A method for constructing a neonatal pain evidence base

Abstract

The application relates to the field of medical data processing and discloses a new-born baby pain evidence base construction method, which converts evidence-based medical texts into structured evidence nodes containing a joint intervention feature vector and an initial sub-item weight vector; clinical operation information is acquired to extract a matching node, an instruction set is issued and a global time anchor point is recorded, continuous physiological data and facial expression events are respectively converted into a standard continuous sequence set and a binary time sequence set in combination with the time anchor point and a discrete time step, the cumulative deviation value of the sequence from a baseline value is calculated, a transient stress inhibition factor and a steady-state recovery factor are obtained by weighting, and are mapped into a comprehensive utility score, and a difference between the comprehensive utility score and a preset utility threshold is used to perform numerical updating on the initial sub-item weight vector. The application provides a unified time alignment benchmark for heterogeneous data, and realizes dynamic closed-loop adjustment of evidence base weights based on real clinical physiological feedback data.

Description

Technical Field

[0001] This invention relates to the field of medical data processing, specifically to a method for constructing a neonatal pain evidence database. Background Technology

[0002] In neonatal intensive care clinical practice, procedures such as heel prick blood sampling and venipuncture are unavoidable. Neonatal pain management currently relies primarily on previously compiled evidence-based medicine guidelines or fixed pain assessment scales. However, because these guidelines are mostly in text or static database form, existing pain management protocols have significant limitations in practical clinical application.

[0003] Following clinical intervention, newborns experience continuous changes in physiological parameters such as heart rate and blood oxygenation, accompanied by discrete event responses such as changes in facial expressions. Existing data processing systems struggle to accurately align these heterogeneous data points, with vastly different sampling frequencies and structures, onto a unified timeline, resulting in a lack of reliable data benchmarks for post-intervention status analysis. Furthermore, traditional clinical assessment methods often rely on data readings at single time points, failing to separate and continuously quantify the transient stress responses during the intervention from the body's recovery process afterward. This makes them susceptible to assessment errors due to occasional fluctuations, hindering a true reflection of the actual effectiveness of the clinical intervention.

[0004] Furthermore, due to the lack of accurate data quantification and feedback pathways, existing neonatal pain evidence bases cannot adjust parameters based on the actual physiological responses of infants. The intervention recommendation weights used to guide clinical decision-making remain fixed, preventing the use of actual clinical feedback data to update and iterate the underlying weight parameters of the evidence base, thus hindering the dynamic optimization of pain management strategies.

[0005] Therefore, this invention proposes a method for constructing a neonatal pain evidence database to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for constructing a neonatal pain evidence database. This method solves the problem that existing neonatal pain management programs mainly rely on static experience guidelines, lack a numerical correction mechanism that incorporates multimodal physiological feedback data, and are difficult to quantitatively assess and dynamically adjust intervention strategies based on individual real-time physiological responses and recovery status.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing a neonatal pain evidence database, comprising the following steps: The evidence-based medicine text is transformed into structured evidence nodes, which include an operation type field, a pain grading threshold field, a joint intervention feature vector, and an initial sub-item weight vector. Acquire clinical operation type information and basic assessment scores, extract the matching structured evidence nodes, and issue the instruction set corresponding to the joint intervention feature vector, recording the instruction execution time as a global time anchor point; Acquire continuous physiological data and facial expression events from clinical monitoring equipment, and combine the discrete time step with the global time anchor point to convert the continuous physiological data and facial expression events into a standard continuous sequence set and a binary time series sequence, respectively. The cumulative deviation of the standard continuous sequence set from the preset physiological baseline value is calculated, and the binarized time series is directly accumulated to obtain the cumulative expression value. The cumulative deviation value and the cumulative expression value are weighted and summed to obtain the transient stress inhibition factor and the homeostasis recovery factor. The transient stress suppression factor and the steady-state recovery factor are linearly mapped to obtain a comprehensive utility score. The difference between the comprehensive utility score and the preset utility threshold is used to update the initial sub-item weight vector.

[0008] Preferably, the step of converting evidence-based medicine text into structured evidence nodes includes: The nursing plan, which includes multiple intervention methods, is broken down into independent dimensions and written into the corresponding positions of the joint intervention feature vector; Evidence-based recommendation level label data is extracted from the evidence-based medicine text, and the evidence-based recommendation level label data is transformed into a set of quantitative constants according to the basic weight mapping table, which serves as the initial sub-item weight vector.

[0009] Preferably, the step of obtaining clinical operation type information and basic assessment scores, and extracting the matching structured evidence nodes, includes: The basic assessment score is converted into a grading identifier through numerical mapping rules, which serves as the status of the pain grading threshold field. The clinical operation type information and the status of the pain grading threshold field are used as the joint retrieval primary key to extract a set of candidate nodes; The activated sub-weights in the initial sub-item weight vectors of each node in the candidate node set are added together, and the node with the highest sum value is selected as the target node.

[0010] Preferably, after the time of recording instruction execution is used as the global time anchor, the method further includes: Monitor the process status, and when a clinical operation end command is triggered, record the clock value as the operation end timestamp; The recovery period duration constant is superimposed on the operation end timestamp to calculate the recovery period end timestamp. The global time anchor, the operation end timestamp, and the recovery period end timestamp together construct an absolute time frame for separating physiological response features. The interval between the global time anchor and the operation end timestamp is defined as the operation period time window, and the interval between the operation end timestamp and the recovery period end timestamp is defined as the recovery period time window.

[0011] Preferably, in the process of combining the discrete time step with the global time anchor to convert the continuous physiological data and the facial expression events into a standard continuous sequence set and a binarized time series, respectively, the conversion of the continuous physiological data into a standard continuous sequence set includes: Calculate and generate a standard discrete-time grid sequence; The true sampling points are found in the continuous physiological data, and linear interpolation is performed based on the time interval between adjacent true sampling points and the difference in physiological values ​​to fill in the numerical values ​​and generate the standard continuous sequence set.

[0012] Preferably, the process of combining the discrete time step with the global time anchor to convert the continuous physiological data and the facial expression events into a standard continuous sequence set and a binarized time series sequence, respectively, further includes: Calculate the time interval between adjacent real sampling points in the continuous physiological data; If the time interval is greater than the preset interruption judgment threshold, the virtual monitoring value at the standard discrete time node of the corresponding segment will be marked as an invalid identifier.

[0013] Preferably, in the process of combining the discrete time step and the global time anchor point to convert the continuous physiological data and the facial expression events into a standard continuous sequence set and a binary time series sequence, respectively, the conversion of the facial expression events into a binary time series sequence includes: The duration segments of the facial expression events are mapped to a standard discrete-time grid sequence; At any standard discrete time node, a state determination is performed. If the node is within the duration of the facial expression event, the sequence value at the node is assigned the value of one. If the node is not within the duration of the facial expression event, the value is assigned to zero.

[0014] Preferably, in the process of calculating the cumulative deviation value of the standard continuous sequence set from the preset physiological baseline value, and directly accumulating the binarized time series to obtain the cumulative expression value, and weighted summing the cumulative deviation value and the cumulative expression value to obtain the transient stress inhibition factor and the homeostasis recovery factor, the process of obtaining the transient stress inhibition factor includes: A standard heart rate sequence is extracted from the set of standard continuous sequences, and a heart rate baseline value is obtained from the preset physiological baseline value. The difference between the standard heart rate sequence value and the heart rate baseline value within the operation period time window is calculated. When the difference is greater than zero, the positive deviation difference is retained, and the positive deviation difference is discretely accumulated to obtain the cumulative heart rate deviation value; The cumulative expression value is obtained by directly summing the binary time sequence values ​​within the same classmate; The transient stress inhibition factor is obtained by applying a preset weighting coefficient to the cumulative deviation of heart rate and the cumulative value of facial expression and then summing them.

[0015] Preferably, in the process of calculating the cumulative deviation value of the standard continuous sequence set from the preset physiological baseline value, and directly accumulating the binarized time series to obtain the cumulative expression value, and weighted summing the cumulative deviation value and the cumulative expression value to obtain the transient stress inhibition factor and the steady-state recovery factor, the process of obtaining the steady-state recovery factor includes: Standard blood oxygen saturation sequence and standard brain oxygen saturation sequence are extracted from the standard continuous sequence set. Blood oxygen saturation line value and brain oxygen saturation line value are obtained from the preset physiological baseline value. The difference between the blood oxygen saturation line value and brain oxygen saturation line value within the recovery period time window and the standard blood oxygen saturation sequence value is calculated. When the difference is greater than zero, the difference is retained and discretely accumulated to obtain the corresponding cumulative blood oxygen deviation value and cumulative brain oxygen deviation value. The cumulative deviation of blood oxygen and the cumulative deviation of brain oxygen are weighted and combined using preset weighting coefficients to generate the steady-state recovery factor.

[0016] Preferably, the step of updating the initial sub-item weight vector using the difference between the comprehensive utility score and a preset utility threshold includes: The difference is used as a feedback driver. Using a preset learning rate constant and the feedback driving term, the sub-weights corresponding to the activated sub-items within the initial sub-item weight vector are numerically overwritten; Verify whether the sub-weights after numerical overwriting are within the preset range, perform boundary constraint judgment, and overwrite the structured evidence node.

[0017] This invention provides a method for constructing a neonatal pain evidence database. It has the following beneficial effects: 1. This invention records a global time anchor point at the moment of issuing intervention instructions and calculates a standard time grid by combining discrete time steps. This converts continuously sampled physiological data into a standard continuous sequence set and converts discrete facial expression events into binary time series. This data processing method provides a unified time reference benchmark for heterogeneous monitoring data, solves the problem of data synchronization and alignment difficulties caused by different sampling frequencies of physiological and behavioral features in clinical practice, and provides a reliable data foundation for subsequent feature calculations.

[0018] 2. This invention extracts transient stress inhibition factors and steady-state recovery factors by calculating the cumulative deviations of the standard sequence and time sequence from the baseline values, thereby achieving a phased quantitative assessment of the intervention effect. Based on the numerical deviation within the time window, the immediate pain stress inhibition state and the subsequent physiological function recovery state are calculated separately, so that the final comprehensive utility score can fully reflect the actual effect of the clinical intervention and avoid the random errors caused by single-time-point assessment.

[0019] 3. This invention uses the numerical difference between the calculated comprehensive utility score and the preset utility threshold as a feedback driver. It uses the learning rate constant to directly overwrite and update the initial sub-item weight vector in the structured evidence node, and uses the physiological feedback data generated in real clinical practice to act on the basic knowledge framework. This allows the originally static medical guideline parameters to undergo numerical iteration based on the actual clinical execution results, and realizes dynamic closed-loop adjustment of the evidence base weights at the underlying logic level. Attached Figure Description

[0020] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the present invention; Figure 3 This is a logical diagram of the underlying data structure and baseline mapping of the evidence library in this invention; Figure 4 This is a logic diagram of instruction dispatch and multi-source data timing alignment processing in this invention; Figure 5 This is the dual-channel feature extraction and adaptive evolution closed-loop control diagram of the present invention; Figure 6 This is a curve showing the changes in dual-channel physiological indicators and the extraction of characteristic factors according to the present invention; Figure 7 This is a graph comparing the dynamic evolution effect of the present invention with the comprehensive utility score of the traditional static solution. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 The present invention provides a neonatal pain evidence database construction system, comprising a terminal layer, an edge layer and a server layer.

[0023] The terminal layer includes a patient monitor, a brain oxygenation meter, a video terminal, and an interactive terminal. The patient monitor collects heart rate and transcutaneous oxygen saturation signals. The brain oxygenation meter acquires local brain oxygen saturation waveform data. The video terminal, mounted above the operating area, records facial movement feature images. The interactive terminal receives pre-operation baseline assessment parameters and outputs intervention commands issued by the system.

[0024] The edge layer deploys a synchronization gateway and an alignment module. The synchronization gateway establishes a communication connection with the terminal layer to mark the input data stream with uniform clock parameters. The alignment module performs discrete-time grid mapping and format conversion calculations on asynchronously transmitted multi-source data.

[0025] The server layer comprises a parsing module, an evidence base module, a feature extraction module, and an evolution control module. The parsing module parses external medical text libraries, generating underlying data relationship nodes. The evidence base module stores intervention feature parameters and their corresponding weight variables. The feature extraction module receives the standard data sequence output by the alignment module and performs temporal operations on physiological signals. The evolution control module reads the temporal operation results and performs overwriting operations on associated fields and table structure reorganization operations within the evidence base module.

[0026] See attached document Figure 2 This invention provides a method for constructing a neonatal pain evidence database, comprising the following steps: S100, the parsing module extracts external evidence-based medicine text, transforms non-drug intervention nursing plans into structured evidence nodes and writes them into the evidence base module. The structured evidence nodes include operation type field, pain grading threshold field, decoupling set joint intervention feature vector and corresponding initial sub-item weight vector. S200: The interactive terminal obtains the current clinical operation instructions and basic assessment scores. The evidence base module retrieves matching nodes based on the scoring results, converts the joint intervention feature vector into an instruction set and sends it to the interactive terminal. When the interactive terminal confirms the execution of the system, the synchronous gateway generates a global time anchor point. S300, the terminal layer device continuously outputs data, the alignment module receives data streams from the monitor, brain oxygen meter and video terminal. Taking the global time anchor point as the reference starting point, the alignment module resamples the continuous physiological data with different sampling frequencies into a standard discrete sequence with a fixed step size, and at the same time converts the discrete facial feature behavior events into a binary time sequence with the same step size. S400, the feature extraction module loads the standardized sequence set, divides the data according to the set time-domain window boundaries and imports it into the transient neural stress channel and the homeostatic metabolic perfusion channel, calculates the discrete cumulative value of the monitoring parameter deviation from the baseline value in each channel, and outputs the transient stress inhibition factor that represents the immediate response and the homeostatic recovery factor that represents the microcirculation recovery. S500, the evolution control module obtains the factor data output by the feature extraction module. Based on the preset orthogonal decoupling mapping matrix, the evolution control module locates the sub-item weight field of the corresponding node in the evidence base module. It uses the feedback driving term to overwrite the weight value with learning rate for the activated valid sub-item and performs boundary constraint judgment to complete the closed-loop update of the database.

[0027] See attached document Figure 3 The method provided by this invention includes the following sub-steps: the parsing module extracts external evidence-based medicine text and transforms non-drug intervention nursing plans into structured evidence nodes, which are then written into the evidence base module. S101, the parsing module retrieves evidence-based clinical practice guidelines and related medical literature from an external medical text database. The parsing module extracts unstructured medical text information into structured data entities and encapsulates these data entities into structured evidence nodes for use by the underlying relational database. In this embodiment, the parsing module uses a named entity recognition network model based on the BERT-CRF architecture to perform text information extraction calculations. The hierarchical structure of this named entity recognition network model includes an input embedding layer, a bidirectional Transformer encoding layer, and a conditional random field decoding layer. In practical applications, the input data is a sequence of character data from evidence-based medical literature that has undergone word segmentation and tokenization preprocessing. The model output corresponds to a stream of medical entity labels with specific business meanings, specifically including operational noun entities, intervention action entities, and evidence level label entities. During the model training and construction phase, the system uses historically manually annotated clinical nursing guideline texts as training samples, manually proofread entity boundaries and category categories as real labels, and uses a cross-entropy loss function combined with a log-likelihood function to optimize the network parameters using gradient descent. The fundamental principle of structuring natural language text into tuples lies in transforming non-standard medical contexts into discrete data arrays that can be directly distributed and invoked by the computer instruction bus.

[0028] The basic data units stored in the structured evidence base module are represented as sets of structured evidence nodes. Structured evidence nodes within the system are defined as tuples. Its mathematical structure is defined as follows: ; In the formula, Represents a tuple of structured evidence nodes; Indicates the operation type field; This represents the pain grading threshold field; Represents the joint intervention feature vector; This represents the initial sub-item weight vector. The operation type field records the specific clinical operation category data that caused the pain. The pain grading threshold field records the estimated pain grading interval status data before the operation.

[0029] S102, the parsing module performs dimensional decoupling and splitting of the combined intervention plan. In routine clinical evidence-based records, nursing intervention plans often exist in the form of a whole combination, making it difficult for the underlying computer system to determine the independent analgesic contribution of individual intervention actions within the combination. To address this logical dead end, this system introduces a multi-dimensional decoupling mechanism. Through a pre-set classification dictionary, the overall nursing plan covering multiple intervention methods in the literature is split into independent sub-dimensional attributes according to the pathway, and written into the corresponding positions of the combined intervention feature vector.

[0030] The decoupling mathematical expression of the joint intervention feature vector is as follows: ; In the formula, Represents the joint intervention feature vector; This indicates the sub-item of taste-suction intervention; This indicates the sub-item of postural tactile intervention; This indicates the sub-item for auditory and olfactory intervention.

[0031] The taste-based sucking intervention sub-item records characteristic parameters such as the administration of sucrose water and the state of non-nutritive sucking operations. The positional-tactile intervention sub-item records characteristic parameters such as kangaroo care and flexed positioning. The auditory-olfactory intervention sub-item records characteristic parameters such as environmental white noise playback parameters and the placement of breast milk gauze. If an intervention protocol extracted from a certain literature contains only a single dimension, inactive sub-items are assigned null values ​​by the system in the joint intervention feature vector. This decoupling method transforms the fuzzy holistic nursing combination into basic vector units that can be independently optimized by multi-channel algorithms.

[0032] S103, the parsing module performs weighted numerical mapping of the evidence-based baseline data. Based on the previous text parsing results, the parsing module extracts the evidence-based recommendation level label data of the corresponding literature and transforms the qualitative level labels into an initial set of quantitative constants within the structured evidence nodes.

[0033] The specific expression of the initial sub-item weight vector is as follows: ; In the formula, This represents the initial sub-item weight vector; Indicates the first sub-weight; Indicates the second sub-weight; This represents the third sub-weight, and the values ​​of the aforementioned sub-weights are set to floating-point numbers within the interval (0,1].

[0034] The first sub-weight establishes a data correspondence with the taste / sucking intervention sub-item. The second sub-weight establishes a data correspondence with the body position / tactile intervention sub-item. The third sub-weight establishes a data correspondence with the hearing / olfactory intervention sub-item. The system has a pre-set basic weight mapping table, and the parsing module executes the assignment logic based on this table. The basic weight mapping table establishes a correlation between the guideline evidence grading standards and numerical ranges. For example, in a specific mapping rule implementation scenario, if the parsing module identifies a literature source as Level 1 strong recommendation evidence (JBI evidence-based level), then all valid sub-items within the corresponding initial sub-item weight vector are assigned an initial weight value between 0.9 and 1.0; if the identification result is Level 3 weak recommendation evidence, then valid sub-items are assigned an initial weight value between 0.4 and 0.6.

[0035] After the parsing module completes the encapsulation and construction of the aforementioned four-tuple fields, it sends the generated structured evidence node data packet to the evidence base module. The evidence base module receives the data packet and establishes a corresponding data index table, forming the foundational knowledge base for the system's adaptive matching and dynamic evolution calculations.

[0036] See attached document Figure 4 In the method provided by this invention, the interactive terminal obtains the current clinical operation instructions and basic assessment scores. The evidence base module retrieves matching nodes based on the scoring results, converts the joint intervention feature vector into an instruction set, and sends it to the interactive terminal. At the moment the interactive terminal confirms the execution, the synchronous gateway generates a global time anchor point. This step may specifically include the following sub-steps: S201, the interactive terminal receives the neonatal pre-operation basic assessment score and the type of clinical operation to be performed, input by medical staff. In a specific implementation scenario, the basic assessment score combines the pre-score of the neonatal pain assessment scale with basic vital sign parameters. The system internally presets a numerical mapping rule for pain grading threshold states, which converts discrete score values ​​into fixed grading identifiers.

[0037] As a specific rule implementation, when the baseline assessment score is in the 0-2 range, the system determines the pain grading threshold status as mild; when the score is in the 3-4 range, it is determined as moderate; and when the score is greater than or equal to 5, it is determined as severe. The evidence base module uses the input clinical operation type and pain grading threshold status as the joint retrieval primary key, performs a traversal search in the underlying structured evidence node pool, and extracts all data entities with the required attributes as a candidate node set.

[0038] For the extracted candidate node set, the evidence base module reads the initial sub-item weight vector configured within each node and calculates the comprehensive recommendation index for each candidate node. The calculation principle of this recommendation index is to sum the intervention weights of each independent dimension that is in an active state, and use the numerical value to intuitively reflect the overall recommendation strength of different combinations of solutions at the evidence-based medicine level. The specific formula for calculating the comprehensive recommendation index is as follows: ; In the formula, This indicates the overall recommendation index; Indicates the index number; This represents the sub-weight. After calculating the values ​​for each candidate node, the evidence base module selects the target node with the highest comprehensive recommendation index as the current matching intervention strategy.

[0039] In step S202, the evidence base module performs reverse decoding on the joint intervention feature vector within the selected target node. Specifically, based on the system's preset dictionary mapping table, it translates the feature vector fields in the underlying relational database into text information that can be recognized and displayed by the front-end hardware, and then sends this instruction set data packet to the interactive terminal. The interactive terminal parses the instruction set, converting the underlying parameters of the taste / sucking intervention sub-item, the body position / tactile intervention sub-item, and the hearing / olfactory intervention sub-item into corresponding nursing execution items, which are then displayed on the screen interface.

[0040] After completing the intervention preparations according to the on-screen prompts, the operator triggers the operation start confirmation command by clicking the controls on the interactive terminal interface. For the touch response mechanism of the interactive terminal interface and the front-end and back-end data communication transmission protocol, those skilled in the art can use a standard long-connection communication network architecture. The specific network handshake and message parsing process are well-known technologies in the field and will not be elaborated upon here.

[0041] S203, the synchronization gateway establishes the system's time calculation boundary. In clinical settings, multi-source heterogeneous monitoring devices exhibit inconsistencies between sampling frequency and data transmission delay during continuous operation. Directly introducing asynchronous data can easily lead to mismatches in subsequent dual-channel feature extraction stages. Therefore, upon receiving the operation start confirmation command from the interactive terminal, the synchronization gateway reads the underlying absolute clock of the system server, records this moment, and marks it as the global time anchor point.

[0042] Set the global time anchor point as This global time anchor point serves as the zero-point starting time reference for all subsequent time-series operation channels.

[0043] After establishing the starting time baseline, the synchronization gateway continuously monitors the process status of the interactive terminal. When the interactive terminal triggers a clinical operation end command, the synchronization gateway records the current clock value and marks it as the operation end timestamp. The system has a pre-set recovery period duration constant. The synchronization gateway adds this recovery period duration constant to the operation end timestamp to calculate the recovery period end timestamp. The calculation formula is as follows: ; In the formula, Indicates the timestamp of the end of the recovery period; Indicates the timestamp indicating the end of the operation; This represents the recovery period duration constant. (For...) The specific value is pre-configured in the system by medical staff based on the expected physiological recovery cycle of different clinical operations. In order to ensure that the steady-state metabolic perfusion channel can capture the complete microcirculation recovery characteristics, its value range is usually set to an integer within the range of 30 seconds to 120 seconds.

[0044] The established global time anchor, operation end timestamp, and recovery period end timestamp together construct an absolute time frame for separating physiological response characteristics at different stages, providing a low-level time alignment reference benchmark to overcome the timing asynchronous defects of heterogeneous devices.

[0045] See attached document Figure 4 In the method provided by this invention, the terminal layer device continuously outputs data, the alignment module receives data streams from the monitor, the brain oxygen meter, and the video terminal, and performs resampling and binarization conversion steps with the global time anchor point as the reference starting point. Specifically, it may include the following sub-steps: S301, after receiving the global time anchor point generated by the synchronization gateway, the alignment module constructs the system's underlying standard time-based operating grid. Clinical data acquisition devices, limited by hardware design, typically have inconsistent sampling refresh rates; for example, the heart rate output frequency of a patient monitor and the refresh rate of a brain oximeter inherently differ. To establish a unified dimension for multi-channel joint computation, the alignment module sets a fixed discrete time step. Combining this discrete time step with the global time anchor point, the alignment module calculates and generates a standard discrete time grid sequence. The specific grid generation formula is as follows: ; In the formula, Represents the standard discrete time node; Indicates the global time anchor point; Represents a natural number index; This represents the discrete time step. In the standard scenario setting for neonatal physiological monitoring, the discrete time step can be configured to 1 second to balance computational accuracy and system data processing load.

[0046] S302, the alignment module performs interpolation and alignment processing on the continuous physiological signals input from the terminal layer. The alignment module continuously acquires raw heart rate data and raw transcutaneous oxygen saturation data collected by the monitor, and raw local cerebral oxygen saturation data collected by the cerebral oxygen analyzer. Since the timestamps of these raw data cannot coincide with standard discrete time nodes, the alignment module calls a linear interpolation algorithm to calculate the virtual monitoring values ​​at each target time node.

[0047] The alignment module searches for two adjacent real sampling points before and after a standard discrete time node in the original data time series. Before performing linear interpolation, the alignment module calculates the time interval between the two adjacent real sampling points. If this time interval is greater than a preset interruption judgment threshold, such as 5 seconds, the system determines that a sensor has detached or that there is an abnormal data interruption in that segment. At this time, the interpolation calculation is stopped and the virtual monitoring values ​​at the standard discrete time node of that segment are marked as invalid.

[0048] If the time interval is less than or equal to a preset interruption threshold, the system calculates and fills in the value at that node based on the time difference between these two points and the physiological value difference, using a linear ratio. After this interpolation operation, the asynchronously transmitted raw data stream is converted into a set of standard continuous sequences that change synchronously on a unified time base. This set specifically includes standard heart rate sequences, standard blood oxygen saturation sequences, and standard cerebral oxygen saturation sequences.

[0049] S303, the alignment module performs binarization temporal transformation on the discrete facial expression events acquired by the video terminal. The video terminal records the video stream of the operating area and outputs facial expression events representing pain responses and their corresponding duration segments through its internally integrated image recognition network. In this embodiment, the image recognition network integrated within the video terminal adopts a convolutional neural network model based on the ResNet architecture. The internal hierarchical structure of this model includes an input convolutional layer, multi-level residual network block groups, and a fully connected classification layer.

[0050] In practical operation, the model receives input data consisting of a sequence of continuous facial video frames that have undergone grayscale and normalization preprocessing. The image dimensions are set to a width-height matrix with preset pixel parameters. After the image data stream passes through multi-level residual network blocks to extract local facial texture features, the output is determined by a fully connected classification layer. The specific state of this output corresponds to the activation probability of specific facial action units in newborns, including the occurrence of frowning, eye squinting, and deepening of nasolabial folds.

[0051] When constructing this network model, the training process uses video frame sequences of newborn pain expressions obtained clinically as training samples, uses pre-annotated facial action unit states as real labels, and applies a binary cross-entropy loss function to perform gradient update optimization on the network weight parameters so that the model outputs the start and end time periods of facial expression events representing pain responses.

[0052] The alignment module reads the start and end times of the aforementioned facial expression events and maps them to a pre-established standard discrete-time grid. At any standard discrete-time node, the system executes state determination logic. If the node is within the duration of the facial expression event, the alignment module assigns a value of 1 to the sequence value at that node; otherwise, it assigns a value of 0. The mathematical expression for the relevant logic is as follows: ; In the formula, Represents the value of a binary time series sequence; This represents a standard discrete time node. Through this transformation calculation, non-periodic asynchronous image recognition events are quantized into digital square wave sequences with the same time base and duration as physiological signals. This preprocessing mechanism unifies the temporal dimension of heterogeneous medical device data, providing a standard format input condition for the subsequent dual-channel feature extraction module.

[0053] See attached document Figure 5 The method provided by this invention includes the following sub-steps: the feature extraction module loads the standardized sequence set, divides the data according to the set time-domain window boundaries and imports it into the transient neural stress channel and the homeostatic metabolic perfusion channel, calculates the discrete cumulative values ​​of the monitoring parameters deviating from the baseline values, and outputs the transient stress inhibition factor and the homeostatic recovery factor. S401, the feature extraction module establishes a stress-free baseline window and calculates the baseline values ​​of each physiological parameter. Before assessing the neonate's physiological stress response to pain, the system obtains its baseline physiological levels under quiet, undisturbed conditions as a reference. The feature extraction module uses a global time anchor point as a reference and extracts a fixed-length data segment along the historical timeline as the stress-free baseline window. In specific implementation scenarios, the length of this stress-free baseline window is configured as an integer within the range of 30 to 60 seconds prior to the operation.

[0054] After determining the stress-free baseline window, the feature extraction module reads the standardized sequence set output by the alignment module and extracts the standard heart rate sequence, standard blood oxygen saturation sequence, and standard brain oxygen saturation sequence within the time window. The feature extraction module calculates the arithmetic mean of the data points of the above three sets of sequences within the stress-free baseline window and calculates the heart rate baseline value, blood oxygen saturation line value, and brain oxygen saturation line value.

[0055] The specific calculation logic for the cerebral oxygen saturation value is as follows: the system extracts all standard cerebral oxygen saturation sequence values ​​within the stress-free baseline window, sums them up, and then divides the sum by the total number of data points within that window. The calculation formula is as follows: ; In the formula, Indicates the brain oxygen line value; Indicates a baseline window without stress; Represents standard discrete time nodes; This represents the standard brain oxygen saturation sequence value at that node; This represents the total number of valid data points within the stress-free baseline window.

[0056] Similarly, the baseline heart rate and serum oxygen levels were also calculated using the arithmetic mean logic described above. These baseline values ​​collectively characterize the stable physiological baseline of the newborn before the current clinical procedure, providing an absolute reference for subsequent deviation calculations.

[0057] S402, the feature extraction module performs data processing within the transient neural stress channel to generate transient stress inhibitory factors. Acute pain stimuli trigger a rapid response of the sympathetic nervous system, typically manifested as a transient increase in heart rate and the appearance of facial pain expressions. The feature extraction module extracts the segment between the global time anchor point and the operation end timestamp as the operation period time window, within which a standard heart rate sequence and a binarized time series sequence are loaded.

[0058] To avoid the algorithmic dead zone caused by the cancellation of positive and negative differences due to fluctuations in heart rate, the feature extraction module calculates the difference between the heart rate value and the heart rate baseline value point by point during the operation period. When the difference is greater than zero, the positive deviation difference is retained, and when the difference is less than or equal to zero, the difference is forcibly set to zero. Then, the filtered values ​​are discretely accumulated to obtain the cumulative heart rate deviation value.

[0059] Meanwhile, the feature extraction module directly sums the binarized time-series values ​​within the same window to obtain the cumulative expression value. The formula for calculating the cumulative heart rate deviation value is as follows: ; In the formula, This indicates the cumulative deviation of heart rate; Represents the standard discrete time node; Indicates the global time anchor point; Indicates the timestamp indicating the end of the operation; This represents the standard heart rate sequence value; This represents the baseline heart rate value; This represents the function that takes the maximum value.

[0060] Subsequently, the feature extraction module uses preset weighting coefficients to perform a weighted summation of the two accumulated values ​​to calculate the transient stress suppression factor. The specific formula is as follows: ; In the formula, Indicates transient stress inhibitory factor; This represents the heart rate weighting coefficient; This represents the facial expression weighting coefficient; This represents the cumulative value of facial expression. In one specific embodiment, the heart rate weighting coefficient is set to a floating-point number within the range of [0.4, 0.6], and the facial expression weighting coefficient is also set to a floating-point number within the range of [0.4, 0.6], with the sum of the two values ​​fixed at 1. The value of the transient stress inhibition factor is used to characterize the inhibitory effect of the intervention on immediate pain nerve conduction.

[0061] S403, the feature extraction module performs data processing within the homeostatic metabolic perfusion channel to generate homeostatic recovery factors. After the clinical procedure, the recovery of local microcirculation and tissue oxygen metabolism is a slow homeostatic reconstruction process. The feature extraction module extracts the segment between the operation end timestamp and the recovery period end timestamp as the recovery period time window, within which standard blood oxygen saturation sequences and standard cerebral oxygen saturation sequences are loaded.

[0062] Since painful stimuli typically increase oxygen consumption in newborns, leading to a decrease in blood oxygen and brain oxygen levels, the feature extraction module performs a one-way filtering judgment. Specifically, the feature extraction module calculates the difference between the baseline value and the current monitored value. If the difference is greater than zero (i.e., the monitored value is lower than the baseline), the data is retained; if the difference is less than or equal to zero, it is set to zero. This allows the module to calculate the cumulative deviation of blood oxygen within the recovery period time window. The formula for calculating the cumulative deviation of blood oxygen is as follows: ; In the formula, This indicates the cumulative deviation of blood oxygen levels; Indicates the timestamp of the end of the recovery period; Indicates blood oxygen level; This represents the standard blood oxygen sequence value.

[0063] Similarly, the feature extraction module uses the same one-way filtering and cumulative calculation logic to obtain the cumulative deviation value of cerebral oxygenation. The feature extraction module then applies another set of preset weighting coefficients to weight and combine the cumulative deviation values ​​of blood oxygenation and cerebral oxygenation to generate a steady-state recovery factor. The specific formula is as follows: ; In the formula, Represents the steady-state recovery factor; This represents the blood oxygen weighting coefficient; This represents the brain oxygen weighting coefficient; This represents the cumulative deviation of cerebral oxygen. In the system configuration, the values ​​of both the blood oxygen weighting coefficient and the cerebral oxygen weighting coefficient are set to floating-point numbers within the range of [0.3, 0.7], and their sum equals 1. The steady-state recovery factor characterizes the physiological metabolic recovery cost of the newborn after the operation. By constructing the above dual-channel feature extraction architecture, the system achieves decoupled extraction of multidimensional physiological signals.

[0064] See attached document Figure 5 The method provided by this invention includes the following sub-steps: the evolution control module acquires transient stress suppression factors and steady-state recovery factors, calculates a comprehensive utility score, and dynamically adjusts the weights of target nodes in the evidence base module accordingly. S501, the evolution control module receives the transient stress inhibition factor and homeostasis recovery factor output by the feature extraction module and constructs a two-dimensional orthogonal evaluation space. The transient stress inhibition factor characterizes the degree of deviation of acute neurological reflexes in newborns during the operation period, while the homeostasis recovery factor characterizes the degree of deviation of local microcirculation and oxygen metabolism during the recovery period. The evolution control module performs a linear mapping combination of these two independent feature factors to calculate the comprehensive utility score of the current clinical intervention plan.

[0065] To avoid an algorithmic dead zone where abnormally high stress values ​​cause the score to drop below the zero lower limit, a bottom-line truncation function is added around the difference operation. The specific formula for calculating the comprehensive utility score is as follows: ; In the formula, This indicates the overall utility score; Represents the ideal physiological constant; Indicates the first mapping coefficient; Indicates the second mapping coefficient; Indicates transient stress inhibitory factor; Represents the steady-state recovery factor; This represents the function that maximizes the value. In the specific parameter configuration of the system, the ideal physiological constant is set to 100 to establish the scoring benchmark. The values ​​of the first and second mapping coefficients are both set to floating-point numbers within the interval (0,1), and their sum is fixed at 1. This calculation logic ensures that the less the physiological parameters deviate from the baseline, the higher the corresponding comprehensive utility score.

[0066] In step S502, the evolution control module performs adaptive evolution driving force calculations based on the comprehensive utility score. The system pre-configures a utility threshold, representing the clinically acceptable baseline intervention effect. The evolution control module calculates the numerical difference between the comprehensive utility score and the utility threshold, using this difference as the feedback driver for adaptive adjustment. If the comprehensive utility score is greater than the utility threshold, the feedback driver is positive, indicating that the current intervention is performing better than the expected baseline; if the comprehensive utility score is less than the utility threshold, the feedback driver is negative, indicating that the current intervention is not effective.

[0067] In S503, the evolution control module performs dynamic updates of the weight vector and database write-back operations for the target node. The evolution control module reads the joint intervention feature vector and the initial sub-item weight vector of the currently executing target node. The system identifies the active sub-items in the current intervention plan by traversing the joint intervention feature vector. For these active and valid sub-items, the evolution control module uses the feedback driving term derived in the previous derivation to update their corresponding sub-weights. The specific weight update formula is as follows: ; In the formula, Indicates the updated sub-weights; Indicates the sub-weights before the update; This represents the learning rate constant; This indicates the overall utility score; This represents the utility threshold. The learning rate constant is used to control the progress of a single evolution and avoid weight oscillations. Its value is usually set to a fixed decimal within the range of [0.01, 0.05]. In a specific implementation application, the utility threshold is configured to be an integer within the range of [75, 85], specifically set by medical staff according to the department's baseline requirements for pain management.

[0068] After completing the above numerical calculations, the evolution control module performs boundary constraint determination. The system verifies whether the updated sub-weight values ​​are within the interval (0,1]. If the calculation result exceeds this valid interval, the system forcibly truncates it according to the nearest principle and assigns a value of 0.01 or 1.0, thereby ensuring the stable operation of the underlying mathematical model. Finally, the evolution control module repackages the updated sub-weights into weight vectors and writes them as new initial sub-item weight vectors into the corresponding target node records in the evidence base module. For the underlying data write-back and transaction lock management mechanism of relational databases, those skilled in the art can use standard data update instructions and concurrency control strategies to implement it. The specific data table operations and log recording processes are well-known technologies in this field and will not be elaborated here.

[0069] Through the above steps, the system completes a full closed-loop iteration. The structured evidence nodes generated statically at the bottom layer, through orthogonal mapping feedback of real physiological characteristic factors, achieve adaptive dynamic weight evolution for specific patient groups.

[0070] Specific application examples and numerical calculations Suppose that the current clinical application scenario is to perform a "heel prick blood sampling" procedure on a newborn.

[0071] 1. Evidence Node Matching and Instruction Dispatch During the clinical preparation phase, the interactive terminal acquires basic assessment scores, and the system determines the pain grading threshold field. The status is "moderate". The evidence base module uses the passed-in operation type field. Using "heel prick" and "moderate" as search criteria, matching structured evidence node tuples were extracted. .

[0072] The joint intervention feature vector of this node Two sub-items were activated: the taste-sucking intervention sub-item. (e.g., sugar water) and postural tactile intervention sub-items (e.g., flexion position), auditory and olfactory intervention sub-items This is a null value. Its corresponding initial sub-item weight vector. Including the first sub-weight With the second sub-weight .

[0073] The system calculates the comprehensive recommendation index. The system determines that the recommended solution has the highest strength and issues the command. Upon confirmation of the operation's commencement, the system records the global time anchor point. .

[0074] 2. Calculation of core feature factors During the operation and recovery periods, the system performs dual-channel data processing. (See attached document.) Figure 6 This diagram visually illustrates the use of a global time anchor point. Based on the baseline, the waveform dynamic process of heart rate, blood oxygen and brain oxygen data deviating from their corresponding baseline values.

[0075] Based on the system's processing results of these continuous waveform data, the feature extraction module obtains the cumulative heart rate deviation value calculated during the operation period. Extract the cumulative facial expression value The system has a preset heart rate weighting coefficient. Facial expression weighting coefficient .

[0076] According to the operational rules, the transient stress inhibitory factor output by the transient neural stress channel The calculation is as follows: .

[0077] Similarly, within the recovery period window, the system calculates the cumulative deviation of blood oxygen based on data truncation and accumulation rules. Brain oxygen cumulative deviation value .

[0078] The system presets the blood oxygen weighting coefficient. Brain oxygen weighting coefficient Homeostatic recovery factors output by the homeostatic metabolic perfusion pathway The calculation is as follows: .

[0079] 3. Evolution of Utility Rating and Weighting The evolution control module performs orthogonal mapping based on the aforementioned dual-channel factors. Ideal physiological constants are then set. First mapping coefficient Second mapping coefficient The system calculates a comprehensive utility score. .

[0080] Overall utility score of the current intervention program The score is 82, assuming the department has pre-set a clinically acceptable utility threshold. Learning rate constant .because The feedback-driven term is positive. The evolution control module performs an overwrite for the activated sub-weights (taking the first sub-weight as an example, updating the sub-weight before). ): Updated sub-weights After the system determines the boundary constraints, it repackages the optimized weights and writes them into the evidence base, completing a personalized adaptive evolution.

[0081] Experimental verification and effect comparison: To verify the effectiveness of the dynamic weight evolution closed-loop mechanism in this invention in practical clinical applications, a comparative verification experiment was conducted on 100 newborns undergoing heel prick blood sampling procedures of the same type. The experiment was divided into a control group (50 cases) and an experimental group (50 cases): the control group adopted the traditional static evidence-based intervention distribution scheme with a fixed initial sub-item weight vector; the experimental group adopted the method of this invention, which allows the system to dynamically update the sub-weights based on feedback.

[0082] See attached document Figure 7 The figure shows the trend of the two sets of overall utility scores as the number of operation executions increases: The trend in the control group: Because the underlying structured evidence nodes remain static, it is impossible to adaptively adjust according to the actual physiological tolerance of different children, and its overall utility score... The scores consistently fluctuated irregularly between 70 and 75, with some individuals showing extremely low scores.

[0083] Experimental group trend: Under the dynamic control mechanism of this invention, the system relies on comprehensive utility scoring. With utility threshold The difference between them drives the continuous adjustment of the updated sub-weights. Make minor adjustments. For example... Figure 7 As shown in the curve, after the first 15 adaptive iterations, the overall utility score of the experimental group was... It showed a clear and steady upward trend; after 40 cases, as the underlying weight data of the evidence base was continuously fitted and optimized, the score of the experimental group was stable and converged to above 85 points.

[0084] The experimental comparison data and curve trends above intuitively and powerfully demonstrate that the method of decoupling multi-source physiological characteristic factors and combining learning rate mechanism for orthogonal mapping feedback in this invention can overcome the limitations of traditional static nursing plans, enabling neonatal pain intervention strategies to have real clinical "self-learning" and "self-optimization" capabilities, and significantly improving the analgesia management effect and the quality of microcirculation recovery.

Claims

1. A method for constructing a neonatal pain evidence database, characterized in that, Includes the following steps: The evidence-based medicine text is transformed into structured evidence nodes, which include an operation type field, a pain grading threshold field, a joint intervention feature vector, and an initial sub-item weight vector. Acquire clinical operation type information and basic assessment scores, extract the matching structured evidence nodes, and issue the instruction set corresponding to the joint intervention feature vector, recording the instruction execution time as a global time anchor point; Acquire continuous physiological data and facial expression events from clinical monitoring equipment, and combine the discrete time step with the global time anchor point to convert the continuous physiological data and facial expression events into a standard continuous sequence set and a binary time series sequence, respectively. The cumulative deviation of the standard continuous sequence set from the preset physiological baseline value is calculated, and the binarized time series is directly accumulated to obtain the cumulative expression value. The cumulative deviation value and the cumulative expression value are weighted and summed to obtain the transient stress inhibition factor and the homeostasis recovery factor. The transient stress suppression factor and the steady-state recovery factor are linearly mapped to obtain a comprehensive utility score. The difference between the comprehensive utility score and the preset utility threshold is used to update the initial sub-item weight vector.

2. The method for constructing a neonatal pain evidence database according to claim 1, characterized in that, The process of converting evidence-based medicine text into structured evidence nodes includes: The nursing plan, which includes multiple intervention methods, is broken down into independent dimensions and written into the corresponding positions of the joint intervention feature vector; Evidence-based recommendation level label data is extracted from the evidence-based medicine text, and the evidence-based recommendation level label data is transformed into a set of quantitative constants according to the basic weight mapping table, which serves as the initial sub-item weight vector.

3. The method for constructing a neonatal pain evidence database according to claim 1, characterized in that, The process of obtaining clinical operation type information and basic assessment scores, and extracting matching structured evidence nodes, includes: The basic assessment score is converted into a grading identifier through numerical mapping rules, which serves as the status of the pain grading threshold field. The clinical operation type information and the status of the pain grading threshold field are used as the joint retrieval primary key to extract a set of candidate nodes; The activated sub-weights in the initial sub-item weight vectors of each node in the candidate node set are added together, and the node with the highest sum value is selected as the target node.

4. The method for constructing a neonatal pain evidence database according to claim 1, characterized in that, Following the recording instruction execution time as the global time anchor point, the following is also included: Monitor the process status, and when a clinical operation end command is triggered, record the clock value as the operation end timestamp; The recovery period duration constant is superimposed on the operation end timestamp to calculate the recovery period end timestamp. The global time anchor, the operation end timestamp, and the recovery period end timestamp together construct an absolute time frame for separating physiological response features.

5. The method for constructing a neonatal pain evidence database according to claim 1, characterized in that, In the process of combining the discrete time step and the global time anchor point to convert the continuous physiological data and the facial expression events into a standard continuous sequence set and a binarized time series, respectively, the conversion of the continuous physiological data into a standard continuous sequence set includes: Calculate and generate a standard discrete-time grid sequence; The true sampling points are found in the continuous physiological data, and linear interpolation is performed based on the time interval between adjacent true sampling points and the difference in physiological values ​​to fill in the numerical values ​​and generate the standard continuous sequence set.

6. The method for constructing a neonatal pain evidence database according to claim 5, characterized in that, The process of combining the discrete time step and the global time anchor to convert the continuous physiological data and the facial expression events into a standard continuous sequence set and a binary time series sequence, respectively, also includes: Calculate the time interval between adjacent real sampling points in the continuous physiological data; If the time interval is greater than the preset interruption judgment threshold, the virtual monitoring value at the standard discrete time node of the corresponding segment will be marked as an invalid identifier.

7. A method for constructing a neonatal pain evidence database according to claim 5, characterized in that, In the process of combining the discrete time step and the global time anchor point to convert the continuous physiological data and the facial expression events into a standard continuous sequence set and a binary time series sequence, respectively, the conversion of the facial expression events into a binary time series sequence includes: The duration segments of the facial expression events are mapped to a standard discrete-time grid sequence; At any standard discrete time node, a state determination is performed. If the node is within the duration of the facial expression event, the sequence value at the node is assigned the value of one. If the node is not within the duration of the facial expression event, the value is assigned to zero.

8. A method for constructing a neonatal pain evidence database according to claim 4, characterized in that, In the process of calculating the cumulative deviation of the standard continuous sequence set from the preset physiological baseline value, and directly accumulating the binarized time series to obtain the cumulative expression value, and then weighted summing the cumulative deviation value and the cumulative expression value to derive the transient stress inhibition factor and the homeostasis recovery factor, the process of deriving the transient stress inhibition factor includes: A standard heart rate sequence is extracted from the set of standard continuous sequences, and a heart rate baseline value is obtained from the preset physiological baseline value. The difference between the standard heart rate sequence value and the heart rate baseline value within the operation period time window is calculated. When the difference is greater than zero, the positive deviation difference is retained, and the positive deviation difference is discretely accumulated to obtain the cumulative heart rate deviation value; The cumulative expression value is obtained by directly summing the binary time sequence values ​​within the same classmate; The transient stress inhibition factor is obtained by applying a preset weighting coefficient to the cumulative deviation of heart rate and the cumulative value of facial expression and then summing them.

9. A method for constructing a neonatal pain evidence database according to claim 8, characterized in that, In the process of calculating the cumulative deviation of the standard continuous sequence set from the preset physiological baseline value, and directly accumulating the binarized time series to obtain the cumulative expression value, and then weighted summing the cumulative deviation value and the cumulative expression value to derive the transient stress inhibition factor and the steady-state recovery factor, the process of deriving the steady-state recovery factor includes: Standard blood oxygen saturation sequence and standard brain oxygen saturation sequence are extracted from the standard continuous sequence set. Blood oxygen saturation line value and brain oxygen saturation line value are obtained from the preset physiological baseline value. The difference between the blood oxygen saturation line value and brain oxygen saturation line value within the recovery period time window and the standard blood oxygen saturation sequence value is calculated. When the difference is greater than zero, the difference is retained and discretely accumulated to obtain the corresponding cumulative blood oxygen deviation value and cumulative brain oxygen deviation value. The cumulative deviation of blood oxygen and the cumulative deviation of brain oxygen are weighted and combined using preset weighting coefficients to generate the steady-state recovery factor.

10. A method for constructing a neonatal pain evidence database according to claim 1, characterized in that, The step of updating the initial sub-item weight vector using the difference between the overall utility score and a preset utility threshold includes: The difference is used as a feedback driver. Using a preset learning rate constant and the feedback driving term, the sub-weights corresponding to the activated sub-items within the initial sub-item weight vector are numerically overwritten; Verify whether the sub-weights after numerical overwriting are within the preset range, perform boundary constraint judgment, and overwrite the structured evidence node.

Images