A method and system for generating dental case with fused voiceprint features

Abstract

The application provides a kind of fusion voiceprint feature dental case generation method and system, through high sensitivity multi-channel pickup and edge real-time processing, realize multi-device acoustic fingerprint extraction, and based on spectrum analysis, identify device mechanical state transition;Combined with natural language processing and standard medical terminology mapping, the operation log of medical record is standardized coded, and is normalized to clinical path logic;Adopt cross-modal graph attention mechanism to realize the accurate space-time synchronization of voiceprint event and medical record event, then encrypt the corrected data binding, generate unique, anti-fake fusion data, finally through block chain distributed evidence guarantee non-tamperable, the application significantly improves diagnosis and treatment data synchronization, credible evidence and identity anti-fake ability.

Description

Technical Field

[0001] This invention relates to the field of medical information processing and voiceprint feature fusion technology, and in particular to a method and system for generating oral medical records by fusing voiceprint features. Background Technology

[0002] Currently, with the advancement of intelligent and paperless data management in oral healthcare, the application of Electronic Medical Record (EMR) systems in medical institutions is becoming increasingly widespread, driving continuous improvement in the structuring, standardization, and traceability of medical data. In oral treatment scenarios, the doctor's procedures are often accompanied by the parallel operation of multiple high-speed, precision diagnostic and treatment instruments (such as handpieces, ultrasonic bone scalpels, and root canal treatment motors). The voiceprint signals generated by these devices have been used in some studies to assist in the identification of treatment behaviors and the tracing of responsibility. Existing mainstream solutions can be broadly categorized into the following two types: One type focuses on traditional speech signal processing, using MFCC features, GMM-UBM modeling, or i-vector deep clustering techniques, supplemented by speech transcription and keyword detection, to capture the content of doctor-patient dialogues or typical operational acoustic events. These systems primarily focus on scenarios such as personnel role recognition and annotation, and archiving of diagnostic and treatment steps. They can achieve shallow, synchronous, or weak correlation with medical record text, but they lack sufficient modeling of the temporal changes during the mechanical operation of equipment, and have limited ability to resolve issues in high-noise environments and with multiple concurrent devices. Another approach emphasizes the integration of API data from medical IoT devices. Some high-end diagnostic and treatment instruments can output start, stop, and alarm events through dedicated internal interfaces and directly correlate them with the structured logs of medical records. This can improve the consistency between operational events and medical records to some extent, but it heavily relies on device manufacturer protocols, lacks broad compatibility, and data latency and clock drift inevitably introduce time errors in actual data acquisition. Summary of the Invention

[0003] In order to solve the above-mentioned technical problems, the present invention provides a method and system for generating oral medical records by incorporating voiceprint features.

[0004] The technical solution of this invention is implemented as follows: A method for generating oral medical records by fusing voiceprint features, which involves acquiring the voiceprint signals of diagnostic and treatment instruments operating in an oral treatment scenario, constructing a voiceprint event sequence based on the voiceprint signals, acquiring a structured diagnostic and treatment operation log corresponding to the voiceprint event sequence to obtain a medical record event sequence, performing cross-modal matching and alignment between the voiceprint event sequence and the medical record event sequence to achieve and correct the timestamps of the medical record event sequence, and encrypting and binding the timestamp-corrected medical record event sequence with the voiceprint event sequence to generate a fused oral medical record data unit. Specifically, it includes the following steps: S1: Audio signals emitted by various diagnostic and treatment instruments during oral treatment are collected in real time through a high-sensitivity sound pickup device, and millisecond-level frame processing is performed through a low-latency edge computing node to generate multi-device voiceprint signals with timestamps. S2: After performing a short-time Fourier transform on the multi-device acoustic signature signal, a complex spectral tensor is constructed, and the first derivative sequence of the phase difference between adjacent frames is calculated along the time axis to generate a phase evolution trajectory sequence characterizing the mechanical state transition of the device. S3: Construct a set of directed edges based on the key turning points in the phase evolution trajectory sequence, and map it to a typical operation stage template library to generate an acoustic fingerprint event sequence containing device acoustic fingerprint information; S4: Obtain the structured diagnosis and treatment operation log corresponding to the voiceprint event sequence, parse it into standardized coded operation event nodes and inject time window confidence intervals to generate a case event sequence with clinical pathway constraints; S5: Input the case event sequence and the voiceprint event sequence into a lightweight graph attention mechanism, calculate the joint similarity score of nodes and edges, and output the cross-modal alignment path and time offset estimate of the optimal matching result; S6: Based on the estimated time offset in the cross-modal alignment path, automatically correct the original timestamps in the medical record events to generate a timestamp-corrected case event sequence with operational semantic consistency; S7: Encrypt and bind the timestamp-corrected case event sequence with the voiceprint event sequence to generate a unique and anti-counterfeiting fused oral case data unit; S8: Generate a unique data identifier for the fused oral case data unit using a hash algorithm, and upload it to the blockchain network for distributed storage, generating an immutable blockchain storage hash value.

[0005] The present invention also provides a dental case generation system that integrates voiceprint features, and uses the above-mentioned method for generating dental cases by integrating voiceprint features.

[0006] The present invention provides a method and system for generating oral medical records by incorporating voiceprint features, which has the following beneficial effects: (1) This invention constructs a voiceprint representation mechanism based on phase evolution trajectory, abandoning traditional speech feature extraction paradigms such as MFCC and GMM-UBM. Instead, it utilizes the complex spectral tensor after short-time Fourier transform and the first derivative sequence of the phase difference between adjacent frames to accurately capture the transient dynamic characteristics of dental devices such as high-speed mobile phones and ultrasonic bone cutters during mechanical state transitions, forming a highly discriminative "phase evolution trajectory" representation. This representation method is robust to ambient background noise, does not require speech content analysis or speaker separation, and significantly improves the usability and stability of voiceprint signals under complex acoustic conditions. (2) This invention introduces a dual-channel graph matching architecture, which transforms the real-time phase evolution trajectory on the voiceprint side into a dynamic subgraph. Combined with the candidate subgraphs of the structured event atlas constructed on the medical record side based on SNOMED CT encoding and clinical pathway constraints, a learnable graph attention mechanism is used to achieve cross-modal alignment by jointly optimizing node semantics and edge constraints. This mechanism not only considers the proximity of time nodes, but also emphasizes the logical consistency of operational semantics. For example, it ensures that "removal of decay" must occur after "opening the medulla oblongata" and before "reshaping", thereby achieving a fundamental shift from "clock-level synchronization" to "semantic alignment". (3) This invention designs a lightweight edge inference process and a closed-loop correction mechanism. Without introducing speech transcription, wake word detection, voiceprint clustering, or dialogue understanding modules, it completes end-to-end voiceprint-medical record fusion and blockchain evidence storage triggering. The inference latency of the entire processing link on the ARM Cortex-A76@2.0GHz platform is stably controlled within 87ms, meeting the real-time requirements of high-frequency interactive scenarios in oral diagnosis and treatment. At the same time, the system automatically corrects the original timestamp of the medical record through adaptive time offset estimation, generates fused oral case data units with operational semantic consistency, and drives hash resignation to achieve tamper-proof on-chain evidence storage update, constructing a trusted closed-loop system of "perception-alignment-correction-evidence storage". This architecture does not require complex parameter tuning, has good interpretability and clinical scalability, and is especially suitable for digital diagnosis and treatment scenarios with multiple devices coexisting, highly personalized operation processes, and regulatory sensitivity, providing solid technical support for the authenticity verification and quality traceability of medical processes. Attached Figure Description

[0007] Figure 1 This is a flowchart of a method for generating oral medical records that integrates voiceprint features according to the present invention; Figure 2 This is a sub-flowchart of a method for generating oral medical records by incorporating voiceprint features according to the present invention. Figure 3 This is another sub-flowchart of a method for generating oral case records that integrates voiceprint features according to the present invention. Detailed Implementation

[0008] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0009] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0010] like Figure 1 As shown, this invention provides a method for generating oral medical records by fusing voiceprint features. The method involves acquiring the voiceprint signals of diagnostic instruments operating in an oral treatment scenario, constructing a voiceprint event sequence based on the voiceprint signals, obtaining a structured diagnostic operation log corresponding to the voiceprint event sequence to obtain a medical record event sequence, performing cross-modal matching and alignment between the voiceprint event sequence and the medical record event sequence to correct the timestamps of the medical record event sequence, and encrypting and binding the timestamp-corrected medical record event sequence with the voiceprint event sequence to generate a fused oral medical record data unit. Specifically, the method includes the following steps: S1: Audio signals emitted by various diagnostic and treatment instruments during oral treatment are collected in real time through a high-sensitivity sound pickup device, and millisecond-level frame processing is performed through a low-latency edge computing node to generate multi-device voiceprint signals with timestamps. S2: After performing a short-time Fourier transform on the multi-device acoustic signature signal, a complex spectral tensor is constructed, and the first derivative sequence of the phase difference between adjacent frames is calculated along the time axis to generate a phase evolution trajectory sequence characterizing the mechanical state transition of the device. S3: Construct a set of directed edges based on the key turning points in the phase evolution trajectory sequence, and map it to a typical operation stage template library to generate an acoustic fingerprint event sequence containing device acoustic fingerprint information; S4: Obtain the structured diagnosis and treatment operation log corresponding to the voiceprint event sequence, parse it into standardized coded operation event nodes and inject time window confidence intervals to generate a case event sequence with clinical pathway constraints; S5: Input the case event sequence and the voiceprint event sequence into a lightweight graph attention mechanism, calculate the joint similarity score of nodes and edges, and output the cross-modal alignment path and time offset estimate of the optimal matching result; S6: Based on the estimated time offset in the cross-modal alignment path, automatically correct the original timestamps in the medical record events to generate a timestamp-corrected case event sequence with operational semantic consistency; S7: Encrypt and bind the timestamp-corrected case event sequence with the voiceprint event sequence to generate a unique and anti-counterfeiting fused oral case data unit; S8: Generate a unique data identifier for the fused oral case data unit using a hash algorithm, and upload it to the blockchain network for distributed storage, generating an immutable blockchain storage hash value.

[0011] Step S1: Audio signals emitted by various diagnostic and treatment instruments during oral treatment are acquired in real time using a high-sensitivity audio pickup device, and then processed in millisecond-level frame segmentation using a low-latency edge computing node to generate multi-device voiceprint signals with timestamps. Specifically, this includes: S1.1: Acquire the mixed acoustic vibrations generated by various diagnostic and treatment instruments such as high-speed handpieces, ultrasonic bone scalpels, and root canal treatment motors during operation in the oral treatment scenario, and use a high-sensitivity broadband audio pickup device to perform acoustic-to-electric conversion processing on the mixed acoustic vibrations to generate an original analog audio signal containing acoustic fingerprint information of multiple devices; When various diagnostic and treatment instruments such as high-speed handpieces, ultrasonic bone scalpels, and root canal treatment motors are in operation in the oral treatment setting, the mixed sound wave vibration signal generated by the sound source is used as the input object, which has high-frequency mechanical vibration and low-frequency background mixed components. A high-sensitivity wideband pickup device (frequency response range: 20Hz~20kHz, equivalent noise level ≤15dBSPL) is used to realize the full-frequency domain acoustic-electric conversion function of the mixed sound wave vibration, and maintain a linear amplitude response inside the transducer module to ensure the amplitude and phase consistency of the acoustic fingerprint. Furthermore, by using a multi-point array pickup arrangement method (array spacing: 50mm), spatial coherence granularity sound source capture is achieved, and a synchronous analog audio signal matrix composed of electrical signals from each channel of the array is generated as the basis for parallel input of subsequent signal processing. Furthermore, a low-distortion preamplifier circuit (total harmonic distortion THD ≤ 0.002%, gain set to 20dB) is used to enhance the amplitude and match the impedance of the analog audio signal matrix, ensuring that the weak device voiceprint signal maintains an effective dynamic range in the subsequent A / D conversion. Furthermore, based on a broadband anti-aliasing filter (cutoff frequency 22kHz, filter order 8), high-frequency suppression processing of the analog audio signal matrix is ​​achieved, and a band-limited analog signal that satisfies the Nyquist sampling theorem is obtained, providing bandwidth control for the subsequent digitization process; Through the above-mentioned acoustic-electric conversion and front-end processing methods, the mixed acoustic wave vibration is converted into a raw analog audio signal containing acoustic fingerprint information of all devices, so as to realize the synchronous capture and structured sound source marking of the operating status of multiple devices. For example, in a dental clinic, when a high-speed handpiece (300,000 rpm), an ultrasonic bone scalpel (25 kHz operating frequency), and a root canal treatment motor (2 kHz operating frequency) are operating simultaneously, a three-channel broadband condenser microphone array (with high pickup sensitivity) is arranged. With a maximum sound pressure level of 120 dB SPL (32 dBV / Pa), the array spacing is 50 mm, forming a spatially coherent optimized pickup layout. Each channel output is processed by a low-noise preamplifier (20 dB gain, equivalent input noise 1.2 nV / √Hz) with JFET input, then enters an eighth-order Butterworth anti-aliasing filter (cutoff frequency 22 kHz, transition bandwidth 2 kHz). The filtered analog signal matrix is ​​then sent to a 24-bit resolution multi-channel A / D conversion module (sampling frequency 48 kHz) for digitization. At this point, the amplitude response of the original analog audio signal maintains a flatness of ±0.1 dB across the entire frequency domain, and the phase difference between array channels does not exceed [a certain value] across the entire frequency domain. This ensures the accuracy of subsequent voiceprint feature extraction and device status recognition. The output raw analog audio signal successfully captures the high-frequency whistling peak of a high-speed mobile phone (approximately 10kHz~12kHz), the harmonic components of an ultrasonic bone scalpel (25kHz fundamental frequency and its harmonics), and the low-frequency periodic waveform of a root canal treatment motor (2kHz), achieving complete recording of acoustic fingerprints from multiple devices; S1.2: Perform analog-to-digital conversion and pre-emphasis filtering on the original analog audio signal, and remove environmental background noise interference based on an adaptive noise suppression algorithm to generate digital multi-channel audio stream data with enhanced signal-to-noise ratio; S1.3: Based on the hardware clock synchronization mechanism of low-latency edge computing nodes, the digital multi-channel audio stream data is subjected to millisecond-level sliding window truncation processing, and an overlapping framing strategy of 20 millisecond frame length and 10 millisecond step length is adopted to generate an audio signal frame sequence with discrete time slice attributes. A sliding window truncation method based on hardware clock synchronization mechanism is adopted for digital multi-channel audio stream data (parameter: node built-in crystal oscillator frequency error ≤1ppm) to delineate the boundaries of continuous time domain segments with millisecond-level precision. Furthermore, by performing temporal segmentation through an overlapping framing strategy (parameters: frame length 20 milliseconds, frame shift 10 milliseconds), a frame sequence sharing half of the sampling points between adjacent frames is generated, resulting in a set of continuous signal segments with enhanced coverage. Furthermore, a cache queue management mechanism (parameter: queue depth = 5 frames) is used to sort the frame sequence after the sliding window is truncated in a first-in-first-out manner, so as to achieve a balance between local cache and real-time output and generate discrete time domain segment data with stable frame order. Furthermore, based on the sampling rate precise matching rule (parameter: sampling frequency 48kHz), the sample index range of each frame is calculated, and the frame start index is determined by the following formula: in, Sampling frequency, Frame shift length (milliseconds); Furthermore, by iteratively calculating the starting index, a complete set of frame start and end indexes is generated, resulting in an audio signal frame sequence with discrete time slice properties. By using this sliding window truncation and overlapping frame processing, the multi-channel digital audio stream data from the previous step is transformed into a frame sequence with precise boundaries, discrete time slices, and can be used for subsequent sound source separation and frequency domain analysis, thus achieving the expected technical effect of temporal feature extraction. For example, in a dental treatment scenario where a high-speed mobile phone and an ultrasonic bone scalpel are operating concurrently, the sampling frequency of the digital audio stream is set to 48kHz, and the crystal oscillator frequency error calibration of the edge computing node is 0.8ppm. The frame length is configured to 20 milliseconds, corresponding to 960 sampling points per frame, and the frame shift is configured to 10 milliseconds, corresponding to 480 frame shift sampling points. According to the formula: The calculated frame start index increment is 480 sample points. With a buffer queue depth of 5 frames, the system can simultaneously retain 5 frames of data for real-time processing, preserving continuity while ensuring output latency remains stable within 87 milliseconds. After this processing step, the generated discrete frame sequence can effectively reduce spectral leakage in the subsequent short-time Fourier transform stage and provide a high-precision time slice input for the spatial filtering algorithm for device sound source separation, significantly improving the stability of voiceprint feature extraction. S1.4: Inject a high-precision global timestamp into each frame of data in the audio signal frame sequence, and use the real-time operating system kernel of the edge computing node to bind and map the nanosecond-level system time with the frame index to generate a set of stamped audio signal frames with absolute time coordinates. For each frame of data in the audio signal frame sequence, a high-precision global timestamp marking algorithm (precision parameter: nanosecond level) is used to implement the function of binding the frame index with the system absolute time; Furthermore, through the highly stable hardware clock synchronization mechanism built into the edge computing node (synchronization source: GPS timing or IEEE1588 precision time protocol), the nanosecond-level system time is periodically corrected in the real-time operating system kernel, and the timing coordinate value that completely corresponds to the sampling start point of the current frame is obtained; Furthermore, the task scheduler of the real-time operating system kernel is used to call the high-resolution time acquisition interface (call interval: ≤1μs) to realize the mapping between frame index and precise time value, and to generate a mapping table structure from frame index to absolute time for subsequent data encapsulation; Furthermore, a consistency check is performed through the time mapping table structure (check standard: time drift is allowed ≤50ns) to realize the logical consistency verification of the binding results, and output the time-stamped dataset of the check-passed frames that meet the accuracy requirements; By using a stamped encapsulation process, the result of the previous step is transformed into a set of stamped audio signal frames with absolute time coordinates, thereby achieving precise synchronization of voiceprint acquisition data from multiple devices under a global time reference. For example, for a digital audio signal frame sequence with a sampling rate of 48kHz, a frame length of 960 sampling points, and a frame step of 480 sampling points, a GPS time source is used to provide an absolute time reference for the hardware clock. The absolute time value of the frame start point is obtained by calling the nanosecond-level time interface through the real-time operating system. For example, if the frame index is 125, the corresponding time value is... Nanoseconds. This time value is bound to the metadata field of frame index 125 and recorded in the mapping table. → The consistency check confirms that the binding drift value is less than [the specified value]. Nanoseconds. During continuous processing, the generated set of stamped audio signal frames covers all frame indices and corresponding absolute times. Subsequent sound source separation and frequency domain analysis can directly utilize this set to achieve high-precision time synchronization and semantic-level matching of voiceprints from multiple devices during diagnosis and treatment. S1.5: Based on the device sound source separation preprocessing logic of the set of stamped audio signal frames, the audio frames corresponding to different diagnostic and treatment instruments are classified and aggregated according to device type to generate multi-device voiceprint signals with timestamps, which serve as the direct input objects for subsequent short-time Fourier transform to construct complex spectral tensors.

[0012] Step S2: After performing a short-time Fourier transform on the multi-device acoustic signature signals, a complex spectral tensor is constructed. The first derivative sequence of the phase difference between adjacent frames is calculated along the time axis to generate a phase evolution trajectory sequence characterizing the mechanical state transitions of the devices. Specifically, this includes: S2.1: The multi-device voiceprint signal with timestamp is subjected to windowing and framing processing. The Hanning window function is used to truncate continuous short-time signal segments to generate a discrete time-domain signal frame sequence with overlapping characteristics, so as to eliminate the spectral leakage effect caused by signal truncation and establish the basic data unit for subsequent frequency domain analysis. S2.2: Perform a fast Fourier transform operation based on the discrete time-domain signal frame sequence to map the time-domain amplitude information to the frequency domain space and generate a complex spectral tensor containing real and imaginary part information, so as to completely preserve the original data of amplitude intensity and phase angle of each frequency component at a specific time. S2.3: Perform inter-frame phase difference calculation on the complex spectrum tensor along the time axis, extract the argument difference between the complex vectors of the two frames at the same frequency bin, and generate a phase difference sequence representing the instantaneous frequency offset to quantify the microscopic change trend of mechanical vibration state during equipment operation. S2.4: Perform first-order derivative operation based on the phase difference sequence to calculate the dynamic gradient of the phase change rate over time, and generate a phase evolution rate sequence that characterizes the acceleration of the mechanical state transition of the device, so as to highlight the drastic phase change characteristics generated by the high-speed mobile phone or ultrasonic bone scalpel at the moment of switching during the operation phase. S2.5: Perform extreme point detection and smoothing filtering on the phase evolution rate sequence to remove high-frequency noise interference and lock the key state transition positions, generating the final phase evolution trajectory sequence that characterizes the mechanical state transition of the equipment, which serves as the basis for constructing the unique acoustic fingerprint of the acoustic event sequence node. Extreme point detection is performed on the phase evolution rate sequence data input (parameter: the detection threshold is set to 0.75 times the peak amplitude of the signal) to achieve preliminary identification of the location of significant phase change during the transition of the mechanical state of the equipment; Furthermore, by using an extreme value classification method based on the first derivative zero-crossing determination (parameter: zero-crossing window length is 3 frames), the detected extreme value points are classified into peak and valley categories, and a classified extreme value dataset containing time index, peak and valley type and phase change rate amplitude is obtained. Furthermore, an adaptive smoothing filter method is used (parameters: sliding window length is 5 frames, weight distribution uses Gaussian kernel). With a value of 2, the amplitude smoothing process is applied to the classification extreme value dataset, and a smooth extreme value trajectory after high-frequency noise suppression is generated; Furthermore, a dynamic threshold update strategy (formula below) is used to remove noise and retain key extrema in smooth extreme value trajectories: in, This is the mean of the current smoothed extreme value trajectory. The standard deviation of the current smooth extreme value trajectory. This is a dynamic threshold value, which is used to eliminate noise extrema and retain only critical state inflection points that exceed the threshold. Furthermore, by reordering the time index and associating the phase change rate amplitude with the encoding method (parameter: amplitude encoding range 0–255), an ordered key point sequence is formed from the retained key inflection points, generating the final phase evolution trajectory sequence that characterizes the transition of the equipment's mechanical state. By detecting extreme points and smoothing filtering, the phase evolution rate sequence results of the previous step are transformed into phase evolution trajectory data with key state transition information and low noise interference, so as to achieve the expected technical effect of constructing unique acoustic fingerprints of voiceprint event sequence nodes. For example, in the voiceprint data processing of high-speed mobile phones, the input phase evolution rate sequence has a frame length of 20 milliseconds, a frame shift of 10 milliseconds, and a detection threshold set to 0.75 times the peak amplitude. An extreme value classification strategy with a zero-crossing window length of 3 frames is used to obtain a total of 38 peak and valley value categories. A Gaussian kernel is applied. An adaptive smoothing filter with a value of 2 is applied, resulting in a smoothed mean. The standard deviation is 12.4. The value is 3.6. Based on the formula, the dynamic threshold T is calculated as follows: Right now After removing extreme points with amplitudes below 17.8, 21 key state inflection points were retained. The amplitudes were mapped to the encoding range of 0–255 to form an ordered key point sequence. The final generated trajectory only contains the significant phase change positions when the high-speed mobile phone switches between the tooth preparation and decorsis stages. This trajectory was used to construct the voiceprint event sequence node and significantly improved the accuracy of device stage recognition in subsequent cross-modal alignment.

[0013] like Figure 2 As shown, step S3 involves constructing a set of directed edges based on key inflection points in the phase evolution trajectory sequence and mapping it to a typical operation stage template library to generate an acoustic fingerprint event sequence containing device acoustic fingerprint information. Specifically, this includes: S3.1: Perform extreme point detection processing on the phase evolution trajectory sequence to extract the set of key turning point coordinates that characterize the sudden change in the mechanical state of the equipment, and generate a candidate node dataset containing timestamps and phase difference derivative amplitudes; For the phase evolution trajectory sequence, a global extremum detection method based on smooth envelope is used (parameters: window length = 5 frames, detection threshold = ). This enables the preliminary identification of locations where amplitude abruptly changes in the trajectory curve; Furthermore, by using a local third-order polynomial fitting method (parameter: fitting interval = ±2 frames), the gradient change rate at the mutation location is accurately quantified, and the precise numerical representation of the mutation amplitude is obtained. Furthermore, a dual-threshold determination method is adopted (high threshold = 0.85 times the amplitude of the main mutation, low threshold = 0.45 times the amplitude of the main mutation) to effectively screen the initially identified extreme points and generate a set of high-confidence extreme points containing timestamp and amplitude information. Furthermore, based on the adjacent difference calculation method of timestamp sequence, the time sparsity of the extreme point set is checked, redundant extreme points with time intervals less than a preset lower limit are removed, and structured key turning point time series data are obtained. By using a structured encapsulation process, amplitude information is paired and mapped with timestamps to generate a candidate node dataset containing timestamps and phase difference derivative amplitudes, enabling accurate node localization for subsequent construction of directed edge sets; For example, in the phase evolution trajectory sequence of a high-speed mobile phone voiceprint signal, the signal sampling rate was set to 48000Hz and the trajectory length was 200 frames. Using a smooth envelope extremum detection method with a window length of 5 frames and a detection threshold of 3 times the standard deviation, 17 amplitude abrupt change locations were initially identified. Gradient change rate was quantified by local third-order polynomial fitting (fitting interval ±2 frames), revealing 10 high-confidence extremum points with amplitudes exceeding 0.85 times the main abrupt change amplitude, and 4 auxiliary extremum points with amplitudes between 0.45 and 0.85 times the main abrupt change amplitude. A dual-threshold judgment algorithm was used to filter the initial extremum points, retaining high-confidence extremum points and necessary auxiliary extremum points. A temporal sparsity test was performed (with a lower time limit set to 15ms), eliminating redundant extremum points with adjacent time intervals less than 15ms, ultimately yielding 12 key inflection points. By encapsulating the timestamps of each key turning point (from the trajectory frame index multiplied by the frame shift time) and the magnitude of the phase difference derivative, a candidate node dataset is formed. Each node in the dataset contains two fields: absolute time (in ms) and magnitude (in rad / s), which can be directly used for template matching in the construction and operation stages of directed edge sets.

[0014] Using a mutation identification formula, the amplitude change rate of the trajectory is analyzed. Make a judgment: in, For phase value, To detect the current phase value within the window, The initial phase value of the window. This is the current timestamp. This is the window start timestamp. The calculation result can be used to determine the threshold of abrupt change magnitude, enabling accurate identification of key turning points; S3.2: Perform temporal adjacency analysis based on the candidate node dataset, calculate the state transition probability between adjacent key inflection points using the sliding window algorithm, and generate a set of directed edges describing the evolution direction of the device's operating state. S3.3: Obtain a pre-built template library of typical operation stages, perform subgraph isomorphic matching operation on the directed edge set to identify the specific diagnosis and treatment operation stage type to which the current voiceprint signal belongs, and generate a preliminary mapping result with stage labels; The subgraph isomorphism detection method (parameters: adjacency matrix encoding, inflection point coordinate threshold) is used on the directed edge set to search for substructures that completely match the template library storage pattern from the current device state transition mode; Furthermore, by using a VF2-based optimization matching method (parameters: node label weight 0.75, edge direction consistency weight 0.25), the joint constraint verification of node features and edge directions during the matching process is realized, and a set of candidate subgraphs that satisfy both topological shape and node feature consistency is obtained. Furthermore, a topological similarity calculation method (formula: Jaccard coefficient) is used to score the similarity between the candidate subgraph set and the typical stage subgraphs in the template library, and a ranking result is generated. The formula is as follows: in, Let the set of edges of the current subgraph be . The set of edges of the template subgraph. Score the similarity. Furthermore, by mapping the stage number corresponding to the highest-scoring template sub-image to the diagnostic operation stage type code through the stage identifier mapping table, the stage to which the voiceprint signal belongs can be identified. Furthermore, a confidence index for the recognition results is calculated using a confidence estimation method (parameter: ratio of the number of matching edges to the total number of template edges), and preliminary mapping data with stage labels and confidence scores is generated. By using subgraph isomorphic matching and similarity scoring, the set of directed edges from the previous step is transformed into a preliminary mapping result containing stage labels and confidence indices, thereby achieving high-precision identification of the diagnostic stage type of voiceprint signals. For example, in the high-speed mobile phone voiceprint signal processing scenario, the template library pre-constructs a typical edge set for the "tooth preparation" stage, with 5 nodes and 6 edges. The node labels are derived from the amplitude range [0.8, 1.2] of the key inflection points of the phase evolution trajectory. During matching, the VF2 algorithm sets a node label weight of 0.75 and an edge direction consistency weight of 0.25 to ensure that the amplitude difference of the matched nodes does not exceed 0.1 and that all edge directions are consistent. When calculating the Jaccard coefficient, the intersection of the candidate subset and the template subgraph has 5 edges, the template has 6 edges, and the candidate has 6 edges. The formula is as follows: The similarity value obtained was 0.833. The mapping table mapped this template stage number to the SNOMED CT code "234567", and calculated the confidence level as 0.833 based on the matching edge ratio. The initial mapping result was stage type "tooth preparation", code "234567", and confidence level of 0.833. This result was bound to the directed edge set in the subsequent acoustic fingerprint feature fusion stage, significantly improving the accuracy of cross-modal alignment. S3.4: Based on the preliminary mapping results, perform acoustic fingerprint feature fusion, bind the matched operation stage type identifier with the topology of the directed edge set, and generate a structured acoustic fingerprint event sequence containing device-unique acoustic fingerprint information; Based on the preliminary mapping results, a topology binding method (parameters: stage label, node index, edge weight) is used to associate and map the diagnostic and treatment operation stage type identifier with the topology of the directed edge set. Furthermore, by using the node attribute injection method (parameters: unique acoustic fingerprint encoding, timestamp precision value), feature fields representing the uniqueness of the device's acoustic fingerprint are embedded in the node structure, and enhanced node data containing stage labels and phase evolution features is obtained. Furthermore, by using the edge attribute weighting operation method (parameters: state transition probability, phase difference change amplitude), a multi-dimensional weight vector is added to each edge in the directed edge set to generate a weighted edge set that simultaneously possesses temporal stability and voiceprint feature sensitivity. Furthermore, by using a graph data structure encapsulation method (parameters: node set, edge set, metadata tag), the enhanced node data and weighted edge set are assembled into a structured graph object, and an acoustic event sequence containing device-unique acoustic fingerprint information is generated. By using topology binding and attribute fusion algorithms, the subgraph isomorphic matching results from the previous step are transformed into a data graph with semantic and device voiceprint unique identifiers for the diagnosis and treatment stage, thereby achieving structural completeness and traceability of voiceprint event sequences. For example, the stage label "Opreparation_stage" in the initial mapping result is bound to the node index sequence {N1, N3, N5}, and a unique acoustic fingerprint code "ACF-98327" and a nanosecond-level timestamp precision value of 10 are added using the node attribute injection method. -9 Within seconds, enhanced node data is generated. For the directed edge set {(N1→N3),(N3→N5)}, the state transition probabilities are calculated to be 0.85 and 0.91, respectively, with phase difference changes of 0.32 radians and 0.47 radians, respectively. Weight vectors are constructed and appended to the edge attributes. Using a graph data structure encapsulation method, the node set {N1,N3,N5}, the weighted edge set, and the stage label metadata are assembled into a voiceprint event sequence object. The final generated graph has a node feature tensor dimension of 128 dimensions and an edge feature tensor dimension of 64 dimensions, achieving a significant improvement in stage recognition accuracy in cross-modal matching tests.

[0015] like Figure 3 As shown, step S4 involves: obtaining the structured diagnostic and treatment operation log corresponding to the voiceprint event sequence, parsing it into standardized coded operation event nodes, injecting time window confidence intervals, and generating a case event sequence with clinical pathway constraints. Specifically, this includes: S4.1: Obtain the structured treatment operation log data stream stored in the oral treatment system, and use a natural language processing interface or device communication protocol parser to extract fields and standardize the format of the structured treatment operation log data stream to generate a non-standardized operation event sequence containing the original operation description, the executing physician identifier and the initial recording time; Data extraction processing is performed on the structured treatment operation log data stream stored in the oral treatment system. The input conditions are a multi-source treatment log set containing information such as physician input, device API reporting, and treatment stage annotation, and have original timestamp and operation description fields. A device communication protocol parser (parameters: supports HL7, DICOM SR and vendor proprietary protocols) is used to implement field-level decomposition of log data streams, transforming raw binary or semi-structured data into an accessible set of key-value pairs; Furthermore, through the natural language processing interface (parameters: using a medical-specific BERT model, word vector dimension 512, context window length 128), multi-layer semantic word segmentation and named entity recognition processing of the operation description field are realized, and a set of semantic tags containing operation verbs, operation objects and stage identifiers are obtained. Furthermore, by using a regular template matching method (parameter: the rule base covers the entire process operation mode such as tooth preparation, root canal cleaning, and filling), a structured pattern mapping of the extracted semantic tag set is realized, and a classified operation type index value is generated to facilitate subsequent encoding mapping; Furthermore, a multi-source timestamp consistency correction method (parameters: time offset tolerance ±1.5 seconds, priority weight physician input > device reporting > system default) is adopted to achieve unified processing of recorded times from different sources and generate a comparable UTC standard time format initial recorded time value; By extracting fields and standardizing formats, the parsing results from the previous step are transformed into a non-standardized sequence of operation events containing the original operation description, the physician's identifier, and the initial recording time, thereby achieving the semantic foundation construction of events on the medical record side. For example, during a root canal treatment, the log data stream generated by the treatment system includes a high-speed mobile phone startup event (reported by the device API, recorded at 14:03:15), a physician manually inputting "Start Corrosion Removal" event (manually entered, recorded at 14:03:16), and a root canal treatment motor switching event (reported by the device API, recorded at 14:05:02). The device communication protocol parser is configured to support HL7 v2.8 and the manufacturer's proprietary JSON format, generating key-value pairs after parsing, such as {"Operation Description":"Start Corrosion Removal","Physician ID":"D102","Timestamp":"2024-05-12T14:03:16Z"}. The natural language processing interface uses a medical BERT model to identify "Start Corrosion Removal" as the verb "removal", the stage identifier "start", and the named entity as the standard surgical step. The regular expression template matching algorithm classifies this event as the "corrosion removal operation" type index value O_03. The timestamp consistency correction algorithm detected a 1-second difference between the high-speed mobile phone startup event and the physician's "start decorpation" record. Within the set tolerance, and given that the physician's entry priority is higher than the device's reporting priority, the timestamps of both were uniformly corrected to the 14:03:16 UTC standard format. The final generated non-standardized operation event sequence includes O_03 type events, bound to physician ID D102 and the initial recording time 14:03:16, serving as input data for subsequent SNOMED CT mapping and time window injection. Verification results show that the event parsing is accurate and the time correction matches the voiceprint stage. S4.2: Based on the original operation description in the non-standardized operation event sequence, the preset SNOMEDCT medical terminology mapping engine is invoked to perform semantic normalization processing, converting the non-standardized operation description into a globally unified standardized coded operation event identifier, so as to generate a set of standardized coded operation event nodes with interoperability. For the original operation descriptions in the non-standardized operation event sequence, a medical semantic embedding method based on bidirectional encoder representations (parameters: medical domain-specific BERT model, word vector dimension 768) is used to extract contextual semantic features of the non-standardized descriptions and generate high-dimensional semantic vectors containing operation keywords, surgical step phrases and related modification information. Furthermore, by using a standardized medical terminology dictionary matching method (parameters: SNOMED CT terminology database version 2023, fuzzy matching threshold 0.85), the cosine similarity between the semantic vector and the terminology database entries is calculated, and a set of candidate standard terms and corresponding similarity scores are obtained for each original description. Furthermore, a multi-condition decision fusion method (parameters: similarity score weight 0.6, character edit distance weight 0.4) is adopted to sort the candidate term set by the comprehensive score of similarity score and edit distance, and generate the standardized coding operation event identifier with the highest comprehensive score; Furthermore, through a semantic ambiguity resolution method (parameters: context-dependent window size 3, event sequence probability model), we can perform context verification on multiple possible mappings for the same original description and output a unique standardized code that conforms to the semantic consistency of the global diagnosis and treatment path. The standardized coding generation module transforms the unique standardized coding operation event identifier into a set of standardized coding operation event nodes with interoperability, thereby achieving semantic consistency of diagnosis and treatment events across systems and institutions. For example, in a dental treatment record, the original procedure description is "root canal dilation". The input non-standardized description is encoded using a medical BERT model, resulting in a 768-dimensional semantic vector. The SNOMED CT v2023 mapping engine is invoked to retrieve candidates with a similarity greater than 0.85 from the terminology database, including "Root canal widening procedure" (similarity 0.92) and "Root canal treatment" (similarity 0.87). The edit distance algorithm calculates the character distance between "root canal dilation" and the standard English terminology, yielding an edit distance score of 0.95 for "Root canal widening procedure". In the multi-condition decision fusion, the similarity score and edit distance score are weighted at 0.6:0.4 to synthesize a total score of 0.932. The context-dependent window contains three operation event nodes before and after the procedure. The probabilistic model determines that this event is adjacent to the existing "root canal cleaning" and must precede "root canal filling". The unique match is the SNOMED CT code "23456002" for "Root canal widening procedure". This set of encoded nodes contains timestamps, physician information, and is stored as case event sequence nodes, enabling a consistent semantic representation of root canal dilation steps across systems; S4.3: For each standardized coded operation event identifier in the set of standardized coded operation event nodes, calculate and allocate the corresponding time window confidence interval parameters according to the standard operation time distribution model defined in the oral clinical diagnosis and treatment guidelines, so as to generate an operation event node with a confidence interval and dynamic temporal fault tolerance boundary. S4.4: Based on the operation event nodes with confidence intervals, the predefined oral treatment clinical pathway constraint rule base is referenced to perform logical relationship deduction, and a set of directed logical edges describing the order of operations, dependencies and mutual exclusion conditions is constructed to generate a clinical pathway constraint relationship graph structure containing strict medical logic norms. A rule-based logical deduction method (parameter: oral treatment clinical pathway constraint rule base) is applied to the set of operation event nodes with confidence intervals to realize the sequential constraint parsing between diagnosis and treatment events; Furthermore, by using a temporal logic verification method (parameters: confidence interval boundaries, event duration), the legality of the arrangement of events on the timeline is judged, and temporal constraint data that meets clinical standards is obtained. Furthermore, by using a dependency analysis method (parameter: a set of preconditions in the rule base), the operation dependency chain between events is constructed, and a directed edge dataset containing dependency weights is generated. Furthermore, by using a mutual exclusion condition detection method (parameter: mutual exclusion operation pairs in the rule base), conflict identification is achieved for events that occur simultaneously or overlap in time, and a set of mutually exclusive relationship edges is generated; By using a topology construction method, sequential constraint data, dependency edge datasets, and mutually exclusive edge sets are merged into a set of directed logical edges, thereby generating a clinical pathway constraint graph structure. For example, in a structured medical record operation event, the event nodes include three categories coded by SNOMED CT: {"tooth preparation"}, {"cartilage removal"}, and {"filling"}, with time window confidence intervals of ±1.0s, ±0.8s, and ±0.9s, respectively. The sequence constraints defined in the rule base are: "tooth preparation" must precede "cartilage removal", "cartilage removal" must precede "filling", the dependency relationship is that "filling" depends on the completion of "cartilage removal" and "tooth preparation", and the mutual exclusion condition is that "cartilage removal" and "root canal cleaning" are mutually exclusive. When performing temporal logic verification on the node set, the formula is used... Calculate the validity of the average interval between events, where To determine the number of events, their sequential relationships were verified to satisfy the rule base constraints. During dependency analysis, for the "filling" node, the weight of the preceding "tooth preparation" was set to 2.0, and the weight of "caries removal" was set to 1.5 using a weight assignment method. No mutual exclusion conflicts were found during mutual exclusion condition detection. Finally, the topology construction generated a clinical path constraint graph containing three nodes, two sequential constraint edges, and one merge dependency edge. This graph was applied to generate case event sequences, and the output graph ensured that the diagnosis and treatment process was complete and conformed to medical logic norms. S4.5: The operation event node with confidence interval is used as a graph node and the clinical path constraint relationship graph structure is used as a graph edge for topological fusion and encapsulation to generate a case event sequence with clinical path constraint relationship, which serves as the standard input object for the medical record side for subsequent cross-modal alignment with the voiceprint event sequence; The set of operational event nodes with confidence intervals is input into the topology fusion algorithm module (parameters: node set capacity limit 1024, edge set capacity limit 2048) to initialize the medical record event nodes as vertices in the graph structure. Furthermore, by using the directed edge inserter in the topology fusion algorithm (parameters: edge weight values ​​range from 0 to 1, weight assignment is based on the clinical pathway priority matrix), the set of directed edges in the clinical pathway constraint relationship graph structure is embedded into the current graph structure, thereby binding the logical dependency relationship between nodes and edges and obtaining a preliminary fusion graph matrix; Furthermore, a graph adjacency matrix normalization method (parameter: normalization factor takes the maximum node degree value) is adopted to optimize the structure of the preliminary fused graph matrix and generate a standardized topological representation that has both time window confidence interval and clinical pathway logic. Furthermore, a graph structure consistency verification method (parameter: mutual exclusion condition weight threshold 0.75) is used to verify the consistency of nodes and connections in the standardized topology representation and generate a valid topology dataset that excludes those that violate clinical pathway rules. By using graph serialization encapsulation (parameters: serialization protocol adopts JSON-LD extended format, time window field precision is in milliseconds), the legitimate topological dataset is transformed into the core structure of case event sequences, achieving standardized input effect of case event sequences in cross-modal alignment; For example, in a dental clinic scenario, the set of operation event nodes with confidence intervals has 512 nodes, and the clinical pathway constraint graph contains 1280 directed edges. The weights of the directed edges are calculated based on the priority matrix and range from 0 to 1. After initializing the graph structure vertices using a topology fusion algorithm, the 1280 directed edges are inserted into the fusion graph to form a preliminary fusion matrix with an adjacency matrix of size 512×512. During normalization, the maximum node degree is set to 16, and the normalization factor is set to 16. Each row element in the adjacency matrix is ​​divided by the normalization factor to adjust the edge weight magnitude. During structural consistency verification, 35 edges that violated the mutual exclusion condition and had weights exceeding 0.75 were detected. These edges were removed, and a valid topology dataset was generated. During serialization and encapsulation, JSON-LD format was used for output, and each time window field retained three milliseconds of precision. The final case event sequence achieved complete semantic matching between the node and edge sequences and the voiceprint event sequence in cross-modal alignment testing, with the time offset controlled within 0.5 seconds, meeting the performance requirements of the real-time evidence storage system.

[0016] Step S5: Input the case event sequence and the voiceprint event sequence into a lightweight graph attention mechanism, calculate the joint similarity score of nodes and edges, and output the cross-modal alignment path and time offset estimate of the optimal matching result. Specifically, this includes: S5.1: Obtain the set of phase evolution trajectory sequence nodes in the voiceprint event sequence and the set of standardized coding operation event nodes in the case event sequence. Perform vector mapping processing on the set of phase evolution trajectory sequence nodes based on the dual-channel graph embedding algorithm to generate a voiceprint node feature tensor, and perform semantic coding processing on the set of standardized coding operation event nodes to generate a medical record node feature tensor. This results in obtaining a voiceprint node feature tensor and a medical record node feature tensor with a unified dimensional space as the input basis for subsequent calculations. Obtain the node set of the phase evolution trajectory sequence in the constructed voiceprint event sequence and the standardized coding operation event node set in the case event sequence as the original input data object of the dual-channel graph embedding algorithm; A multi-dimensional temporal feature embedding method (embedding dimension: 64, normalization method: Z-score) is adopted to perform vector mapping processing on the set of nodes of the phase evolution trajectory sequence, so as to encode the phase difference derivative amplitude, abrupt frequency bin index and timestamp information of each voiceprint node into a feature vector of uniform length, so as to eliminate the inconsistency of the dimension of different nodes in the feature space. Furthermore, by using the local phase gradient coding method (window length: 3 frames, gradient threshold: 0.15 rad / frame), the local phase change rate between nodes is introduced as an additional feature component during the mapping process, which improves the sensitivity of the embedded vector to the semantics of stage transitions and generates the voiceprint node feature tensor. A standardized medical semantic embedding method (encoding method: SNOMED CT one-hot encoding + embeddinglookup, embedding dimension: 64) is used to perform semantic encoding processing on the set of standardized coded operation event nodes in the case event sequence, so as to uniformly map the international standard code, clinical time parameters and confidence interval boundary values ​​of each type of diagnosis and treatment operation to a continuous vector space. Furthermore, by using the clinical pathway sequential position encoding method (position encoding period: 512, encoding amplitude range: 0~1), the encoding vector of the relative position of the node in the clinical pathway is added during the semantic embedding process to strengthen the logical order constraint ability of the medical record node feature tensor in the graph matching process. By using dual-channel feature dimension alignment processing (feature alignment strategy: linear transformation matrix, transformation dimension: 64×64), the embedding dimension of the voiceprint node feature tensor and the medical record node feature tensor is completely unified, ensuring the consistency of the operation matrix structure for subsequent cross-modal computation. Through the above processing of the dual-channel graph embedding algorithm, the original voiceprint node set and medical record event node set are transformed into voiceprint node feature tensors and medical record node feature tensors with a unified dimensional space, respectively, realizing the structural standardization of cross-modal input and providing a stable data foundation for the subsequent calculation of node edge joint similarity. For example, in a scenario involving a mixed operation of tooth preparation and root canal cleaning, the voiceprint event sequence includes a high-speed handpiece node, an ultrasonic bone scalpel node, and a root canal treatment motor node. The phase difference derivative amplitudes of each node are 0.27 rad, 0.31 rad, and 0.18 rad, respectively, with timestamp precision at the nanosecond level. A 64-dimensional embedded temporal feature mapping is used to map the phase difference amplitude, mutation frequency bin index (bin 52), and timestamp of the high-speed handpiece node into a vector containing normalized gradient encoding. If the gradient threshold is set to a certain value, the additional gradient component value is 0.21. The "tooth preparation" node in the case event sequence is encoded using SNOMED CT to obtain the international standard code 456789, with a one-hot vector length of 1024. An embedding lookup generates a 64-dimensional semantic vector, and a positional encoding value of 0.35 (the second position of this node in the clinical pathway) is added. The dual-channel alignment matrix is... After performing matrix multiplication, the dimensions of the two types of feature tensors are guaranteed to be consistent. In this embodiment, both the voiceprint node feature tensor and the medical record node feature tensor are 64-dimensional and are uniformly normalized to a distribution with a mean of 0 and a variance of 1. The numerical stability of the joint similarity calculation of nodes and edges during subsequent cross-modal graph matching is significantly improved, and the calculation time is controlled within 27ms, which meets the real-time alignment requirements. S5.2: Based on the voiceprint node feature tensor, extract the phase difference evolution pattern edge features between adjacent nodes, and based on the medical record node feature tensor, extract the clinical path constraint relationship edge features. Use the graph structure fusion mechanism to perform heterogeneous alignment processing on the phase difference evolution pattern edge features and the clinical path constraint relationship edge features to generate a joint edge feature matrix containing temporal invariance and logical constraints, providing a structured topological association basis for the attention mechanism. S5.3: Utilize a learnable multi-head graph attention network to receive the voiceprint node feature tensor, medical record node feature tensor, and joint edge feature matrix, perform cross-modal node-edge joint attention weight calculation processing, and transform the calculated original attention score into a probability distribution using the softmax normalization function to generate a dynamic attention weight matrix that characterizes the semantic association strength between the voiceprint topology sequence and the case event sequence. For the voiceprint node feature tensor, medical record node feature tensor and joint edge feature matrix, a learnable multi-head graph attention network (parameters: number of heads h≥4, embedding dimension d=128) is used to realize the function of cross-modal node-edge joint attention weight calculation; Furthermore, through an adaptive query key-value mapping mechanism (parameter: the mapping matrix weights are initialized using Xavier random distribution), a unified query space is constructed for the feature vectors of the voiceprint side nodes and the feature vectors of the medical record side nodes, and the original attention scores of each node pair under specific edge feature constraints are obtained. Furthermore, by performing head-by-head grouping normalization mapping (parameter: the number of groups equals the number of heads h), the sensitivity of different perceptual channels to the original attention scores is enhanced, and multiple sets of original attention score matrices are generated; Furthermore, the softmax normalization function (parameter: temperature coefficient) is used. =1.0) Perform probability distribution transformation on the original attention score matrix to normalize the attention score of each candidate node, so that the total weight of the same candidate set is 1; Furthermore, the probability distribution matrices of each channel are merged through multi-head attention fusion processing (parameter: weighting method is equal weight superposition) to generate a dynamic attention weight matrix that represents the semantic association strength between the voiceprint topology sequence and the case event sequence; By using a learnable multi-head graph attention network, the feature tensor and joint edge feature matrix results from the previous step are transformed into a dynamic attention weight matrix, achieving interpretable quantification of cross-modal node-edge associations. For example, in a single medical scenario, the voiceprint node feature tensor is a 128-dimensional floating-point array, the medical record node feature tensor is a 128-dimensional floating-point array, and the joint edge feature matrix is ​​a 128-dimensional matrix containing 64-dimensional temporal constraints and 64-dimensional logical relationship features. The learnable multi-head graph attention network has 8 heads (h=8), an embedding dimension (d=128), and the mapping matrix weights are initialized using Xavier. After adaptive query key-value mapping, each node pair outputs a RAW_SCORE matrix on 8 channels, with each matrix having a dimension of (number of nodes × number of nodes). Temperature coefficient during softmax normalization is also considered. =1.0, the calculated probability distribution satisfies the normalization condition. Finally, the weight matrices of each channel are superimposed with equal weights, and the output is a dynamic attention weight matrix with the same dimension as the input. Its maximum semantic association value is concentrated on the voiceprint event "high-speed mobile phone shutdown" node and the medical record event "dental preparation completed" node pair. The dynamic weight value is greatly improved, which verifies the significant improvement in cross-modal semantic matching efficiency and accuracy. S5.4: Based on the dynamic attention weight matrix, perform weighted aggregation operation on the voiceprint node feature tensor and the medical record node feature tensor, calculate the node edge joint similarity score of each pair of candidate node combinations, and select the matching sequence with the highest cumulative similarity score from all candidate combinations through the maximum likelihood estimation strategy to generate the initial cross-modal alignment path representing the best semantic correspondence. S5.5: Based on the timestamp difference data of each matching node pair in the initial cross-modal alignment path, the dynamic time warping optimization algorithm is applied to iteratively correct the local time deviation, calculate the globally optimal time offset compensation amount, and finally output the cross-modal alignment path containing the precise matching node sequence and the corresponding time offset estimate, so as to complete the precise spatiotemporal synchronization of voiceprint events and medical record events. Based on the timestamp difference data of each matching node pair in the initial cross-modal alignment path, a dynamic time warping optimization algorithm (parameters: window radius = 5, step size = 1) is adopted to realize the iterative correction function of local time deviation. Furthermore, the residual minimization method (parameters: number of iterations = 50, convergence threshold = ...) is used. This enables synchronous optimization of the time difference sequence of matching nodes and obtains the locally dynamically adjusted time difference matrix as an intermediate calculation result. Furthermore, by using a weighted path constraint method (parameter: weight allocation is based on node similarity scores), the local deviation in the time difference matrix is ​​globally correlated and corrected, and a global time difference mapping table that conforms to cross-modal matching topology constraints is generated. Furthermore, based on the global time difference mapping table, the optimal time offset compensation amount for each matching node pair is calculated using the compensation amount solution formula. The formula is as follows: in, Let be the similarity weight of the i-th matching node pair. This corresponds to the time difference. Furthermore, the compensation amount calculated by the above formula is updated iteratively until the global time deviation converges to the target threshold, generating the final time offset compensation amount sequence; By using a dynamic time warping optimization algorithm, the initial cross-modal alignment path result from the previous step is transformed into data containing a sequence of precisely matched nodes and corresponding time offset estimates, thereby achieving precise spatiotemporal synchronization between voiceprint events and medical record events. For example, in a multi-device parallel oral diagnosis and treatment scenario, the acquired voiceprint event sequence and case event sequence are processed by S5.4 to obtain an initial alignment path, with a total of 120 matched node pairs, an average time difference of 3.8 seconds, and a standard deviation of approximately 1.5 seconds. When executing this sub-step, the dynamic time warping window radius is set to 5 frames, the step size is 1 frame, the residual minimization iteration count is 50, and the convergence threshold is... The weights of each node were set to fluctuate between 0.4 and 1.0 using a weighted path constraint algorithm. The initial compensation value was 2.0 seconds, which converged to 1.1 seconds after the 27th iteration. The time difference and weights of each matching node were then substituted into the above formula. The global time compensation was calculated to be 1.1 seconds, and the deviation was stabilized within ±0.42 seconds through iterative optimization. The final output cross-modal alignment path contains a precisely matched node sequence and the corresponding global time offset estimate, providing a high-precision timing reference for the timestamp correction stage of S6.

[0017] Step S6: Based on the estimated time offset in the cross-modal alignment path, automatically correct the original timestamps in the medical record events to generate a timestamp-corrected case event sequence with operational semantic consistency. Specifically, this includes: S6.1: Obtain the time offset estimation sequence of the cross-modal alignment path output and the corresponding case event sequence node identifiers. Use a linear interpolation algorithm to smooth the discrete time offset estimates and generate a dynamic time series compensation curve in the continuous time domain to eliminate time series jump noise caused by sampling jitter. S6.2: Based on the dynamic time-series compensation curve, extract the instantaneous compensation amount that matches each standardized coded operation event node in the case event sequence, and use vector addition to superimpose the instantaneous compensation amount onto the numerical representation of the original timestamp to generate a preliminary corrected candidate timestamp set. S6.3: For the preliminary corrected candidate timestamp set, call the clinical pathway constraint relationship rule base to perform logical consistency verification, identify and remove abnormal timestamps that violate the order of diagnosis and treatment constraints, and generate a subset of legal timestamps that conform to medical logic norms; For the preliminary corrected candidate timestamp set, a clinical pathway constraint logic test method based on rule base references (parameters: oral treatment sequence rule set, mutual exclusion condition mechanism) is adopted to realize the order legality judgment of the treatment events corresponding to each timestamp; Furthermore, the position verification of candidate timestamp sequences in the clinical pathway constraint graph is achieved by using a topological sorting method (parameters: event node set, directed edge order relationship), and node identification data that violate topological constraints are obtained; Furthermore, through a conflict detection algorithm (parameters: conflict type enumeration values ​​include inversion, concurrent conflict, and mutual exclusion violation), abnormal timestamps that violate the order of diagnosis and treatment constraints in the timestamp set are identified, and an abnormal timestamp index list is generated; Furthermore, through anomaly removal processing (parameters: list of anomaly timestamp indices, set of candidate timestamps), all timestamps identified as invalid are removed from the set, and a clean set of valid timestamp candidates is generated; Furthermore, the removed timestamp set is mapped back to the standardized coded operation event node set through a set reconstruction method, generating a valid timestamp subset that conforms to medical logic specifications; By using a clinical pathway constraint-based logical verification and conflict elimination method, the candidate timestamp set from the previous step is transformed into legitimate timestamp data that conforms to the order of diagnosis and treatment, thus achieving the expected technical effect of medical logical consistency verification of medical record time information. For example, for a dental treatment process comprising three stages—high-speed handpiece tooth preparation, root canal cleaning, and heated gutta-percha filling—the preliminarily corrected candidate timestamp set is as follows: [The text abruptly ends here, so the translation stops as well.] Seconds, root canal cleaning node at Seconds, the hot gutta-percha filling node is at Seconds. The order constraint defined in the rule base is: tooth preparation → root canal cleaning → heated gutta-percha filling. A topological sorting algorithm is used to sort the event nodes according to directed edge relationships to obtain the ideal time sequence: seconds → seconds → Within seconds, a conflict was detected between the timestamp of the heated gutta-percha filling node and the root canal cleaning node sequence. Further, the heated gutta-percha filling node timestamp was marked as an anomaly using a conflict detection algorithm, and the anomaly index list contained this node. It was then removed through anomaly removal processing. The time value in seconds is used to reconstruct the remaining set of timestamps, and the valid subset of timestamps is used for tooth preparation. Seconds, root canal cleaning Seconds. In this scenario, the subset of valid timestamps output after this sub-step is executed completely conforms to the order of diagnosis and treatment, effectively eliminating time records in reverse order across stages, and significantly improving the reliability and consistency of medical record time information; S6.4: Based on the aforementioned subset of legitimate timestamps, a confidence-weighted fusion strategy is used to perform optimal calculations on time points with multiple conflicting voiceprint feature mappings, determine a unique final corrected time value, and generate a timestamp-corrected case event sequence with operational semantic consistency. For a subset of valid timestamps, a confidence-weighted fusion method (parameters: confidence weight calculation model, conflict detection threshold) is adopted to achieve the optimal solution for conflicting time points of multiple voiceprint feature mappings. Furthermore, the multi-source confidence of conflicting timestamps is integrated using the Bayesian confidence update method (parameters: prior confidence, likelihood function, posterior normalization coefficient), and a multi-source fused confidence matrix is ​​obtained. Furthermore, a minimum cost matching method is adopted (parameters: cost function is the square of time deviation, constraint is the logical consistency of clinical pathway) to achieve the optimal matching selection of conflict time points in the confidence matrix and generate candidate final time value sequences; Furthermore, based on the candidate final time value sequence, a global consistency optimization method (parameters: global smooth decay factor λ, time series continuity constraint) is applied to achieve global time correction of conflicting time points and generate a uniquely determined final correction time value; The following weighted fusion calculation formula is used to uniquely determine the conflict time point output: in, For the final correction time value, Let be the confidence weight for the i-th voiceprint feature mapping. These are the corresponding timestamp candidate values; Through the above confidence-weighted fusion and optimization process, the results of the legitimate timestamp subset from the previous step are transformed into a uniquely determined final corrected time value, thereby achieving a timestamp-corrected case event sequence with operational semantic consistency. For example, in a multi-device concurrent oral treatment scenario, there are three conflicting time points in the valid timestamp subset that map to the same operation event, namely... Second, Seconds and seconds, and their corresponding confidence weights are configured as follows: , , When calculating the final time value based on the above formula, the parameters are plugged in as follows: The final correction time value can be obtained through calculation. The time correction was validated within seconds and passed under a conflict detection threshold of 0.15 seconds. This result significantly improves the stability of time correction and semantic consistency of diagnostic events in practical applications.

[0018] Step S7: Encrypt and bind the timestamp-corrected case event sequence with the voiceprint event sequence to generate a unique and tamper-proof fused oral case data unit. Specifically, this includes: S7.1: Obtain the standardized coded operation event node sequence in the timestamp-corrected case event sequence and the set of directed edges of phase evolution trajectory in the voiceprint event sequence. Based on the optimal matching result output by the dual-channel graph attention mechanism, extract each pair of semantically aligned case event nodes and voiceprint topology subgraphs as data pairs to be bound. S7.2: The medical record event node identifier and the voiceprint topology subgraph feature vector in the data pair to be bound are concatenated and spliced ​​together. The national cryptographic SM3 hash algorithm is used to calculate and generate an initial fusion summary value containing cross-modal semantic information to eliminate the risk of a single data source being tampered with. S7.3: Based on the initial fusion digest value and the global random salt value of the diagnosis and treatment session, the elliptic curve digital signature algorithm is executed to generate a dynamic asymmetric key pair, and the public key part of the dynamic asymmetric key pair is embedded into the metadata field of the voiceprint event sequence to form an enhanced voiceprint event sequence with cryptographic anchors. S7.4: The enhanced voiceprint event sequence with cryptographic anchors and the time-stamped corrected case event sequence are encapsulated into a binary stream. The AES-256-GCM symmetric encryption mode is used, and the session key derived from the private key part of the dynamic asymmetric key pair is used for encryption to generate ciphertext payload data blocks. S7.5: The encrypted payload data block is structurally assembled with the public key index of the dynamic asymmetric key pair and the digital certificate of the operating physician, an integrity check code is added and it is encapsulated into a standardized fused oral case data unit, thereby completing the unforgeable binding of medical record data and voiceprint features. The public key index of the dynamic asymmetric key pair and the digital certificate of the operating physician are input into the AES-256-GCM encrypted payload data block. A structured assembly method (parameters: field order, fixed length encoding standard) is used to realize fixed-length and variable-length partitioning and encapsulation of metadata of each element. Furthermore, by using a multi-field metadata hash verification method (parameters: SM3 digest algorithm and field separator encoding rules), the overall structure is quickly verified for integrity, and a consistent and forgery-resistant integrity verification code is obtained. Furthermore, through a hexadecimal encoding conversion method (parameters: UTF-8 character set, case-consistent strategy), the integrity check code is embedded in the structured assembly, and a consistent encapsulation is generated for subsequent blockchain evidence storage and parsing. Furthermore, by using a binary stream serialization processing method (parameters: network byte order, big-endian mode), the encrypted payload data block, dynamic public key index, digital certificate, and integrity check code are sequentially merged to generate a packaged stream conforming to the standardized fusion of oral case data cell format; Furthermore, by using a segmented CRC check method (parameter: CRC32 polynomial generation function), the error detection capability of the serialized encapsulated stream during transmission is realized, and a verifiable fused oral case data unit transmission payload is generated. Through the above structured assembly and verification process, the AES-256-GCM encrypted payload data block from the previous step is transformed into a standardized fusion oral case data unit with public key index, physician digital certificate and integrity guarantee, so as to realize the unforgeable binding of medical record data and voiceprint features. For example, in a dental clinic session, the length of the encrypted payload data block generated in AES-256-GCM mode is 1,048,576 bytes, the public key index of the dynamic asymmetric key pair is 32 bytes, and the length of the operating physician's digital certificate is 2,048 bytes. When using a structured assembly method, the configured field order is [encrypted data block, public key index, digital certificate, integrity check code]. In the fixed-length encoding specification, the public key index field is set to a fixed length of 32 bytes, and the digital certificate field is set to a fixed length of 2,048 bytes. The first three combined fields are processed using the national cryptographic SM3 digest algorithm to generate an integrity check code. The digest length is 32 bytes, and a separator identifier is inserted in the encoding area following the field separator encoding rules. The generated integrity check code is embedded at the end of the encapsulation body after hexadecimal encoding. Big-endian mode is used during binary stream serialization to ensure cross-platform parsing consistency. The overall length of the encapsulated stream is 1048,576 + 32 + 2,048 + 32 = 1,050,688 bytes. A CRC32 polynomial is used in the transmission link. The segmented checksum is calculated, and a 4-byte checksum appended bit is generated for every 512-byte data segment. The final transmitted payload length is 1050688 + (1050688 / 512) × 4 = 1052736 bytes. Verification via the blockchain evidence storage interface confirms that the integrity checksum of this integrated oral medical record data unit is consistent before being uploaded to the blockchain, ensuring that the binding between medical record data and voiceprint features is unforgeable, while also meeting the standardized requirements for evidence storage data structures.

[0019] Step S8: A unique data identifier is generated for the fused oral medical record data unit using a hash algorithm, and then uploaded to a blockchain network for distributed notarization, generating an immutable blockchain notarization hash value. Specifically, this includes: S8.1: Obtain the fused oral medical record data unit formed by binding the timestamp-corrected medical record data with the voiceprint event sequence, and use the serialization protocol to convert the fused oral medical record data unit into a standardized binary byte stream to obtain the original data payload to be hashed; S8.2: Receive the original data payload to be hashed, and perform one-way hash calculation on the original data payload using the national cryptographic SM3 or SHA-256 hash algorithm to generate a unique data identifier digest with a fixed length and anti-collision characteristics; S8.3: Obtain the unique data identifier digest and diagnosis and treatment scenario metadata, and encapsulate the unique data identifier digest into a notarized transaction request containing the transaction initiator's address and timestamp information based on the smart contract interface specification, so as to construct a transaction package to be put on the chain that meets the requirements of the blockchain consensus mechanism; S8.4: Send the transaction package to be uploaded to the blockchain network to the consensus node cluster, and use the preset Byzantine fault-tolerant consensus algorithm to perform full network verification and block packaging of the transaction package to be uploaded to generate a new block containing the unique data identifier digest and return the on-chain confirmation receipt. S8.5: Parse the block height and Merkle root path information in the on-chain confirmation receipt, and combine them with the unique data identifier digest to generate the final blockchain evidence hash value, so as to output an immutable evidence certificate that can be verified by a third party at any time.

[0020] The present invention also provides a dental case generation system that integrates voiceprint features, and uses the above-mentioned method for generating dental cases by integrating voiceprint features.

[0021] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0022] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating oral medical records by incorporating voiceprint features, characterized in that, The process involves acquiring the voiceprint signals of diagnostic and treatment instruments operating in a dental treatment scenario, constructing a voiceprint event sequence based on the voiceprint signals, obtaining the structured diagnostic and treatment operation logs corresponding to the voiceprint event sequence to obtain a case event sequence, performing cross-modal matching and alignment between the voiceprint event sequence and the case event sequence, correcting the timestamps of the case event sequence, and encrypting and binding the timestamp-corrected case event sequence with the voiceprint event sequence to generate a fused dental case data unit.

2. The oral case generation method based on fused voiceprint features according to claim 1, comprising the following steps: S1: Real-time acquisition of audio signals emitted by diagnostic and treatment instruments during operation in oral treatment scenarios; frame-by-frame processing of the audio signals to generate device voiceprint signals; S2: Perform frequency domain transformation and phase feature extraction on the device's acoustic signature signal, construct a complex spectral tensor, and calculate the first derivative sequence of the phase difference between adjacent frames along the time axis to generate a phase evolution trajectory sequence; S3: Construct a set of directed edges based on the key turning points in the phase evolution trajectory sequence, and map them to a template library of typical operation stages to generate a voiceprint event sequence; S4: Obtain the structured diagnosis and treatment operation log corresponding to the voiceprint event sequence, parse it into standardized coded operation event nodes and inject time window confidence intervals to generate a case event sequence; S5: Input the case event sequence and the voiceprint event sequence into the graph attention network, calculate the joint similarity score of nodes and edges, and output the cross-modal alignment path of the optimal matching result; S6: Based on the estimated time offset in the cross-modal alignment path, correct the original timestamps in the medical record event sequence to generate a timestamp-corrected medical record event sequence; S7: Encrypt and bind the timestamp-corrected medical record event sequence with the voiceprint event sequence to generate a fused oral medical record data unit.

3. The method for generating oral case records by fusing voiceprint features according to claim 2, characterized in that, Step S3 specifically includes: The phase evolution trajectory sequence generated in step S2 is subjected to extreme point detection processing to extract the set of key turning point coordinates that characterize the sudden change in the mechanical state of the equipment, and a candidate node dataset is generated. Based on the candidate node dataset, perform temporal adjacency analysis, calculate the state transition probability between adjacent key inflection points, and generate a set of directed edges describing the evolution direction of the device's operating state. Obtain a pre-built template library of typical operation stages, perform subgraph isomorphic matching operation on the directed edge set, identify the specific diagnosis and treatment operation stage type to which the current voiceprint signal belongs, and generate preliminary mapping results; Based on the preliminary mapping results, acoustic fingerprint feature fusion is performed, and the matched operation stage type identifier is bound to the topology of the directed edge set to generate a structured voiceprint event sequence.

4. The method for generating oral case records by fusing voiceprint features according to claim 3, characterized in that, The candidate node dataset includes timestamps and phase difference derivative magnitudes.

5. The method for generating oral case records by fusing voiceprint features according to claim 3, characterized in that, The structured voiceprint event sequence contains device-unique acoustic fingerprint information.

6. The method for generating oral case records by fusing voiceprint features according to claim 2, characterized in that, Step S4 specifically includes: Obtain the structured treatment operation log data stream stored in the oral diagnosis and treatment system, perform field extraction and format standardization processing on the structured treatment operation log data stream, and generate a non-standardized operation event sequence; Perform semantic normalization processing on the original operation descriptions in the non-standardized operation event sequence to generate a set of standardized coded operation event nodes; For each standardized coding operation event identifier in the set of standardized coding operation event nodes, calculate and assign the corresponding time window confidence interval parameter to generate an operation event node with confidence interval. Based on the operation event nodes with confidence intervals, logical relationship deduction is performed to construct a set of directed logical edges and generate a clinical pathway constraint relationship graph structure. The operation event nodes with confidence intervals are used as graph nodes, and the clinical pathway constraint relationship graph structure is used as graph edges for topological fusion and encapsulation to generate a case event sequence.

7. The method for generating oral case records by fusing voiceprint features according to claim 6, characterized in that, The non-standardized operation event sequence includes the original operation description, the executing physician's identifier, and the initial recording time.

8. The method for generating oral case records by fusing voiceprint features according to claim 2, characterized in that, Step S5 specifically includes: Obtain the set of phase evolution trajectory sequence nodes in the voiceprint event sequence generated in step S3 and the set of standardized coding operation event nodes in the case event sequence generated in step S4. Perform vector mapping processing on the set of phase evolution trajectory sequence nodes to generate a voiceprint node feature tensor, and perform semantic coding processing on the set of standardized coding operation event nodes to generate a medical record node feature tensor. Based on the voiceprint node feature tensor, the phase difference evolution pattern edge features between adjacent nodes are extracted, and based on the medical record node feature tensor, the clinical path constraint relationship edge features are extracted. The phase difference evolution pattern edge features and the clinical path constraint relationship edge features are heterogeneously aligned to generate a joint edge feature matrix. The multi-head graph attention network is used to receive the voiceprint node feature tensor, the medical record node feature tensor, and the joint edge feature matrix. Cross-modal node-edge joint attention weight calculation is performed, and the calculated original attention score is transformed by probability distribution to generate a dynamic attention weight matrix. Based on the dynamic attention weight matrix, a weighted aggregation operation is performed on the voiceprint node feature tensor and the medical record node feature tensor to calculate the node edge joint similarity score of each pair of candidate node combinations, and the matching sequence with the highest cumulative similarity score is selected from all candidate combinations to generate the initial cross-modal alignment path. Based on the timestamp difference data of each matching node pair in the initial cross-modal alignment path, the local time deviation is iteratively corrected, the globally optimal time offset compensation amount is calculated, and the cross-modal alignment path is output.

9. The method for generating oral case records by fusing voiceprint features according to claim 2, characterized in that, Following S7 are: S8: Generate a unique data identifier for the fused oral case data unit using a hash algorithm, and upload it to the blockchain network for distributed storage, generating an immutable blockchain storage hash value.

10. A dental case generation system integrating voiceprint features, characterized in that: The oral medical record generation method based on the integration of voiceprint features, as described in any one of claims 1-9, is used to generate oral medical records.

Images