This case study involved an experienced RAS surgeon (> 250 RAS procedures). We aimed to explore the feasibility of subject-specific predictive modeling in assessing surgical performance. Focusing on one participant allowed us to examine intra-individual performance patterns. This design supports the development of accurate, personalized models and represents an initial step toward individualized surgical support systems. Prior to the study, the participant provided written informed consent and was explicitly informed that task difficulty and environmental stressors would be systematically manipulated as part of the experimental design.

The experimental setup is illustrated in Figure 1A. The participant performed two simulation tasks, Suture Sponge 1 (SS1) and Energy Switching 1 (ES1), using the da Vinci Skills Simulator (dVSS) connected to a da Vinci Xi surgeon console (Intuitive Surgical, Inc., Sunnyvale, CA, USA). Three independent variables, each with two levels, were applied to induce variability in stress and workload: OR noise level (low vs. high), surgical posture (expert-like vs. novice-like), and task type (SS1 vs. ES1).

High OR noise ranged from 55 to 65 dB(A), while low OR noise was set at 40–50 dB(A). Both conditions were played through speakers using a pre-recorded surgical environment soundtrack (Carillo et al., 2019). For surgical posture, the expert-like condition allowed the participant to adjust the console ergonomic controls for optimal comfort. In the novice-like condition, the armrest was set lower to reduce forearm support, replicating suboptimal ergonomics that may increase muscle tension (Franklin et al., 2003; Takayasu et al., 2018; Carillo et al., 2019).

The SS1 task involved manipulating a curved needle through two designated points on adjacent sponge faces using both hands. This task simulates essential suturing operations and demands a high degree of precision and focus. The ES1 task required the participant to adjust the camera view to identify and deal with multiple targets using appropriate instruments. It simulated multitask coordination in dynamic surgical scenario that requires rapid decision-making and frequent instrument switching. These two tasks were selected to represent distinct surgical skills. This approach helps reduce task-specific bias and improves the generalizability of the findings.

This section describes the extracted physiological and subjective features. Physiological measures included EEG, EOG, EMG, EDA, and HR. Subjective workload was assessed using the NASA-TLX and SURG-TLX. All acquired data were preprocessed and analyzed to derive relevant metrics as features for subsequent modeling.

Physiological signals were continuously recorded throughout the experiment. The placement of electrodes and sensors for EEG, EOG, EMG, EDA, and HR measurements is illustrated in Figure 1B. EEG, EOG, EMG, and EDA signals were recorded at 1,000 Hz using the BIOPAC MP 160 system (BIOPAC Systems Inc., Goleta, CA, USA) with wireless transmitters and AcqKnowledge 5.0 (BIOPAC Systems Inc., Goleta, CA, USA). HR was measured with the Polar Verity Sense (Polar Electro Oy, Kempele, Finland) which uses photoplethysmography (PPG) and samples at 135 Hz for pulse detection. Subjective workload was assessed using paper-based versions of the NASA-TLX and SURG-TLX.

Figure 1C illustrates the Experimental Procedure. Two days prior to the experiment, the participant practiced both tasks to ensure familiarity and reduce performance variability. At the start of each session, the participant’s skin was cleaned, and conductive paste was applied before attaching the EEG electrodes. MVC was measured, followed by a five-minute rest period before a 10-min baseline recording of EEG, EDA, and HR.

The participant then performed the SS1 and ES1 tasks in a randomized order (three repetitions per task). Subjective evaluations were conducted after each task. A 15-min break was provided before the next trial to minimize fatigue and carryover effects. Over four consecutive days, the participant completed six unique task–posture–noise combinations (each with three trials) per day. Task order was counterbalanced to minimize order effects. All sessions started at consistent times to control for circadian variability. In total, the participant completed 72 trials (2 tasks × 2 noise levels × 2 postures × 9 repetitions), providing a comprehensive dataset for modeling.

Task performance was evaluated by dVSS as a composite score ranging from 0 to 100. This score incorporated completion time, master workspace range, and task-specific precision and error metrics (Havemann et al., 2019). Performance scores were dichotomized into high and low categories using a median split. The task performance prediction (high vs. low) was formulated as a supervised classification problem using the extracted features as input. Four classifiers were trained and compared: CatBoost, random forest (RF), support vector machine (SVM), and logistic regression (LR). CatBoost was chosen for its robustness in small datasets, leveraging ordered boosting to reduce overfitting, and its symmetric tree-growing strategy, which enhances training efficiency and model consistency. The model was trained with 355 trees and a learning rate of 0.015. These hyperparameters were manually selected to achieve a better balance between training and testing performance, rather than solely maximizing test AUC (Prokhorenkova et al., 2019). The RF model used 100 trees. The SVM model applied a radial basis function (RBF) kernel. The hyperparameters C and gamma were selected based on empirical performance across logarithmically spaced values. Logistic regression with L2 regularization was used as the baseline linear model. All models were implemented in Python. Figure 1D illustrates the overall workflow of the machine learning approach used in this study.

Model performance was evaluated using repeated 10-fold cross-validation. Performance metrics were averaged across folds to ensure reliability. We used the area under the curve (AUC) of the receiver operating characteristic (ROC) curve and accuracy, as well as additional confusion matrix metrics such as precision, recall, and F1-score to provide a comprehensive assessment.

After classification, SHAP was used to interpret the best model (Lundberg and Lee, 2017). Feature importance was quantified by averaging the absolute SHAP values across all trials. The distribution of SHAP values for each feature was further analyzed to evaluate its impact on the model’s predictions.

As shown in Figure 2, ROC curves were used to evaluate the performance of the four ML models. The CatBoost model achieved the highest test AUC of 0.807, outperforming random forest (0.796), logistic regression (0.789), and SVM (0.777), indicating its discriminative capability.

Additional performance metrics, including accuracy, precision, recall, and F1 score, are listed in Supplementary Table 1. CatBoost attained the highest accuracy of 0.795, along with similarly high precision (0.796), recall (0.793), and F1 score (0.795), suggesting its stable performance across multiple evaluation dimensions. Although random forest had the highest recall (0.818), its precision (0.770) and F1 score (0.793) were slightly lower than CatBoost. SVM showed moderately strong performance, particularly in recall (0.815), but had lower precision (0.749) and F1 score (0.781) compared to CatBoost and random forest. In contrast, logistic regression underperformed, especially in precision (0.696) and F1 score (0.726), indicating its limited effectiveness in this classification task. Overall, these results highlight CatBoost’s robustness across multiple evaluation metrics, supporting its role as the most effective model in this study.

As shown in Figure 3, the SHAP feature importance plot (bar chart) ranks input features based on their average absolute SHAP values, reflecting their overall contribution to the model’s predictions, while the SHAP summary plot (bee swarm) visualizes the distribution of SHAP values for each feature, capturing both the direction and magnitude of their effects on individual predictions. The NASA-TLX and SURG-TLX workload scores emerged as the top two most important predictors, suggesting a strong association between subjective workload assessments in predicting task performance. Additionally, the HR_mean and MVE% of several muscles (especially MVE%_Trap and BB_MVE%) were also identified as important features contributing to the model’s performance predictions. Conversely, Fz_BAR, Cz_BAR, and Pz_BAR exhibited relatively lower importance, suggesting that these EEG-derived features were less predictive of task performance in this context. Among EDA-related features, nSCR/s had moderate importance, while SCL_mean showed relatively low importance. Finally, SDHR and TA_MVE% demonstrated the lowest contribution to the model’s predictions. While the low importance of TA_MVE% reflects limited contribution from this specific muscle group, the result for SDHR should be interpreted with caution, as it is a simplified proxy that may not fully represent HRV.

The SHAP analysis indicated that NASA-TLX and SURG-TLX were among the most influential features in the model’s performance predictions, with higher workload scores associated with lower predicted performance in this modeling context. NASA-TLX appeared slightly more influential in the model’s output than SURG-TLX, which may reflect its broader sensitivity within this task context, though such interpretation should be drawn cautiously given SHAP rankings are model-specific. Given the simplified nature of the simulated RAS task, NASA-TLX may better capture workload variations. Some dimensions of SURG-TLX, such as task complexity and situational stress, may be less sensitive in this setting than in real surgical procedures. However, in real surgeries with greater procedural complexity and teamwork demands, SURG-TLX may demonstrate stronger predictive power (Ma et al., 2021). Nevertheless, since subjective evaluations cannot be conducted during actual surgeries, further research is needed to explore the relationship between subjective assessments and physiological indicators.

Among the autonomic nervous system metrics, mean HR exhibited the highest contribution to model predictions, consistent with prior research linking increased HR to higher cognitive workload and reduced performance (Mansikka et al., 2016; Sazuka et al., 2024). While the SHAP analysis showed a pattern where higher SDHR values tended to co-occur with better predicted performance, its overall contribution to the model output was relatively small. SDHR, calculated from PPG-derived HR rather than ECG-based R-R intervals, may lack robustness in reflecting autonomic modulation. Future studies should consider ECG to obtain more reliable HRV metrics such as RMSSD and HF/LF for a more accurate assessment of physiological stress and surgical performance (The et al., 2020).

nSCR/s ranked eighth in feature importance. Higher values were predominantly observed in low-performance trials. This finding supports the view that increased electrodermal activity reflects elevated physiological stress and cognitive load, which may interfere with motor execution. Excessive sympathetic activation has been associated with impaired motor control and cognitive overload during high-precision tasks (Awtry et al., 2025). In contrast, the mean SCL showed low feature importance and exhibited inconsistent patterns. This may be attributed to its sensitivity to chronic stress rather than acute task-related demands (Visnovcova et al., 2024). Further research is needed to clarify its relevance in predicting performance in real-world surgical environments.

The SHAP analysis revealed notable variability in the importance of MVE% across different muscles. The trapezius, biceps brachii, erector spinae, and extensor digitorum communis emerged as the most influential, while the flexor digitorum superficialis and tibialis anterior contributed less to model predictions. Higher MVE% in the erector spinae and extensor digitorum communis was associated with lower performance, which may reflect a pattern related to postural instability and increased hand exertion. Conversely, greater trapezius activation correlated with better performance, which may suggest an association with upper limb stabilization during surgical tasks (Rodrigues Armijo et al., 2020). Interestingly, the biceps brachii exhibited a U-shaped SHAP pattern. Both low and high activation levels were linked to improved performance predictions, suggesting that this muscle may activate differently depending on task demands. This pattern implies that distinct motor strategies may be associated with optimal performance under varying conditions (Takatoku and Fujiwara, 2010). It should be noted that SHAP provides model interpretability rather than causal inference. Therefore, further studies integrating EMG pattern analysis with kinematic data are required to clarify the functional contributions of these muscles.

BAR exhibited low overall predictive importance but displayed distinct SHAP patterns across the electrode sites. At Fz, a U-shaped relationship was observed, with extreme BAR values associated with poorer performance and moderate values linked to better outcomes. This trend aligns with the Yerkes–Dodson law, which suggests that moderate arousal may enhance performance, while excessive stress impairs it (Khazaei et al., 2021). At Cz, BAR did not demonstrate a clear pattern, possibly due to its primary role in motor control rather than cognitive workload (Shaw et al., 2019). At Pz, higher BAR correlated with improved performance, potentially indicating enhanced task monitoring or sensory integration under increased cognitive demands on Kahya et al. (2022). The use of only three EEG channels provided a practical, low-interference setup, but it limits the ability to analyze region-specific or lateralized brain dynamics involved in surgical tasks. Although BAR was not among the top predictors in this study, EEG remains a valuable modality for investigating neural mechanisms in surgical performance. Future research should explore additional EEG-derived metrics and refine task design to support more comprehensive analyses.

While several physiological indicators demonstrated predictive value, their application in real surgeries may be limited by practical constraints. For instance, certain EMG and EDA electrode placements may hinder precise surgical maneuvers, as noted by the participant. Although these measures offered valuable insights, future studies should carefully assess their feasibility in surgical environments. Our findings also highlight the potential of explainable ML in surgical performance assessment. SHAP analysis facilitates the identification of features that the model considered influential in its predictions, thereby enhancing interpretability. This kind of transparency is crucial for clinical adoption. If ML models can explain their feedback (e.g., showing that high EEG-based cognitive workload was associated with lower model-predicted performance), surgeons are more likely to trust and integrate them into surgical practice.

This study has several limitations. Most notably, it involved a single participant, which limits the generalizability of the findings. While the results offer initial insights, they cannot be extended to broader surgeon populations without replication in larger and more diverse samples. In particular, ML models trained on a single participant data may overfit to individual physiological or behavioral traits, limiting their applicability to other surgeons. Physiological indicators such as HR and EMG activation patterns can vary with fitness level, posture, and coping strategies. Additionally, subjective workload ratings like NASA-TLX may be influenced by individual interpretation bias. To minimize interference during task execution, we used a simplified EEG montage (Fz, Cz, Pz) and estimated HRV using PPG-derived HR instead of ECG. While both choices improved experimental feasibility, they limit the spatial resolution of EEG data and the robustness of autonomic stress measurements. Additionally, as SHAP values reflect the internal logic of a specific model, their interpretability is limited by model quality and data representativeness. Finally, our study involved a controlled environment, whereas real surgeries present dynamic stressors, such as time pressure, communication errors, or patient instability, which were not fully replicated.

Future research should compare stress-response patterns across surgeons with different experience levels, as key physiological indicators may vary (e.g., experts relying more on HRV, while novices exhibit elevated EDA associated with anxiety). Sensor design should also be optimized to minimize interference, exploring alternatives such as dry EEG electrodes integrated into surgical caps or other wearable devices. In addition, to obtain more precise autonomic stress analysis, future studies should use ECG rather than PPG, as ECG provides more reliable HRV metrics. Reducing the number of sensors to those with the highest predictive value, as identified in our SHAP analysis, could enhance the feasibility of real-time stress monitoring tools in surgical practice.