Defining surgical strokes in lumbar laminectomy: toward objective skill assessment and robotic integration.

PURPOSE: To characterize discrete drilling stroke phenotypes during lumbar laminectomy using quantitative motion features, and to assess how phenotype usage varies by procedural stage, training level, and navigation assistance. METHODS: Eleven orthopedic trainees each performed 10 lumbar laminectomies (110 total) on a CT-based VR simulator (CAPTAiN), split between CAPTAiN-assisted and non-navigated conditions. Burr strokes were quantified by length, velocity, acceleration, jerk, bone removed, and scaled proximity to the dura. Stroke feature vectors were embedded using UMAP and clustered via k-means (k = 3) into phenotypes. Phenotype distributions were compared across PGY level, procedural time quartiles (Q1-Q4), and navigation status using Bonferroni-corrected chi-squared tests (p < 0.05). RESULTS: Clustering identified three stroke phenotypes: Exploratory (long, fast, high jerk, near dura), Debulking (moderate length/speed, greatest bone removal, farthest from dura), and Refinement (short, controlled, minimal bone removal). Analysis of procedural time quartiles showed a shift from Exploratory strokes in Q1 to Debulking dominance in Q2-Q4, with Refinement stable across time. Exploratory strokes decreased from 53.1% at 1 year to 19.3% at 4 years (p < 1.17 × 10- 8), while Debulking rose from 27.0% to > 51% (p < 2.74 × 10- 5), and Refinement increased from 19.7% to 37.3% (p = 0.0052). Navigation reduced Exploratory strokes (28.9% vs. 39.2%, p = 0.0046) and increased Debulking (51.6% vs. 45.3%) and Refinement (20.1% vs. 16.4%, p = 0.032). CONCLUSIONS: VR-based motion analysis identifies three drilling stroke phenotypes for lumbar laminectomy that can guide surgical skill assessment, trainee feedback, and future applications in automated robotic surgery and simulation-based training.