Acoustic Cues into a Surgeon-assist Physical AI for Detecting Bone Penetration During Spinal Surgery

Acoustic Cues into a Surgeon-assist Physical AI for Detecting Bone Penetration During Spinal Surgery | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Acoustic Cues into a Surgeon-assist Physical AI for Detecting Bone Penetration During Spinal Surgery Hideaki Fujiwara, Takahito Fujimori, Yuya Kanie, Masayuki Furuya, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8680502/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 19 Apr, 2026 Read the published version in Scientific Reports → Version 1 posted 14 You are reading this latest preprint version Abstract Purpose Skilled surgeons can detect subtle changes in bone-cutting sounds to recognize bone penetration, but acoustic cues are subjective and dependent on surgical experience. The objective of this study was to develop an artificial intelligence (AI) model capable of detecting bone penetration from intraoperative percussion sounds. Methods A total of 1,236 chisel strikes were identified from intraoperative recordings obtained during lumbar and thoracic spinal decompression surgeries performed using a chisel, and were labeled as penetration or non-penetration. Acoustic features were extracted per strike and expanded across 3-hit sliding windows to capture dynamic temporal changes. A gradient boosting machine learning classifier (LightGBM) was trained, with 10% of data as an independent test set. Model performance was primarily evaluated using the area under the precision–recall curve (PR-AUC) and receiver operating characteristic AUC (ROC-AUC). Results On the independent test set, the model yielded a ROC-AUC of 0.838 and a PR-AUC of 0.604. In a sensitivity analysis, the model achieved a sensitivity of 0.828, specificity of 0.767, accuracy 0.784, F1-score 0.686, and precision 0.585. Feature importance analysis revealed that dynamic changes between consecutive strikes, such as mel-frequency cepstral coefficients (MFCCs), zero-crossing rate, and spectral contrast, were the most influential predictors. Conclusion An AI model analyzing intraoperative percussion sounds demonstrated a robust ability to detect bone penetration, with consistent ROC performance and clinically acceptable PR-AUC. Quantifying acoustic cues traditionally interpreted subjectively by surgeons may support intraoperative decision-making and surgical training. Physical sciences/Engineering Health sciences/Health care Health sciences/Medical research acoustic analysis bone penetration spinal decompression machine learning Physical AI surgeon-assist intraoperative sound diagnostic accuracy gradient boosting precision–recall analysis surgical skill quantification Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction Sound is an important source of information, especially when visual inspection is not possible. Auscultation is a representative example, where physicians assess internal organ conditions by listening to breath and heart sounds. In orthopedics as well, it is known that before the invention of X-rays, fractures were diagnosed by percussion sounds.(1,2) Laminectomy using chisel is one of the fundamental procedures in spinal decompression surgery.(3) During percussion with a chisel, the surgeon integrates multiple sensory cues—such as resistance of the chisel, speed, and progression—to determine whether the bone has been cut.(4) One of the key pieces of information for this judgment is the change in the percussion sound. This principle is similar to that observed in arthroplasty, where the sound changes as the prosthesis becomes firmly seated in the joint.(1,5,6) Typically, perceiving these sound changes requires surgical experience; however, recent studies have begun to explore the possibility of using artificial intelligence (AI) to replicate human ability.(7–10) In industry, acoustic AI has been applied to detecting delamination or cracks in concrete and tiles and to identifying machine tool faults.(11) In medicine, similar approaches have been used to analyze respiratory and heart sounds for abnormality detection.(12,13) This paradigm—learning and interpreting physical signals such as sound and vibration using sensors and AI—is known as Physical AI and is an emerging research field.(14,15) We aimed to extend this concept to the surgical domain, envisioning the development of an AI system that can support surgeons— Surgeon-assist Physical AI (SPAI) . The objective of this study was to construct an AI model that can determine whether bone has been cut based on percussion sounds during laminectomy and to evaluate its performance. Methods This study was approved by our institution's ethics committee. This was a retrospective diagnostic accuracy study conducted in accordance with the Standards for Reporting Studies of Diagnostic accuracy for artificial intelligence (STARD-AI) guidelines.(16) Eligibility criteria were consecutive patients among those for whom intraoperative audio–video recordings were successfully obtained, who underwent lumbar or thoracic spine surgery with chisel-based bone resection at two spine centers between July 3, 2022, and September 12, 2024. (Fig. 1 ). All surgeries were performed by three experienced spine surgeons (with 21, 19, and 14 years of experience, respectively). Percussion strikes with the chisel were recorded intraoperatively using an iPhone or iPad at a sampling frequency of 44.1 kHz. The chisels were made of stainless steel with resin handles (Fig. 2 and Supplementary video_1). In each percussion sequence, the chisel was tapped multiple times to gradually advance through the bone. When the chisel penetrated the cortical bone, the sound of the strike changed. In some procedures, surgeons avoided full cortical penetration by leaving a thin cortical layer and completing bone removal using levering or prying maneuvers. Because these techniques do not involve percussion-based cortical breakthrough or its characteristic acoustic change, such cases were excluded a priori from acoustic analysis. Each strike was labeled as penetration (p) or non-penetration (np) (Fig. 3 ). Because no objective physical gold standard exists for identifying the exact moment of bone penetration, surgeon consensus review of the complete operative video served as the reference standard. Labeling was based on visual confirmation of cortical breakthrough and the transition to bone removal and/or the next surgical step, with acoustic changes considered only as supportive context. Initial labeling was performed by a spine surgeon with 6 years of experience and subsequently reviewed and confirmed by a senior spine surgeon with 21 years of experience, with final labels determined by consensus. Acoustic feature extraction For each individual strike, 30 base acoustic features were extracted, including measures of energy, frequency distribution, mel-frequency cepstral coefficients (MFCC), and spectral patterns (Table 1 ). To better capture relative changes and scaling effects across consecutive strikes, additional derived features were computed, including differences, ratios, and log-transformed differences between strikes (Supplementary 1). Table 1 Univariate analysis of 30 basic acoustic features. Feature Non-penetration Penetration p-value MFCC_2 87.6 ± 34.8 68.4 ± 34.5 < 0.001 Mean Zero Crossing Rate 0.055 ± 0.02 0.051 ± 0.03 < 0.001 Mean Spectral Centroid (Hz) 5342 ± 1381 5870 ± 1395 < 0.001 MFCC_3 -53.1 ± 21.8 -61.1 ± 24.1 < 0.001 Contrast_7 (dB) 28.8 ± 4.5 30.3 ± 3.8 < 0.001 MFCC_6 17.1 ± 15 21.7 ± 13.3 < 0.001 Mean Amplitude 0.12 ± 0.05 0.11 ± 0.05 < 0.001 Mean Spectral Roll-off (Hz, 85%) 10099 ± 2825 10805 ± 2810 0.002 Contrast_1 (dB) 10.9 ± 4.4 9.9 ± 3.9 0.002 Mean Spectral Flatness 0.03 ± 0.03 0.02 ± 0.04 0.006 MFCC_1 -71.6 ± 35.2 -77.5 ± 36.3 0.007 MFCC_12 -4.7 ± 14.3 -2.4 ± 13.4 0.035 MFCC_8 22.1 ± 12.6 24.3 ± 13.3 0.059 Energy 25.1 ± 20.5 22.8 ± 17.1 0.084 MFCC_13 -11.2 ± 11.9 -10 ± 11.3 0.12 Contrast_4 (dB) 11 ± 3.3 11.2 ± 2.9 0.15 Duration (s) 0.02 ± 0.1 0.02 ± 0.01 0.18 Contrast_2 (dB) 4.6 ± 2.7 4.8 ± 2.9 0.20 MFCC_7 -14.8 ± 13.9 -13.7 ± 13.7 0.23 Mean RMS 0.1 ± 0.03 0.1 ± 0.04 0.24 Mean Spectral Bandwidth (Hz) 4295 ± 809 4360 ± 823 0.27 MFCC_5 -20.2 ± 17.8 -18.8 ± 16.5 0.39 MFCC_11 -7.1 ± 11.3 -8 ± 11.2 0.43 MFCC_9 -16.2 ± 12.3 -17.2 ± 11 0.44 Contrast_5 (dB) 12.4 ± 2.8 12.6 ± 2.6 0.48 Peaks Detected 1 ± 0.4 1 ± 0 0.65 Contrast_3 (dB) 8.4 ± 3.2 8.5 ± 3.4 0.79 MFCC_10 11.5 ± 11.8 11.5 ± 12 0.79 MFCC_4 32.1 ± 20.5 32 ± 21.1 0.81 Contrast_6 (dB) 14.8 ± 2.3 14.8 ± 2.3 0.89 The rows are sorted in descending order of p-values from top to bottom. MFCC; mel-frequency cepstral coefficients. A sliding window of three consecutive strikes (3-hit window, stride = 1) was then constructed (Fig. 4 ). For each 3-hit window, acoustic features were expanded to capture both the properties of individual strikes and the changes that occur between consecutive strikes. This process yielded a total of 570 features per window, enabling the model to learn not only how each strike sounds in isolation but also how the acoustic pattern evolves as the chisel approaches bone penetration. AI model development A machine learning classifier was developed using LightGBM, trained on an expanded feature set comprising 570 features per 3-hit window. Data splitting was performed at the recording level to prevent information leakage between training and test sets. Model selection and hyperparameter optimization Model Selection and hyperparameter optimization was conducted using 5-fold cross-validation on the training set, with grouping by recording identifier to ensure that all windows from a given recording session appeared exclusively in either the training or validation fold. The primary evaluation metric was the area under the precision–recall curve (PR-AUC), selected for its superior sensitivity to model performance on the minority (positive) class in imbalanced datasets. Model performance was summarized as mean ± standard deviation across cross-validation folds. Secondary metrics included the area under the receiver operating characteristic curve (ROC-AUC) and threshold-dependent measures (accuracy, F1-score, precision, recall) computed at an optimized threshold that achieves sensitivity (recall) ≥ 80%. Early stopping (patience = 100 rounds) based on validation set PR-AUC was applied during cross-validation to identify the optimal number of boosting iterations and prevent overfitting. Operating threshold determination The operating threshold for binary classification was determined using a clinically-driven approach prioritizing sensitivity. Given the importance of minimizing false negatives (missed bone breakthrough events) to prevent inadvertent injury to underlying structures, we targeted a minimum sensitivity of 80% based on clinical safety requirements. Using out-of-fold (OOF) predictions aggregated across all cross-validation folds, we analyzed the precision–recall curve and selected the highest probability threshold that achieved sensitivity (recall) ≥ 80%. Model training and evaluation The final model was trained on 154 recordings using optimal hyperparameters identified during cross-validation, with 18 recordings reserved for internal validation. Training iterations were fixed to the mean optimal count across folds to ensure consistency with cross-validation results. To maintain consistency with cross-validation results, the model was trained for a fixed number of iterations equal to the mean optimal iteration count across folds. The internal validation set was used for monitoring purposes only and did not influence model training process or stopping criteria. The final model was evaluated on the independent hold-out test set (20 recordings, 102 windows) using both threshold-independent metrics (PR-AUC, ROC-AUC) and threshold-dependent classification performance at the predetermined operating threshold. Model performance was reported separately for cross-validation results reflecting training set generalization, independent test set performance at the clinically-optimized operating threshold, and threshold sensitivity analysis showing performance trade-offs across multiple operating points. Statistical analysis Univariate comparisons of individual acoustic features between penetration (p) and non-penetration (np) strikes were performed using the Mann–Whitney U test, given the non-normal distribution of the data. A significance threshold of p < 0.05 (two-sided) was applied. All analyses were conducted using Python 3.10.14 with LightGBM 4.5.0, scikit-learn 1.5.2, pandas 2.2.2, and NumPy 2.1.3 (Supplementary 2). Results Demographics A total of 37 patients were included (mean age 61 ± 26 years, 25 with lumbar spinal stenosis and 12 with spinal deformity). From these patients, 739 surgical video clips were obtained. Among them, 192 audio records containing penetration sounds were extracted (Fig. 1 ). Each audio record contained an average of 6 ± 2 strikes (range 3–15), resulting in a total of 1236 labeled strikes: 215 penetration and 1021 non-penetration. Single-strike features Comparison of acoustic features between penetration (n = 215) and non-penetration (n = 1021) strikes revealed significant differences in multiple spectral characteristics, including MFCC_2, MFCC_3, MFCC_6, MFCC_1, and MFCC_12, mean zero-crossing rate, mean spectral centroid, spectral roll-off, spectral flatness, and spectral contrast (Table 1 ). Most of these features represent tonal brightness or sharpness of the sound. Three-strike windows From the 1236 strikes, 852 three-strike windows were generated (215 penetration windows, 637 non-penetration windows; 25% vs. 75%) (Fig. 4 ). Among these, 20 audio records containing 102 windows (29 penetration windows, 73 non-penetration windows; 28% vs. 72%) were reserved as an independent test set. The remaining 172 audio records (750 windows: 186 penetration windows, 564 non-penetration windows; 25% vs. 75%) was further split into an internal training subset (154 recordings) and an internal validation subset (18 recordings). Model performance Using the optimal hyperparameter configuration, the model demonstrated stable performance across 5-fold cross-validation, achieving a mean PR-AUC of 0.645 ± 0.071 and a mean ROC-AUC of 0.823 ± 0.049 (Table 2 ). Early stopping identified an average optimal training length of 106 iterations. Based on out-of-fold predictions, an operating threshold of 0.2305 was selected to achieve the prespecified target sensitivity of 0.80 on the training set, corresponding to an accuracy of 0.807, F1-score of 0.551, specificity of 0.635, and precision of 0.420. When evaluated on the independent hold-out test set, the model maintained robust discrimination, yielding a PR-AUC of 0.604 and a ROC-AUC of 0.838. Table 2 Model performance of LightGBM at an operating threshold (0.2305). ROC-AUC PR-AUC Sensitivity (Recall) Precision (Positive predictive value) Accuracy Specificity F1-score Threshold Cross validation (5-fold) 0.822 0.645 0.801 0.420 0.807 0.635 0.551 0.2305 ROC-AUC; receiver operating characteristic curve – area under the curve PR-AUC; precision–recall – area under the curve Threshold sensitivity analysis Using the pre-specified operating threshold derived from out-of-fold predictions (0.2305), test-set performance was sensitivity 0.655, specificity 0.877, and precision 0.679 (Table 3 ). To illustrate the operating trade-offs, we additionally report test-set performance across alternative probability thresholds (Table 4 ). In an exploratory post-hoc assessment, we identified that a threshold of 0.15 would yield sensitivity ≥ 0.80 on this specific test set; at this threshold, sensitivity was 0.828, specificity 0.767, accuracy 0.784, and precision 0.585. Table 3 Threshold sensitivity analysis of test set. Threshold Sensitivity (recall) Precision (Positive predictive value) Accuracy Specificity F1-Score 0.1 0.931 0.466 0.676 0.575 0.621 0.15* 0.828 0.585 0.784 0.767 0.686 0.20 0.724 0.656 0.814 0.849 0.688 0.2305† 0.655 0.679 0.814 0.877 0.667 0.25 0.621 0.692 0.814 0.890 0.655 0.30 0.586 0.680 0.804 0.890 0.628 *Selected operating threshold (target sensitivity ≥ 80%) † Training-derived threshold from out of fold (OOF) predictions Table 4 Confusion matrix on the independent test set at the threshold of 0.15. AI prediction Penetration Non-penetration Total Reference standard Penetration 24 (TP) 5 (FN) 29 Non-penetration 17 (FP) 56 (TN) 73 Total 41 61 102 TP; true positive, FN; false negative, FP; false positive, TN; true negative . Accuracy = 80/102 = 0.784, Recall (Sensitivity) = 24/29 = 0.828, Specificity = 56/73 = 0.767, Precision (Positive predictive value) = 24/41 = 0.585, F1-score = 0.686 Feature importance Feature importance analysis identified the most influential predictors as temporal dynamics of acoustic features across strikes (Table 5 and Supplementary 3). The top-ranked feature was MFCC_3__slope3, representing the slope of the 3rd mel-frequency cepstral coefficient across the three-strike window. This was followed by MFCC_2__slope3 and MFCC_2__range3. Other highly ranked features included Contrast_7 (dB)__t3, MFCC_3__r21, and mean Zero Crossing Rate____std3. Table 5 Top 10 feature importances of the model. Ranking Feature Importance 1 MFCC_3__slope3 1294 2 MFCC_2__slope3 501 3 MFCC_2__range3 310 4 Contrast_7 (dB)__t3 280 5 MFCC_3__r21 261 6 Mean Zero Crossing Rate__std3 176 7 MFCC_10__range3 146 8 MFCC_5__t1 144 9 Contrast_1 (dB)__median3 136 10 MFCC_3__d21 134 Discussion In this study, we developed and validated an AI model capable of detecting bone penetration during spinal decompression surgery by analyzing chisel percussion sounds. The model achieved a PR-AUC of 0.645 in cross-validation and 0.604 in an independent test set, with ROC-AUC values consistently above 0.80. Given that the prevalence of penetration events in our dataset was relatively low (≈ 17%), these PR-AUC values represent meaningful discriminative performance well above chance level, where a random classifier would be expected to achieve only ~ 0.17. These findings indicate that the approach generalizes reasonably well to unseen data despite the limited sample size.(17,18) Importantly, feature importance analysis revealed that the model primarily relied on changes, slopes, and ranges of acoustic features across sequential strikes, rather than on static single-strike values. This finding underscores the importance of capturing temporal dynamics when distinguishing penetration from non-penetration sounds. This suggests that the algorithm mimics the way surgeons perceive the brightening or sharpening of sound when penetration occurs, providing an interpretable link between acoustic dynamics and surgical decision-making. Novelty of the study To our knowledge, this is the first report to apply machine learning to analyze intraoperative percussion sounds in spinal surgery. Previous research in medicine has primarily focused on respiratory sounds,(12) cardiac auscultation,(13,19) or industrial quality control tasks,(11) but not on surgical sound analysis. While experienced surgeons have long relied on subtle acoustic cues during procedures such as laminectomy or arthroplasty, these perceptions have remained subjective and dependent on years of training (1,2,5,6). Our study demonstrates that these acoustic cues can be objectively captured, quantified, and classified using AI. By focusing on three-strike windows and emphasizing sound changes across beats, our approach goes beyond static sound analysis and more closely reflects the way surgeons perceive “patterns of change” rather than isolated signals. This methodological innovation may open a new research field—surgical acoustics—as part of the broader domain of Physical AI. Clinical relevance The clinical implications of this work are considerable. First, the ability to objectively recognize bone penetration can support less experienced surgeons during spinal decompression by providing real-time feedback, thereby helping them learn to identify subtle acoustic cues that are otherwise acquired only through prolonged practice. Such a system may also reduce inter-surgeon variability and improve safety, particularly in procedures requiring precise bony removal. For example, when a junior surgeon performs bone cutting, an experienced supervisor can recognize from sound alone when the bone has already been penetrated and intervene to prevent excessive chiseling. By making this auditory judgment explicit and accessible, the proposed AI model has the potential to replicate this supervisory role and enhance both safety and training. Our results demonstrate that different probability thresholds produce predictable trade-offs between sensitivity and specificity, suggesting that the model can be flexibly tuned according to clinical context. Second, as surgical robotics continues to advance,(15) integration of acoustic AI could provide robots with an additional sensory modality, complementing force or visual feedback and enabling safer autonomous or semi-autonomous bone cutting.(20,21) In this sense, our work represents an early step toward what we propose as Surgeon-assist Physical AI (SPAI)—an AI framework designed to emulate and augment the sensory judgments traditionally made by surgeons. Why the sound changes The acoustic change that occurs when the chisel penetrates bone can be explained by basic principles of material vibration.(6,22–24) When a chisel strikes cortical bone, the sound is generated by the vibration of both the chisel and the bone structure. Before penetration, the chisel is resisted by solid bone, producing a duller and more uniform tone. Once penetration occurs, the structural integrity of the bone at that point is lost, resulting in a sharper, higher-pitched sound.(2) This is analogous to the difference between striking a solid wall versus a hollow structure: the resonance shifts, and the resulting sound becomes brighter and more distinct. Acoustic features and their interpretation The model’s reliance on features such as mel-frequency cepstral coefficients (MFCCs), zero-crossing rate, and spectral contrast reflects the acoustic properties that change most prominently during penetration.(25,26) Unlike static single-strike features, the top predictors were predominantly measures of temporal dynamics—including slopes, ranges, and ratios across consecutive strikes—indicating that the model captured not only the absolute tonal quality of each strike but also how it evolved over time. MFCCs are widely used in speech and audio recognition to represent the timbre or tonal quality of a sound. In our analysis, slope- and range-based MFCC features (e.g., MFCC_3__slope3, MFCC_2__slope3, MFCC_2__range3) ranked among the most important predictors, suggesting that penetration was associated with systematic shifts in tonal patterns across the three-strike window. Taken together, these findings provide an objective explanation of the subjective impression surgeons describe—namely, that the sound becomes “higher” or “brighter” when the bone gives way. Importantly, the model emphasized how features change across sequential strikes, mirroring the clinical reality that surgeons detect penetration not by an isolated sound but by recognizing the evolving acoustic pattern across a series of hammer strikes. Limitations This study has some limitations. First, the dataset was relatively small and derived exclusively from procedures performed by experienced spine surgeons. Although the number of individual strikes was sufficient for feature extraction, the diversity of surgical contexts was limited. Second, the present approach relied on manually segmented individual strikes and the construction of fixed three-strike windows. While effective as a proof of concept, this design does not yet analyze the full continuous acoustic stream as it would occur intraoperatively. Conclusion In summary, this study demonstrates the feasibility of using machine learning to recognize bone penetration during spinal decompression by analyzing acoustic signals from chisel strikes. This approach could enhance surgical training, provide real-time decision support for less experienced surgeons, and ultimately serve as a foundation for multimodal feedback systems in robotic or computer-assisted surgery. Declarations Institution ethics committee approval: The study protocol was approved by the University of Osaka Hospital Ethical Review Board (protocol number: 22099), and informed consent was waived due to the retrospective nature of the study. The study was conducted in accordance with the Declaration of Helsinki. Declaration of competing interest: The authors declare that they have no conflict of interest. Acknowledgement: This work was supported by JSPS KAKENHI No. JP21K20966, NSK Nakanishi Foundation, and J&J Medical Research Grant. Fundings: This work was supported by JSPS KAKENHI No. JP21K20966, NSK Nakanishi Foundation, and J&J Medical Research Grant. Author contribution : H.F. and T.F. contributed to conception, design, and drafting of the manuscript; T.F. also contributed to data acquisition, data analysis, statistical analysis, and funding. Y.K., M.F., K.H., and K.S. contributed to data acquisition. H.I. and K.K. (Author 1) contributed to data analysis, with K.K. (Author 1) also performing critical revision. K.K. (Author 2) contributed to data acquisition, manuscript drafting, and statistical analysis. Y.U. contributed to funding. S.O. contributed to supervision. All authors critically revised the manuscript for important intellectual content and approved the final version of the manuscript. Availability of data and materials All data generated or analyzed during this study are available upon request from the corresponding author. References HernigouP. History of bone acoustic in fracture diagnosis: crepitus in antiquity; bone percussion with Auenbrugger; bone auscultation with Laennec and Lisfranc; monitoring cementless hip arthroplasty fixation with acoustic and sensor. Int. Orthop. 461657–1666 (2022). AlizadA. et al. Vibrational characteristics of bone fracture and fracture repair: application to excised rat femur. J. Biomech. Eng. 128300–308 (2006). FujimoriT. et al. Cost-effectiveness of lumbar fenestration surgery in the Japanese universal health insurance system. J. Orthop. Sci. 23889–894 (2018). CaesarendraW. Bone drilling: review with lab case study of bone layer classification using vibration signal and deep learning methods. Eng. 583 (2024). HommaY. et al. Artificial intelligence for distinguishment of hammering sound in total hip arthroplasty. Sci. Rep. 129826 (2022). WeiJ. C. J. et al. Using acoustic vibrations as a method for implant insertion assessment in total hip arthroplasty. Sensors 221609 (2022). FujimoriT. et al. Development of artificial intelligence for automated measurement of cervical lordosis on lateral radiographs. Sci. Rep. 1215732 (2022). KitaK. et al. Automated entry of paper-based patient-reported outcomes: applying deep learning to the Japanese Orthopaedic Association Back Pain Evaluation Questionnaire. Comput. Biol. Med. 172108197 (2024). KitaK. et al. Use of artificial intelligence to identify data elements for the Japanese Orthopaedic Association National Registry from operative records. J. Orthop. Sci. 281392–1399 (2023). NakajimaN. et al. A comparison between GPT-3.5GPT-4and GPT-4V: can the large language model (ChatGPT) pass the Japanese Board of Orthopaedic Surgery Examination? Cureus 16e56402 (2024). ItoA. et al. Hammering test for tile wall using deep learning. Appl. Sci. 151500 (2025). KandaswamyA. et al. Neural classification of lung sounds using wavelet coefficients. Comput. Biol. Med. 34523–537 (2004). AbbasS. et al. Artificial intelligence framework for heart disease classification from audio signals. Sci. Rep. 143123 (2024). DewiR. S. et al. A systematic review of physical artificial intelligence (physical AI): conceptsapplicationschallengesand future directions. J. Artif. Intell. Eng. Appl. 42246–2253 (2025). TafatW. et al. Artificial intelligence in orthopaedic surgery: a comprehensive review of current innovations and future directions. Comput. Struct. Biotechnol. Rep. 1100006 (2024). SounderajahV. et al. Developing a reporting guideline for artificial intelligence-centred diagnostic test accuracy studies: the STARD-AI protocol. BMJ Open 11e047709 (2021). MartinB. et al. Use of the area under the precision-recall curve to evaluate prediction models of rare critical illness events. Pediatr. Crit. Care Med. 26e855–e859 (2025). SalmiM. et al. Handling imbalanced medical datasets: review of a decade of research. Artif. Intell. Rev. 57273 (2024). ZhaoQ. et al. Deep learning in heart sound analysis: from techniques to clinical applications. Health Data Sci. 40182 (2024). IshidaH. et al. Haptic-assisted collaborative robot framework for improved situational awareness in skull base surgery. arXiv 2401.11709 (2024). TorunY. & PazarciÖ. Parametric power spectral density estimation-based breakthrough detection for orthopedic bone drilling with acoustic emission signal analysis. Acoust. Aust. 48221–231 (2020). GuanF. et al. State recognition of bone drilling based on acoustic emission in pedicle screw operation. Sensors 181484 (2018). SeiboldM. et al. Real-time acoustic sensing and artificial intelligence for error prevention in orthopedic surgery. Sci. Rep. 119638 (2021). García-VilanaS. et al. Acoustic emission in bone biomechanics: a comprehensive review of mechanical properties and predictive damage modeling. Sensors 25598 (2025). DengM. et al. Heart sound classification based on improved MFCC features and convolutional recurrent neural networks. Neural Netw. 13022–32 (2020). SinghA.AroraV. & SinghM. Heart sound classification using harmonic and percussive spectral features from phonocardiograms with a deep ANN approach. Appl. Sci. 1410201 (2024). Additional Declarations No competing interests reported. Supplementary Files supplementaryvideo1.mp4 Supplementary_video_1 This video shows actual bone chisel operation and the corresponding audio waveform. The surgeon determined the bone was cut and stopped percussion. The final strike caused a change in the audio (red box). Supplementary1.docx Supplementary2.docx Supplementary3.docx StardAI.docx Cite Share Download PDF Status: Published Journal Publication published 19 Apr, 2026 Read the published version in Scientific Reports → Version 1 posted Editorial decision: Revision requested 24 Mar, 2026 Reviews received at journal 23 Mar, 2026 Reviewers agreed at journal 22 Mar, 2026 Reviews received at journal 20 Mar, 2026 Reviewers agreed at journal 20 Mar, 2026 Reviews received at journal 16 Mar, 2026 Reviewers agreed at journal 16 Mar, 2026 Reviewers agreed at journal 16 Mar, 2026 Reviewers agreed at journal 11 Mar, 2026 Reviewers invited by journal 10 Mar, 2026 Editor invited by journal 04 Feb, 2026 Editor assigned by journal 25 Jan, 2026 Submission checks completed at journal 25 Jan, 2026 First submitted to journal 23 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8680502","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":604423081,"identity":"72921106-b5ab-46c9-bde3-c920448fcb87","order_by":0,"name":"Hideaki Fujiwara","email":"","orcid":"","institution":"Hoshigaoka Medical Center","correspondingAuthor":false,"prefix":"","firstName":"Hideaki","middleName":"","lastName":"Fujiwara","suffix":""},{"id":604423082,"identity":"3cf859cb-9b4a-4f97-b0b7-8b1d89a943a5","order_by":1,"name":"Takahito Fujimori","email":"data:image/png;base64,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","orcid":"","institution":"The University of Osaka","correspondingAuthor":true,"prefix":"","firstName":"Takahito","middleName":"","lastName":"Fujimori","suffix":""},{"id":604423083,"identity":"cbabc4de-4b44-4abb-abfd-15b009c05f68","order_by":2,"name":"Yuya Kanie","email":"","orcid":"","institution":"The University of Osaka","correspondingAuthor":false,"prefix":"","firstName":"Yuya","middleName":"","lastName":"Kanie","suffix":""},{"id":604423084,"identity":"9ad05dc6-116d-4f5f-be2f-62be9c09aa38","order_by":3,"name":"Masayuki Furuya","email":"","orcid":"","institution":"The University of Osaka","correspondingAuthor":false,"prefix":"","firstName":"Masayuki","middleName":"","lastName":"Furuya","suffix":""},{"id":604423085,"identity":"501f119b-3da4-4a9b-8456-df4a99057863","order_by":4,"name":"Hirotatsu Imai","email":"","orcid":"","institution":"The University of Osaka","correspondingAuthor":false,"prefix":"","firstName":"Hirotatsu","middleName":"","lastName":"Imai","suffix":""},{"id":604423086,"identity":"71a8256a-ff9d-44ec-985b-bd2363efe2f8","order_by":5,"name":"Koki Hosozawa","email":"","orcid":"","institution":"The University of Osaka","correspondingAuthor":false,"prefix":"","firstName":"Koki","middleName":"","lastName":"Hosozawa","suffix":""},{"id":604423087,"identity":"30db6820-a05e-47d4-9c2c-ef8f9cf22bf0","order_by":6,"name":"Kosuke Kita","email":"","orcid":"","institution":"The University of Osaka","correspondingAuthor":false,"prefix":"","firstName":"Kosuke","middleName":"","lastName":"Kita","suffix":""},{"id":604423089,"identity":"4243f920-3a07-49cf-a682-092549d6f869","order_by":7,"name":"Koki Kishimoto","email":"","orcid":"","institution":"The University of Osaka","correspondingAuthor":false,"prefix":"","firstName":"Koki","middleName":"","lastName":"Kishimoto","suffix":""},{"id":604423091,"identity":"420be0ae-043b-45fb-b048-78153f490481","order_by":8,"name":"Kei Shinyashiki","email":"","orcid":"","institution":"The University of Osaka","correspondingAuthor":false,"prefix":"","firstName":"Kei","middleName":"","lastName":"Shinyashiki","suffix":""},{"id":604423092,"identity":"c13d266b-de1b-4c5f-94a5-341a122038ca","order_by":9,"name":"Yuichiro Ukon","email":"","orcid":"","institution":"The University of Osaka","correspondingAuthor":false,"prefix":"","firstName":"Yuichiro","middleName":"","lastName":"Ukon","suffix":""},{"id":604423094,"identity":"774b8e86-d099-442c-a0c2-e42e9809736b","order_by":10,"name":"Seiji Okada","email":"","orcid":"","institution":"The University of Osaka","correspondingAuthor":false,"prefix":"","firstName":"Seiji","middleName":"","lastName":"Okada","suffix":""}],"badges":[],"createdAt":"2026-01-23 15:08:38","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8680502/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8680502/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1038/s41598-026-48857-w","type":"published","date":"2026-04-19T15:56:59+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":104594088,"identity":"31f01ea8-ae24-4a1a-b5ad-5ea1bc9f2ffd","added_by":"auto","created_at":"2026-03-13 17:47:37","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":235676,"visible":true,"origin":"","legend":"\u003cp\u003eFlow diagram of data selection and partitioning.\u003c/p\u003e\n\u003cp\u003eA total of 739 video clips from 37 patients were collected. Clips without penetrating sounds were excluded, leaving 192 audio records. From these, 852 three-strike windows were generated. The dataset was divided into an independent test set (102windows from 20 audio records) and a training dataset (750 windows from 172 audio records). The training dataset was further split into an internal training set (669 windows from 154 audio records) and an internal validation set (81 windows from 18 audio records).\u003c/p\u003e","description":"","filename":"Fig1.png","url":"https://assets-eu.researchsquare.com/files/rs-8680502/v1/d6a65e77af48d050985ff154.png"},{"id":104594090,"identity":"041b5477-cfcb-4578-ade4-7fd7a22b5073","added_by":"auto","created_at":"2026-03-13 17:47:37","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1812208,"visible":true,"origin":"","legend":"\u003cp\u003ePerforming a laminectomy with a chisel.\u003c/p\u003e\n\u003cp\u003eThe surgeon performed laminectomy or facetectomy using a hammer-and-chisel technique.\u003c/p\u003e","description":"","filename":"Fig2.png","url":"https://assets-eu.researchsquare.com/files/rs-8680502/v1/40b55430d06a8923a31188aa.png"},{"id":104594092,"identity":"f939c329-0fcb-4204-835a-c0a3fc507207","added_by":"auto","created_at":"2026-03-13 17:47:37","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":2212918,"visible":true,"origin":"","legend":"\u003cp\u003eExample of annotation process for chisel percussion sounds.\u003c/p\u003e\n\u003cp\u003eThe upper panel shows a waveform segmented into individual strikes. Each strike was labeled as either non-penetration (np, blue dashed boxes) or penetration (p, red box) based on surgical video review and sound change. The lower panel illustrates two representative acoustic features (MFCC_2 in yellow and mean spectral centroid in green), highlighting that values typically shift when penetration occurs. These feature plots are shown to demonstrate the acoustic differences between np and p strikes.\u003c/p\u003e","description":"","filename":"Fig3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8680502/v1/05ca7004568d63fda2b41d60.jpg"},{"id":104594095,"identity":"1d844b2f-fbea-453f-8f17-a87fb6709902","added_by":"auto","created_at":"2026-03-13 17:47:37","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1704444,"visible":true,"origin":"","legend":"\u003cp\u003eConstruction of 3-hit windows for acoustic analysis.\u003c/p\u003e\n\u003cp\u003eTo capture temporal changes, consecutive strikes were grouped into 3-hit windows with a stride of 1. Each window inherited the label of the final strike within the sequence. For example, windows ending in 2np, 3np, and 4np were labeled as non-penetration, whereas the window ending in 5p was labeled as penetration.\u003c/p\u003e","description":"","filename":"Fig4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-8680502/v1/2fbd562c328578afa8b15f5c.jpg"},{"id":107351087,"identity":"9a582d7b-1cf8-4dd9-bd55-ea7b38e73ac4","added_by":"auto","created_at":"2026-04-20 16:09:20","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6466231,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8680502/v1/dd9ca4a1-778f-4890-b926-72c7d1718c80.pdf"},{"id":104781631,"identity":"257b9f04-e46f-480a-b345-9e3dba726a3f","added_by":"auto","created_at":"2026-03-17 07:56:03","extension":"mp4","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":6630307,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary_video_1\u003c/p\u003e\n\u003cp\u003eThis video shows actual bone chisel operation and the corresponding audio waveform. The surgeon determined the bone was cut and stopped percussion. The final strike caused a change in the audio (red box).\u003c/p\u003e","description":"","filename":"supplementaryvideo1.mp4","url":"https://assets-eu.researchsquare.com/files/rs-8680502/v1/1333a8e830e3d10d2c745750.mp4"},{"id":104781692,"identity":"566c7624-89f1-4fba-928e-2b29738b2c7c","added_by":"auto","created_at":"2026-03-17 07:56:10","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":75659,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementary1.docx","url":"https://assets-eu.researchsquare.com/files/rs-8680502/v1/daed7a5b4119f1f73a90b500.docx"},{"id":104594089,"identity":"082d7bbd-d41d-49c7-b70a-ba0a6aa7b4f5","added_by":"auto","created_at":"2026-03-13 17:47:37","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":38021,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementary2.docx","url":"https://assets-eu.researchsquare.com/files/rs-8680502/v1/9ec3cc6cf09f00246f504213.docx"},{"id":104594091,"identity":"1aed8a22-f46a-47b3-84c8-092d9fe5ed35","added_by":"auto","created_at":"2026-03-13 17:47:37","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":54587,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementary3.docx","url":"https://assets-eu.researchsquare.com/files/rs-8680502/v1/00f2948ed9836aa3dd4f5979.docx"},{"id":104594094,"identity":"084c9d64-5170-40c6-aa76-a1bb15cdccb2","added_by":"auto","created_at":"2026-03-13 17:47:37","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":24120,"visible":true,"origin":"","legend":"","description":"","filename":"StardAI.docx","url":"https://assets-eu.researchsquare.com/files/rs-8680502/v1/c6e913805fe60d2534306535.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Acoustic Cues into a Surgeon-assist Physical AI for Detecting Bone Penetration During Spinal Surgery","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSound is an important source of information, especially when visual inspection is not possible. Auscultation is a representative example, where physicians assess internal organ conditions by listening to breath and heart sounds. In orthopedics as well, it is known that before the invention of X-rays, fractures were diagnosed by percussion sounds.(1,2)\u003c/p\u003e \u003cp\u003eLaminectomy using chisel is one of the fundamental procedures in spinal decompression surgery.(3) During percussion with a chisel, the surgeon integrates multiple sensory cues\u0026mdash;such as resistance of the chisel, speed, and progression\u0026mdash;to determine whether the bone has been cut.(4) One of the key pieces of information for this judgment is the change in the percussion sound. This principle is similar to that observed in arthroplasty, where the sound changes as the prosthesis becomes firmly seated in the joint.(1,5,6) Typically, perceiving these sound changes requires surgical experience; however, recent studies have begun to explore the possibility of using artificial intelligence (AI) to replicate human ability.(7\u0026ndash;10)\u003c/p\u003e \u003cp\u003eIn industry, acoustic AI has been applied to detecting delamination or cracks in concrete and tiles and to identifying machine tool faults.(11) In medicine, similar approaches have been used to analyze respiratory and heart sounds for abnormality detection.(12,13) This paradigm\u0026mdash;learning and interpreting physical signals such as sound and vibration using sensors and AI\u0026mdash;is known as Physical AI and is an emerging research field.(14,15)\u003c/p\u003e \u003cp\u003eWe aimed to extend this concept to the surgical domain, envisioning the development of an AI system that can support surgeons\u0026mdash;\u003cem\u003eSurgeon-assist Physical AI (SPAI)\u003c/em\u003e. The objective of this study was to construct an AI model that can determine whether bone has been cut based on percussion sounds during laminectomy and to evaluate its performance.\u003c/p\u003e"},{"header":"Methods","content":"\u003cp\u003e This study was approved by our institution's ethics committee. This was a retrospective diagnostic accuracy study conducted in accordance with the Standards for Reporting Studies of Diagnostic accuracy for artificial intelligence (STARD-AI) guidelines.(16) Eligibility criteria were consecutive patients among those for whom intraoperative audio\u0026ndash;video recordings were successfully obtained, who underwent lumbar or thoracic spine surgery with chisel-based bone resection at two spine centers between July 3, 2022, and September 12, 2024. (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). All surgeries were performed by three experienced spine surgeons (with 21, 19, and 14 years of experience, respectively). Percussion strikes with the chisel were recorded intraoperatively using an iPhone or iPad at a sampling frequency of 44.1 kHz. The chisels were made of stainless steel with resin handles (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e and Supplementary video_1). In each percussion sequence, the chisel was tapped multiple times to gradually advance through the bone. When the chisel penetrated the cortical bone, the sound of the strike changed. In some procedures, surgeons avoided full cortical penetration by leaving a thin cortical layer and completing bone removal using levering or prying maneuvers. Because these techniques do not involve percussion-based cortical breakthrough or its characteristic acoustic change, such cases were excluded a priori from acoustic analysis.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eEach strike was labeled as penetration (p) or non-penetration (np) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). Because no objective physical gold standard exists for identifying the exact moment of bone penetration, surgeon consensus review of the complete operative video served as the reference standard. Labeling was based on visual confirmation of cortical breakthrough and the transition to bone removal and/or the next surgical step, with acoustic changes considered only as supportive context. Initial labeling was performed by a spine surgeon with 6 years of experience and subsequently reviewed and confirmed by a senior spine surgeon with 21 years of experience, with final labels determined by consensus.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eAcoustic feature extraction\u003c/h2\u003e \u003cp\u003eFor each individual strike, 30 base acoustic features were extracted, including measures of energy, frequency distribution, mel-frequency cepstral coefficients (MFCC), and spectral patterns (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). To better capture relative changes and scaling effects across consecutive strikes, additional derived features were computed, including differences, ratios, and log-transformed differences between strikes (Supplementary 1).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e Univariate analysis of 30 basic acoustic features.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFeature\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNon-penetration\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePenetration\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003ep-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e87.6\u0026thinsp;\u0026plusmn;\u0026thinsp;34.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e68.4\u0026thinsp;\u0026plusmn;\u0026thinsp;34.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean Zero Crossing Rate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.055\u0026thinsp;\u0026plusmn;\u0026thinsp;0.02\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.051\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean Spectral Centroid (Hz)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e5342\u0026thinsp;\u0026plusmn;\u0026thinsp;1381\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e5870\u0026thinsp;\u0026plusmn;\u0026thinsp;1395\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-53.1\u0026thinsp;\u0026plusmn;\u0026thinsp;21.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-61.1\u0026thinsp;\u0026plusmn;\u0026thinsp;24.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContrast_7 (dB)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28.8\u0026thinsp;\u0026plusmn;\u0026thinsp;4.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e30.3\u0026thinsp;\u0026plusmn;\u0026thinsp;3.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e17.1\u0026thinsp;\u0026plusmn;\u0026thinsp;15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e21.7\u0026thinsp;\u0026plusmn;\u0026thinsp;13.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean Amplitude\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.12\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.11\u0026thinsp;\u0026plusmn;\u0026thinsp;0.05\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean Spectral Roll-off (Hz, 85%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10099\u0026thinsp;\u0026plusmn;\u0026thinsp;2825\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e10805\u0026thinsp;\u0026plusmn;\u0026thinsp;2810\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContrast_1 (dB)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e10.9\u0026thinsp;\u0026plusmn;\u0026thinsp;4.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e9.9\u0026thinsp;\u0026plusmn;\u0026thinsp;3.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.002\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean Spectral Flatness\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.03\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.02\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.006\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-71.6\u0026thinsp;\u0026plusmn;\u0026thinsp;35.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-77.5\u0026thinsp;\u0026plusmn;\u0026thinsp;36.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.007\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-4.7\u0026thinsp;\u0026plusmn;\u0026thinsp;14.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-2.4\u0026thinsp;\u0026plusmn;\u0026thinsp;13.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.035\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e22.1\u0026thinsp;\u0026plusmn;\u0026thinsp;12.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e24.3\u0026thinsp;\u0026plusmn;\u0026thinsp;13.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.059\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEnergy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e25.1\u0026thinsp;\u0026plusmn;\u0026thinsp;20.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e22.8\u0026thinsp;\u0026plusmn;\u0026thinsp;17.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.084\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-11.2\u0026thinsp;\u0026plusmn;\u0026thinsp;11.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-10\u0026thinsp;\u0026plusmn;\u0026thinsp;11.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.12\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContrast_4 (dB)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11\u0026thinsp;\u0026plusmn;\u0026thinsp;3.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11.2\u0026thinsp;\u0026plusmn;\u0026thinsp;2.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.15\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDuration (s)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.02\u0026thinsp;\u0026plusmn;\u0026thinsp;0.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.02\u0026thinsp;\u0026plusmn;\u0026thinsp;0.01\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.18\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContrast_2 (dB)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4.6\u0026thinsp;\u0026plusmn;\u0026thinsp;2.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4.8\u0026thinsp;\u0026plusmn;\u0026thinsp;2.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.20\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-14.8\u0026thinsp;\u0026plusmn;\u0026thinsp;13.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-13.7\u0026thinsp;\u0026plusmn;\u0026thinsp;13.7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.23\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean RMS\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.1\u0026thinsp;\u0026plusmn;\u0026thinsp;0.03\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.1\u0026thinsp;\u0026plusmn;\u0026thinsp;0.04\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.24\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean Spectral Bandwidth (Hz)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e4295\u0026thinsp;\u0026plusmn;\u0026thinsp;809\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4360\u0026thinsp;\u0026plusmn;\u0026thinsp;823\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.27\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-20.2\u0026thinsp;\u0026plusmn;\u0026thinsp;17.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-18.8\u0026thinsp;\u0026plusmn;\u0026thinsp;16.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-7.1\u0026thinsp;\u0026plusmn;\u0026thinsp;11.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-8\u0026thinsp;\u0026plusmn;\u0026thinsp;11.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.43\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e-16.2\u0026thinsp;\u0026plusmn;\u0026thinsp;12.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e-17.2\u0026thinsp;\u0026plusmn;\u0026thinsp;11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContrast_5 (dB)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12.4\u0026thinsp;\u0026plusmn;\u0026thinsp;2.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e12.6\u0026thinsp;\u0026plusmn;\u0026thinsp;2.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeaks Detected\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u0026thinsp;\u0026plusmn;\u0026thinsp;0.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u0026thinsp;\u0026plusmn;\u0026thinsp;0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContrast_3 (dB)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8.4\u0026thinsp;\u0026plusmn;\u0026thinsp;3.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.5\u0026thinsp;\u0026plusmn;\u0026thinsp;3.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11.5\u0026thinsp;\u0026plusmn;\u0026thinsp;11.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e11.5\u0026thinsp;\u0026plusmn;\u0026thinsp;12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMFCC_4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e32.1\u0026thinsp;\u0026plusmn;\u0026thinsp;20.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e32\u0026thinsp;\u0026plusmn;\u0026thinsp;21.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eContrast_6 (dB)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.8\u0026thinsp;\u0026plusmn;\u0026thinsp;2.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e14.8\u0026thinsp;\u0026plusmn;\u0026thinsp;2.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"4\" nameend=\"c4\" namest=\"c1\"\u003e \u003cp\u003eThe rows are sorted in descending order of p-values from top to bottom.\u003c/p\u003e \u003cp\u003eMFCC; mel-frequency cepstral coefficients.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eA sliding window of three consecutive strikes (3-hit window, stride\u0026thinsp;=\u0026thinsp;1) was then constructed (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). For each 3-hit window, acoustic features were expanded to capture both the properties of individual strikes and the changes that occur between consecutive strikes. This process yielded a total of 570 features per window, enabling the model to learn not only how each strike sounds in isolation but also how the acoustic pattern evolves as the chisel approaches bone penetration.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eAI model development\u003c/h3\u003e\n\u003cp\u003eA machine learning classifier was developed using LightGBM, trained on an expanded feature set comprising 570 features per 3-hit window. Data splitting was performed at the recording level to prevent information leakage between training and test sets.\u003c/p\u003e\n\u003ch3\u003eModel selection and hyperparameter optimization\u003c/h3\u003e\n\u003cp\u003eModel Selection and hyperparameter optimization was conducted using 5-fold cross-validation on the training set, with grouping by recording identifier to ensure that all windows from a given recording session appeared exclusively in either the training or validation fold. The primary evaluation metric was the area under the precision\u0026ndash;recall curve (PR-AUC), selected for its superior sensitivity to model performance on the minority (positive) class in imbalanced datasets. Model performance was summarized as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation across cross-validation folds. Secondary metrics included the area under the receiver operating characteristic curve (ROC-AUC) and threshold-dependent measures (accuracy, F1-score, precision, recall) computed at an optimized threshold that achieves sensitivity (recall)\u0026thinsp;\u0026ge;\u0026thinsp;80%. Early stopping (patience\u0026thinsp;=\u0026thinsp;100 rounds) based on validation set PR-AUC was applied during cross-validation to identify the optimal number of boosting iterations and prevent overfitting.\u003c/p\u003e\n\u003ch3\u003eOperating threshold determination\u003c/h3\u003e\n\u003cp\u003eThe operating threshold for binary classification was determined using a clinically-driven approach prioritizing sensitivity. Given the importance of minimizing false negatives (missed bone breakthrough events) to prevent inadvertent injury to underlying structures, we targeted a minimum sensitivity of 80% based on clinical safety requirements. Using out-of-fold (OOF) predictions aggregated across all cross-validation folds, we analyzed the precision\u0026ndash;recall curve and selected the highest probability threshold that achieved sensitivity (recall)\u0026thinsp;\u0026ge;\u0026thinsp;80%.\u003c/p\u003e\n\u003ch3\u003eModel training and evaluation\u003c/h3\u003e\n\u003cp\u003eThe final model was trained on 154 recordings using optimal hyperparameters identified during cross-validation, with 18 recordings reserved for internal validation. Training iterations were fixed to the mean optimal count across folds to ensure consistency with cross-validation results. To maintain consistency with cross-validation results, the model was trained for a fixed number of iterations equal to the mean optimal iteration count across folds. The internal validation set was used for monitoring purposes only and did not influence model training process or stopping criteria. The final model was evaluated on the independent hold-out test set (20 recordings, 102 windows) using both threshold-independent metrics (PR-AUC, ROC-AUC) and threshold-dependent classification performance at the predetermined operating threshold.\u003c/p\u003e \u003cp\u003eModel performance was reported separately for cross-validation results reflecting training set generalization, independent test set performance at the clinically-optimized operating threshold, and threshold sensitivity analysis showing performance trade-offs across multiple operating points.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eUnivariate comparisons of individual acoustic features between penetration (p) and non-penetration (np) strikes were performed using the Mann\u0026ndash;Whitney U test, given the non-normal distribution of the data.\u003c/p\u003e \u003cp\u003eA significance threshold of \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05 (two-sided) was applied. All analyses were conducted using Python 3.10.14 with LightGBM 4.5.0, scikit-learn 1.5.2, pandas 2.2.2, and NumPy 2.1.3 (Supplementary 2).\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eDemographics\u003c/h2\u003e \u003cp\u003eA total of 37 patients were included (mean age 61\u0026thinsp;\u0026plusmn;\u0026thinsp;26 years, 25 with lumbar spinal stenosis and 12 with spinal deformity). From these patients, 739 surgical video clips were obtained. Among them, 192 audio records containing penetration sounds were extracted (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Each audio record contained an average of 6\u0026thinsp;\u0026plusmn;\u0026thinsp;2 strikes (range 3\u0026ndash;15), resulting in a total of 1236 labeled strikes: 215 penetration and 1021 non-penetration.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eSingle-strike features\u003c/h2\u003e \u003cp\u003eComparison of acoustic features between penetration (n\u0026thinsp;=\u0026thinsp;215) and non-penetration (n\u0026thinsp;=\u0026thinsp;1021) strikes revealed significant differences in multiple spectral characteristics, including MFCC_2, MFCC_3, MFCC_6, MFCC_1, and MFCC_12, mean zero-crossing rate, mean spectral centroid, spectral roll-off, spectral flatness, and spectral contrast (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Most of these features represent tonal brightness or sharpness of the sound.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eThree-strike windows\u003c/h2\u003e \u003cp\u003eFrom the 1236 strikes, 852 three-strike windows were generated (215 penetration windows, 637 non-penetration windows; 25% vs. 75%) (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Among these, 20 audio records containing 102 windows (29 penetration windows, 73 non-penetration windows; 28% vs. 72%) were reserved as an independent test set. The remaining 172 audio records (750 windows: 186 penetration windows, 564 non-penetration windows; 25% vs. 75%) was further split into an internal training subset (154 recordings) and an internal validation subset (18 recordings).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eModel performance\u003c/h2\u003e \u003cp\u003eUsing the optimal hyperparameter configuration, the model demonstrated stable performance across 5-fold cross-validation, achieving a mean PR-AUC of 0.645\u0026thinsp;\u0026plusmn;\u0026thinsp;0.071 and a mean ROC-AUC of 0.823\u0026thinsp;\u0026plusmn;\u0026thinsp;0.049 (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Early stopping identified an average optimal training length of 106 iterations. Based on out-of-fold predictions, an operating threshold of 0.2305 was selected to achieve the prespecified target sensitivity of 0.80 on the training set, corresponding to an accuracy of 0.807, F1-score of 0.551, specificity of 0.635, and precision of 0.420. When evaluated on the independent hold-out test set, the model maintained robust discrimination, yielding a PR-AUC of 0.604 and a ROC-AUC of 0.838.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eModel performance of LightGBM at an operating threshold (0.2305).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eROC-AUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePR-AUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003cp\u003e(Recall)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003cp\u003e(Positive predictive value)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eF1-score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eThreshold\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCross validation (5-fold)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.822\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.645\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.801\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.420\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.807\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.635\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e0.551\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e \u003cp\u003e0.2305\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"9\" nameend=\"c9\" namest=\"c1\"\u003e \u003cp\u003eROC-AUC; receiver operating characteristic curve \u0026ndash; area under the curve\u003c/p\u003e \u003cp\u003ePR-AUC; precision\u0026ndash;recall \u0026ndash; area under the curve\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eThreshold sensitivity analysis\u003c/h2\u003e \u003cp\u003eUsing the pre-specified operating threshold derived from out-of-fold predictions (0.2305), test-set performance was sensitivity 0.655, specificity 0.877, and precision 0.679 (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). To illustrate the operating trade-offs, we additionally report test-set performance across alternative probability thresholds (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). In an exploratory post-hoc assessment, we identified that a threshold of 0.15 would yield sensitivity\u0026thinsp;\u0026ge;\u0026thinsp;0.80 on this specific test set; at this threshold, sensitivity was 0.828, specificity 0.767, accuracy 0.784, and precision 0.585.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eThreshold sensitivity analysis of test set.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eThreshold\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSensitivity \u003c/p\u003e \u003cp\u003e(recall)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003cp\u003e(Positive predictive value)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eF1-Score\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.931\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.466\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.676\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.575\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.621\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.15*\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.828\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.585\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.784\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.767\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.686\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.724\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.656\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.814\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.849\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.688\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.2305\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.655\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.679\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.814\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.877\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.667\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.621\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.692\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.814\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.890\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.655\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.586\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.680\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.804\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.890\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.628\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003e*Selected operating threshold (target sensitivity\u0026thinsp;\u0026ge;\u0026thinsp;80%)\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"6\"\u003e\u0026dagger; Training-derived threshold from out of fold (OOF) predictions\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eConfusion matrix on the independent test set at the threshold of 0.15.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e \u003cp\u003eAI prediction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePenetration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNon-penetration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e\u003cb\u003eReference standard\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePenetration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e24 (TP)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e5 (FN)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e29\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNon-penetration\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17 (FP)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e56 (TN)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e73\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e102\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eTP; true positive, FN; false negative, FP; false positive, TN; true negative .\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"5\"\u003eAccuracy\u0026thinsp;=\u0026thinsp;80/102\u0026thinsp;=\u0026thinsp;0.784, Recall (Sensitivity)\u0026thinsp;=\u0026thinsp;24/29\u0026thinsp;=\u0026thinsp;0.828, Specificity\u0026thinsp;=\u0026thinsp;56/73\u0026thinsp;=\u0026thinsp;0.767, Precision (Positive predictive value)\u0026thinsp;=\u0026thinsp;24/41\u0026thinsp;=\u0026thinsp;0.585, F1-score\u0026thinsp;=\u0026thinsp;0.686\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eFeature importance\u003c/h2\u003e \u003cp\u003eFeature importance analysis identified the most influential predictors as temporal dynamics of acoustic features across strikes (Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e and Supplementary 3). The top-ranked feature was MFCC_3__slope3, representing the slope of the 3rd mel-frequency cepstral coefficient across the three-strike window. This was followed by MFCC_2__slope3 and MFCC_2__range3. Other highly ranked features included Contrast_7 (dB)__t3, MFCC_3__r21, and mean Zero Crossing Rate____std3.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eTop 10 feature importances of the model.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRanking\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFeature\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eImportance\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMFCC_3__slope3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1294\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMFCC_2__slope3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e501\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMFCC_2__range3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e310\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eContrast_7 (dB)__t3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e280\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMFCC_3__r21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e261\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMean Zero Crossing Rate__std3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e176\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMFCC_10__range3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e146\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMFCC_5__t1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e144\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eContrast_1 (dB)__median3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e136\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMFCC_3__d21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e134\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we developed and validated an AI model capable of detecting bone penetration during spinal decompression surgery by analyzing chisel percussion sounds. The model achieved a PR-AUC of 0.645 in cross-validation and 0.604 in an independent test set, with ROC-AUC values consistently above 0.80. Given that the prevalence of penetration events in our dataset was relatively low (\u0026asymp;\u0026thinsp;17%), these PR-AUC values represent meaningful discriminative performance well above chance level, where a random classifier would be expected to achieve only\u0026thinsp;~\u0026thinsp;0.17. These findings indicate that the approach generalizes reasonably well to unseen data despite the limited sample size.(17,18) Importantly, feature importance analysis revealed that the model primarily relied on changes, slopes, and ranges of acoustic features across sequential strikes, rather than on static single-strike values. This finding underscores the importance of capturing temporal dynamics when distinguishing penetration from non-penetration sounds. This suggests that the algorithm mimics the way surgeons perceive the brightening or sharpening of sound when penetration occurs, providing an interpretable link between acoustic dynamics and surgical decision-making.\u003c/p\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eNovelty of the study\u003c/h2\u003e \u003cp\u003eTo our knowledge, this is the first report to apply machine learning to analyze intraoperative percussion sounds in spinal surgery. Previous research in medicine has primarily focused on respiratory sounds,(12) cardiac auscultation,(13,19) or industrial quality control tasks,(11) but not on surgical sound analysis. While experienced surgeons have long relied on subtle acoustic cues during procedures such as laminectomy or arthroplasty, these perceptions have remained subjective and dependent on years of training (1,2,5,6). Our study demonstrates that these acoustic cues can be objectively captured, quantified, and classified using AI. By focusing on three-strike windows and emphasizing sound changes across beats, our approach goes beyond static sound analysis and more closely reflects the way surgeons perceive \u0026ldquo;patterns of change\u0026rdquo; rather than isolated signals. This methodological innovation may open a new research field\u0026mdash;surgical acoustics\u0026mdash;as part of the broader domain of Physical AI.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eClinical relevance\u003c/h2\u003e \u003cp\u003eThe clinical implications of this work are considerable. First, the ability to objectively recognize bone penetration can support less experienced surgeons during spinal decompression by providing real-time feedback, thereby helping them learn to identify subtle acoustic cues that are otherwise acquired only through prolonged practice. Such a system may also reduce inter-surgeon variability and improve safety, particularly in procedures requiring precise bony removal. For example, when a junior surgeon performs bone cutting, an experienced supervisor can recognize from sound alone when the bone has already been penetrated and intervene to prevent excessive chiseling. By making this auditory judgment explicit and accessible, the proposed AI model has the potential to replicate this supervisory role and enhance both safety and training.\u003c/p\u003e \u003cp\u003eOur results demonstrate that different probability thresholds produce predictable trade-offs between sensitivity and specificity, suggesting that the model can be flexibly tuned according to clinical context.\u003c/p\u003e \u003cp\u003eSecond, as surgical robotics continues to advance,(15) integration of acoustic AI could provide robots with an additional sensory modality, complementing force or visual feedback and enabling safer autonomous or semi-autonomous bone cutting.(20,21) In this sense, our work represents an early step toward what we propose as Surgeon-assist Physical AI (SPAI)\u0026mdash;an AI framework designed to emulate and augment the sensory judgments traditionally made by surgeons.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eWhy the sound changes\u003c/h2\u003e \u003cp\u003eThe acoustic change that occurs when the chisel penetrates bone can be explained by basic principles of material vibration.(6,22\u0026ndash;24) When a chisel strikes cortical bone, the sound is generated by the vibration of both the chisel and the bone structure. Before penetration, the chisel is resisted by solid bone, producing a duller and more uniform tone. Once penetration occurs, the structural integrity of the bone at that point is lost, resulting in a sharper, higher-pitched sound.(2) This is analogous to the difference between striking a solid wall versus a hollow structure: the resonance shifts, and the resulting sound becomes brighter and more distinct.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eAcoustic features and their interpretation\u003c/h2\u003e \u003cp\u003eThe model\u0026rsquo;s reliance on features such as mel-frequency cepstral coefficients (MFCCs), zero-crossing rate, and spectral contrast reflects the acoustic properties that change most prominently during penetration.(25,26) Unlike static single-strike features, the top predictors were predominantly measures of temporal dynamics\u0026mdash;including slopes, ranges, and ratios across consecutive strikes\u0026mdash;indicating that the model captured not only the absolute tonal quality of each strike but also how it evolved over time.\u003c/p\u003e \u003cp\u003eMFCCs are widely used in speech and audio recognition to represent the timbre or tonal quality of a sound. In our analysis, slope- and range-based MFCC features (e.g., MFCC_3__slope3, MFCC_2__slope3, MFCC_2__range3) ranked among the most important predictors, suggesting that penetration was associated with systematic shifts in tonal patterns across the three-strike window.\u003c/p\u003e \u003cp\u003eTaken together, these findings provide an objective explanation of the subjective impression surgeons describe\u0026mdash;namely, that the sound becomes \u0026ldquo;higher\u0026rdquo; or \u0026ldquo;brighter\u0026rdquo; when the bone gives way. Importantly, the model emphasized how features change across sequential strikes, mirroring the clinical reality that surgeons detect penetration not by an isolated sound but by recognizing the evolving acoustic pattern across a series of hammer strikes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eLimitations\u003c/h2\u003e \u003cp\u003eThis study has some limitations. First, the dataset was relatively small and derived exclusively from procedures performed by experienced spine surgeons. Although the number of individual strikes was sufficient for feature extraction, the diversity of surgical contexts was limited.\u003c/p\u003e \u003cp\u003eSecond, the present approach relied on manually segmented individual strikes and the construction of fixed three-strike windows. While effective as a proof of concept, this design does not yet analyze the full continuous acoustic stream as it would occur intraoperatively.\u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn summary, this study demonstrates the feasibility of using machine learning to recognize bone penetration during spinal decompression by analyzing acoustic signals from chisel strikes. This approach could enhance surgical training, provide real-time decision support for less experienced surgeons, and ultimately serve as a foundation for multimodal feedback systems in robotic or computer-assisted surgery.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eInstitution ethics committee approval:\u0026nbsp;\u003c/strong\u003eThe study protocol was approved by the University of Osaka Hospital Ethical Review Board (protocol number: 22099), and informed consent was waived due to the retrospective nature of the study. The study was conducted in accordance with the Declaration of Helsinki.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of competing interest:\u0026nbsp;\u003c/strong\u003eThe authors declare that they have no conflict of interest.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgement:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by JSPS KAKENHI No. JP21K20966, NSK Nakanishi Foundation, and J\u0026amp;J Medical Research Grant.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFundings:\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work was supported by JSPS KAKENHI No. JP21K20966, NSK Nakanishi Foundation, and J\u0026amp;J Medical Research Grant.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contribution\u003c/strong\u003e: H.F. and T.F. contributed to conception, design, and drafting of the manuscript; T.F. also contributed to data acquisition, data analysis, statistical analysis, and funding. Y.K., M.F., K.H., and K.S. contributed to data acquisition. H.I. and K.K. (Author 1) contributed to data analysis, with K.K. (Author 1) also performing critical revision. K.K. (Author 2) contributed to data acquisition, manuscript drafting, and statistical analysis. Y.U. contributed to funding. S.O. contributed to supervision. All authors critically revised the manuscript for important intellectual content and approved the final version of the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAll data generated or analyzed during this study\u0026nbsp;\u003c/strong\u003eare available upon request from the corresponding author.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eHernigouP. History of bone acoustic in fracture diagnosis: crepitus in antiquity; bone percussion with Auenbrugger; bone auscultation with Laennec and Lisfranc; monitoring cementless hip arthroplasty fixation with acoustic and sensor. Int. Orthop. 461657\u0026ndash;1666 (2022).\u003c/li\u003e\n\u003cli\u003eAlizadA. et al. Vibrational characteristics of bone fracture and fracture repair: application to excised rat femur. J. Biomech. Eng. 128300\u0026ndash;308 (2006).\u003c/li\u003e\n\u003cli\u003eFujimoriT. et al. Cost-effectiveness of lumbar fenestration surgery in the Japanese universal health insurance system. J. Orthop. Sci. 23889\u0026ndash;894 (2018).\u003c/li\u003e\n\u003cli\u003eCaesarendraW. Bone drilling: review with lab case study of bone layer classification using vibration signal and deep learning methods. Eng. 583 (2024).\u003c/li\u003e\n\u003cli\u003eHommaY. et al. Artificial intelligence for distinguishment of hammering sound in total hip arthroplasty. Sci. Rep. 129826 (2022).\u003c/li\u003e\n\u003cli\u003eWeiJ. C. J. et al. Using acoustic vibrations as a method for implant insertion assessment in total hip arthroplasty. Sensors 221609 (2022).\u003c/li\u003e\n\u003cli\u003eFujimoriT. et al. Development of artificial intelligence for automated measurement of cervical lordosis on lateral radiographs. Sci. Rep. 1215732 (2022).\u003c/li\u003e\n\u003cli\u003eKitaK. et al. Automated entry of paper-based patient-reported outcomes: applying deep learning to the Japanese Orthopaedic Association Back Pain Evaluation Questionnaire. Comput. Biol. Med. 172108197 (2024).\u003c/li\u003e\n\u003cli\u003eKitaK. et al. Use of artificial intelligence to identify data elements for the Japanese Orthopaedic Association National Registry from operative records. J. Orthop. Sci. 281392\u0026ndash;1399 (2023).\u003c/li\u003e\n\u003cli\u003eNakajimaN. et al. A comparison between GPT-3.5GPT-4and GPT-4V: can the large language model (ChatGPT) pass the Japanese Board of Orthopaedic Surgery Examination? Cureus 16e56402 (2024).\u003c/li\u003e\n\u003cli\u003eItoA. et al. Hammering test for tile wall using deep learning. Appl. Sci. 151500 (2025).\u003c/li\u003e\n\u003cli\u003eKandaswamyA. et al. Neural classification of lung sounds using wavelet coefficients. Comput. Biol. Med. 34523\u0026ndash;537 (2004).\u003c/li\u003e\n\u003cli\u003eAbbasS. et al. Artificial intelligence framework for heart disease classification from audio signals. Sci. Rep. 143123 (2024).\u003c/li\u003e\n\u003cli\u003eDewiR. S. et al. A systematic review of physical artificial intelligence (physical AI): conceptsapplicationschallengesand future directions. J. Artif. Intell. Eng. Appl. 42246\u0026ndash;2253 (2025).\u003c/li\u003e\n\u003cli\u003eTafatW. et al. Artificial intelligence in orthopaedic surgery: a comprehensive review of current innovations and future directions. Comput. Struct. Biotechnol. Rep. 1100006 (2024).\u003c/li\u003e\n\u003cli\u003eSounderajahV. et al. Developing a reporting guideline for artificial intelligence-centred diagnostic test accuracy studies: the STARD-AI protocol. BMJ Open 11e047709 (2021).\u003c/li\u003e\n\u003cli\u003eMartinB. et al. Use of the area under the precision-recall curve to evaluate prediction models of rare critical illness events. Pediatr. Crit. Care Med. 26e855\u0026ndash;e859 (2025).\u003c/li\u003e\n\u003cli\u003eSalmiM. et al. Handling imbalanced medical datasets: review of a decade of research. Artif. Intell. Rev. 57273 (2024).\u003c/li\u003e\n\u003cli\u003eZhaoQ. et al. Deep learning in heart sound analysis: from techniques to clinical applications. Health Data Sci. 40182 (2024).\u003c/li\u003e\n\u003cli\u003eIshidaH. et al. Haptic-assisted collaborative robot framework for improved situational awareness in skull base surgery. arXiv 2401.11709 (2024).\u003c/li\u003e\n\u003cli\u003eTorunY. \u0026amp; Pazarci\u0026Ouml;. Parametric power spectral density estimation-based breakthrough detection for orthopedic bone drilling with acoustic emission signal analysis. Acoust. Aust. 48221\u0026ndash;231 (2020).\u003c/li\u003e\n\u003cli\u003eGuanF. et al. State recognition of bone drilling based on acoustic emission in pedicle screw operation. Sensors 181484 (2018).\u003c/li\u003e\n\u003cli\u003eSeiboldM. et al. Real-time acoustic sensing and artificial intelligence for error prevention in orthopedic surgery. Sci. Rep. 119638 (2021).\u003c/li\u003e\n\u003cli\u003eGarc\u0026iacute;a-VilanaS. et al. Acoustic emission in bone biomechanics: a comprehensive review of mechanical properties and predictive damage modeling. Sensors 25598 (2025).\u003c/li\u003e\n\u003cli\u003eDengM. et al. Heart sound classification based on improved MFCC features and convolutional recurrent neural networks. Neural Netw. 13022\u0026ndash;32 (2020).\u003c/li\u003e\n\u003cli\u003eSinghA.AroraV. \u0026amp; SinghM. Heart sound classification using harmonic and percussive spectral features from phonocardiograms with a deep ANN approach. Appl. Sci. 1410201 (2024).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"acoustic analysis, bone penetration, spinal decompression, machine learning, Physical AI, surgeon-assist, intraoperative sound, diagnostic accuracy, gradient boosting, precision–recall analysis, surgical skill quantification","lastPublishedDoi":"10.21203/rs.3.rs-8680502/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8680502/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cb\u003ePurpose\u003c/b\u003e\u003c/p\u003e \u003cp\u003eSkilled surgeons can detect subtle changes in bone-cutting sounds to recognize bone penetration, but acoustic cues are subjective and dependent on surgical experience. The objective of this study was to develop an artificial intelligence (AI) model capable of detecting bone penetration from intraoperative percussion sounds.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMethods\u003c/b\u003e\u003c/p\u003e \u003cp\u003eA total of 1,236 chisel strikes were identified from intraoperative recordings obtained during lumbar and thoracic spinal decompression surgeries performed using a chisel, and were labeled as penetration or non-penetration. Acoustic features were extracted per strike and expanded across 3-hit sliding windows to capture dynamic temporal changes. A gradient boosting machine learning classifier (LightGBM) was trained, with 10% of data as an independent test set. Model performance was primarily evaluated using the area under the precision\u0026ndash;recall curve (PR-AUC) and receiver operating characteristic AUC (ROC-AUC).\u003c/p\u003e\u003cp\u003e\u003cb\u003eResults\u003c/b\u003e\u003c/p\u003e \u003cp\u003eOn the independent test set, the model yielded a ROC-AUC of 0.838 and a PR-AUC of 0.604. In a sensitivity analysis, the model achieved a sensitivity of 0.828, specificity of 0.767, accuracy 0.784, F1-score 0.686, and precision 0.585. Feature importance analysis revealed that dynamic changes between consecutive strikes, such as mel-frequency cepstral coefficients (MFCCs), zero-crossing rate, and spectral contrast, were the most influential predictors.\u003c/p\u003e\u003cp\u003e\u003cb\u003eConclusion\u003c/b\u003e\u003c/p\u003e \u003cp\u003eAn AI model analyzing intraoperative percussion sounds demonstrated a robust ability to detect bone penetration, with consistent ROC performance and clinically acceptable PR-AUC. Quantifying acoustic cues traditionally interpreted subjectively by surgeons may support intraoperative decision-making and surgical training.\u003c/p\u003e","manuscriptTitle":"Acoustic Cues into a Surgeon-assist Physical AI for Detecting Bone Penetration During Spinal Surgery","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-13 17:47:32","doi":"10.21203/rs.3.rs-8680502/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-24T09:45:44+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-23T12:51:02+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"101468723336950423369622651292685166099","date":"2026-03-22T04:58:02+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-20T15:32:47+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"14084022826859245997351223984239189573","date":"2026-03-20T15:27:46+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-16T14:59:45+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"150171957167349137958971494132233707816","date":"2026-03-16T05:05:13+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"175604440477565475257115441188362427417","date":"2026-03-16T04:15:34+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"338850034169028437397589158154840735908","date":"2026-03-11T11:46:44+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-03-11T03:42:16+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-02-04T14:07:16+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-26T03:10:57+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-26T03:10:54+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-01-23T14:50:02+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"b580989a-f09b-478f-9a20-4faa05fa0baf","owner":[],"postedDate":"March 13th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[{"id":64320010,"name":"Physical sciences/Engineering"},{"id":64320011,"name":"Health sciences/Health care"},{"id":64320012,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2026-04-20T16:07:08+00:00","versionOfRecord":{"articleIdentity":"rs-8680502","link":"https://doi.org/10.1038/s41598-026-48857-w","journal":{"identity":"scientific-reports","isVorOnly":false,"title":"Scientific Reports"},"publishedOn":"2026-04-19 15:56:59","publishedOnDateReadable":"April 19th, 2026"},"versionCreatedAt":"2026-03-13 17:47:32","video":"","vorDoi":"10.1038/s41598-026-48857-w","vorDoiUrl":"https://doi.org/10.1038/s41598-026-48857-w","workflowStages":[]},"version":"v1","identity":"rs-8680502","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8680502","identity":"rs-8680502","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}