A multi-modal non-invasive cancer pain flare monitoring and warning system and method

Abstract

The present application relates to the field of medical data monitoring and processing, and discloses a multi-modal non-invasive cancer pain burst pain monitoring and early warning system and method, which comprises a wearable terminal, a mobile application module, a cloud engine and a medical workstation. The method relies on the above system to perform: collecting patient pulse wave, skin electricity and surface electromyography signal and pain score data; time aligning and slicing the physiological signals to construct a multi-modal physiological feature vector; extracting a medical history weight factor based on the medical record, generating a state mask for interference data cleaning through consistency comparison; calculating the relative change amount of the features, inversely calibrating the weight parameters with the pain score as the true value label and outputting the estimated pain index; calculating the trend factor according to the estimated pain index sequence and calculating the predicted pain index; comparing the predicted pain index with the threshold value to perform early warning push or linkage with analgesic equipment. The present application can effectively eliminate physiological data interference, realize individualized evaluation and prospective safety warning.

Description

Technical Field

[0001] This invention relates to the field of medical data monitoring and processing technology, specifically to a multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system and method. Background Technology

[0002] Cancer pain is a common clinical symptom in patients with malignant tumors. Breakthrough pain, in particular, is characterized by its sudden onset and severe intensity, requiring timely intervention with analgesics. Current routine methods for monitoring cancer pain primarily rely on patients' subjective descriptions or manual assessment using digital pain scales. This manual assessment method increases the workload of medical staff during ward rounds and has a time lag; by the time patients can clearly call out or express pain, their pain perception has often already reached a high level. Furthermore, this method cannot cover the continuous monitoring needs of patients who are asleep or unconscious.

[0003] To achieve objective, automated, and continuous monitoring, existing technologies have introduced non-invasive physiological monitoring devices to indirectly assess pain. However, current monitoring schemes are typically limited to collecting single-dimensional physiological indicators. In complex clinical environments, routine changes in patient position, fever, or emotional fluctuations can cause sudden changes in single signals such as heart rate or skin surface conductivity, making it difficult for the system to effectively distinguish between physiological stress caused by actual pain and interference from daily activities, resulting in a high false alarm rate. Existing assessment algorithms often use fixed statistical models and universal thresholds, ignoring individual differences in patients' autonomic nervous system baseline levels and stress sensitivity. The models cannot be calibrated based on patients' actual feedback, leading to discrepancies between assessment results and actual pain levels. Current monitoring systems primarily rely on immediate alarms after data exceeds limits, lacking the ability to calculate and prospectively predict the rate of pain evolution by combining continuous time-series data, and failing to allow healthcare personnel sufficient lead time for intravenous injections or drug onset. In the technical solutions for automated linkage between the system and devices such as self-controlled analgesia pumps, the lack of physical safety interlocking logic that matches the pharmacokinetics of analgesic drugs poses a risk of triggering multiple dosing when faced with continuous high-level warning signals, making it difficult to ensure clinical medication safety. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system and method, which solves the problems of existing cancer pain monitoring schemes that cannot effectively eliminate interference from daily physiological signals, resulting in a high false alarm rate, fixed assessment models lacking individual dynamic adaptability, delayed early warning, and automated linkage analgesia devices lacking a safety interlock mechanism.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The first aspect of the present invention provides a multimodal non-invasive cancer pain breakthrough monitoring and early warning system, including a wearable terminal, a mobile application module, a cloud engine and a medical workstation.

[0006] The wearable terminal collects the patient's pulse wave, skin conductance, and surface electromyography signals and sends them to the mobile application module; The mobile application module is used to receive pain score data and transmit the received physiological signals and pain score data to the cloud engine. The cloud engine includes: The preprocessing module is used to perform time alignment and slicing of physiological signals, calculate variability features, and construct multimodal physiological feature vectors. The discrimination module extracts prior medical history weight factors based on medical record data, performs consistency comparison on variable features to generate a state mask, uses the state mask to clean up interfering data and outputs an effective multimodal physiological feature vector. The calibration module is used to calculate the relative change of the effective multimodal physiological feature vector relative to the individual physiological baseline vector, construct a loss function using pain score data as ground truth labels, inversely calibrate the weight parameters, and output the estimated pain index. The trend analysis module is used to calculate the trend factor of the estimated pain index sequence and extrapolate the predicted pain index. The early warning module is used to compare the predicted pain index with the set early warning threshold to generate an early warning command, which is then pushed to the medical workstation.

[0007] Preferably, the preprocessing module constructs multimodal physiological feature vectors including: Extract the time interval sequence of adjacent peak points of the pulse wave signal, calculate the average value, and generate pulse wave feature values; The skin electromyography (EMG) response component of the isolated EMG signal is extracted to generate EMG characteristic values ​​by extracting the pulse peak amplitude; power frequency interference of the surface EMG signal is filtered out and the root mean square amplitude value is calculated to generate EMG characteristic values. The dimensions of pulse wave feature value, skin conductance feature value and electromyography feature value are combined to construct a multimodal physiological feature vector.

[0008] This technical solution extracts the variability features of cardiovascular, sympathetic nerve and somatic muscle tension, and transforms discrete heterogeneous time-series signals into a unified matrix data structure, thus establishing a data foundation for subsequent multidimensional index fusion calculation.

[0009] Furthermore, the discrimination module generates a state mask including: Configure numerical prior medical history weighting factors based on the disease label categories included in the patient's medical record data; Each feature value is compared and mapped to a dynamically determined physiological fluctuation threshold as an activation state indicator variable; The consistency score is calculated by combining the activation status indicator variables with the prior medical history weight factor; the consistency score is compared with the mask judgment threshold and the status mask is output. Multiply the state mask with the multimodal physiological feature vector to output the effective multimodal physiological feature vector.

[0010] This technical solution utilizes the medical history weighting system to assess the stress sensitivity of specific lesion areas. It uses the synchronous activation judgment logic of multi-dimensional signals to screen out single-dimensional environmental or action interference. It then uses the generated mask matrix to directly perform matrix multiplication operations, thereby achieving physical layer isolation between real pain physiological data and non-pain interference data.

[0011] In a preferred embodiment of the present invention, the calibration module reverse calibrates the weight parameters including: Extract continuous and effective multimodal physiological feature vectors from individuals in a pain-free resting state, calculate their average values, and generate an individual physiological baseline vector. The relative change in features is obtained by calculating the difference between the effective multimodal physiological feature vector of the current time window and the individual physiological baseline vector. The estimated pain index is obtained by calculating the relative changes of features and the weight parameters corresponding to each feature. A loss function is constructed using pain score data, the mean squared error of the predicted pain index, and a regularization term for the weight parameters. The gradient of the loss function is calculated using the gradient descent algorithm, and the weight parameters are updated.

[0012] This technical solution introduces real digital scale scores from patients to construct a closed-loop assessment model with regularization constraints, eliminating the impact of differences in the group's physiological baseline on the accuracy of the algorithm, and enabling the system to have the ability to adaptively correct for individual autonomic neural response patterns.

[0013] Preferably, the calibration module is also configured with zero-order hold logic: When the state mask indicates that the current data is interference data, the effective multimodal physiological feature vector is set to zero, the calibration module skips the calculation of the relative change of features and loss function within the current time window, and the output predicted pain index directly uses the value of the previous effective time window.

[0014] This technical solution avoids negative data jumps caused by directly subtracting the zero vector generated by the cleaning action from the reference vector, thus maintaining the continuity of the model at the mathematical operation level.

[0015] Furthermore, the trend analysis module calculates and predicts the pain index, including: A linear fitting algorithm was used to fit the predicted pain index sequence within the sliding time window, and the slope of the fitted line was extracted as a trend factor. When the trend factor is greater than or equal to zero, the current estimated pain index is superimposed on the product of the trend factor and the forward forecast time span to obtain the predicted pain index. When the trend factor is less than zero, the predicted pain index is directly assigned the current estimated pain index.

[0016] This technical solution uses historical time-series change rate to quantify the rate of pain evolution, reserving time for pre-intervention to allow for drug efficacy; at the same time, it sets a one-way numerical truncation logic for pain relief trends to prevent the formula derivation from producing underestimations that have no clinical reference significance.

[0017] In another preferred embodiment of the present invention, the early warning module generates early warning instructions including: Set moderate and severe pain thresholds and map the predicted pain index to an indicator variable for the warning level; Extract the predicted pain index from adjacent historical time windows, calculate the percentage of sampling points that exceed the moderate pain threshold, and generate alarm confidence. When the alarm confidence level is greater than or equal to the confidence threshold, the warning status is confirmed to be valid, and a regular attention instruction or an emergency intervention instruction is generated based on the warning level indicator variable.

[0018] This technical solution constructs a time delay fault-tolerant mechanism based on time-series data to filter out transient high-frequency interference caused by device touch or occasional patient movements, thereby reducing alarm fatigue caused by frequent false alarms.

[0019] Preferably, the early warning module is configured with security lockout time verification logic based on emergency intervention commands: Get the time difference between the current system timestamp and the last recorded timestamp of the self-controlled analgesia pump administering additional medication; When the time difference is greater than or equal to the safety lockout time parameter, a drive signal is generated to control the additional dose of the self-controlled analgesia pump. When the time difference is less than the safety lockout time parameter, the control of the self-controlled analgesia pump remains locked.

[0020] This technical solution forcibly introduces a time-physical boundary determined by pharmacokinetics into the automated drug delivery process, thereby avoiding the risk of over-infusion caused by continuous triggering of high-level warning states.

[0021] Furthermore, the preprocessing module extracts a globally unified timestamp based on system clock information and uses an interpolation algorithm to align asynchronously sampled pulse waves, electrodermal activity (EDA), and electromyography (EMG) signals onto the same time axis for slicing. This technical solution eliminates the inherent sampling frequency time difference of multiple types of hardware sensors from the data source, ensuring the consistency of temporal alignment of multimodal physiological features in subsequent time window slicing and matrix operations.

[0022] A second aspect of the present invention provides a multimodal non-invasive cancer pain breakthrough pain monitoring and early warning method, comprising the following steps: Acquire patient pulse wave, skin conductance and surface electromyography signals, as well as pain score data; Physiological signals are time-aligned and sliced, variability features are calculated, and multimodal physiological feature vectors are constructed. Based on medical record data, prior medical history weight factors are extracted, consistency comparison of variable features is performed to generate state masks, and the state masks are used to clean up interfering data and output effective multimodal physiological feature vectors. Calculate the relative change of effective multimodal physiological feature vectors relative to individual physiological baseline vectors, construct a loss function using pain score data as ground truth labels, back-calibrate weight parameters, and output the predicted pain index; Calculate the trend factor of the estimated pain index sequence to extrapolate the predicted pain index; The predicted pain index is compared with the warning threshold to generate a warning command for information push or device linkage.

[0023] This invention provides a multimodal, non-invasive cancer pain breakthrough pain monitoring and early warning system and method. It has the following beneficial effects: 1. This invention extracts prior medical history weighting factors through a discrimination module and generates a state mask by comparing the multimodal variability features with dynamic thresholds. The state mask is then used to perform a multiplication operation on the physiological feature vector to isolate non-pain interference data. This enables the system to effectively distinguish between the combined physiological stress caused by the patient's actual pain sensation and the single signal fluctuation caused by routine turning or abnormal body temperature, reducing the false alarm rate of the system in complex clinical environments.

[0024] 2. This invention extracts individual physiological baseline vectors through a calibration module to calculate the relative changes in features, and introduces patient-entered pain score data as ground truth labels. It then uses a loss function with regularization and a gradient descent algorithm to inversely update the weight parameters of each physiological feature. This eliminates the influence of differences in autonomic nervous system baseline levels between individual patients on the calculation results, enabling the pain assessment formula to dynamically adjust assessment parameters based on real feedback, thus improving the accuracy of the system's predicted pain index.

[0025] 3. This invention uses a trend analysis module to extract the fitting slope of the predicted pain index sequence to calculate and predict the pain index. Furthermore, it introduces a safety lock-in time verification logic into the device linkage link of the early warning module. By comparing the difference between the system timestamp and the previous injection timestamp, the locking state of the self-controlled analgesia pump is controlled. This achieves proactive early warning to reserve time for clinical intervention while establishing a mandatory physical boundary determined by pharmacokinetic parameters. This prevents the system from issuing over-infusion commands under continuously exceeding limits, ensuring the safety of automated drug delivery. Attached Figure Description

[0026] Figure 1 This is a system module architecture diagram of the present invention; Figure 2 This is a flowchart of the method steps of the present invention; Figure 3 The following is a multimodal physiological signal and trend prediction evolution curve of the cancer pain breakthrough pain prodromal period of the present invention. Among them, sub-figure (a) is the evolution curve of pulse wave and skin electrophysiology characteristics, sub-figure (b) is the evolution curve of surface electromyography characteristics, and sub-figure (c) is the estimated pain index and prospective prediction trend. Figure 4 This is a graph showing the convergence verification of the mean absolute error in the personalized calibration process of the pain assessment model of the present invention. Figure 5 This is a comparison chart of receiver operating characteristics between the experimental group and the control group in breakthrough pain monitoring of the present invention. Sub-figure (a) is a bar chart comparing the false alarm rate of the clinical system, and sub-figure (b) is a ROC curve comparing the breakthrough pain detection performance. Detailed Implementation

[0027] 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.

[0028] Please see the appendix Figure 1 This invention provides a multimodal non-invasive cancer pain breakthrough monitoring and early warning system, comprising: a wearable terminal, a mobile application module, a cloud engine, and a medical workstation.

[0029] Wearable devices include pulse wave sensors, skin conductance sensors, electromyography (EMG) sensors, and communication modules.

[0030] The pulse wave sensor is configured to acquire the patient's cardiovascular signals. The skin conductance sensor is configured to acquire the patient's skin conductance signals. The electromyography (EMG) sensor is configured to acquire surface EMG signals reflecting changes in the patient's muscle tension.

[0031] The communication module is electrically connected to the pulse wave sensor, skin conductance sensor, and electromyography sensor. The communication module establishes a communication link with the mobile application module and transmits the collected physiological signals to the mobile application module in real time.

[0032] The mobile application module runs on smart devices. It features a data interaction interface to display the device's connection status and operating parameters. The module also includes a tag input interface to receive pain score data (such as digital pain scale scores) and analgesic pump press event data input by the patient.

[0033] The mobile application module communicates with the cloud engine. The mobile application module transmits the received physiological signals and tag data to the cloud engine.

[0034] The cloud engine is configured with multiple logical modules that execute data processing algorithms. The cloud engine includes a preprocessing module, a discrimination module, a calibration module, a trend analysis module, and an early warning module.

[0035] The healthcare workstation establishes a data connection with the cloud engine. The workstation includes a terminal display screen and mobile nursing devices. The workstation is configured to receive and display alerts from the cloud engine.

[0036] See attached document Figure 2 This invention provides a multimodal, non-invasive method for monitoring and early warning of cancer pain breakthrough pain, comprising the following steps: S1, the wearable terminal acquires the patient's continuous pulse wave signal, skin conductance signal and electromyography signal, and transmits the physiological signals to the cloud engine through the mobile application module; S2, the cloud engine calls the preprocessing module to establish a unified timestamp to perform time alignment processing on physiological signals, and performs time-domain slicing with a preset length to calculate the variability characteristics of physiological indicators within each analysis time window; S3, the cloud engine calls the discrimination module to retrieve the patient's past medical history records to establish judgment weights, performs consistency comparison on the variability characteristics of various physiological indicators within the same time window, outputs a state mask based on the comparison results, separates the real physiological stress data and removes interfering data; S4, the cloud engine calls the calibration module to extract the physiological characteristic variability indicators of the patient in a painless resting state to establish a personal physiological baseline, calculate the change of real-time monitoring characteristics and personal physiological baseline, use the digital pain scale score data uploaded by the mobile application module as the true value label, use the loss function to build a model to update and calibrate the weight parameters corresponding to each physiological indicator in the pain assessment formula, and output the estimated pain index of the current time window. S5, the cloud engine calls the trend analysis module to extract the continuous estimated pain index sequence within the sliding time window, calculates the change rate trend factor of the pain index, and uses the trend factor to estimate the predicted pain index at the set future time point. In S6, the cloud engine calls the early warning module to compare the predicted pain index with the graded early warning thresholds set in the system memory, and generates corresponding early warning instructions to be pushed to the medical workstation.

[0037] The following detailed description of the various system modules and algorithm logic of the above-described technical solution of the present invention, in conjunction with specific embodiments, provides a detailed explanation.

[0038] The cloud engine invokes the preprocessing module to perform time synchronization and feature dimensionality reduction on the physiological signals collected by the wearable terminal. The specific data processing procedure performed by the preprocessing module is as follows: S201. In practical non-invasive monitoring environments, the sampling frequencies of pulse wave sensors, electrodermal sensors, and electromyography (EMG) sensors typically have inherent differences. To eliminate this hardware-level asynchrony, the preprocessing module extracts a globally unified timestamp based on the system clock information attached to the communication module and uses interpolation algorithms such as linear interpolation or cubic spline interpolation to align the acquired cardiovascular, electrodermal, and surface EMG signals onto the same time axis. The preprocessing module sets a fixed-length time window on the time axis to slice the data, dividing it into continuous analysis intervals. Optionally, the length of this time window can be set to 10s-60s to ensure that the system can sensitively capture the transient physiological mutations caused by acute breakthrough pain.

[0039] Since sympathetic nerve excitation directly affects heart rhythm, manifesting as changes in heart rate variability, this application extracts the variability characteristics of pulse wave signals to reflect this physiological stress. The preprocessing module performs bandpass filtering on the pulse wave signal, typically set between 0.5Hz and 5Hz, to remove baseline drift and high-frequency noise. Within a time window, the preprocessing module identifies peak points in the pulse wave signal, calculates the time interval sequence between adjacent peak points, and then calculates the heart rate variability characteristic value. The formula for calculating the heart rate variability characteristic value by the preprocessing module is as follows: ; In the formula, For the first Pulse wave characteristic values ​​for each time window; For the first The total number of pulse wave peaks within a time window; The count variable is the peak value of the pulse wave; For the first The time interval between the peak of each pulse wave and the previous peak; For the first The arithmetic mean of the peak time intervals of all pulse waves within a given time window. To ensure the completeness and stability of the algorithm logic, when... hour, The default value is 0.

[0040] S202. When the body is under stress, activation of the sympathetic nervous system increases sweat gland secretion, leading to a short-term abrupt change in skin conductivity. The preprocessing module performs convex optimization decomposition on the skin electrical signal, separating the original signal into a low-frequency baseline component of skin electrical conductivity and a high-frequency component of skin electrical response. This signal decomposition process can be implemented by those skilled in the art using known algorithms such as nonnegative matrix factorization, and will not be elaborated here. The preprocessing module statistically analyzes the number of independent pulses generated by the separated skin electrical response component within the time window and extracts the peak amplitude data of each pulse. The formula for calculating the sympathetic nervous system variability eigenvalues ​​in the preprocessing module is as follows: ; In the formula, For the first Skin electrophysiological characteristics within a time window; For the first The total number of skin conductance response pulses extracted within each time window; The counting variable for skin conductance response pulses; For the first The highest amplitude value of each skin conductance pulse. If no valid skin conductance pulse is detected within the current time window, i.e. hour, The value is set to 0.

[0041] Patients experiencing severe pain often display a pained expression or are forced into a fixed posture. These involuntary muscle movements can be quantified by changes in the amplitude of surface electromyography (EMG) signals. The preprocessing module uses a notch filter to remove 50Hz or 60Hz power frequency interference from the EMG signals and calculates the root mean square value reflecting the degree of body tension based on the amplitude of the filtered signal. The formula for calculating the muscle tension characteristic value by the preprocessing module is as follows: ; In the formula, For the first Electromyographic characteristics of each time window; For the first The total number of discrete sampling points of electromyography signals within a time window; The count variable for the sampling points; For the first The digital amplitude of the electromyographic signal after filtering for each sampling point.

[0042] S203, the preprocessing module merges the pulse wave feature values, skin conductance feature values, and electromyography feature values ​​calculated within the same time window. This merging operation transforms the discrete parameters of the sympathetic nervous system and somatic motion dimensions into a unified data structure that can be used for matrix operations by subsequent logic modules. The formula for constructing the multimodal physiological feature vector by the preprocessing module is as follows: ; In the formula, For the first The multimodal physiological feature vectors output by each time window; For the first Pulse wave characteristic values ​​for each time window; For the first Skin electrophysiological characteristics within a time window; For the first The electromyographic feature values ​​for each time window are calculated. The preprocessing module then transmits the calculated multimodal physiological feature vectors to the discrimination module.

[0043] The cloud engine invokes the discrimination module to verify the authenticity of the multimodal physiological feature vectors input from the preprocessing module. The specific data processing procedure performed by the discrimination module is as follows: S301, the discrimination module retrieves the patient's electronic medical record data to construct prior medical history weights. Different underlying disease states can lead to differences in the stress sensitivity of local nerves and muscles to external stimuli. The discrimination module extracts specific disease labels from the medical record data and maps them to numerical prior medical history weight factors. When the medical record data contains labels related to bone metastasis or nerve compression, it indicates that the patient often experiences mechanical breakthrough pain when changing position. In this case, simple muscle movement signals have higher pain indicative reference value.

[0044] Based on this, the discrimination module assigns a compensation constant greater than 1 to the prior medical history weight factor for electromyography dimension features, for example, it can be set in the range of 1.5-2.5; when there is no corresponding label in the medical record data, the weight factor is assigned a default value of 1.

[0045] S302, the discrimination module performs threshold comparison and binarization mapping on the multimodal physiological feature vector. The discrimination module compares the pulse wave feature value, skin conductance feature value, and electromyography feature value in the multimodal physiological feature vector with the various physiological fluctuation thresholds set in the system memory.

[0046] In a preferred implementation, the aforementioned physiological fluctuation threshold can be dynamically determined by collecting physiological data of the patient in a stable resting state, calculating its statistical mean, and adding 2-3 times the standard deviation to eliminate background noise interference. The discrimination module uses a step function to map continuous feature values ​​into discrete activation state indicator variables. The formulas for calculating each activation state indicator variable by the discrimination module are as follows: ; ; ; In the formula, For the first The pulse wave activation status indicator variable for each time window; For the first Pulse wave characteristic values ​​for each time window; The threshold for physiological fluctuations in the pulse wave; For the first Skin electrophysiological activation status indicator variable for each time window; For the first Skin electrophysiological characteristics within a time window; The threshold for skin electrophysiological fluctuations; For the first Electromyographic activation state indicator variables for each time window; For the first Electromyographic characteristics of each time window; This represents the threshold for electromyographic fluctuations.

[0047] S303, the discrimination module constructs a multi-signal consistency logic check matrix and generates a state mask. Non-pain interference factors exist in the clinical environment.

[0048] For example, fever in patients usually leads to increased heart rate and sweating, which is reflected in the signal as a synchronous increase in pulse wave and skin conductance, but without the electromyographic mutations that characterize a painful face or forced posture. For example, routine turning over in patients without bone metastases usually only triggers a mutation in electromyographic signals, without being accompanied by sympathetic sweating impulses.

[0049] The discrimination module performs a linear combination of the activation state indicator variable and the prior medical history weighting factor to calculate a consistency score, and compares this score with a decision threshold to output a binary state mask. The formula for calculating the state mask by the discrimination module is as follows: ; ; In the formula, For the first Consistency score for each time window; For the first The pulse wave activation status indicator variable for each time window; For the first Skin electrophysiological activation status indicator variable for each time window; Prior medical history weighting factors for electromyography (EMG) dimension features; For the first Electromyographic activation state indicator variables for each time window; For the first The state mask output for each time window; This is the masking threshold. To effectively eliminate single interferences while also considering specific medical histories, the masking threshold... The default setting is 2. Under this configuration, for typical patients ( ( ), at least two dimensions of physiological signals must be synchronously activated to trigger the state mask setting; however, for patients with bone metastases, if set In this case, a single, intense physical movement by the patient can cause the consistency score to reach the masking threshold, thus ensuring that mechanical burst pain is not missed.

[0050] S304, the discrimination module performs data cleaning and isolation using a state mask. The discrimination module multiplies the generated state mask (containing 0s or 1s) with the multimodal physiological feature vector input from the preprocessing module. When the system determines that the current data fluctuation is due to interference caused by changes in body position or environmental heat, the state mask is set to 0, the multimodal physiological feature vector is set to zero, and interfering data is prevented from entering the subsequent pain assessment model. When the system determines that the multimodal physiological features represent a genuine combined sympathetic and motor nerve stress response, the state mask is set to 1, and the genuine data is allowed to pass. The specific formula for the cleaned feature vector output by the discrimination module is as follows: ; In the formula, For the first Effective multimodal physiological feature vectors after cleaning within a time window; For the first The state mask output for each time window; For the first The multimodal physiological feature vectors are input within a time window. The discrimination module transmits the cleaned and effective multimodal physiological feature vectors to the calibration module.

[0051] The cloud engine calls the calibration module to receive the valid multimodal physiological feature vectors output by the discrimination module, and uses a dynamic adaptive algorithm to personalize the pain assessment model. Specifically, the data processing procedure performed by the calibration module is as follows: S401, the calibration module establishes the patient's personal physiological baseline vector and calculates the relative changes in features. Significant differences exist in autonomic baselines such as baseline heart rate and baseline skin impedance among individuals; directly using absolute physiological values ​​for assessment can easily lead to misjudgment. In one feasible implementation, the system collects multiple consecutive effective multimodal physiological feature vectors during the patient's stable resting period upon admission and when the patient's reported digital pain scale score is 0 (i.e., pain-free resting state), calculates the arithmetic mean, and generates the personal physiological baseline vector. During actual monitoring, the calibration module calculates the relative change between the current effective multimodal physiological feature vector and the personal physiological baseline vector. The formula for calculating the relative change in the calibration module is as follows: ; ; In the formula, For the first A vector of relative changes in multimodal physiological characteristics over a time window; For the first Effective multimodal physiological feature vectors after cleaning within a time window; This represents the individual's physiological baseline vector. This represents the individual's physiological baseline constant for pulse waves. This represents the individual's baseline skin electrophysiological constant. This represents the individual's baseline electromyographic constant.

[0052] S402, the calibration module constructs the basic objective pain assessment formula. The calibration module uses the relative changes calculated above as input parameters, combined with the weight parameters corresponding to each physiological dimension and the bias constant, to calculate the estimated pain index for the current time window. The formula for calculating the estimated pain index by the calibration module is as follows: ; In the formula, For the first The estimated pain index output by the system within each time window; This represents the weight vector for each physiological dimension; This is the transpose of the weight vector; For the first A vector of relative changes in multimodal physiological characteristics over a time window; This is a bias constant, which is usually set to 0 during system initialization; These are the characteristic weighting parameters for the pulse wave; For the first The relative changes in pulse wave characteristics within each time window; For skin electrophysiological characteristics weight parameters; For the first The relative change in skin electrophysiological characteristics within a time window; These are the weighting parameters for electromyographic features; For the first The relative changes in electromyographic features over each time window. The weight vector in the system initialization state. Statistical empirical preset values ​​based on population clinical data were adopted.

[0053] It should be noted that when the discrimination module outputs the state mask... A value of 0 (i.e., the current time window is determined to be non-pain-affected data) results in an effective multimodal physiological feature vector after cleaning. When the zero vector is used, the calibration module will skip the calculation of the relative change of features in the current time window and the formula for the estimated pain index. The estimated pain index output by the system will directly use the value of the previous effective time window (i.e., zero-order hold logic is used). This avoids negative abnormal fluctuations caused by directly subtracting the zero vector from the baseline vector, and ensures the completeness of the data cleaning logic at the mathematical operation level.

[0054] S403, the calibration module receives dynamically anchored truth labels and constructs a loss function. To adapt the pain assessment model to individual patient differences, the system requires real-world feedback data for calibration. The calibration module obtains the label data via the mobile application module.

[0055] To prevent the model from generating erroneous reverse gradients under extreme high pain conditions, preferably, the mapping value is simultaneously forced by the system to be no less than the estimated pain index output by the system in the current time window, so as to ensure that the model maintains or positively increases the weight sensitivity of the corresponding physiological characteristics when the patient takes emergency analgesia actions.

[0056] The calibration module combines the predicted pain index with the true value label to construct a mean squared error loss function including a regularization term. Introducing a regularization term into the loss function aims to limit drastic fluctuations in the weight parameters and prevent the model from overfitting to individual, highly subjective, abnormal patient scores. The formula for the loss function constructed by the calibration module is as follows: ; In the formula, This represents the mean squared error loss function value for the current calculation period; For the first The estimated pain index output by the system within each time window; For the first The true value label received within each time window; The regularization coefficient is set and typically ranges from 0.01 to 0.1. Let L be the L2 norm of the weight vector.

[0057] S404, the calibration module performs inverse self-calibration of personalized weight parameters. The calibration module uses the gradient descent algorithm to calculate the gradient of the loss function with respect to the weight vector and inversely fine-tunes the weight parameters in the pain assessment formula. The algorithm iteratively finds the parameter combination that minimizes the error between the true label and the system's predicted value. The formula for updating the weight vector by the calibration module is as follows: ; ; In the formula, This is the gradient vector of the loss function with respect to the weight vector; The estimated pain index output by the system for the k-th time window; For the first The true value label received within each time window; For the first A vector of relative changes in multimodal physiological characteristics over a time window; The set regularization coefficient; This is the old weight vector before the update; This is the updated weight vector; The learning rate parameter is set, and its value is usually between 0.001 and 0.05 to ensure smooth iteration.

[0058] It should be noted that during the monitoring time window when no truth labels are uploaded from the mobile application module, the system does not trigger the above reverse update process, but maintains the current weight parameter combination and continues to output the objective estimated pain index according to the formula.

[0059] Let's take a clinically common example of an anxiety-prone patient to illustrate the principle behind this algorithm. When faced with mild discomfort, anxiety-prone patients often experience overactivity of their sympathetic nervous system, leading to an abnormally high relative change in skin conductance characteristics. This, in turn, causes the system to initially calculate an overestimated pain index. When the patient enters a lower value reflecting their actual mild pain as the true value label, the difference between the estimated value and the true value label in the formula becomes a positive number.

[0060] Based on the gradient calculation and update formulas described above, the calibration module generates negative gradient compensation during reverse updates, automatically reducing the values ​​of the skin conductance feature weight parameters. This closed-loop adaptive update mechanism helps evaluate how the model dynamically adapts to the unique physiological response patterns of different patients, completing the calibration from a general population model to a personalized model.

[0061] The cloud engine calls the trend analysis module to receive the estimated pain index for the current time window from the calibration module and performs a forward-looking prediction by combining it with historical time-series data. Specifically, the data processing procedure performed by the trend analysis module is as follows: S501, the trend analysis module extracts a continuous sequence of predicted pain indices within a sliding time window. Because cancer breakthrough pain is usually accompanied by continuous physiological changes, a single static assessment often cannot reflect the rate of pain worsening in a timely manner; therefore, dynamic analysis based on time-series data is necessary. The trend analysis module sets a fixed-length historical backtracking sliding time window in system memory.

[0062] As monitoring progresses, the trend analysis module continuously collects the continuously estimated pain index output by the calibration module to construct a time-series data sequence. In cases where historical time-series data is insufficient during the initial system startup or due to data cleaning, the trend analysis module employs zero-order hold logic to fill the sequence with the most recent valid data to ensure the integrity of the data dimensions within the sliding time window.

[0063] In specific clinical applications, the length of the retrospective sliding time window is usually set between 3 and 10 minutes. This setting range is beneficial for the system to capture data on physiological indicators during the period of sudden changes in the pre-acute breakthrough pain.

[0064] S502, the trend analysis module calculates the time-series change rate and trend factor. The trend analysis module calculates the change rate of the pain index in the historical sequence to quantify the rate of deterioration of autonomic nervous system and somatic tension during the prodromal phase of breakthrough pain. The trend analysis module uses the least squares method to linearly fit the estimated pain index sequence within the retrospective sliding time window, extracting the slope of the fitted line as the trend factor. In the specific system implementation, the polynomial expansion and matrix solution process of the least squares method can be implemented using standard data fitting algorithms by those skilled in the art, and will not be elaborated here. The formula for calculating the trend factor in the trend analysis module is as follows: ; In the formula, The trend factor is calculated for the current time window; This represents the total number of data points contained within the backtracking sliding time window; For counting data points; For the first The timestamps corresponding to each data point; This is the arithmetic mean of all timestamps within the backtracking sliding time window; For the first Each data point corresponds to a time window in the system's estimated pain index. This is the arithmetic mean of all predicted pain indices within the retrospective sliding time window.

[0065] S503, the trend analysis module uses a prospective prediction model. In clinical practice, intervention for cancer breakthrough pain usually requires allowing sufficient time for drug onset. The trend analysis module sets a prospective prediction time span, superimposing the estimated pain index of the current time window with a numerical increment calculated based on trend factors to obtain the predicted pain index. The prospective prediction time span is typically set to 5-15 minutes to allow medical staff time for intravenous injections or adjustments to analgesic pump parameters. The formula for calculating the predicted pain index in the trend analysis module is as follows: ; In the formula, The predicted pain index obtained through calculation; The estimated pain index output by the system for the current time window; The trend factor is calculated for the current time window; The set forward forecast time span.

[0066] To ensure the completeness of the algorithm logic and the physical meaning of the output values, the trend analysis module introduces boundary judgment logic before performing the inference calculation. When the trend factor is less than zero, it indicates that the patient's pain is in a relief or decline trend. If the value is directly substituted into the formula at this time, it may produce abnormally low predicted values ​​or negative values ​​with no clinical reference significance.

[0067] In this scenario, the trend analysis module directly assigns the predicted pain index to the estimated pain index for the current time window. When the trend factor is greater than or equal to zero, indicating that the pain is intensifying, the trend analysis module performs an accumulation calculation according to the aforementioned deduction formula. Furthermore, if the accumulated predicted pain index exceeds the maximum threshold of the system's pain rating scale (e.g., a maximum score of 10 on the Digital Nucleic Acid Scale), the trend analysis module truncates the output to this maximum threshold to prevent out-of-bounds anomalies. After calculation, the trend analysis module transmits the generated predicted pain index to the early warning module.

[0068] The cloud engine calls the early warning module to receive the predicted pain index output by the trend analysis module, and then performs multi-level early warning assessments and terminal linkage control. Specifically, the data processing procedure performed by the early warning module is as follows: S601, the early warning module performs a step-by-step threshold comparison and level classification on the predicted pain index. Clinically, differentiated interventions are required for different degrees of pain; therefore, the system needs to discretize the continuous predicted values ​​into corresponding early warning levels. The early warning module sets moderate and severe pain thresholds in the system configuration. In the mapping standard of the Clinical Digital Pain Scale, the moderate pain threshold is usually set to 4, and the severe pain threshold is usually set to 7. The formula for calculating the early warning level indicator variable by the early warning module is as follows: ; In the formula, This is the warning level indicator variable for the current time window, where 0 represents a mild or painless state that does not require intervention, 1 represents a moderate pain state that requires attention, and 2 represents a severe burst pain state that requires emergency intervention. The predicted pain index is calculated by the trend analysis module. The set threshold for moderate pain warning; The set threshold for severe pain warning.

[0069] S602, the early warning module constructs a time-delay fault-tolerance mechanism to calculate alarm confidence. Because physiological signals may be subject to transient high-frequency interference caused by accidental coughing or instrument contact during acquisition, immediately triggering an alarm for a single instance of exceeding the limit could easily lead to frequent false alarms and alarm fatigue due to decreased attention among medical staff. The early warning module extracts the predicted pain index corresponding to the current time window and multiple adjacent historical time windows, calculating the proportion of time spent within the alarm range. The formula for calculating alarm confidence by the early warning module is as follows: ; ; In the formula, The calculated alarm confidence level; The set confidence level verification window length is typically set to 3-5 consecutive time windows to balance the timeliness and smoothness of the judgment. This is the window offset variable; It is a step activation function; For the first The predicted pain index corresponding to each time window; The set moderate pain warning threshold is used. The warning module determines the confidence level of the alarm. If the confidence threshold is greater than or equal to the preset confidence threshold (e.g., 0.6 based on clinical experience, representing that more than 60% of the sampling points are in an over-limit state during the time period under investigation), the current warning status is confirmed to be valid, and the subsequent issuance process is executed; if it is less than the confidence threshold, it is regarded as transient interference and filtered out, and the system remains silent.

[0070] S603, the early warning module generates corresponding network dispatch instructions based on the early warning level. When the early warning status is confirmed to be valid and the early warning level indicator variable is 1, the early warning module generates a routine attention instruction and encapsulates this instruction and chart data containing multimodal physiological characteristics into a standardized network data packet, which is then pushed to the mobile application module equipped by medical staff and patients' families via a wireless communication link to prompt them to conduct routine rounds.

[0071] In practical implementation, the encapsulation and routing of network data packets can be achieved using standard communication protocols such as TCP / IP by those skilled in the art. The underlying network transmission is a well-known technology in this field and will not be elaborated upon here. When the warning status is confirmed to be valid and the warning level indicator variable is 2, the warning module generates an emergency intervention command and triggers the highest priority audio-visual vibration alert at the interface of the mobile application module.

[0072] S604, the early warning module executes automated analgesia device linkage based on safety lockout period foolproof logic. For critically ill patients with authorization to link with the self-controlled analgesia pump, the early warning module directly converts emergency intervention commands into drive level signals that control the self-controlled analgesia pump to release a single additional dose.

[0073] To prevent continuous high-level alarms from causing analgesic drug overdose and resulting in respiratory depression or other medical accidents, the early warning module configures a safety lockout time check before issuing the drive level signal. By introducing a time-dimensional lockout state, multiple invalid triggers within a short period are limited. The formula for the early warning module to generate the analgesic pump drive signal is as follows: ; ; ; In the formula, The Boolean drive signal output to the self-controlled analgesia pump is 1, which represents triggering additional drug administration, and 0 represents maintaining the lockout. This is the indicator variable for the warning level output in the current time window; The confidence level is satisfied by the flag bit; Set the safety lock release flag; The calculated alarm confidence level; The set confidence threshold; This is the current system timestamp; The timestamp recorded in system memory of the last time the patient-controlled analgesia pump administered an additional dose; This is the set clinical safety lockout time parameter. This safety lockout time parameter... The preparation time is determined based on the half-life and pharmacokinetic characteristics of the analgesic drug used, and is usually between 15 min and 30 min.

[0074] As a preferred implementation, when the system is first started and initialized or when the self-controlled analgesia pump is first connected, the system memory will... The initial value is assigned as the current system timestamp minus the security lock time parameter. The difference. The above physical safety boundary verification logic ensures that the drug administration operation of the closed-loop analgesia system remains within the allowable range of clinical safe drug use guidelines even when high-level alarms are continuously triggered.

[0075] Specific application examples: Please see the appendix Figure 3 -Appendix Figure 5 To enable those skilled in the art to better understand the execution process of this invention, the following is a detailed explanation of the collaborative operation of the various logic modules of this system, using a specific clinical application example and incorporating a set of actual monitoring values.

[0076] Clinical scenario setup and system initialization (S401-S402): Assume the patient is a 65-year-old male with bone metastasis in his medical record. The system uses his prior medical history as a weighting factor. The value was assigned to 2.0. During the system initialization phase, the individual physiological baseline vector was collected during the patient's stable resting period when they experienced mild pain (approximately 2 minutes). The system's preset initial value for the weight vector is... Bias constant The system's set threshold for physiological fluctuations is: =30, , Mask determination threshold .

[0077] Interference data cleaning phase (simulating a normal rolling motion) (S302-S304) during the monitoring During this time window, the patient underwent a routine, painless repositioning. The system collected the following characteristic values: After assessment: ; .

[0078] Calculate the consistency score: .because State mask (Note: Because turning over in patients with bone metastases easily induces mechanical breakthrough pain, the system assigns high weight to electromyographic mutations and allows them to proceed without interference, reflecting the role of prior medical history in preventing missed diagnoses.) If the patient has no bone metastases ( If so, then the score is... State mask The cleaned vector will be set to zero, thus successfully intercepting the overturning interference.

[0079] Breakthrough pain onset and feature extraction (S203-S402) during the monitoring period Within a time window, the patient experiences a sudden onset of acute pain. The preprocessing module calculates: Since all three features exceeded the threshold, and the indicator variables were all 1, the consistency score was [not specified]. Output mask .

[0080] The calibration module calculates the relative change: .

[0081] Calculate the current estimated pain index: The system's initial assessment score is 6.5.

[0082] Pain Model Adaptive Calibration (Patient Presses Pain Pump) (S403-S404) At this point, the patient, experiencing unbearable pain, presses the self-controlled analgesia (PCA) pump. The system receives the event command and maps it to a truth label. The system performs backpropagation of the loss function (assuming a learning rate). Regularization coefficient (ignoring minor regularization terms): gradient .

[0083] Update weights: .

[0084] Logical significance: Because the system previously underestimated the patient's pain (6.5 < 8), the algorithm automatically generated positive compensation, increasing the weight parameters of various physiological characteristics. In subsequent monitoring, the model will better reflect the patient's sensitive physiological responses.

[0085] Trend Analysis and Linked Early Warning (S502-S604): The trend analysis module extracts continuous pain indices within a retrospective sliding time window (set to 5 minutes) and obtains trend factors through least squares fitting. (This indicates that the pain is rapidly intensifying.) Set a forward prediction span. Calculate the predicted pain index: .

[0086] Since 12.5 exceeds the scale's upper limit, the system executes a truncation logic to prevent errors, ultimately outputting... The early warning module compares the threshold ( Since 10 ≥ 7, the indicator variable for the warning level is... System verification alarm confidence level (Greater than the threshold of 0.6), and the time difference since the last administration (e.g., 45 minutes) is greater than the set safety lockout time. Minutes), lock flag released ( Finally, the system generates... The drive signal automatically controls the analgesic pump to administer additional medication.

[0087] Experimental verification and effect comparison: To further verify the clinical effectiveness of the multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system provided by this invention, the applicant conducted a 6-month clinical comparative verification experiment at the oncology department of a tertiary hospital. The experiment included 120 hospitalized patients with a history of cancer pain breakthrough pain, who were randomly divided into a control group (60 patients, using traditional single-modal vital sign threshold alarm and routine on-demand medication) and an experimental group (60 patients, using the multimodal adaptive system described in the embodiments of this invention).

[0088] The experimental results are as follows: Improved accuracy of early warnings and ability to prevent false alarms: Because the control group lacked multimodal consistency comparison and state mask isolation logic, its false alarm rate was as high as 28.4% due to the influence of patient turning over, coughing, and environmental fever. The experimental group of this invention, by executing the multimodal consistency logic screening matrix in step S3, successfully eliminated 91% of non-pain interference data, and the overall false alarm rate significantly decreased to 5.2% (specific comparison results are shown below). Figure 5 (as shown in subgraph (a)); at the same time, combined with Figure 5 As can be seen from the ROC curve of the breakthrough pain detection performance in subplot (b), this invention significantly reduces the false positive rate while maintaining a high detection sensitivity of 96.8% for breakthrough pain (corresponding to an AUC of 0.92 in the experimental group, which is better than 0.74 in the control group).

[0089] Forecast lead time assessment: This invention achieves prospective prediction by extracting the trend factor of the time-series change rate through the trend analysis module in step S5. Experimental data statistics show that the experimental group's system issued a high-level warning (level 2) on average 8.4 minutes earlier than the time when the patient's subjective pain reached unbearable levels and requested manual medication. Figure 3 For example, subgraphs (a) and (b) fully capture the multimodal physiological signal rise and abrupt change process during the prodromal phase of breakthrough pain, while subgraph (c) visually demonstrates how the trend analysis module deduces the prospective prediction trend (shaded area) based on the current estimated index (step line), and triggers the severe pain threshold (dot mark) before the actual breakthrough pain peak (asterisk mark) arrives. This predictive lead time provides sufficient time window for the onset of intravenous analgesics, effectively avoiding the severe pain experienced by patients during the peak of breakthrough pain.

[0090] Personalized convergence analysis of pain assessment models: Error analysis was performed by extracting background log data from the experimental group. During the initial monitoring phase (0-12 hours) after the system was first connected to the patient, due to the use of population-based general weights, the mean absolute error (MAE) between the system's output estimated pain index and the patient's subjective NRS score was 1.75 points (e.g., ...). Figure 4 (As shown in the initial zero-point position); with the intervention of the dynamic adaptive backcorrection algorithm in step S4, after approximately 3-5 truth label anchoring iterations, the system completes personalized learning. For example... Figure 4 As shown in the MAE convergence curve, during the 24-48 hour monitoring period (corresponding to the later iterations in the figure), the pain assessment MAE converged and stabilized at around 0.68 points, fully demonstrating the extremely high reliability of the loss function and gradient descent update logic in clinical individualized calibration.

[0091] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multimodal, non-invasive cancer pain breakthrough pain monitoring and early warning system, characterized in that, This includes wearable devices, mobile application modules, cloud engines, and healthcare workstations; The wearable terminal collects the patient's pulse wave, skin conductance, and surface electromyography signals and sends them to the mobile application module; The mobile application module is used to receive pain score data and transmit the received physiological signals and the pain score data to the cloud engine. The cloud engine includes: The preprocessing module is used to perform time alignment and slicing of the physiological signals, calculate variability features, and construct multimodal physiological feature vectors. The discrimination module extracts prior medical history weight factors based on medical record data, performs consistency comparison on the variability features to generate a state mask, uses the state mask to clean up interference data, and outputs an effective multimodal physiological feature vector. The calibration module is used to calculate the relative change of the effective multimodal physiological feature vector relative to the individual physiological baseline vector, construct a loss function using the pain score data as the ground truth label, back-calibrate the weight parameters, and output the estimated pain index. The trend analysis module is used to calculate the trend factor of the estimated pain index sequence and extrapolate the predicted pain index. The early warning module is used to compare the predicted pain index with the set early warning threshold to generate an early warning command, which is then pushed to the medical workstation.

2. The multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system according to claim 1, characterized in that, The preprocessing module constructs multimodal physiological feature vectors, including: Extract the time interval sequence of adjacent peak points of the pulse wave signal, calculate the average value, and generate pulse wave feature values; The skin conductance response component of the skin conductance signal is separated, and the peak amplitude of the pulse is extracted to generate skin conductance characteristic values; Power frequency interference of surface electromyography signals is filtered out and the root mean square amplitude value is calculated to generate electromyography characteristic values; The pulse wave feature value, skin conductance feature value, and electromyography feature value dimensions are combined to construct the multimodal physiological feature vector.

3. The multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system according to claim 2, characterized in that, The discrimination module generates a state mask, including: Configure numerical prior medical history weighting factors based on the disease label categories contained in the patient's medical record data; Each feature value is compared and mapped to a dynamically determined physiological fluctuation threshold as an activation state indicator variable; The consistency score is obtained by combining the activation status indicator variables and the prior medical history weighting factor. The consistency score is compared with the mask determination threshold, and the state mask is output. The state mask is multiplied by the multimodal physiological feature vector to output the effective multimodal physiological feature vector.

4. The multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system according to claim 1, characterized in that, The calibration module reverse-calibrates the weight parameters, including: Extract the continuous effective multimodal physiological feature vectors under pain-free resting state, calculate the average value to generate the personal physiological baseline vector; The relative change of the feature is obtained by calculating the difference between the effective multimodal physiological feature vector and the individual physiological baseline vector in the current time window. The estimated pain index is obtained by calculating the relative change of the features and the weight parameters corresponding to each feature; The loss function is constructed using the mean squared error between the pain score data and the estimated pain index, as well as the regularization term for the weight parameters; The gradient of the loss function is calculated using the gradient descent algorithm, and the weight parameters are updated.

5. A multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system according to claim 4, characterized in that, The calibration module is also configured with zero-order hold logic: When the state mask indicates that the current data is interference data, the effective multimodal physiological feature vector is set to zero, the calibration module skips the calculation of the relative change of the features and the loss function within the current time window, and the output of the estimated pain index directly uses the value of the previous effective time window.

6. The multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system according to claim 1, characterized in that, The trend analysis module calculates and predicts the pain index, including: A linear fitting algorithm is used to fit the estimated pain index sequence within the sliding time window, and the slope of the fitted line is extracted as the trend factor. When the trend factor is greater than or equal to zero, the predicted pain index is obtained by adding the product of the trend factor and the prospective prediction time span to the current estimated pain index. When the trend factor is less than zero, the predicted pain index is directly assigned the current estimated pain index.

7. The multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system according to claim 1, characterized in that, The early warning module generates early warning commands, including: Set moderate pain threshold and severe pain threshold, and map the predicted pain index as an early warning level indicator variable; Extract the predicted pain index from adjacent historical time windows, calculate the percentage of sampling points exceeding the moderate pain threshold, and generate an alarm confidence level. When the alarm confidence level is greater than or equal to the confidence threshold, the warning status is confirmed to be valid, and a regular attention instruction or an emergency intervention instruction is generated based on the warning level indicator variable.

8. A multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system according to claim 7, characterized in that, The early warning module is configured with security lock time verification logic based on the emergency intervention command: Get the time difference between the current system timestamp and the last recorded timestamp of the self-controlled analgesia pump administering additional medication; When the time difference is greater than or equal to the safety lockout time parameter, a drive signal is generated to control the additional dose of the self-controlled analgesia pump. When the time difference is less than the safety lock time parameter, the control of the self-controlled analgesia pump remains locked.

9. A multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system according to claim 1, characterized in that, The preprocessing module also extracts a globally unified timestamp based on the system clock information and uses an interpolation algorithm to align the asynchronously sampled pulse wave, skin conductance, and surface electromyography signals to the same time coordinate axis for slicing.

10. A multimodal non-invasive cancer pain breakthrough pain monitoring and early warning method, applied to a multimodal non-invasive cancer pain breakthrough pain monitoring and early warning system as described in any one of claims 1-9, characterized in that, Includes the following steps: Acquire patient pulse wave, skin conductance and surface electromyography signals, as well as pain score data; The physiological signals are time-aligned and sliced, variability features are calculated, and a multimodal physiological feature vector is constructed. Based on medical record data, prior medical history weight factors are extracted, consistency comparison is performed on the variability features to generate a state mask, the state mask is used to clean up interference data and output an effective multimodal physiological feature vector. Calculate the relative change of the effective multimodal physiological feature vector relative to the individual physiological baseline vector, construct a loss function using the pain score data as the ground truth label, calibrate the weight parameters in reverse, and output the estimated pain index. Calculate the trend factor of the estimated pain index sequence to extrapolate the predicted pain index; The predicted pain index is compared with the warning threshold to generate a warning command for information push or device linkage.

Images