Can fundamental kinematic indicators predict surgical expertise in robot-assisted surgery?

Hand-controller coordinates during one anastomosis step

The study is a retrospective observational analysis of consecutive robot-assisted radical prostatectomies performed with the Hinotori Surgical Robot System at Kobe University Hospital between December 2020 and November 2023. Of 132 consecutive procedures, 98 (74.2%) had complete kinematic logs during vesicourethral anastomosis; the other 34 were excluded because of technical logging failures, mainly network connectivity problems during early adoption.

The kinematic window is narrow and specific: it runs from the first needle grasp to final knot completion of the vesicourethral anastomosis. The system recorded bilateral hand-controller coordinates (x, y, z in millimetres) at 1 Hz. Missing data under 0.1% were filled bidirectionally, trajectories shorter than 100 samples (100 s) were excluded, and left and right trajectories were truncated to equal length.

Expertise was defined by credential, not by outcome. Surgeons were labelled Proctor (expert) or Non-Proctor according to formal Japanese Proctor certification for robot-assisted urologic surgery, which requires board certification, society membership, at least 40 cases as primary surgeon in radical prostatectomy and/or cystectomy, at least one publication or conference presentation on robot-assisted surgery, and completion of an official Proctor education seminar. Certified Proctors performed 68 procedures (69.4%) and non-Proctors 30 (30.6%); all 16 surgeons were right-handed. Patient demographics did not differ significantly between the two groups, which the authors use to argue that outcome differences reflect technique rather than case selection.

Twenty-five features built from position data alone

The feature pipeline converts raw coordinates into 25 kinematic parameters across five domains, all derived solely from 3D position, which is why the authors describe them as theoretically extractable from any robotic system with position tracking. Movement smoothness is captured by dimensionless jerk and mean jerk. Speed characteristics include mean velocity, coefficient of variation, and percentile distributions. Movement patterns include dwell-time percentage (the share of samples with speed below 2 mm/s) and peak density, the number of peaks in the speed profile per unit time. Workspace utilization uses a three-dimensional bounding-box volume and range of motion. Path efficiency is the ratio of straight-line distance between start and end to actual path length. Bimanual coordination covers interhand speed correlation, mean hand distance, motion synchrony over 50-sample windows, and bilateral dwell synchrony.

Two of these definitions matter for reading the results. Peak density is a frequency count of local maxima in the hand-speed profile, measured in peaks per second — not a smoothness score and not a direct measure of surgical skill. Dwell time is the proportion of samples below 2 mm/s, so a lower value means less time spent nearly stationary. The paper is explicit that higher peak density should not be read as skilful rhythm in isolation, because repetitive velocity peaks could also come from mechanical or inefficient motion.

Experts moved smoother, faster, and with denser peaks

Proctor surgeons showed a left peak density of 0.087 ± 0.005 peaks/s versus 0.079 ± 0.005 in non-Proctors (Cohen's d = 1.618, p < 0.001), a 10.1% higher frequency of velocity peaks and the largest effect size in the study. Dimensionless jerk was lower for Proctors in both hands — (4.3 ± 2.3) × 10^5 versus (7.8 ± 3.3) × 10^5 on the left and (4.2 ± 2.2) × 10^5 versus (7.3 ± 3.0) × 10^5 on the right — which the authors describe as 44% smoother movement. Anastomosis time was 45.5 ± 16.1 min for Proctors versus 67.8 ± 18.8 min for non-Proctors (d = 1.313, p < 0.001), a 32.9% reduction.

Fifteen of the 25 features (60%) remained statistically significant after false discovery rate correction. The left and right hands did not contribute equally: left-hand features discriminated better on movement-pattern metrics and path length, while right-hand features discriminated better on speed metrics and dwell time. The authors interpret the left-hand dominance as consistent with the non-dominant hand's role in holding tissue tension and controlling the needle during anastomosis — an interpretation, not a measured mechanism.

One feature needs a caution attached to it. Workspace volume is a bounding-box measure and is sensitive to spatial outliers; under a 5th-to-95th percentile robust estimate, the left-hand between-group difference attenuated and lost significance (Cohen's d falling from 0.57 to 0.29) and the right-hand difference shrank (d from 0.65 to 0.52). Workspace findings should therefore be read as less stable than the peak-density and jerk results.

Case-wise accuracy hides a surgeon-level weakness

The Random Forest classifier reached a mean AUC of 0.892 ± 0.075 across 20 independent train-test splits, 0.882 under five-fold cross-validation, and an average precision of 0.940 despite the 68-versus-30 class imbalance. Feature importance ranked left peak density highest at 11.9%, followed by right speed coefficient of variation at 11.1% and anastomosis time at 9.0%.

Those numbers are case-wise: each procedure is a data point, and the same surgeon appears in both training and test folds. Because expertise is a property of a surgeon rather than a case, the authors also ran leave-one-surgeon-out and surgeon-grouped validation. Here the picture changes sharply. The full 25-feature Random Forest fell to a leave-one-surgeon-out AUC of 0.571, against a case-wise estimate of 0.888 — evidence that much of its case-wise performance reflected surgeon-specific patterns rather than transferable skill signal.

Restricting the model to the three highest-ranked features — left peak density, right speed coefficient of variation, and anastomosis time — produced leave-one-surgeon-out AUCs of 0.788 for Random Forest and 0.844 for logistic regression, with surgeon-grouped AUCs of 0.877 and 0.883. Left peak density modelled alone reached 0.869. The authors argue that a single feature cannot encode multivariable surgeon-specific patterns, making this the most leakage-resistant evidence that the expertise signal generalizes across surgeons. Sensitivity analysis supported stability, with AUCs from 0.86 to 0.91 across surgeon inclusion/exclusion scenarios, including 0.86 when the highest-volume non-Proctor surgeon (16 of 30 non-Proctor cases) was excluded.

Peak density is not simply a proxy for speed

A natural objection is that peak density just measures how fast a surgeon works. The paper tests this directly. Peak density correlated with task time (Spearman rho = −0.53), so the two overlap, but peak density was not reducible to task time: after adjustment for task time it retained discrimination across unseen surgeons, with a residualized leave-one-surgeon-out AUC of 0.771 and a partial correlation with expertise of r = 0.45 (p < 0.001). Combining peak density with task time improved unseen-surgeon discrimination from 0.869 to 0.885.

The authors' reading is that repeated movement structure and efficiency should be interpreted together rather than as rhythm alone. They speculate that the combination of lower dwell time and more consistent speed in experts may reflect what expertise research calls a 'flow state', but they explicitly label this conjecture requiring dedicated psychological assessment. Coordinate-only data also cannot segment the individual gestures — needle driving, suture handling, hand repositioning — that the velocity peaks plausibly correspond to.

Clinical outcome links are weak and exploratory

Partial correlation analysis controlling for anastomosis time found two time-independent features associated with estimated blood loss: left movement sharpness (r = 0.202, p = 0.046) and left workspace volume (r = 0.204, p = 0.044). These are small correlations, and the workspace measure is the same one that weakened under robust estimation. An exploratory analysis of the five patients with anastomotic leaks reported larger effect sizes for movement-related features than in patients without leaks, but with n = 5 this cannot support a predictive claim.

Group-level clinical differences were mixed. Proctors had shorter console times (227.6 ± 78.4 versus 281.4 ± 76.9 min, p = 0.008) with comparable blood loss (150.6 ± 187.3 versus 203.0 ± 305.1 mL, p = 0.975). Continence recovery at 3 months was 46.2% for Proctors versus 23.3% for non-Proctors (p = 0.058), a difference that did not reach conventional significance and that diminished by 6 months (58.5% versus 53.3%, p = 0.805). The abstract's statement that kinematic patterns were associated with clinical outcomes should therefore be read as covering these modest, partly exploratory associations, not as evidence that kinematics predict patient outcomes.

Two experts, two adaptation phenotypes

The learning-curve analysis is the most concrete demonstration that expertise is not a single trajectory. Two certified Proctors transitioning onto the Hinotori platform behaved differently. Surgeon A (n = 18) improved progressively: a 25.7% reduction in anastomosis time, an 18.5% increase in speed correlation, and a 50.8% decrease in left workspace volume. Surgeon B (n = 14) maintained consistent performance metrics from the first case.

The largest divergence was in bimanual coordination. Surgeon A improved bilateral dwell time by 8.7% and speed correlation by 18.5%, while Surgeon B showed a 4.8% increase in bilateral dwell time and a 7.4% decrease in speed correlation. With only two surgeons, these are described as distinct phenotypes with implications for personalized training — a hypothesis-generating observation, not a validated classification of how experts adapt.

What these indicators can and cannot support

The paper's own framing is that coordinate-derived indicators are complementary to, not replacements for, expert video-based tools such as GEARS, which scores depth perception, bimanual dexterity, efficiency, force sensitivity, autonomy, and robotic control through expert review. The authors state they do not claim GEARS-level convergent validity for this retrospective coordinate-only dataset, and they call for prospective studies with blinded multi-rater GEARS assessment and inter-rater reliability analysis.

Scope limits are equally specific. The data come from one institution, one robotic platform, one procedure type, and the vesicourethral anastomosis step only, in 16 right-handed surgeons; the authors note that long-term oncological outcomes and mixed-handedness cohorts require prospective external validation before any credentialing or training use. Raw coordinate availability also varies across platforms, so cross-platform transfer remains untested.

The practical priority that follows is a comparison, not a deployment. If the goal is a model that works on a surgeon it has never seen, the evidence favours a small, interpretable set — left peak density, right speed coefficient of variation, and anastomosis time, or left peak density alone — over the full 25-feature model, whose case-wise AUC of 0.892 collapses to 0.571 on unseen surgeons. The paper's most useful contribution is that distinction: high within-dataset accuracy is not the same claim as generalizable skill assessment, and only the parsimonious indicators survived the surgeon-level test.

Sources used in this answer