 and May 2021 were analyzed. Patient demographics and perioperative outcomes in the validation videos were compared using Fisher's exact test and the Mann–Whitney U test. All p‐values were two‐sided, and statistical significance was set at p < 0.05. Furthermore, all statistical analyses were performed using EZR (Jichi Medical University Saitama Medical Center), a graphical user interface for R version 2.13.0 (R Project for Statistical Computing) [19].

Individual surgeon identifiers were not available; hence, clustering by surgeon could not be explicitly modeled, and separation at the surgeon level between the training and validation sets was not feasible. Therefore, additional sensitivity analyses were performed to mitigate clustering‐related bias (Figures [Link], [Link]).

The institutional distribution of cases in the training and test datasets is summarized in Table S4. Videos from the same institutions were included in both datasets; however, the test set comprised cases from multiple centers and was not dominated by a single institution.

3. Results

3.1. Dataset

We constructed a phase classification model using 137 RDG videos, comprising 109 training and 28 test cases. Patient characteristics included in the training and test datasets are summarized in Table 2. No differences were observed in patients' backgrounds between the training and test data. Overall, 113 patients (83.7%) had clinical stage I disease. D1+/D2 dissection was performed in 50 (37.0%) and 81 (60.0%) patients, respectively. Regarding reconstruction methods, Billroth I, Billroth II, and Roux‐en‐Y were performed in 69 (51.1%), 33 (24.4%), and 33 (24.4%) cases, respectively. The median operative time and blood loss were 304 (range: 201–560) min and 15 (range: 0–250) mL, respectively. Postoperative complications ≥ Clavien–Dindo Grade IIIa were observed in only two patients (1.5%).

TABLE 2.

Baseline characteristics describing the collected training and test data set (n = 137).

	Training (n = 109)	Test (n = 28)	p
Age, years a	71 (27–89)	71.5 (37–84)	0.72
Sex
Male	69 (64.5%)	15 (53.6%)	0.38
Female	38 (35.5%)	13 (46.4%)
BMI, kg/m [2] a	22.7 (10.7–33.8)	21.7 (17.8–31.9)	0.76
ASA‐PS
1	22 (20.6%)	5 (17.9%)	0.94
2	75 (70.1%)	21 (75.0%)
3	10 (9.3%)	2 (7.1%)
Preoperative chemotherapy	0	1 (3.6%)	0.2
UICC‐TNM stage (8th edition)			0.28
I	90 (84.1%)	23 (82.1%)
IIA	2 (1.9%)	2 (7.1%)
IIB	9 (8.4%)	3 (10.7%)
III	6 (5.6%)	0
IVA	0	0
IVB	0	0
Experience of surgeons, years a	22 (14–38)	21 (9–28)	0.32
Experience in robotic surgery, cases a			0.5
0–1	1 (0.9%)	0
2–10	17 (15.6%)	3 (10.7%)
11–40	34 (31.2%)	11 (39.3%)
41–100	44 (40.4%)	8 (28.6%)
101 >	13 (11.9%)	6 (21.4%)
Da Vinci model
Si	47 (43.9%)	9 (32.1%)	0.28
Xi	60 (56.1%)	19 (67.9%)
Operation time, min a	304 (201–560)	303 (225–485)	0.59
Console time, min a	227 (89–414)	233 (155–420)	0.19
Blood loss, mL a	15 (0–250)	9.5 (0–156)	0.63
Conversion to open surgery	0	0	NA
Drain amylase, U/L	171 (0–4200)	171.5 (0–3050)	0.98
Postoperative complication ≥ C‐D Grade IIIa	2 (1.9%)	0	1
Postoperative hospital stay, day a	10 (7–48)	10 (7–16)	0.37
Lymph node dissection			0.12
D1+	44 (41.1%)	6 (21.4%)
D2	60 (56.1%)	21 (75.0)
Others	3 (2.8%)	1 (3.6%)
Reconstruction			0.36
B‐I	58 (54.2%)	6 (21.4%)
B‐II	25 (23.4%)	21 (75.0%)
RY	24 (22.4%)	1 (3.6%)

Abbreviations: ASA‐PS, American Society of Anesthesiologists; BMI, body mass index; UICC, Union for International Cancer Control.

^a Median (range).

Baseline clinicopathological characteristics and procedural details stratified by surgeon expertise are presented in Table S2. No clinically relevant imbalance was observed in the extent of lymphadenectomy between the two groups.

3.2. Phase Recognition

All test data were extracted and analyzed at 1 fps. The results of the phase recognition from the test dataset are presented in Table 3 and Figure 1. Phase 3 (right greater curvature lymphadenectomy) had the highest F1‐score (0.91), whereas Phase 7 (transection of the stomach) had the lowest (0.72). The overall accuracy of the nine‐phase recognition model was 87.0%. The average precision, recall, and F1‐score were 87.0%, 87.0%, and 87.0%, respectively. A weighted average was applied to account for phase frequency, rather than a macro average. Number of frames during each phase was shown in Figure 2. A timeline‐based visualization demonstrating the temporal alignment between surgeon‐annotated ground‐truth phases and AI‐predicted phases across representative test cases was presented to enhance clinical interpretability (Figure 3).

TABLE 3.

Precision, recall, and F1 scores for surgical phase recognition.

	Precision	Recall	F1 score
Phase 1. Preparation	0.88	0.92	0.90
Phase 2. Lymphadenectomy on the left greater curvature	0.90	0.89	0.89
Phase 3. Lymphadenectomy on the right greater curvature	0.91	0.92	0.91
Phase 4. Transection of the duodenum	0.71	0.73	0.72
Phase 5. Suprapancreatic lymphadenectomy	0.91	0.90	0.90
Phase 6. Lymphadenectomy on the lesser curvature	0.83	0.79	0.81
Phase 7. Transection of the stomach	0.76	0.68	0.72
Phase 8. Reconstruction	0.76	0.77	0.77
Phase 9. After reconstruction, to the completion of surgery	0.75	0.80	0.78
Overall	0.87	0.87	0.87

FIGURE 1.

Confusion matrix. The vertical and horizontal axes represent the ground‐truth and AI‐predicted phases, respectively. Each cell shows the proportion of frames for a given ground‐truth phase classified into each predicted phase (row‐wise normalization). Diagonal values represent the per‐phase true positive rate (68%–92%).

FIGURE 2.

Number of frames using predicted labels during each phase.

FIGURE 3.

Transitions of the surgical phase during a representative robotic distal gastrectomy case. Representative cases were selected to reflect typical procedural flow without major intraoperative events, and discrepancies were most frequently observed at the phase transition boundaries.

3.3. Relationships Between Frame Count and the Surgeon's Experience

The number of frames per surgical phase and the relationships between phase duration and surgeon experience are presented in Table 4. Phase duration varied significantly with the surgeon's experience level, particularly in Phase 3. The higher case‐volume group had a lower frame count than the trainee group in the labeled (3614 vs. 2614, p = 0.01) and model‐predicted phases (3789 vs. 2547, p < 0.01) (Table 4). After excluding the highest‐volume center (n = 50), the difference in Phase 3 duration between the experience groups remained directionally consistent (3614 vs. 2614 frames; p = 0.01) (Figure S3).

TABLE 4.

Relationship between surgical experience and duration of each phase.

Phase	Labeled phase			Predicted phase
	Trainee	Higher case‐volume group	p	Trainee	Higher case‐volume group	p
1	1113 (556–2051)	1677 (1296–1919)	< 0.01	1127 (744–2199)	1641 (1285–2102)	0.01
2	1214 (878–2109)	1297 (758–2565)	0.72	1258 (852–2222)	1191 (716–2668)	0.81
3	3614 (1491–7024)	2614 (2394–3362)	0.01	3789 (1570–6929)	2547 (2227–3096)	< 0.01
4	219 (95–1126)	369 (152–533)	0.52	274 (109–891)	449 (236–590)	0.11
5	3923 (1855–9338)	3234 (2813–5202)	0.35	4039 (1766–8236)	3262 (2940–5304)	0.53
6	878 (337–2014)	959 (726–1737)	1	935 (378–2052)	901 (678–1477)	0.84
7	640 (177–1236)	716 (292–989)	0.81	536 (222–1138)	621 (372–1069)	0.57
8	1904 (1013–4362)	1638 (1062–3684)	0.64	2157 (1104–2808)	1601 (1234–2174)	0.11
9	1565 (708–2234)	1452 (933–1846)	0.36	1512 (917–3516)	1617 (1047–2422)	0.89

4. Discussion

We developed a large, annotated, multicenter dataset comprising 137 RDG videos and constructed a deep learning model for automatic surgical phase recognition in RDG. The model achieved an overall classification accuracy of 87.0%, demonstrating its ability to recognize surgical phases with high accuracy. Notably, surgeons with higher robotic case volume demonstrated shorter and less variable Phase 3 durations, which is an oncologically critical step in RDG. These findings indicate gradual stabilization of phase‐specific performance rather than a discrete threshold for surgical proficiency. Accordingly, our focus was not to provide a comprehensive assessment of surgical skill; rather, AI‐based surgical phase analysis was used to characterize phase‐specific performance profiles during RDG, particularly in this oncologically critical step.

Previous studies have reported that the accuracy of surgical phase recognition using surgical videos from multiple centers may be marginally lower than that from a single center, owing to the variety of approaches and forceps used [20, 21]. In this study, no difference in accuracy was observed compared with that in single‐center studies. Takeuchi et al. developed a surgical phase‐recognition model for RDG, achieving 87% accuracy using single‐institution data [21], which aligns with our multicenter result of 87%. This consistency may be because RDG is a surgical procedure consisting of clearly defined surgical phases and is standardized in Japan with relatively few variations [22, 23]. Therefore, a good recognition accuracy was achieved even with a multicenter video.

The durations of certain stages are significantly associated with a surgeon's learning curve [24]. In gastric cancer surgery, the overall surgical time decreases with increased experience; however, reports identifying which specific surgical phase contributes most significantly to this reduction are limited [25]. While robotic surgery has been associated with lower complication rates [6], many complications reportedly occur during the learning curve [25], underscoring the need for effective skill acquisition. Understanding the components of the operation that become more efficient with experience may facilitate enhanced learning outcomes.

Reportedly, pancreatic compression for more than a few minutes during minimally invasive gastrectomy increases the incidence of postoperative complications [26]. In Phase 3 (right‐sided lymphadenectomy of the greater curvature), the pancreas may be compressed; therefore, completing the procedure in a shorter time may be associated with the incidence of pancreatic fistulas. Moreover, in RDG, Phase 3 is time‐consuming, and identifying the optimal dissection layer is difficult. This variation is significant and depends on factors such as the presence of adhesions, mesocolon suspension, and the shape of the pancreatic head. Owing to such marked anatomical differences, experience and skills are required to shorten the operation time in Phase 3. Consequently, variations in the duration of Phase 3 may be due to the surgeon's level of experience.

Surgical skill assessment is essential for education, and it is always better to systematically assess surgical skills from multiple perspectives using the Objective Structured Assessment of Technical Skill (OSATS) and Global Operative Assessment of Laparoscopic Skills (GOALS). Conversely, it is time‐consuming and labor‐intensive [27, 28]. This issue needs to be addressed, and AI‐based methods offer a potential solution by automating skill evaluation. Notably, several reports have been published on surgical skills assessment using AI‐driven metrics and various evaluation indices, including the AI Confidence Score and measures, such as the ratio of the total procedure time to instrument visibility time [13, 14]. These metrics consider factors, including appropriate surgical field exposure with adequate traction, effective collaboration with assistants, precision of the dissection layer, and clear landmark identification [29].

Because surgical procedures consist of multiple distinct phases, each involving different tasks and requirements, we hypothesized that the parameters used to evaluate surgical skills vary depending on the specific phase of the procedure. Furthermore, we hypothesized that certain phases are strongly associated with skill proficiency than others. We focused on identifying such distinctions, intending to highlight the varying importance of individual phases in skill assessment. In addition, we hope that our findings encourage further consideration of phase‐specific evaluations during surgical training and analysis.

This study has some important limitations. First, the absence of surgeon‐level identifiers precluded full hierarchical modeling, and residual clustering effects at the surgeon and institutional levels could not be excluded completely. In addition, including videos from the same institutions in the training and validation sets may have led to optimistic performance estimates, and external validation using strict institution‐level separation is warranted. Second, a direct or comprehensive assessment of surgical skill was not performed. Rather, surgeon experience was analyzed using case volume as a surrogate measure, and differences in phase‐specific performance trends were examined. Therefore, the findings should not be interpreted as definitive measures of surgical proficiency. The observed differences in phase‐specific performance indicate gradual stabilization with increasing experience, rather than evidence of objective skill acquisition. Residual confounding owing to differences in case mix, including ASA‐PS and reconstruction methods, cannot be completely excluded. Third, surgeon experience was dichotomized using a threshold of 100 robotic cases. This cutoff was informed by prior literature on learning curves in robotic gastrectomy and other robotic procedures [25, 30, 31, 32, 33]; however, it remains an imperfect surrogate and does not imply a discrete threshold for surgical mastery. Fourth, this exploratory proof‐of‐concept study was not designed to evaluate clinical outcomes. In the present cohort, we could not use patient outcomes, such as postoperative complications and recurrence rates as surrogates for surgical skill due to the low incidence of events and the lack of long‐term follow‐up. Therefore, potential associations between phase duration and specific complications, such as pancreatic fistula, should be interpreted with caution, and future outcome‐oriented studies are required for validation. Fifth, potential overfitting may have occurred, as the model was not validated using an external cohort. Standard performance metrics were reported, but additional indices such as receiver operating characteristic curves and area under the curve could further strengthen model evaluation. Moreover, the relatively low F1 scores observed in certain phases likely reflect class imbalance and limited frame counts. Sixth, this was a multicenter study, but all videos were obtained from institutions in Japan. The relatively low body mass index of the patient population may limit generalizability to other populations. Nevertheless, preserving model performance across multiple centers suggests a degree of procedural standardization in RDG within Japan. Future larger studies with more diverse datasets, external validation, and outcome‐based and framework‐based assessments, such as OSATS, GEARS, and GOALS, are required to further establish the clinical use of AI‐based surgical phase analysis [27, 28, 34]. All participating institutions in this study were relatively high‐volume centers with standardized RDG procedures. Hence, this setting was appropriate for establishing the feasibility of AI‐based surgical phase recognition; however, applicability to more heterogeneous environments, such as low‐volume centers with greater procedural variability, remains uncertain. Given the potential educational application of surgical phase analysis, future studies should validate the model across institutions with varying case volumes and workflow patterns.

In conclusion, we developed a deep learning–based surgical phase recognition model for RDG using a multicenter video dataset. The model demonstrated robust performance in phase classification and identified phase‐specific differences in operative duration across experience levels. These findings suggest that surgical phase analysis may provide an exploratory framework for performance profiling and educational support in robotic gastrectomy.

Author Contributions

Masaru Komatsu: conceptualization, data curation, investigation, formal analysis, writing – original draft. Daichi Kitaguchi: conceptualization, data curation, investigation, formal analysis. Mitsumasa Yoshida: conceptualization, data curation, investigation, formal analysis. Yoichi Ninomiya: conceptualization, data curation, investigation, formal analysis. Yuya Takabe: conceptualization, data curation, investigation, formal analysis. Masaaki Ito: conceptualization, data curation, investigation, formal analysis, supervision. Nobuyoshi Takeshita: conceptualization, data curation, investigation, formal analysis. Masahiro Yura: conceptualization, data curation, investigation, formal analysis. Takahiro Kinoshita: conceptualization, data curation, investigation, formal analysis.

Funding

This study was funded by the National Cancer Center, Japan Research and Development Fund (grant number: 2021‐A‐9).

Ethics Statement

The Institutional Review Board (IRB) of the National Cancer Center Hospital East (Registration No. 2020–315) and the respective IRBs of all participating institutions approved the study protocol. The study protocol can be freely accessed via the web pages of the participating organizations found at the URL: https://www.s‐access.ncc.go.jp/policy/#from2019‐subheadline. The research project protocol complies with the provisions of the Declaration of Helsinki (as amended in Fortaleza, Brazil, in October 2013).

Consent

All participants provided written informed consent before participation.

Conflicts of Interest

Dr. Nobuyoshi Takeshita was funded by the National Cancer Center, Japan Research and Development Fund (grant number 2021‐A‐9). Drs. Masaru Komatsu, Daichi Kitaguchi, Yoichi Ninomiya, Yuya Takabe, Mitsumasa Yoshida, Masahiro Yura, Takahiro Kinoshita, and Masaaki Ito have no conflicts of interest or financial ties to disclose.

Supporting information

Figure S1: Sensitivity analyses of Phase 3 duration according to alternative surgeon experience thresholds. Box plots compare Phase 3 duration between surgeons stratified by robotic case volume using cutoffs of ≤ 10 versus > 10, ≤ 40 versus > 40, and ≤ 100 versus > 100 cases. A significant difference was observed at the > 100‐case threshold, whereas no significant differences were detected at lower cutoffs, suggesting gradual stabilization of phase‐specific performance with increasing experience rather than a discrete proficiency threshold.

Figure S2: (a) Relationship between surgeon robotic case volume categories and Phase 3 duration. Each point represents an individual case, stratified by predefined surgeon experience categories (0–10, 11–40, 41–100, and > 100 robotic cases). A gradual directional trend toward shorter phase 3 duration with increasing experience is observed across ordinal categories, without assuming a discrete proficiency threshold. (b) Relationship between surgeon robot case volume category and total of all phase duration. Directional trend toward a gradual decrease in duration with increasing experience was not observed.

Figure S3: Comparison of phase 3 duration between surgeons with ≤ 100 and > 100 robotic cases after exclusion of the highest‐volume center.

Table S1: Details of the participating facilities.

Table S2: Baseline characteristics describing the collected trainee and higher case‐volume group data set (n = 137).

Table S3: Learning conditions.

Table S4: Distribution of training and test data sets for each participating facility.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.