Comparative Vibroacoustic Analysis of Spinal Needle Insertion Actions: Quincke vs. Sprotte

Abstract

Tactile feedback during needle placement is often the only guiding sense, yet it remains subjective and strongly operator-dependent. We investigated whether friction-induced vibrations captured proximally at the needle hub can provide complementary information by distinguishing two spinal needle designs, Quincke (cutting bevel tip) and Sprotte (noncutting pencil-point tip). Audio signals were recorded at 48 kHz using a micro-electro-mechanical systems (MEMS) microphone during manual insertions into dry foam. A dataset of 207 recordings (103 Sprotte, 104 Quincke), acquired across two sessions, was segmented into four sequential actions, yielding 828 segments of 300 ms each. From each segment, 372 acoustic features were extracted, including Mel-frequency cepstral coefficients (MFCCs) with temporal derivatives, Melspectrogram statistics, spectral descriptors, and short-time Fourier transform (STFT) features. We evaluated classical machine learning models and convolutional neural networks (CNNs) under a strict recording-level split to avoid data leakage. The best needle classification achieved 95.2% accuracy on held-out data and 93.3% ± 2.1% in grouped crossvalidation. The top CNN reached 86.9%. Four-class action recognition achieved 90.5%, and per-action Quincke-vs.- Sprotte classification ranged from 88.1% to 95.2%. These findings suggest that proximal vibroacoustic sensing can capture needle-dependent interaction signatures even when cleaner and noisier recordings are analysed together, supporting its potential as a complementary feedback modality in needle procedures.