2025 Proffered Presentations
S174: IDENTIFYING SURGICAL PHENOTYPES OF A MASTOIDECTOMY USING UNSUPERVISED CLUSTERING
Oren Wei, BS1; Andy S Ding, MD, MSE2; Nimesh Nagururu, MSE1; Adnan Munawar, PhD3; Manish Sahu, PhD3; Russell H Taylor, PhD3; Francis X Creighton, MD2; 1Johns Hopkins School of Medicine; 2Department of Otolaryngology - Head and Neck Surgery, Johns Hopkins; 3Department of Computer Science, Johns Hopkins University
Background: Temporal bone surgery requires a three-dimensional understanding of complex anatomy and their relationships to surrounding structures. Given this complexity, surgical trainees must learn how to modulate the angle, force, and direction of the surgical drill during temporal bone procedures to facilitate safe and efficient bone removal. While many descriptive surgical guides and textbooks exist to aid in instructing proper drilling techniques, few studies have objectively studied drilling techniques to quantitatively understand drill stroke patterns during key parts of temporal bone procedures. The goal of this study is to define and describe “phenotypes” of strokes during mastoidectomies using measurable instrument stroke characteristics. These phenotypes not only aid in describing drilling patterns for surgical training but also provide crucial kinematic information for future robotic and computer-assistance systems that are being developed for temporal bone surgery.
Methods: Forty mastoidectomies were performed by four operators (two attendings and two neurotology fellows) on the Fully Immersive Virtual Reality System (FIVRS) temporal bone surgical simulator, a virtual surgical environment with 3D haptic training modules. Drill position was tracked for each trial, and drill strokes were recorded by the simulator. Stroke metrics, such as stroke length, velocity, acceleration, jerk, and voxels removed, were extracted for analysis. Multi-dimensional unsupervised clustering of instrument strokes across these metrics was performed using k-means with k = 5 clusters as identified using the elbow method. Each cluster was defined as a distinct stroke phenotype. Stroke phenotypes were then analyzed for prevalence at various timepoints throughout the surgery and characterized by the average magnitude of their stroke metric components.
Results: Five stroke phenotypes were identified after k-means clustering of stroke characteristics (Figure 1).
Figure 1. Average Magnitude of Stroke Metric Components by Phenotype
By analyzing stroke length, velocity, acceleration, jerk, and voxels removed for each stroke phenotype, distinct differences in phenotypes were identified (Table 1).
Table 1. Qualitative Descriptions of Instrument Stroke Phenotypes
Tracking the prevalence of each phenotype with respect to time (Figure 2), we see a higher proportion of high velocity strokes belonging to phenotypes 1 and 3 near the beginning of the procedure and a higher proportion of short, slow strokes belonging to phenotype 0 near the end of the procedure. Phenotype 4, which corresponds to slow, moderately long strokes, were more prevalent in the middle quartiles of the procedure.
Figure 2. Distribution of Stroke Phenotypes for each Quartile of Procedure Time
.jpg)
Conclusion: We classified all instrument strokes performed during simulated mastoidectomies into five phenotypes and have distinguished these phenotypes via quantifiable stroke characteristics. This analysis technique lays an important foundation for precisely describing surgical technique for trainees. These phenotypes also provide a basis for future efforts to automate surgical tasks by translating data into a machine-interpretable form. Future work will focus on analyzing strokes with respect to nearby critical anatomy and associating stroke phenotypes with surgical phases rather than time quartiles.