2026-09-30
On 98 Hinotori prostatectomies, a 25-feature Random Forest scored AUC 0.89 by case but 0.57 with surgeons held out; left-hand peak density alone still reached 0.87.
Robotic surgery is no longer a single-vendor market. Next to da Vinci sit Japan's Hinotori, China's Micro Hand S, and Europe's Versius and Hugo. Each vendor's proprietary event log can score skill, and none of those scores transfer. Every platform still has to log 3D end-effector positions to close the control loop. The Kobe group asked whether those coordinates alone can separate experts from non-experts and line up with clinical outcomes.
If they can, skill assessment does not have to live inside one company's API.
Consecutive robot-assisted radical prostatectomies on Hinotori at Kobe University Hospital, December 2020 to November 2023. Of 132 cases, 98 had complete kinematic logs during vesicourethral anastomosis; 34 dropped out mainly to early network logging failures. Sixteen right-handed surgeons, split by Japanese urologic-robotics Proctor certification: 68 proctor cases, 30 non-proctor. Age, BMI, PSA, prostate volume, Gleason, and stage did not differ between groups.
The anastomosis window ran from first needle grasp to final knot. Hand-controller coordinates were sampled at 1 Hz and turned into 25 features in five buckets: smoothness (dimensionless jerk), speed statistics, movement pattern (dwell time, peak density), workspace volume, and bimanual coordination. The classifier was a balanced Random Forest in scikit-learn. The test that matters is leave-one-surgeon-out: every case from one person held out, because expertise belongs to the surgeon, not the case.
Twenty case-wise 80/20 splits gave AUC 0.892 ± 0.075, five-fold 0.882, average precision 0.940. Hold out whole surgeons and the same 25-feature forest falls to 0.571 against a case-wise 0.888. The high-dimensional model is recognizing people.
The transferable signal sits in a handful of features. Importance: left peak density 11.9%, right speed coefficient of variation 11.1%, anastomosis time 9.0%. Those three alone: leave-one-surgeon-out AUC 0.788 (forest) and 0.844 (logistic regression); surgeon-grouped cross-validation 0.877 and 0.883. Left peak density by itself reaches 0.869. After residualizing for task duration it still separates unseen surgeons (AUC 0.771); adding duration back lifts 0.869 to 0.885.
| metric | Proctor | non-Proctor | Cohen's d |
| left peak density (peaks/s) | 0.087±0.005 | 0.079±0.005 | 1.618 |
| left dimensionless jerk | 4.3×10⁵ | 7.8×10⁵ | 1.314 |
| anastomosis time (min) | 45.5±16.1 | 67.8±18.8 | 1.313 |
| right speed CV | 1.02±0.08 | 1.11±0.09 | 1.041 |
Experts ran 10.1% more left-hand speed peaks, about 44% smoother motion, and 32.9% shorter anastomosis. Fifteen of 25 features survived FDR. Console time 227.6 vs 281.4 min (p=0.008); blood loss 150.6 vs 203.0 mL, not significant. Pad-free rates 46.2% vs 23.3% at 3 months (p=0.058), essentially tied at 6 months. Anastomotic leak n=5: large effect sizes, unusable sample.
Two certified experts adapted differently to Hinotori. Surgeon A (18 cases) cut anastomosis time 25.7% and left workspace 50.8%. Surgeon B (14 cases) was stable from case one. Experts do not share one transfer curve.
For anyone building intraoperative scoring or training systems, the warning is about validation. Case-level splits leak surgeon identity. Half the reported accuracy of the 25-feature forest is person recognition. What travels with the person is a few readable motor stats: how often the non-dominant hand peaks, how stable the dominant-hand speed is, how long the task takes.
Hinotori offers coordinates, not da Vinci's 40-plus automated performance metrics (camera, clutch, energy). Hung and colleagues separated experts with those proprietary events; this paper says position is enough, provided the whole feature bag is not dumped into a model that will memorize the operator. The suggested cutoffs (left peak density above 0.082, dimensionless jerk below 5.5×10⁵, speed CV below 1.10) are labeled hypothesis-generating. They are not a credentialing line.
This is measurement work, not a new surgical model. Random Forest plus 25 hand-built features is not a methods contribution. The contribution is the validation design that caught the identity leak.
Single suturing phase, one hospital, one Hinotori. No dissection, energy devices, or intraoperative judgment. 1 Hz is far below the 30-60 Hz needed for third-order derivatives, so jerk is a coarse proxy; the main claims rest on peak density, task time, and speed variability. Workspace volume is a bounding box and outlier-sensitive: a 5th-to-95th percentile estimate dropped the left-hand group gap from d=0.57 to 0.29 and lost significance. Expertise is Proctor certification, not blinded GEARS. All 16 surgeons are right-handed. Leak analysis n=5. Code is promised at publication; raw trajectories can identify surgeons and stay on request.
Cross-platform transfer is an extrapolation. Nothing here was rerun on da Vinci or Versius. "Any system with coordinates can extract these features" is a design claim, not a result.