Machine Learning Detection of Peritoneal Entry from Vibroacoustic Signals During Veress Needle Insertion

Background/Objectives: Establishing access is a hazardous step of laparoscopic surgery, and the widely used Veress needle offers only subjective tactile and auditory cues for peritoneal entry. Proximal vibroacoustic (VA) sensing captures needle-tissue interaction non-invasively and is known to contain a characteristic peritoneal puncture signature, which could help surgeons to precisely identify peritoneal entry. However, the VA signature has been evaluated only by retrospective, video-anchored and manual signal analysis. We investigate whether peritoneal entry can instead be detected automatically. Methods: Cavity entry is framed as a binary classification of VA signal segments, and four model families are benchmarked under a single unified protocol: a one-dimensional convolutional neural network (1D CNN), a multilayer perceptron, a bidirectional long short-term memory network (LSTM), and a support vector machine (SVM) with a radial basis function kernel. All models were trained, selected, and threshold-tuned on a synthetic abdominal phantom (996 segments) and tested once on an independent porcine cadaver dataset (154 cavity, 462 non-cavity signals) so that every test result measures generalisation across a phantom-to-porcine domain shift. Results: The 1D CNN performed the best (accuracy 0.869, AUROC 0.904, macro F1 0.833, cavity recall 0.818), closely followed by the SVM (AUROC 0.900, AUPRC 0.878). Conclusions: These results establish the feasibility of automatically detecting peritoneal entry across a phantom-to-porcine domain shift and extend the descriptive cadaveric characterisation of the puncture signature toward the automated machine learning-based detection of peritoneal entry in laparoscopic access with a Veress needle.