Predictive reliability assessment method for wearable physiological parameter monitoring device, serial training method for confidence classification network in cloud server, and wearable physiological parameter monitoring device

Abstract

The application relates to the field of physiological signal intelligent processing of cooperation between a cloud server and a wearable physiological parameter monitoring device, and discloses a prediction reliability evaluation method, a cloud server, a monitoring device and a storage medium. The method extracts, by the cloud server, a target query vector from a cross attention sublayer of a prediction network decoder, calculates a discrete degree measure of an attention weight distribution of a calibration memory sequence, inputs the measure, a prediction value and a prior anchor value into an independent confidence classification network in a three-way combination, outputs a reliability level, and issues a control instruction to drive an output management control module of the monitoring device to perform a device-side action. The application helps to add a reliability indication signal under a single forward inference cost.

Description

Technical Field

[0001] This application relates to the field of intelligent physiological signal processing technology that integrates cloud servers and wearable physiological parameter monitoring devices. Background Technology

[0002] With the rapid development of wearable physiological parameter monitoring technology, wearable or portable monitoring devices such as wrist-worn non-invasive blood glucose estimation devices that use photoplethysmography (PPG) signals as input, continuous blood pressure estimation devices that use multi-channel pulse waves collected simultaneously from the fingertips and wrist as input, and finger clip-on blood oxygen saturation estimation devices that use red and infrared dual-wavelength PPG signals as input are being increasingly deployed in daily continuous physiological monitoring scenarios. In these application scenarios, the monitoring devices typically upload the collected physiological signals to a cloud server after preprocessing. An encoder-decoder type prediction network deployed in the cloud infers the physiological signals and returns the predicted values ​​of the target physiological parameters to the monitoring device. The predicted values ​​are then further processed and used by the monitoring device's display module, human-computer interaction module, data upload module, and linked external modules.

[0003] However, existing technologies have the following limitations in the aforementioned scenario: First, existing prediction systems that use physiological signals such as photoplethysmography (PPG) pulse waves as input typically output only a single scalar prediction value in each single-sample inference, without addressing a mechanism to output a reliability indicator signal for that inference. Downstream device modules can only process all prediction values ​​with the same level of confidence. Second, methods based on multiple random forward propagation require performing multiple complete forward inferences in each inference, placing a significant burden on the throughput and concurrent service capabilities of cloud servers. Methods based on multi-model ensembles require the parallel deployment of multiple independent model replicas, with storage and computing power costs increasing almost linearly with the number of models. Methods based on post-processing calibration are mainly geared towards classification tasks and struggle to reflect the real-time fluctuations of the network's internal state relative to the current sample in a single inference during regression prediction. Methods based on evidence-based deep learning require modification of the original prediction network's output head and training loss, resulting in low compatibility with already deployed regression prediction networks. Third, none of the above methods address how to ensure that the reliability indicator signal ultimately lands on the hardware output management actions of the monitoring device. Therefore, a new technical solution is urgently needed to address these issues. Summary of the Invention

[0004] The purpose of this application is to provide a predictive reliability assessment method for wearable physiological parameter monitoring devices, a serial training method for confidence classification networks in cloud servers, and a wearable physiological parameter monitoring device, so as to solve the problems mentioned in the background art.

[0005] This application discloses a predictive reliability assessment method for wearable physiological parameter monitoring devices, executed on a cloud server communicatively connected to the wearable physiological parameter monitoring device, comprising: Step 1: Acquire the time series or characteristic sequence of physiological signals collected by the physiological sensors of the monitoring device, preprocessed, and uploaded; Step 2: Input the physiological signal time series or its feature sequence together with the calibration memory data into the trained encoder-decoder type prediction network. The decoder of the prediction network performs cross-attention calculation on the calibration memory data and outputs the predicted value of the target physiological parameter. Step 3: Extract the attention weight distribution of the target query vector to the calibration memory sequence from at least one cross-attention sublayer of the decoder described in Step 2; Step 4: Based on the attention weight distribution obtained in Step 3, calculate a scalar metric that characterizes its degree of dispersion to obtain the degree of dispersion metric; Step 5: The dispersion measure obtained in Step 4, the predicted value of the target physiological parameter obtained in Step 2, and the prior anchor value are used as three features and input together into the confidence classification network that does not share parameters with the prediction network. Step Six: The confidence classification network outputs the reliability level corresponding to the predicted value of the target physiological parameter; Step 7: Based on the reliability level output in Step 6, the cloud server generates control commands and sends them to the monitoring device to control the output management control module of the monitoring device to perform at least one device-side action on the predicted value of the target physiological parameter, including output marking, display suppression, resampling prompt, recalibration prompt, or data transmission strategy adjustment.

[0006] In a preferred embodiment, the dispersion measure in step four is performed using at least one of the following calculation methods: Shannon's information entropy: ; Normalized entropy: or ; Impurity of the gin: ; Valid number of entries: ; Top-k attention quality: ; The KL divergence between the attention weight distribution and the uniform distribution: ; Coefficient of variation: ; in, The length of the calibration memory sequence, The number of valid memory entries that were not masked in the calibration memory sequence. For the attention weight distribution corresponding to the first The weight value of each memory location, This is a numerical stability constant used to avoid zero values ​​in logarithmic operations. Represents the natural logarithm function. It is a preset positive integer. This indicates the top ranking in the attention weight distribution. The set of position indices of the maximum weight value and These are the standard deviation function and the mean function of the attention weight distribution, respectively.

[0007] In a preferred embodiment, step five, which involves inputting the discreteness measure obtained in step four, the predicted value of the target physiological parameter obtained in step two, and the prior anchor value as three features into the confidence classification network, includes: after normalizing the dimensions of the prior anchor value, the predicted value of the target physiological parameter, and the discreteness measure, concatenating them into a three-dimensional feature vector. The three-dimensional feature vector is projected through two fully connected layers and connected by a nonlinear activation function to map it into a high-dimensional embedding vector. The high-dimensional embedding vector is added to the learnable token embedding to initialize the confidence classification query vector of the confidence classification network decoder. The prior anchor value is obtained in one of the following ways: the reference value obtained from the most recent offline physical calibration, the sliding weighted average of the predicted values ​​at the most recent preset number of times, the low-frequency physiological trend estimate extracted by the low-frequency physiological trend estimation module, the weighted average of the aforementioned preset number of reference values ​​with the highest similarity to the current input in the calibration memory sequence, or the baseline estimate output by an independent baseline estimation network.

[0008] In a preferred embodiment, in step seven, when the reliability level is in the low confidence category, the device-side action controlled by the generated control command to control the monitoring device to perform includes at least one of the following: reducing the data upload priority flag; Report the abnormal status indicator to the cloud server; Switch to the backup prediction model within the cloud server; The predicted values ​​of the target physiological parameters belonging to the low-confidence category are restricted from entering the downstream closed-loop control module of the monitoring device.

[0009] In a preferred embodiment, in step two, each record in the calibration memory data consists of the following four features: a historical physiological signal waveform segment, the corresponding reference value, the similarity between the current target physiological signal and the historical physiological signal in the low-dimensional feature space, and the relative acquisition time interval between the acquisition time of the record and the current inference time.

[0010] This application also discloses a serial training method for confidence classification networks in cloud servers, including: Phase 1: Using training samples containing time series of physiological signals and target reference values, train the encoder-decoder type prediction network through regression loss until convergence, and then freeze the network parameters of the prediction network. Second stage: With the network parameters frozen and not participating in gradient updates, the training samples are input into the prediction network to perform a single forward inference, obtain fixed target physiological parameter prediction values, and extract the attention weight distribution from at least one cross-attention sublayer of the prediction network decoder to calculate the degree of dispersion measure. Based on the absolute or relative error between the predicted value of the fixed target physiological parameter and the target reference value, and at least two preset error thresholds... The numerical comparison relationship is used to generate stable discrete classification labels for the training samples offline across training rounds; The fixed target physiological parameter prediction value, the dispersion measure and the prior anchor value are jointly input into the confidence classification network as three features. The parameters of the confidence classification network are updated with weighted cross-entropy loss, wherein the discrete classification labels of the low-confidence categories are assigned higher class weights than those of the high-confidence categories.

[0011] In a preferred embodiment, the weighted cross-entropy loss takes the form of: in, For a small batch of samples, The size of the mini-batch sample set. The number of categories for reliability levels. For the first The discrete classification labels of each sample For the confidence classification network, the first The output of the nth sample Class probability value, For the first Class weights This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. The function is the natural logarithm; the category weights satisfy... This allows the low-confidence category to receive the highest category weight.

[0012] This application also discloses a wearable physiological parameter monitoring device, which is connected to a cloud server, including: a physiological signal sensor configured to collect physiological signals; The signal preprocessing circuit is configured to preprocess the physiological signal and convert it into a physiological signal time series or its characteristic sequence. The communication module is configured to send the physiological signal time series or its characteristic sequence to the cloud server and receive control instructions issued by the cloud server, wherein the control instructions are generated by the cloud server after executing the evaluation method as described in any one of claims 1 to 5 based on the received data; The output management control module is configured to perform at least one device-side action on the predicted value of the target physiological parameter, including output marking, display suppression, resampling prompt, recalibration prompt, or data transmission strategy adjustment, based on the received control command.

[0013] In a preferred embodiment, the physiological signal sensor is a photoplethysmography (PPG) sensor, and the physiological signal time series is a PPG signal time series; the predicted value of the target physiological parameter is one of the following: a non-invasive blood glucose estimate, a continuous blood pressure estimate based on pulse wave conduction time or multi-channel pulse wave, or a blood oxygen saturation estimate based on dual-wavelength PPG.

[0014] The technical solutions provided by the embodiments of this application, under the edge-cloud collaborative architecture, through the synergistic cooperation of the aforementioned technical features, help to obtain corresponding technical effects in multiple dimensions simultaneously, which will be described in detail below.

[0015] Since the attention weight distribution is used as a reused object of the internal intermediate state from a single forward inference, the calculation of the discreteness measure only involves the summation and logarithm of the already calculated attention weight matrix, without introducing additional complete forward inference. Therefore, it helps the cloud server to synchronously produce the target physiological parameter prediction value and the corresponding reliability level at the cost of a single forward inference in each sampling period. Compared with the solution that requires repeated forward inference or parallel integration of multiple models, it helps to reduce the overall computing power and storage overhead of the server in the high-concurrency request scenario in the cloud. At the same time, since the discreteness measure is anchored in the physical meaning of "the degree of concentration of similar reference samples that can be located in the historical calibration memory of the current physiological state", the generated reliability indication signal has interpretable meaning at the engineering level.

[0016] Furthermore, since the proposed scheme initializes the confidence classification query vector by normalizing and jointly embedding the three features—the discreteness measure, the predicted value of the target physiological parameter, and the prior anchor value—and the independently trained confidence classification network learns the nonlinear mapping from the three features to the discrete reliability level, when a non-monotonic or nonlinear correspondence occurs between the discreteness measure and the actual prediction error, the scheme helps to mitigate the instability problem caused by direct thresholding of a single measure by introducing two auxiliary information sources: the predicted value and the prior anchor value. When the discreteness measure is further adopted in various optional forms (including but not limited to Shannon information entropy, normalized entropy, Gini impurity, effective memory count, Top-k attention quality, and uniformly distributed KL),... When considering divergence and coefficient of variation, the proposed scheme can be flexibly adapted to the instruction set support and numerical computation characteristics of the cloud-based inference nodes. For example, Gini impurity is preferred on processors that do not support efficient logarithmic operations, and normalized entropy is preferred in scenarios with high requirements for cross-sample comparability. When a normalized exponential function with a mask applied to the filling position is used, it helps to maintain the cross-sample comparability of the metric under different effective memory counts. When a multi-layer weighted fusion or multi-head aggregation method is used, it helps the cross-attention sublayer closer to the decoder output to obtain higher feature weights because the information it aggregates is closer to the final prediction, thereby further improving the stability of the reliability level determination. When the prior anchor value is further flexibly selected from multiple sources such as offline physical calibration value, sliding weighted average, low-frequency physiological trend estimate, weighted average of nearest neighbor reference value and independent baseline estimation network output value, the scheme can still operate stably under the difference in the availability of prior information in different application scenarios, thus expanding the application and adaptation range of the scheme.

[0017] Furthermore, since the confidence classification network in the proposed scheme is structurally independent of the prediction network and their parameters are not shared, and since the proposed scheme adopts a two-stage serial training framework—in which the second stage trains the confidence classification network with weighted cross-entropy loss while the prediction network parameters are frozen and do not participate in any gradient updates—it helps to structurally isolate the gradient interference between the regression prediction target and the reliability classification target, avoiding the prediction network from distorting the prediction distribution or fabricating high-confidence attention signals to reduce classification loss. This helps to maintain the regression accuracy of the original prediction network while introducing reliability assessment capabilities. At the same time, this scheme can be integrated into the existing cloud inference system in an incremental upgrade manner without modifying the structure and parameters of the deployed prediction network. Furthermore, since the offline-generated discrete classification labels are generated solely based on the numerical comparison between fixed predicted values ​​and target reference values, they remain stable across training rounds without drifting, contributing to the optimization and stability of the training process. When directional overwriting rules and directional correctness exemption rules are added, the training objective and the final output management objective on the device side correspond in an engineering sense, which helps improve the engineering consistency between reliability level determination and device-side actions in application scenarios sensitive to directional bias. When the first stage further employs asymmetric Huber regression loss and imposes heavier penalties on samples with low reference values ​​that are directionally biased and high reference values ​​that are directionally biased, it helps the trained prediction network to have a relatively small prediction bias in the stated direction. The weighted cross-entropy loss, which assigns the highest class weight to the low-confidence category, helps improve the recall level of the confidence classification network for the low-confidence category and reduces the probability of low-reliability predictions being misclassified as high-reliability predictions.

[0018] Regarding the calibration memory data, this scheme clearly states that each record consists of four features: a historical physiological signal waveform segment, the corresponding reference value, the similarity between the current target physiological signal and the historical physiological signal in the low-dimensional feature space, and the relative acquisition time interval between the acquisition time of the record and the current inference time. The introduction of the similarity allows the prediction network to obtain a preliminary matching indication between each historical record and the current sample before performing cross-attention, which helps to improve the efficiency of memory retrieval and the targeting of cross-attention calculation.

[0019] Furthermore, since the proposed solution sends the reliability level to the monitoring device in the form of control commands and directly drives its output management control module to perform hardware actions such as output marking, display suppression, resampling prompts, recalibration prompts, or data transmission strategy adjustments, the output of the reliability assessment ultimately affects the device hardware level. This helps reduce the probability that low-reliability prediction values ​​are processed as high-confidence prediction values ​​by downstream device modules, thereby improving the overall output reliability management capability of the monitoring device. When the reliability level is a low-confidence category, the cloud server can also perform actions such as reducing data upload priority, reporting abnormal status indicators, and switching within the cloud server. Actions such as switching to a backup prediction model or restricting low-confidence prediction values ​​from entering the downstream closed-loop control module make the differentiated execution of equipment-side actions more targeted. When the reliability level jumps near the category critical threshold, a hysteresis threshold mechanism, a time window majority voting smoothing mechanism, or a first-order hysteresis filter smoothing process is further applied, it helps to avoid frequent switching of equipment-side actions. When the input of the confidence classification network is additionally combined with signal quality indicators, the number of calibration samples, the acquisition time interval, the memory library coverage, or historical prediction fluctuation indicators, it helps to provide a more robust reliability judgment in the case of signal quality fluctuations or insufficient calibration library coverage.

[0020] Furthermore, since the overall inference calculation of the proposed scheme is centrally deployed on a cloud server, the monitoring device only undertakes the functions of physiological signal acquisition, preprocessing, data uploading, and driving hardware according to control commands. This helps to keep the monitoring device within a range suitable for continuous wear in terms of power consumption, size, and cost. Since the attention weight distribution extraction process, the three-way joint embedding process, and the dual-network serial training process do not impose strict restrictions on specific physiological parameter objects, the proposed scheme can be applied to various scenarios such as non-invasive blood glucose estimation with PPG signal as input, continuous blood pressure estimation with multi-channel pulse wave as input, and blood oxygen saturation estimation with dual-wavelength PPG as input. Only the sensor type, sampling parameters, prior anchor value definition, and error threshold value need to be adjusted accordingly to reuse the same technical framework, which has good versatility and scalability.

[0021] It should be noted that the achievement of the technical effect is based on the premise that the various technical features of this solution are interconnected and cooperate in accordance with the aforementioned internal context. In extreme cases such as when the physiological sensor fails completely or the distribution of the input physiological signal drifts significantly relative to the training set, the reliability level output by this solution may also be inaccurate. In such boundary cases, the independent sensor self-test module or the external detection module on the monitoring device side can be used as a fallback, and the reliability level output by the solution can be logically combined with the reliability level output by the solution in the output management and control module to act on the downstream device.

[0022] The specification of this application contains numerous technical features distributed across various technical solutions. Listing all possible combinations of these technical features (i.e., technical solutions) would make the specification excessively lengthy. To avoid this problem, the various technical features disclosed in the above-described invention, the various technical features disclosed in the following embodiments and examples, and the various technical features disclosed in the accompanying drawings can be freely combined to form various new technical solutions (all of which are considered to have been described in this specification), unless such a combination of technical features is technically infeasible. For example, one example discloses feature A+B+C, and another example discloses feature A+B+D+E. Features C and D are equivalent technical means that serve the same function, and technically only one needs to be used; they cannot be used simultaneously. Feature E can technically be combined with feature C. Therefore, the solution A+B+C+D should not be considered as described because it is technically infeasible, while the solution A+B+C+E should be considered as described. Attached Figure Description

[0023] Figure 1 is a schematic diagram of the overall architecture of the dual-network cascade according to an embodiment of the present application, showing the structural relationship of the prediction network and the confidence classification network being executed serially on the cloud server side, and the flow of the dispersion measure of the attention weight distribution between the two.

[0024] Figure 2 is a schematic diagram of the calculation process for the discreteness measure of the attention weight distribution according to an embodiment of this application, and compares the typical forms of the attention weight distribution under the low uncertainty case of highly concentrated attention and the high uncertainty case of approximately uniform attention distribution.

[0025] Figure 3 is a schematic diagram of the input feature fusion of the confidence classification network according to an embodiment of this application. It shows the process by which the prior anchor value, the predicted value of the prediction network and the attention entropy are normalized in terms of dimensions, concatenated and linearly projected to a high-dimensional embedding and initialized into a confidence classification query vector. Then, the confidence classification network decoder and the three-class output head are mapped to the probability distribution of each reliability level category.

[0026] Figure 4 is a schematic diagram of a two-stage serial training process according to an embodiment of this application, showing the training process of the first stage training the prediction network with asymmetric Huber regression loss and the second stage training the confidence classification network with weighted cross-entropy loss based on offline generated stable discrete classification labels when the parameters of the prediction network are frozen and do not participate in gradient updates.

[0027] Figure 5 This is a flowchart illustrating a predictive reliability assessment method for wearable physiological parameter monitoring devices according to an embodiment of this application. Detailed Implementation

[0028] In the following description, numerous technical details are presented to enable those skilled in the art to better understand this application. However, those skilled in the art will understand that the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0029] Explanation of some concepts: An encoder-decoder predictive network refers to a neural network structure in which an encoder encodes the input physiological signal time series and calibration memory data, a decoder uses the target query vector as the initial query and performs cross-attention calculation on the calibration memory sequence, and finally outputs the predicted value of the target physiological parameter through a regression mapping head.

[0030] Calibration memory data refers to a set of data consisting of several historical calibration records, which are used as key-value sequences for the prediction network to query when performing cross-attention calculation. Each record consists of four features: a historical physiological signal waveform segment, the corresponding reference value, the similarity between the current target physiological signal and the historical physiological signal in the low-dimensional feature space, and the relative acquisition time interval between the acquisition time of the record and the current inference time.

[0031] The target query token refers to the initial query used by the predictive network decoder during attention computation, initialized by adding the learnable parameters and prior anchor values. The vector.

[0032] Attention weight distribution refers to the probability vector obtained by linearly projecting the target query vector in the decoder's cross-attention sublayer, scaling the dot product with the key matrix of the calibration memory sequence, and then processing it with a normalized exponential function. Each component has a non-negative value, and the sum of all components is one. .

[0033] Discreteness measures refer to scalar indicators that reflect the degree to which the attention weights are concentrated or dispersed on the calibration memory sequence, including but not limited to Shannon information entropy, normalized entropy, Gini impurity, number of effective memories, Top-k attention quality, KL divergence from a uniform distribution, and coefficient of variation.

[0034] Shannon entropy refers to one implementation of the measure of discreteness, whereby the attention weight distribution is... The calculated scalar value, where It is a numerical stability constant used to avoid zero values ​​in logarithmic operations.

[0035] The prior anchor value refers to the scalar value used as the third input feature of the confidence classification network to characterize the current baseline level of physiological parameters. It can be obtained by any of the following methods: the reference value of the most recent offline physical calibration, the sliding weighted average of the predicted values ​​of the most recent preset number of time moments, the low-frequency physiological trend estimate, the weighted average of several reference values ​​with the highest similarity in the calibration memory sequence, or the output of an independent baseline estimation network.

[0036] A confidence classification network is an independent neural network that does not share parameters with the prediction network. It initializes the confidence classification query vector by jointly embedding three features: the discreteness measure, the predicted value of the target physiological parameter, and the prior anchor value. The vector is then mapped to the probability distribution of each reliability level category by its decoder.

[0037] The confidence classification query token is the query vector used as the initial query during attention calculation by the confidence classification network decoder. It is initialized by adding the high-dimensional embedding vector obtained by normalizing, concatenating, and projecting the three features with a fully connected layer, and the learnable token embedding. It is an independent vector that acts on different networks, just like the target query vector.

[0038] The reliability level refers to the discrete category identifier, characterized by the credibility of the predicted value of the target physiological parameter, output by the confidence classification network. Its output granularity can be binary, triadic, or quinary, etc.; in a triadic implementation, it is exemplarily divided into a high-confidence category. , medium confidence category With low confidence category .

[0039] The output management and control module refers to a functional module located on one side of the wearable physiological parameter monitoring device and configured to perform at least one device-side hardware action on the predicted value of the target physiological parameter according to control instructions from the cloud server, such as output marking, display suppression, resampling prompt, recalibration prompt, or data transmission strategy adjustment.

[0040] Serial training refers to a two-stage training process executed sequentially: the first stage trains the prediction network with regression loss until convergence and then freezes its parameters; the second stage trains the confidence classification network with weighted cross-entropy loss while the prediction network parameters are frozen and do not participate in gradient updates.

[0041] Directional overwrite rule refers to the rule that, when generating training labels for the confidence classification network, if the reference value is in a preset low or high reference value range and the predicted value deviates from the preset threshold in the corresponding direction, the training label is forcibly overwritten as a low confidence category.

[0042] The orientation-correct exemption rule refers to the rule that, when generating the training label, if the reference value is in a preset low or high reference value range and the predicted orientation is correct, the training label is kept as a high-confidence category.

[0043] The following is a brief summary of some of the innovative aspects of this application: In summary, the technical scenario addressed in this application is an end-to-cloud collaborative physiological monitoring scenario where wearable physiological parameter monitoring devices (including but not limited to non-invasive blood glucose estimation devices based on PPG, continuous blood pressure estimation devices based on multi-channel pulse waves, and blood oxygen saturation estimation devices based on dual-wavelength PPG) serve as the front end, and a cloud server connected to them serves as the inference entity. The specific technical challenge this application seeks to overcome lies in how to simultaneously attach a reliability indication signal, strictly corresponding to the inference, to each single-sample forward inference without introducing multiple complete forward inferences, without modifying the structure and parameters of the deployed prediction network, and without causing the reliability assessment target to back-interfere with the regression prediction accuracy, and how to ensure that the indication signal ultimately lands on the hardware output management action of the monitoring device.

[0044] To resolve the aforementioned contradictions, the technical solution of this application reconstructs the overall causal chain of "single inference—internal intermediate state—reliability measurement—device-side action" through the inherent relationship of mutual nesting and constraint among several technical features: First, attention weight distribution It is not taken from any arbitrary location in the network, but is specifically selected from the prediction network. Figure 1 In the final cross-attention sublayer of the decoder in network A), the target query vector... The weight distribution of the calibration memory sequence (see Figure 2) is physically anchored to "the degree to which the current physiological state can be located in the historical calibration memory as a cluster of similar reference samples," and the resulting measure of dispersion is... (For example, Shannon information entropy) fully reuses intermediate states already existing in a single forward inference without repeated inference; secondly, given the inventor's discovery There is a non-monotonic and non-linear correspondence between the actual prediction error and the actual prediction error. This application deliberately avoids [the following]: Instead of the simplified path of directly mapping thresholds to reliability levels, we will instead... Predicted values Compared with prior anchor value The three features, after being normalized to their dimensions, are jointly embedded and used to initialize the confidence classification query vector. (See Figure 3) The confidence classification network (network B in Figure 1), which is independent of network A and does not share parameters, learns three-way features to discrete levels. The nonlinear mapping; third, a two-stage serial training framework is adopted (see Figure 4), in the second stage, the network A parameters The network B parameters are frozen and updated only with weighted cross-entropy. This structurally isolates the gradient interference between the two targets. It is precisely the aforementioned nested feature associations and synergistic cooperation that enable the proposed solution to simultaneously produce predicted values ​​and reliability levels without increasing the computational budget of the complete forward inference. The reliability level is then transmitted to the monitoring equipment via the cloud in the form of control commands, driving its output management and control module to perform hardware actions such as display suppression, resampling prompts, or data transmission strategy adjustments. Thus, the overall concept is grounded in the non-obvious composite technical path of "internal intermediate state - cross-network migration - device hardware control".

[0045] Furthermore, the inventors of this application, through long-term and in-depth research, have discovered that the aforementioned technical problems are not isolated from each other, but rather intertwined and mutually constrained under the end-cloud collaborative deployment architecture where the cloud server undertakes prediction and inference and the wearable device only undertakes signal acquisition and hardware output management, forming a set of complex technical contradictions that need to be solved simultaneously.

[0046] The inventors first systematically analyzed the deployment costs of existing uncertainty quantification methods in high-concurrency cloud scenarios and found that, whether based on multiple random forward propagation schemes or multi-model integration schemes, they essentially require cloud servers to perform computational overhead far exceeding that of a single forward inference by the original prediction network in each sampling period. This overhead is significantly amplified when facing concurrent inference requests from large-scale wearable terminals, thereby affecting the overall service throughput and end-to-end response latency. The fundamental reason is that such schemes externalize the source of uncertainty signals to a single forward inference, which inevitably introduces additional complete forward inference times or additional model copies. Secondly, after meticulously observing the internal intermediate states of the encoder-decoder type prediction network during the inference process, the inventors realized that the distribution of attention weights of the target query vector to the calibration memory sequence in the cross-attention sublayer of the network decoder itself reflects the key information of "whether the current physiological state can locate a sufficiently similar reference sample in the historical calibration memory," thus containing reusable uncertainty signals from within a single forward inference. However, through repeated experiments and theoretical derivation, the inventors further realized that there is a non-monotonic and non-linear correspondence between the discreteness measure of the attention weight distribution and the actual prediction error—the actual error at the same discreteness level may differ significantly due to differences in input distribution and prior bias. This makes the simplified approach of directly mapping the measure to a reliability level using threshold comparison unstable in engineering. More importantly, after conducting a mechanistic analysis of the scheme of merging the uncertainty quantification target and the regression prediction target into a single multi-task loss for joint training, the inventors found that the joint training can cause the prediction network to distort its prediction distribution or fabricate high-confidence signals to reduce the uncertainty quantification loss due to gradient coupling between targets, thereby impairing the accuracy of the original regression prediction. Based on the above in-depth research, the inventors of this application creatively propose a technical solution with the following core concept: The dispersion measure of the attention weight distribution is used as an endogenous uncertainty signal from a single forward inference, and this signal is jointly input into an independently trained confidence classification network along with the predicted value and the prior anchor value. A two-stage serial training method is employed to isolate the mutual interference between the regression and classification objectives. The implementation process of this application is described in detail below through specific embodiments.

[0047] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. It should be noted that the numerical values, network scale, thresholds, and hardware parameters listed below are merely exemplary, and those skilled in the art can make adaptive adjustments without departing from the concept of this application.

[0048] Example 1: Predictive Reliability Assessment Method Based on Cloud Server This example provides a predictive reliability assessment method for wearable physiological parameter monitoring devices, the overall architecture of which is as follows: Figure 1 As shown, the entire method is executed on a cloud server side that maintains a communication connection with the wearable physiological parameter monitoring device via a wireless communication network. It is particularly important to note that in this embodiment, the forward inference calculations of both the prediction network and the confidence classification network are completed on the cloud server side. The wearable physiological parameter monitoring device side only undertakes the functions of physiological signal acquisition, preliminary preprocessing, data uploading, and receiving control commands to drive its output management and control module to perform corresponding actions. This end-to-cloud collaborative deployment method helps to concentrate computing power requirements on the cloud, keeping the wearable device's hardware resources, power consumption, and size within a range suitable for continuous wear.

[0049] Specifically, the method described in this embodiment includes the following steps.

[0050] Step 100: The cloud server acquires the time series or characteristic sequence of physiological signals collected by the physiological sensors of the wearable physiological parameter monitoring device, preprocessed on the device side, and uploaded through the communication network.

[0051] Furthermore, the physiological sensor can be, exemplarily, at least one of a photoplethysmogram (PPG) sensor, a continuous glucose monitoring current-type sensor, or a multi-wavelength optical sensor; the device-side preprocessing can, exemplarily, include, but is not limited to, bandpass filtering, baseline drift removal, signal windowing, and zero-mean unit variance normalization. Optionally, the wearable physiological parameter monitoring device can also perform lightweight feature extraction (e.g., window-by-window statistical features or time-frequency features) on the device side before uploading to reduce wireless transmission bandwidth usage. In this case, the cloud server receives the feature sequence of the physiological signal time series. Optionally, the cloud server performs integrity verification and necessary secondary preprocessing on the received data before proceeding to subsequent steps.

[0052] Step 200: The cloud server inputs the physiological signal time series or its feature sequence along with the calibration memory data into a trained and deployed encoder-decoder type prediction network (hereinafter referred to as Network A). The decoder of the prediction network performs cross-attention calculation on the calibration memory data and outputs the predicted value of the target physiological parameter. .

[0053] Specifically, network A can employ an encoder-decoder Transformer architecture. Its encoder encodes the calibration memory data to form a key matrix of the calibration memory sequence. AND-value matrix The decoder uses the target query token (denoted as...) As the initial query, self-attention computation is performed sequentially on the target physiological signal token sequence, followed by cross-attention computation on the calibration memory sequence. Optionally, the target query vector is initialized by adding the embeddings of learnable parameters and prior anchor values.

[0054] More specifically, each record in the calibration memory data consists of four features: a historical physiological signal waveform segment, a corresponding reference value, the similarity between the current target physiological signal and the historical physiological signal in the low-dimensional feature space, and the relative acquisition time interval between the acquisition time of the historical record and the current inference time. The similarity can be characterized by cosine similarity, normalized Euclidean distance, or embedding distance output by an independent lightweight encoder. Its function is to enable network A to obtain a preliminary matching indication between each historical record and the current sample before performing cross-attention, thereby improving memory retrieval efficiency. The four features, after their respective projections or encodings, are concatenated into a unified-dimensional memory token, forming the key matrix. AND-value matrix .

[0055] The network A ultimately maps the output vector of the last layer target query vector position of the decoder to a scalar predicted value via a regression mapping head located at the decoder output. That is, the predicted value of the target physiological parameter.

[0056] Step 300: The cloud server extracts the attention weight distribution of the target query vector on the calibration memory sequence from at least one cross-attention sublayer of the decoder described in step 200 (for comparison of its calculation process and typical case, see [link to relevant documentation]). Figure 2 ).

[0057] For example, in this embodiment, the cross-attention sublayer of the last layer of the decoder is selected as the extraction location. The engineering consideration for selecting the last layer is that the target query vector of the last layer of the decoder has been iteratively refined through layer-by-layer self-attention and cross-attention, which aggregates comprehensive information from the target physiological signal and historical calibration memory. The engineering consideration for selecting the cross-attention sublayer instead of the self-attention sublayer is that the source of the uncertainty mainly lies in whether the current physiological state can locate a sufficiently similar reference sample in the historical calibration memory bank, and the physical meaning of the cross-attention sublayer corresponds precisely to this process.

[0058] Furthermore, the attention weight distribution is obtained through the following sub-steps: Step 310: Obtain the query vector by linear projection of the target query vector. The key matrix is ​​obtained by linearly projecting the calibration memory sequence. Step 320: Take the dot product of the query vector and the transpose of the key matrix, and divide by the square root of each head dimension as a scaling factor to obtain the attention score vector: in, The query vector is the result of linear projection of the target query vector. The key matrix is ​​the result of linear projection of the calibration memory sequence. This indicates that the key matrix is ​​transposed. For each dimension of attention head, The scaling factor is used to avoid excessively large dot product values ​​that could saturate the subsequent normalization exponential function; Step 330: When there are padding positions in the calibration memory sequence, before executing the normalization exponential function, a very small negative number (essentially negative infinity) is applied to the attention score of the padding position as a mask, so that the weight corresponding to the padding position is approximately zero after processing by the normalization exponential function, thereby avoiding the padding position from interfering with the substantive content of the attention weight distribution; Step 340: The attention score vector is processed by the normalization exponential function to obtain the attention weight distribution: in, For normalized exponential functions, , and , The length of the calibration memory sequence.

[0059] Optionally, as various implementation methods: when the cross-attention sublayer includes a multi-head attention mechanism, the single-head attention weight distribution is extracted for each attention head according to steps 310 to 340 above, and the subsequent dispersion measure is calculated for each head. Then, the weights are aggregated using methods such as averaging, weighted summation, taking the maximum value, or concatenation to form the final feature for subsequent steps. When a multi-layer extraction method is used, the attention weight distribution is extracted from each cross-attention sublayer of the decoder, and the total dispersion measure is obtained using methods such as cross-layer averaging, cross-layer weighted summation, taking the maximum value across layers, or cross-layer concatenation. In the implementation of cross-layer weighted summation, the cross-attention sublayer closer to the decoder output corresponds to a larger aggregation weight. The engineering implication is that the information aggregated by the cross-attention sublayer closer to the decoder output is closer to the final prediction and better reflects the uncertainty of this inference.

[0060] Step 400: The cloud server calculates a scalar metric representing the degree of dispersion of the attention weight distribution obtained in step 300, and obtains the degree of dispersion metric.

[0061] For example, the measure of dispersion can be Shannon information entropy as a preferred implementation: in, The Shannon information entropy of the attention weight distribution; For the attention weight distribution corresponding to the first The weight value of each memory location; The length of the calibration memory sequence; To avoid the logarithmic operation from producing a zero value, a relatively small positive number can be chosen; This represents the natural logarithm function. This embodiment uses the convention... This is to handle the singularity of logarithmic operations when the weight values ​​approach zero.

[0062] More specifically, the physical meaning of Shannon's information entropy is: The smaller the value, the more the attention weight distribution is concentrated on a few historical records, which corresponds to the network finding a reference sample with high similarity to the current physiological state in this inference, and the internal basis of a single inference is more sufficient; The larger the value, the closer the attention weight distribution is to a uniform distribution, which corresponds to a lack of references in the calibration memory that are highly similar to the current sample, and relatively insufficient internal basis for a single inference.

[0063] Optionally, as an alternative implementation of the dispersion measure, this embodiment also describes the following measurement forms. All of these forms can reflect the degree of concentration or dispersion of the attention weight distribution and fall within the scope of the "dispersion measure" described in this application.

[0064] Normalized entropy: in, The normalized entropy normalizes the entropy value to the number of valid unmasked memories in the calibration memory sequence. Intervals facilitate cross-sample comparisons under different effective memory counts.

[0065] It should be noted that the calculation of the normalized entropy is based on... ≥ 2 is a prerequisite; when When = 1, A value of 0 results in a denominator of zero, a degradation scenario corresponding to the existence of only one valid memory record in the calibration memory sequence. In this scenario, the normalized entropy can be defined as 1 (i.e., maximum uncertainty), or the unnormalized Shannon information entropy can be used as an alternative metric. In engineering implementation, it is typically required that the calibration memory sequence contain at least two valid records before the inference process can be initiated.

[0066] Impurity of the gin: in, The Gini impurity is an indicator that does not involve logarithmic operations, making it particularly suitable for processors that do not support efficient logarithmic operations.

[0067] Valid number of entries: in, This intuitively represents the number of memory locations that are actually "effectively utilized" on the calibrated memory sequence by which the attention weights are distributed.

[0068] Top-k attention quality: in, This indicates the top ranking in the attention weight distribution. The set of position indices of the maximum weight value It is a preset positive integer; Reflects the focus on the front Attention quality at the most similar memory location.

[0069] The KL divergence between the attention weight distribution and the uniform distribution: in, This reflects the degree of deviation of the attention weight distribution from a uniform distribution. The greater the deviation, the more the weight distribution deviates from uniformity, i.e., the more it tends to be concentrated.

[0070] Coefficient of variation: in, and These are the standard deviation function and the mean function of the attention weight distribution, respectively.

[0071] It should be noted that this application does not impose strict limitations on the specific calculation form of the discreteness measure; furthermore, other internal intermediate states from a single forward inference (such as the variance of the output layer of network A, the norm of the last hidden state of the decoder, the embedding vector distance, etc.) can also be used as alternative implementation methods in subsequent steps.

[0072] It should be noted that the value of the dispersion measure is not strictly monotonic with the actual prediction error. For example, when the attention is incorrectly focused on a few historical locations that are not actually similar to the current sample, the measure value may be small, but the corresponding prediction error may be large. Based on this non-monotonic and non-linear correspondence, this embodiment does not directly use the threshold comparison result of the dispersion measure as the basis for determining the reliability level. Instead, in subsequent steps, it is used as an input feature and fed into an independently trained confidence classification network (hereinafter referred to as Network B) to learn the relevant non-linear mapping, as detailed in steps 500 to 600.

[0073] Step 500: The cloud server uses the dispersion measure obtained in step 400, the predicted value of the target physiological parameter obtained in step 200, and the prior anchor value as three features, and inputs them into network B, which does not share parameters with network A.

[0074] Specifically, this step includes the following sub-steps: like Figure 3 As shown, step 510: After dimensional normalization of the prior anchor point value, the predicted value of the target physiological parameter, and the dispersion measure, they are concatenated into a three-dimensional feature vector in a preset order: in, The prior anchor value, The predicted value for the target physiological parameter. This is a measure of the degree of dispersion (using Shannon information entropy as an example here, but other measures can be applied similarly). This is a dimensional normalization constant. Its function is to normalize the prior anchor point value and the predicted value of the target physiological parameter to a numerical range that is approximately of the same order of magnitude as the dispersion measure, so as to avoid numerical imbalance in subsequent projection layers due to dimensional differences.

[0075] Step 520: Project the three-dimensional feature vector through two fully connected layers and connect the two layers with a non-linear activation function to map it into a high-dimensional embedding vector: in, , , , These are the weight matrix and bias vector of two fully connected layers, respectively. It is a nonlinear activation function (in alternative implementations, other nonlinear activation functions such as ReLU or SiLU may be used). The high-dimensional embedding vector is used as the high-dimensional embedding vector. Optionally, the high-dimensional embedding vector is added to a learnable token embedding to serve as the confidence classification query token of the network B decoder, to strictly distinguish it from the target query vector used in the network A decoder. It should be noted that the confidence classification query vector and the target query vector are two independent vectors, operating on different networks, and their parameters are not shared.

[0076] Furthermore, the prior anchor value can be obtained from any of the following methods: a reference value obtained from the most recent offline physical calibration (e.g., finger-prick blood calibration used in blood glucose applications); a sliding weighted average of predicted values ​​at the most recent preset number of moments (the weights can be given in an exponential decay form); a low-frequency physiological trend estimate extracted from the historical prediction sequence by the low-frequency physiological trend estimation module; a weighted average of the aforementioned preset number of reference values ​​in the calibration memory sequence that have the highest similarity to the current input; or a baseline estimate output by an independent baseline estimation network. All of these multiple prior anchor definition methods fall within the scope of this application and can be selected based on the specific application scenario and data availability.

[0077] Step 530: The confidence classification query vector is used as the initial query, and together with the tokenized representation of the physiological signal time series and the calibration memory sequence, it is input into the decoder of network B. Self-attention calculation and cross-attention calculation on the calibration memory sequence are performed sequentially. It should be noted that, in one embodiment, network B may adopt an encoder-decoder type architecture similar to that of network A, but the two do not share parameters, have different training objectives, and have different output structures.

[0078] Optionally, as part of various implementations, the architecture of network B can employ an encoder-decoder type Transformer network (consistent with this embodiment), a multilayer perceptron, a lightweight structure that shares encoder features with network A but trains only independent classification heads, or a gradient boosting decision tree, etc. In deployment variants with limited computing power or stringent requirements for inference latency, the decoder portion of network B can be replaced by a multilayer perceptron, retaining only the nonlinear mapping function from three features to the reliability level.

[0079] Furthermore, as an implementation of input expansion, in addition to the three core features mentioned above, at least one of the following additional features can be jointly input into the network B: signal quality indicators (e.g., signal-to-noise ratio, motion artifact indicators), number of calibration samples, acquisition time interval, memory coverage features (e.g., nearest neighbor distance of the current sample in the memory feature space), and historical prediction fluctuation indicators (e.g., standard deviation of predicted values ​​at recent times). The introduction of these additional features helps to provide more robust reliability determination in cases of signal quality fluctuations or insufficient calibration library coverage. These additional features can be dimensionally normalized and then concatenated with the three core features.

[0080] Step 600: The network B outputs the reliability level corresponding to the predicted value of the target physiological parameter.

[0081] Specifically, the output vector of the last layer of the decoder of network B at the confidence classification query vector position is normalized and projected onto the classification head to obtain the score vectors (i.e., logits vectors) of each category, and then normalized by an exponential function to obtain the probability distribution of each reliability level: in, For each category's score vector, The probability distribution for each reliability level is shown below. This represents the total number of reliability level categories. The final reliability level is determined by the category with the highest probability. in, This is the final reliability level output in this step. This indicates the category index that makes the expression within the parentheses reach its maximum value.

[0082] Optionally, the output granularity of the reliability level can be in different forms, such as two-class, three-class, or five-class classification. For example, in a three-class implementation, the reliability level includes a high-confidence category. , medium confidence category and low confidence categories In the two-category implementation, the reliability level is simplified to two categories: "reliable" and "unreliable". In the five-category implementation, the reliability level is further subdivided into five levels: "high", "relatively high", "medium", "relatively low" and "low", which is convenient for use in more refined device-side action mapping.

[0083] Step 700: Based on the reliability level output in step 600, the cloud server generates control commands and sends them to the wearable physiological parameter monitoring device via the communication network to control the output management control module of the monitoring device to perform at least one device-side action on the predicted value of the target physiological parameter.

[0084] It should be noted that the generation and distribution of the control commands are completed by the cloud server, and the monitoring device only executes the corresponding hardware actions based on the received control commands. This end-to-cloud collaborative approach centralizes the complex calculations for reliability assessment in the cloud, thereby reducing the implementation complexity of the monitoring device.

[0085] Specifically, the device-side actions include, but are not limited to: outputting a marker on the display module of the monitoring device for the predicted value of the target physiological parameter; triggering the prediction value display suppression circuit of the monitoring device to suppress the display of the predicted value of the target physiological parameter; triggering a resampling prompt through the human-computer interaction module of the monitoring device; triggering a recalibration prompt through the human-computer interaction module of the monitoring device; and adjusting the data transmission strategy from the monitoring device to the cloud server (including but not limited to upload frequency, upload priority, or upload content).

[0086] For example, in a three-class classification implementation, when the reliability level is a high-confidence category... At that time, the control commands generated by the cloud server cause the monitoring device to output the predicted values ​​of the target physiological parameters according to the normal path; when the reliability level is of medium confidence category When the control command causes the monitoring device to trigger a display label or suggest supplementary sampling; when the reliability level is low confidence category When the control command causes the monitoring device to perform at least one device-side action, options include: lowering the data upload priority flag; reporting an abnormal status flag to the cloud server; switching to a backup prediction model (e.g., a backoff model based on linear extrapolation or historical moving average) within the cloud server; and restricting the predicted values ​​of the target physiological parameters belonging to the low-confidence category from entering the downstream closed-loop control module of the monitoring device. It should be noted that the above-described implementation involving the "downstream closed-loop control module" is only an optional embodiment, and this application does not involve specific treatment or drug administration decisions.

[0087] Furthermore, to avoid frequent jumps in the reliability level near the category critical threshold, which could lead to frequent switching of device actions, in an optional embodiment, the cloud server also performs smoothing processing on the reliability level sequence before generating control commands. The smoothing processing can employ any of the following methods or combinations: a hysteresis threshold mechanism, i.e., maintaining a preset difference between the critical threshold for transitioning from a low-risk level to a high-risk level and the critical threshold for reverse retreat; a time window majority voting smoothing mechanism, i.e., taking the level that appears most frequently in the most recent inference results as the current level; and a first-order hysteresis filtering mechanism, i.e., applying weighted smoothing to the logits of each category in the time dimension. The first-order hysteresis filtering can be exemplarily expressed as: in, For a moment The smoothed logits vector, For a moment The smoothed logits vector, For a moment The original output logits vector of network B, This is an adjustable smoothing coefficient, and its value range is... The larger the value, the more historical information is retained and the stronger the smoothing effect. The smoothing parameters can be flexibly selected according to the dynamic characteristics of the application scenario.

[0088] The above embodiments have the following technical effects: First, since the discreteness metric directly reuses the attention weight distribution calculated by network A in a single forward inference, it does not require the introduction of multiple complete forward inferences, which helps to add a reliability indicator signal to single-sample inference without significantly increasing computational overhead. Second, the three-way joint embedding enables the final reliability level to comprehensively consider the internal intermediate state, the current predicted value, and the prior anchor information, which helps to overcome the instability problem caused by the non-monotonic and nonlinear correspondence between a single discreteness metric and the prediction error. Third, since the method is deployed on a cloud server, the monitoring device does not need to bear the storage and computing power overhead of network A and network B, which facilitates integration in an incremental upgrade manner on wearable devices with limited computing power. Fourth, the reliability level is issued to the monitoring device in the form of control commands and directly drives its output management and control module to perform hardware actions, so that the output ultimately acts on the device hardware level, which helps to reduce the probability that low-reliability prediction values ​​are processed as high-reliability prediction values ​​by downstream device modules.

[0089] Example 2: Serial Training Method for Two Networks This example provides a serial training method for the aforementioned network B, with the following two-stage process: Figure 4As shown, the entire process is executed on the training server cluster, and the obtained parameters are then deployed to the cloud inference server for use. The training method adopts a two-stage serial execution training process, allowing network A and network B to be trained independently based on their respective loss targets, structurally isolating the mutual interference between their training targets.

[0090] Step 800 (First stage: Training the network A): Using a training sample set containing time series of physiological signals paired with target reference values, train the network A through regression loss until convergence, and then freeze the network parameters of the network A.

[0091] Specifically, let the training set be... ,in For the first The input features of each training sample (including the physiological signal time series and the corresponding calibration memory data). For the corresponding target reference value, This represents the total number of training samples. Optionally, the regression loss used in this stage is the asymmetric Huber loss. in, This indicates that network A has parameters Below is the input The given predicted value, The expectation operator represents the expectation of the distribution of training samples (at the implementation level, it corresponds to the sample average on the training set). For asymmetric weighting functions that depend on reference values, For The Huber loss function is the loss function for the inflection point. The Huber loss function is defined as follows: in, Indicates the predicted residual. represents the inflection point parameter of the Huber loss.

[0092] Furthermore, the inflection point The weights can be dynamically adjusted according to the range of the target reference value. For example, a smaller inflection point can be used for the low reference value range to strengthen the L1 linear penalty, while a larger inflection point can be used for the high reference value range to appropriately relax the penalty. The configuration applies heavier penalties to specific deviation directions—for example, relatively heavier penalties are applied to both overestimation in the direction of low reference values ​​and underestimation in the direction of high reference values. The engineering motivation for this asymmetric weighting is that if the predicted values ​​corresponding to these two types of directional deviations are directly read by downstream device modules, it is more likely to trigger inappropriate device-side actions. Therefore, applying heavier penalties to the regression errors in the corresponding directions during the training phase helps the trained network A to have a smaller prediction bias in those directions.

[0093] This training phase employs a standard deep learning training workflow, including but not limited to adaptive learning rate optimizers, early stopping strategies based on training and validation set partitioning, and mini-batch stochastic gradient descent. After training, the network A with the best regression performance on the validation set is selected. Save and freeze, and maintain in the second stage. It remains unchanged and does not participate in any gradient updates.

[0094] Step 900 (Second Phase: Training Network B): In the parameters of Network A In a frozen state and not participating in gradient updates, the training samples are input into network A to perform a single forward inference, obtain fixed target physiological parameter prediction values ​​and dispersion measures, and train network B accordingly.

[0095] Specifically, this stage further includes the following sub-steps: Step 910: For each training sample in the training sample set The network A, with its parameters frozen, performs a forward inference to obtain the predicted value of the target physiological parameter. The aforementioned dispersion measure and together with the aforementioned prior anchor value and the corresponding target reference value Also cached offline. It should be noted that, due to... It has been frozen, the aforementioned and For each training sample, it is a fixed quantity that can be cached offline once before the start of the second phase and remain unchanged throughout the training process of the network B. That is, it is "stable for each training sample but different across samples" and does not dynamically drift between different training rounds.

[0096] Step 920: Predict the value based on the fixed target physiological parameter. Compared with the target reference value The absolute or relative error between them, and at least two preset error thresholds. The numerical comparison relationship is used to generate stable discrete classification labels for the training samples offline across training rounds. : in, The prediction residual of the training samples, and The preset error threshold, Corresponding to high confidence categories , medium confidence category and low confidence categories Optionally, the threshold can also be given in the form of relative error or in the form of quantiles based on the residual distribution of the training set. It should be emphasized that the label generation process is based only on the numerical comparison between the absolute (or relative) error and the preset threshold, and does not involve any rules related to diagnosis or treatment.

[0097] Step 930: Optionally, directional overwriting rules and directional correctness exemption rules are applied to the discrete classification labels obtained in step 920. Specifically, the directional overwriting rules are configured as follows: when the target reference value is in a preset low reference value range and the fixed target physiological parameter prediction value is too high and exceeds a first directional threshold, or when the target reference value is in a preset high reference value range and the fixed target physiological parameter prediction value is too low and exceeds a second directional threshold, the discrete classification labels are forcibly overwritten. Overwrite as low confidence category The correct orientation exemption rule is configured as follows: when the target reference value is in the low reference value range and the predicted orientation is correct (i.e., ... When, or when the target reference value is in the high reference value range and the prediction direction is correct (i.e., When doing so, always maintain the discrete classification label. For high confidence category .

[0098] It should be further noted that the engineering motivation for the directional overwriting rule and the direction-correct exemption rule is that, in the two cases where the prediction is too high in the direction of the low reference value range or too low in the direction of the high reference value range, if the predicted value is directly read by the downstream device module, it will be more likely to trigger an inappropriate output action. Therefore, a heavier penalty is applied to the corresponding direction when the tag is generated. Conversely, maintaining a high confidence class on the direction-correct sample helps to reduce the unnecessary device-side action triggering rate under normal operating conditions. The directional overwriting rule and the direction-correct exemption rule can be enabled individually, in combination, or all disabled in different implementations.

[0099] Step 940: Cache the data from step 910. , and As described in step 500 of Embodiment 1, the three-way feature vectors are input into network B to obtain probability vectors for each category. .

[0100] Step 950: Update the parameters of network B using weighted cross-entropy loss: in, For a small batch of sample sets, The number of samples in the mini-batch sample set. The number of categories for reliability levels. For the first The discrete classification labels of each sample For network B, the first The output of the nth sample Class probability value, For the first Class weights This is an indicator function (it takes the value 1 if the condition inside the parentheses is true, otherwise it takes the value 0). It is a natural logarithm function. Optionally, the class weights satisfy... This imbalanced weighting results in the low-confidence category receiving the highest category weight. The engineering motivation for this imbalanced weighting is to ensure that the network B maintains a high recall rate for the low-confidence category, thereby reducing the probability that a low-reliability prediction is misclassified as a high-reliability prediction.

[0101] This training phase follows a similar standard deep learning training process to the first phase. Optionally, the early stopping metric can be a weighted combination of low-confidence class recall and overall accuracy. After training, the B parameters of the network that performs best on the validation set are saved. The final result and They are jointly deployed to a cloud-based inference server for the inference process described in Example 1.

[0102] It should be noted that the serial training method described in this embodiment is relative to the regression loss of network A. The classification loss of network B Weighted merging into The joint training scheme has structural differences. If joint training is used, the parameter updates of network A will be simultaneously affected by… The gradient effect can distort the prediction distribution of the prediction network in an effort to reduce classification loss—for example, by tending to output "safe predictions" closer to the label mean on uncertain samples to match high-confidence labels, or by distorting the attention distribution to fake high-confidence signals—thus impairing the prediction network's performance. The corresponding regression accuracy. This embodiment uses two-stage serial training, so that the parameters of network A are only changed by... The parameters of network B are determined solely by... The decision is to structurally isolate the mutual interference between the two targets; at the same time, this application's solution can integrate reliability level assessment capabilities in an incremental upgrade manner without modifying the already deployed predictive network structure and parameters.

[0103] Optionally, regarding the training data, the training samples may be derived from publicly available datasets or self-built datasets that have been legally and compliantly obtained and passed the corresponding ethical review. The human physiological data used must be from legal and compliant sources, and the annotation rules must not contain discriminatory or illegal content. The dataset is exemplarily split by subject number to avoid data leakage from the same subject across datasets.

[0104] The above embodiments have the following technical effects: First, since the network A is trained with pure regression loss in the first stage, and the network B is trained after freezing the network A in the second stage, the mutual interference between the regression target and the classification target is structurally isolated, which helps to maintain the regression accuracy of the original prediction network while introducing reliability assessment; Second, the labels generated based on the numerical comparison relationship between the fixed predicted value and the target reference value have stability across training rounds, avoiding optimization instability caused by label drift during training; Third, the settings of asymmetric Huber regression loss, directional overwriting rules, and directional correct exemption rules make the training target and the final device-side output management target consistent in engineering, which helps to improve the output reliability of the final system in boundary cases.

[0105] Example 3: Cloud Server Device This example provides a cloud server device for implementing the evaluation method described in Example 1 and / or the training method described in Example 2. It includes a processor and a memory communicatively connected to the processor. The processor is exemplarily one or a combination of a general-purpose central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), or a tensor processing unit (TPU). The memory exemplarily includes read-only memory, random access memory, and non-volatile storage media (e.g., solid-state drive, hard disk drive) for storing computer program instructions, model parameters (including parameters of network A and network B), a calibration memory database, and training logs. When the computer program instructions are executed by the processor, steps 100 to 700 of the evaluation method described in Example 1 are implemented, or steps 800 to 950 of the training method described in Example 2 are implemented.

[0106] Furthermore, the cloud server also includes a communication module configured to maintain a communication connection with one or more wearable physiological parameter monitoring devices via a wired or wireless communication network, receive physiological signal time series or their characteristic sequences uploaded by the monitoring devices, and issue control commands to the monitoring devices. Optionally, the cloud server also includes a model management module and a backup prediction module. The model management module is configured to perform version management, online updates, and rollback processing for network A and network B; the backup prediction module is activated when the reliability level is in the low confidence category, and the cloud server internally switches to the backup prediction model to generate rollback prediction values.

[0107] The above embodiments have the following technical effects: the cloud server has sufficient computing power and storage resources, which can support a large-scale network A and network B as well as a complete calibration memory bank, thus greatly reducing the implementation complexity of the wearable physiological parameter monitoring device on the edge; at the same time, centralized deployment in the cloud facilitates the continuous updating and monitoring of the model.

[0108] Example 4: Wearable Physiological Parameter Monitoring Device This example provides a wearable physiological parameter monitoring device that communicates with the aforementioned cloud server. The monitoring device includes the following hardware modules: a physiological signal sensor configured to collect raw physiological signals from the wearer's body surface. For example, the physiological signal sensor is a photoplethysmography (PPG) sensor, and the physiological signal it outputs is a PPG signal; in other embodiments, the physiological signal sensor may be a continuous glucose monitoring sensor, an electrocardiogram (ECG) sensor, a dual-wavelength optical blood oxygenation sensor, or a combination thereof.

[0109] The signal preprocessing circuit is electrically connected to the physiological signal sensor and is configured to perform preprocessing operations such as bandpass filtering, analog-to-digital conversion, motion artifact suppression, and normalization on the raw physiological signal, and output the physiological signal time series or its characteristic sequence.

[0110] The communication module is configured to transmit the physiological signal time series or its characteristic sequence to the cloud server via a wireless communication network, and to receive the control commands issued by the cloud server. The control commands are generated by the cloud server after executing the evaluation method described in Embodiment 1 based on the received data. The communication module exemplarily supports at least one wireless communication method among Bluetooth, Wi-Fi, or cellular networks.

[0111] The output management control module is configured to perform at least one device-side action on the predicted value of the target physiological parameter, including output marking, display suppression, resampling prompt, recalibration prompt, or data transmission strategy adjustment, based on the received control command. Specifically, the output management control module may include at least one of the following sub-circuits or sub-modules: a display driving circuit, a display suppression circuit, an audio-visual prompt sub-circuit, a resampling command generator, a recalibration prompt circuit, a data upload priority marking circuit, and a backup prediction model switching interface, to respond to different types of control commands respectively.

[0112] Furthermore, as a typical application, the physiological signal sensor can be a photoplethysmography (PPG) sensor, and the physiological signal time series is a PPG signal time series; the predicted value of the target physiological parameter is one of the following: a non-invasive blood glucose estimate, a continuous blood pressure estimate based on pulse wave conduction time or multi-channel pulse waves, or a blood oxygen saturation estimate based on dual-wavelength PPG. In different application scenarios, the specific form of the monitoring device can be a wrist-worn, finger-clip, patch, or cuff device.

[0113] The above embodiments have the following technical effects: the wearable physiological parameter monitoring device only undertakes the functions of signal acquisition, preprocessing, communication and hardware output management, which makes it easy to implement in a low-power, small-size and low-cost form; and obtains a high level of predictive reliability assessment capability through collaboration with the cloud server.

[0114] Example 5: Typical Application Scenarios and Implementation Methods This example further provides specific implementation methods of this application in three typical application scenarios to illustrate the feasibility of the methods, devices, and training methods described in this application under different physiological parameter prediction objects.

[0115] Application Scenario 1: Non-invasive blood glucose estimation device based on PPG signal.

[0116] In this application scenario, the physiological signal sensor of the monitoring device is a photoplethysmography (PPG) sensor, and the physiological signal time series is a sampling sequence of PPG waveform signals within a certain time window. The reference value for each record in the calibration memory data is the reference blood glucose concentration obtained from fingertip blood calibration at the corresponding time. Network A outputs a non-invasive blood glucose estimate. The units are the same as the target reference values. Prior anchor point values. The reference blood glucose value can be obtained from the most recent finger-prick blood glucose calibration, or it can be a moving weighted average of the predicted values ​​from the most recent preset number of time points. The error threshold for the three-class classification of the network B... and The tolerance for prediction errors is configured by those skilled in the art according to the application. When network B outputs a low-confidence category... At that time, the control command issued by the cloud server controls the monitoring device to trigger at least one of the following actions: predicted value display suppression circuit, resampling command, recalibration prompt circuit, or data upload priority flag.

[0117] Application Scenario 2: Continuous blood pressure estimation device based on multi-channel pulse waves.

[0118] In this application scenario, the physiological signal sensor of the monitoring device is exemplarily a dual-channel PPG sensor that simultaneously collects data from two different body surface locations (e.g., fingertips and wrists) of the same subject. Device-side preprocessing exemplarily includes bandpass filtering and normalization of the two PPG signals, and time alignment of the zero-crossing points of the two pulse waves using cross-correlation. The reference values ​​for each record in the calibration memory data are the systolic and diastolic blood pressure reference values ​​obtained from cuff-type blood pressure monitor calibration; network A simultaneously outputs two scalar predicted values ​​for systolic and diastolic blood pressure. Prior anchor value. For example, the systolic blood pressure value obtained from the most recent cuff blood pressure monitor calibration is used. When the network B outputs a low confidence category, the control command instructs the monitoring device to trigger the cuff calibration prompt circuit and pause the continuous blood pressure value display.

[0119] Application Scenario 3: Blood oxygen saturation estimation device based on dual-wavelength PPG.

[0120] In this application scenario, the physiological signal sensor of the monitoring device is exemplarily a dual-channel PPG sensor with red and infrared light. Device-side preprocessing exemplarily includes bandpass filtering and normalization of the two-channel PPG, and calculating the AC component ratio sequence of the two channels as additional input features. The reference value for each record in the calibration memory data is the blood oxygen saturation reference value obtained from the calibration of the finger-clip pulse oximeter. The network A outputs the blood oxygen saturation (…). Scalar predicted value. Prior anchor value. For example, a moving average of the predicted values ​​over a preset time period is taken. When network B outputs a low-confidence category, the control command instructs the monitoring device to trigger a re-acquisition command and mark the current predicted value with a specific marker on the display module.

[0121] It should be noted that the attention weight distribution extraction process, the three-way joint embedding process, and the dual-network serial training process in the above application scenarios one to three are consistent with those in embodiments one and two. The only differences are in the type of physiological sensor used, the sampling rate, the window length, the physical meaning of the target parameters, the definition of the prior anchor point value, and the specific value of the error threshold. This application does not impose strict limitations on the specific types of physiological parameters, sensor types, sampling rates, and window lengths of the application objects.

[0122] The above embodiments have the following technical effects: The prediction reliability assessment method provided in this application can be implemented in a unified framework on a variety of encoder-decoder type prediction devices that take physiological signals as input. It only requires adjustment of the sensor type, the definition of the prior anchor point value and the error threshold value according to the specific application, and has good application adaptability and scalability.

[0123] Example 6: Computer-readable storage medium and computer program product. This example provides a computer-readable storage medium storing computer program instructions thereon; and a computer program product including computer program instructions. When the computer program instructions are executed by a processor, they implement steps 100 to 700 of the evaluation method described in Example 1, or steps 800 to 950 of the serial training method described in Example 2.

[0124] Furthermore, the computer-readable storage medium includes, but is not limited to, any form of non-volatile storage medium or combination thereof, such as read-only memory, random access memory, flash memory, solid-state drive, hard disk drive, optical disk, semiconductor memory, etc. The computer program product can be provided in a form that facilitates deployment, such as program code, binary executable files, container images, software packages, application programs, firmware images, or cloud service interfaces.

[0125] The above embodiments have the following technical effects: the computer-readable storage medium and computer program product facilitate the deployment and distribution of the method described in this application on different hardware platforms, and expand the implementation forms of this application.

[0126] It should be specifically noted that when the physiological sensor completely fails (e.g., remains at zero or saturated for an extended period) or the distribution of the input physiological signal drifts significantly relative to the training set (i.e., exceeds the distribution scenario, out-of-distribution), the reliability level output by this application may itself be inaccurate. In such extreme cases, the monitoring device should preferably be used in conjunction with an independently deployed sensor self-test module or an out-of-distribution detection module (e.g., a detection circuit based on signal amplitude thresholds and spectral energy thresholds) for fallback processing; the reliability level output by this application and the determination of the fallback module can be combined in an "OR" logic combination in the output management and control module of the monitoring device to affect the downstream device operation.

[0127] In summary, based on embodiments one through six, the solution described in this application has the following overall technical effects: First, since the dispersion measure directly reuses the attention weight distribution calculated by network A in a single forward inference, only one complete forward inference is required in each sampling period. This allows for the output of the corresponding reliability level while simultaneously outputting the predicted value of the target physiological parameter, which helps to add a reliability indication signal to the single-sample inference without significantly increasing computational overhead.

[0128] Secondly, since the network B uses the discreteness measure, the predicted value of the target physiological parameter and the prior anchor value as input through three-way joint embedding, and the independently trained network learns the nonlinear mapping from the three-way features to the reliability level, it helps to overcome the instability problem of judgment caused by the non-monotonic and nonlinear correspondence between a single discrete measure and the prediction error.

[0129] Third, since the serial training method freezes all parameters of network A in the second stage, the regression prediction target and the reliability classification target are structurally isolated from each other, which helps to avoid the target coupling interference caused by joint training, and enables this scheme to integrate reliability assessment capabilities in an incremental upgrade manner without modifying the deployed prediction network.

[0130] Fourth, since the reliability level is issued to the monitoring equipment in the form of control commands and directly drives its output management control module to perform corresponding equipment-side actions, the output of this application solution ultimately acts on the equipment hardware level, which helps to improve the overall output reliability management capability of the monitoring equipment.

[0131] Fifth, since the cloud deployment method concentrates the main computing power requirements on the cloud server, the monitoring device does not need to bear the storage and computing power overhead of the network A and the network B, which helps to reduce the computing power, power consumption and size constraints of the wearable device, and facilitates the implementation and expansion of this solution in large-scale wearable continuous monitoring application scenarios.

[0132] Sixth, the method, equipment and training method described above can reuse the same technical framework for different physiological parameter prediction objects (including but not limited to non-invasive blood glucose, continuous blood pressure, blood oxygen saturation, etc.) by only adjusting the sensor type, sampling parameters and threshold values, which has good versatility and scalability.

[0133] The embodiments of this application are further explained below by way of example.

[0134] For example, a two-stage communication architecture can be adopted between the cloud server and the wearable physiological parameter monitoring device—the monitoring device connects to a local gateway (e.g., a smartphone or dedicated gateway device) via a low-power Bluetooth short-range communication link, and then the local gateway accesses the cloud server via Wi-Fi or a cellular mobile communication network; alternatively, the monitoring device can directly access the cloud server via a cellular IoT module (e.g., NB-IoT or Cat.1). The upload frequency of the physiological signal time series is configured according to the specific application scenario: in continuous monitoring scenarios, a periodic upload method can be used, uploading the time series in each sampling window in segments; in event-triggered scenarios, a threshold-based triggering method can be used, uploading only when a specific event occurs. The cloud server can set up a receiving buffer queue to cope with network jitter. The end-to-end round-trip latency of the communication link under normal operating conditions is within seconds, meeting the real-time requirements of continuous physiological monitoring scenarios measured in minutes; communication encryption is exemplarily based on TLS 1.3.

[0135] In one specific implementation, both network A and network B adopt an encoder-decoder type Transformer structure, and the embedding dimension is exemplarily taken as follows: The encoder and decoder each have 3 layers, and the number of multi-head attention heads is exemplarily taken as follows: Dimensions per head Hidden Dimensions of Feedforward Sublayer The activation function is GELU with a parameter size of approximately 0.5 M; in an alternative implementation, the number of layers can be 2 to 6, and the number of attention heads... The number of values ​​can be 2, 4, 8, or 16, depending on the embedded dimension. The maximum number of dimensions can be 64, 128, or 256; the hidden dimension of the feedforward sublayer can be 256 or 512; and the calibration memory length can be adjusted. Options include 32, 64, or 128. The specific value of the above parameters can be chosen by weighing the cloud server's computing power budget against the target response latency requirements.

[0136] In the implementation of blood glucose applications, the PPG sampling rate is exemplarily taken as follows: The sampling window length is, for example, 10 seconds, and the input sequence length is... The tokenization layer uses a one-dimensional convolutional layer, with an example kernel size of 16 and a stride of 8, outputting a token length. In blood pressure applications, exemplarily, one can take... Window time 8 seconds; this can be exemplarily taken in blood oxygenation applications. Window for 5 seconds.

[0137] Take an example , can also to Values ​​within a range; blood glucose applications For example, 100 (mg / dL) is used to normalize the original blood glucose scalar value to... Magnitude; Error threshold in blood glucose applications and Take, for example, respectively and (Approximately corresponding to relative errors of 10% and 20%), or values ​​can be taken based on the quantiles of the training set residual distribution (e.g., 60% and 90% quantiles); in blood pressure applications, this can be exemplarily taken as... , ; Blood oxygenation can be exemplarily taken , .

[0138] In blood glucose applications, the low reference range is exemplarily defined as follows: The first direction threshold is exemplarily taken as follows: The high reference value range is exemplarily shown as follows: The second direction threshold is exemplarily taken as follows: In other applications such as blood pressure and blood oxygen monitoring, the directional threshold can be configured by those skilled in the art based on the sensitivity of the specific application to directional deviation.

[0139] Category weight ratio is taken as an example , can also to The recall rate is adjusted within the range based on the low confidence category requirement. In the asymmetric Huber regression loss for blood glucose applications, the low reference range ( )Pick Take the middle interval High reference value range ( )Pick Asymmetric weights Samples with directional bias of 1.5 times were penalized for both low-value regions and high-value regions.

[0140] Both network A and network B are trained using the Adam optimizer; the learning rate of network A is set to an exemplary value. The learning rate of network B is exemplarily taken as follows: The batch size is 64 for example; the early stop patience is 20 epochs for example; the early stop metric of network B is exemplarily a 1:1 weighted combination of low-confidence category recall and overall precision. The training data comes from publicly available physiological signal datasets (e.g., the publicly available dataset OhioT1DM containing synchronously collected PPG and continuous blood glucose monitoring data) and self-built datasets that have been legally and compliantly obtained and approved by the relevant ethical review agencies; the dataset is divided into training set, validation set, and test set (subject-wise split) according to the subject number, with an exemplary split ratio of 7:1:2 or 7:1.5:1.5; the human physiological data sources used are legal and compliant, and the annotation rules do not contain discriminatory or illegal content.

[0141] Hysteresis threshold is taken as an example The time window length is taken as an example. One inference cycle; continuous At least in the next vote The device state machine is updated only when all outputs are of the same category; first-order hysteresis filter coefficient. For example, 0.3 is used, but it can also be adjusted between 0.1 and 0.5 according to the dynamic characteristics.

[0142] The sliding weighted average window length is exemplarily taken as follows: At any given moment, the weights are... Decay; time decay constant Take an example The low-pass filter cutoff frequency can also be adjusted within the range of 10-60 minutes. Take an example .

[0143] Network A and Network B can be deployed on cloud inference nodes equipped with graphics processing units or dedicated AI accelerators. The inference nodes can employ model quantization (e.g., INT8 quantization) to reduce inference time. In concurrent service scenarios, the cloud server can perform dynamic batching inference on requests from different wearable physiological parameter monitoring devices to improve throughput, and distribute the requests to idle inference nodes through an inference service load balancer. The cloud inference nodes can also be configured with a model version canary release mechanism to support smooth switching during online model upgrades. Under the aforementioned hardware platform and concurrent conditions, the single forward inference time of Network A is in the millisecond range. The method described in this application only additionally introduces the summation and logarithmic operation of the attention weight distribution and a small-scale forward inference of Network B.

[0144] The backup prediction model can be, for example, a low-order extrapolation model based on a moving weighted average of historical prediction values, a retrieval-based backoff model based on a weighted average of nearest neighbor reference values ​​in the calibration memory, or a simplified prediction model based on a lightweight one-dimensional convolutional neural network. The switching triggering conditions can be exemplarily included: the reliability level outputting a low-confidence category for a number of consecutive times; the cloud server detecting a persistently abnormal attention distribution in network A (e.g., continuously approaching a uniform distribution for more than a preset period); or the signal quality index reported by the monitoring device being lower than a preset threshold. After switching to the backup prediction model, the cloud server can still run the original network A in parallel to monitor its subsequent performance, and switch back to the original network A when the original network A returns to normal.

[0145] In blood pressure estimation applications, the similarity can be exemplarily calculated based on the cosine similarity of pulse wave morphology feature vectors (e.g., vectors composed of manually constructed features such as pulse wave amplitude, rise time, and diastolic descent slope); the relative acquisition time interval is in minutes and decays exponentially. Normalization to The interval, in which This is a preset time constant. In blood oxygen estimation applications, the similarity can be exemplarily calculated based on the normalized Euclidean distance of the red / infrared AC component ratio sequence of dual-wavelength PPG.

[0146] For example, in the event of a communication interruption, the monitoring device can activate lightweight fallback logic locally: perform a simple signal quality check on the physiological signal time series; if the signal quality is lower than a preset threshold, trigger local resampling and recalibration prompts; simultaneously, temporarily store the physiological signal time series to be uploaded in the monitoring device's local memory, and upload them in batches to the cloud server for supplementary inference after communication is restored. This local fallback logic does not involve deep learning inference calculations and does not overlap with network A and network B in the cloud. The technical solutions provided in the above embodiments, through the synergistic cooperation of the aforementioned technical features under the edge-cloud collaborative architecture, help to achieve corresponding technical effects in multiple dimensions simultaneously, which will be explained below.

[0147] Since the attention weight distribution is used as a reused object of the internal intermediate state from a single forward inference, the calculation of the discreteness measure only involves the summation and logarithmic operations of the already calculated attention weight matrix, without introducing additional complete forward inference. Therefore, it helps the cloud server to synchronously produce the target physiological parameter prediction value and the corresponding reliability level at the cost of a single forward inference in each sampling period. Compared with the solution that requires repeated forward inference or parallel integration of multiple models, it helps to reduce the overall computing power and storage overhead of the server in the high-concurrency request scenario in the cloud. At the same time, since the discreteness measure is anchored in the physical meaning of "the degree of concentration of similar reference samples that can be located in the historical calibration memory of the current physiological state", the generated reliability indication signal has interpretable meaning at the engineering level.

[0148] Furthermore, since the proposed scheme initializes the confidence classification query vector by normalizing and jointly embedding the three features—the discreteness measure, the predicted value of the target physiological parameter, and the prior anchor value—and the independently trained confidence classification network learns the nonlinear mapping from the three features to the discrete reliability level, when a non-monotonic or nonlinear correspondence occurs between the discreteness measure and the actual prediction error, the scheme helps to mitigate the instability problem caused by direct thresholding of a single measure by introducing two auxiliary information sources: the predicted value and the prior anchor value. When the discreteness measure is further adopted in various optional forms (including but not limited to Shannon information entropy, normalized entropy, Gini impurity, effective memory count, Top-k attention quality, and uniformly distributed KL),... When considering divergence and coefficient of variation, the proposed scheme can be flexibly adapted to the instruction set support and numerical computation characteristics of the cloud-based inference nodes. For example, Gini impurity is preferred on processors that do not support efficient logarithmic operations, and normalized entropy is preferred in scenarios with high requirements for cross-sample comparability. When a normalized exponential function with a mask applied to the filling position is used, it helps to maintain the cross-sample comparability of the metric under different effective memory counts. When a multi-layer weighted fusion or multi-head aggregation method is used, it helps the cross-attention sublayer closer to the decoder output to obtain higher feature weights because the information it aggregates is closer to the final prediction, thereby further improving the stability of the reliability level determination. When the prior anchor value is further flexibly selected from multiple sources such as offline physical calibration value, sliding weighted average, low-frequency physiological trend estimate, weighted average of nearest neighbor reference value and independent baseline estimation network output value, the scheme can still operate stably under the difference in the availability of prior information in different application scenarios, thus expanding the application and adaptation range of the scheme.

[0149] Furthermore, since the confidence classification network in the proposed scheme is structurally independent of the prediction network and their parameters are not shared, and since the proposed scheme adopts a two-stage serial training framework—in which the second stage trains the confidence classification network with weighted cross-entropy loss while the prediction network parameters are frozen and do not participate in any gradient updates—it helps to structurally isolate the gradient interference between the regression prediction target and the reliability classification target, avoiding the prediction network from distorting the prediction distribution or fabricating high-confidence attention signals to reduce classification loss. This helps to maintain the regression accuracy of the original prediction network while introducing reliability assessment capabilities. At the same time, this scheme can be integrated into the existing cloud inference system in an incremental upgrade manner without modifying the structure and parameters of the deployed prediction network. Furthermore, since the offline-generated discrete classification labels are generated solely based on the numerical comparison between fixed predicted values ​​and target reference values, they remain stable across training rounds without drifting, contributing to the optimization and stability of the training process. When directional overwriting rules and directional correctness exemption rules are added, the training objective and the final output management objective on the device side correspond in an engineering sense, which helps improve the engineering consistency between reliability level determination and device-side actions in application scenarios sensitive to directional bias. When the first stage further employs asymmetric Huber regression loss and imposes heavier penalties on samples with low reference values ​​that are directionally biased and high reference values ​​that are directionally biased, it helps the trained prediction network to have a relatively small prediction bias in the stated direction. The weighted cross-entropy loss, which assigns the highest class weight to the low-confidence category, helps improve the recall level of the confidence classification network for the low-confidence category and reduces the probability of low-reliability predictions being misclassified as high-reliability predictions.

[0150] Regarding the calibration memory data, this scheme clearly states that each record consists of four features: a historical physiological signal waveform segment, the corresponding reference value, the similarity between the current target physiological signal and the historical physiological signal in the low-dimensional feature space, and the relative acquisition time interval between the acquisition time of the record and the current inference time. The introduction of the similarity allows the prediction network to obtain a preliminary matching indication between each historical record and the current sample before performing cross-attention, which helps to improve the efficiency of memory retrieval and the targeting of cross-attention calculation.

[0151] Furthermore, since the proposed solution sends the reliability level to the monitoring device in the form of control commands and directly drives its output management control module to perform hardware actions such as output marking, display suppression, resampling prompts, recalibration prompts, or data transmission strategy adjustments, the output of the reliability assessment ultimately affects the device hardware level. This helps reduce the probability that low-reliability prediction values ​​are processed as high-confidence prediction values ​​by downstream device modules, thereby improving the overall output reliability management capability of the monitoring device. When the reliability level is a low-confidence category, the cloud server can also perform actions such as reducing data upload priority, reporting abnormal status indicators, and switching within the cloud server. Actions such as switching to a backup prediction model or restricting low-confidence prediction values ​​from entering the downstream closed-loop control module make the differentiated execution of equipment-side actions more targeted. When the reliability level jumps near the category critical threshold, a hysteresis threshold mechanism, a time window majority voting smoothing mechanism, or a first-order hysteresis filter smoothing process is further applied, it helps to avoid frequent switching of equipment-side actions. When the input of the confidence classification network is additionally combined with signal quality indicators, the number of calibration samples, the acquisition time interval, the memory library coverage, or historical prediction fluctuation indicators, it helps to provide a more robust reliability judgment in the case of signal quality fluctuations or insufficient calibration library coverage.

[0152] Furthermore, since the overall inference calculation of the proposed scheme is centrally deployed on a cloud server, the monitoring device only undertakes the functions of physiological signal acquisition, preprocessing, data uploading, and driving hardware according to control commands. This helps to keep the monitoring device within a range suitable for continuous wear in terms of power consumption, size, and cost. Since the attention weight distribution extraction process, the three-way joint embedding process, and the dual-network serial training process do not impose strict restrictions on specific physiological parameter objects, the proposed scheme can be applied to various scenarios such as non-invasive blood glucose estimation with PPG signal as input, continuous blood pressure estimation with multi-channel pulse wave as input, and blood oxygen saturation estimation with dual-wavelength PPG as input. Only the sensor type, sampling parameters, prior anchor value definition, and error threshold value need to be adjusted accordingly to reuse the same technical framework, which has good versatility and scalability.

[0153] It should be noted that the achievement of the technical effect is based on the premise that the various technical features of this solution are interconnected and cooperate in accordance with the aforementioned internal context. In extreme cases such as when the physiological sensor fails completely or the distribution of the input physiological signal drifts significantly relative to the training set, the reliability level output by this solution may also be inaccurate. In such boundary cases, the independent sensor self-test module or the external detection module on the monitoring device side can be used as a fallback, and the reliability level output by the solution can be logically combined with the reliability level output by the solution in the output management and control module to act on the downstream device.

[0154] It should be noted that in this patent application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this patent application, if it refers to performing an action according to an element, it means performing the action at least according to that element, including two cases: performing the action only according to that element, and performing the action according to that element and other elements. Expressions such as "multiple," "repeatedly," and "various" include two, two times, two kinds, and more than two, more than two times, and more than two kinds.

[0155] All documents mentioned in this application are considered to be incorporated in their entirety into the disclosure of this application so that they can serve as a basis for modifications if necessary. Furthermore, it should be understood that after reading the foregoing disclosure of this application, those skilled in the art can make various alterations or modifications to this application, and these equivalent forms also fall within the scope of protection claimed in this application.

Claims

1. A method for predictive reliability assessment of wearable physiological parameter monitoring devices, characterized in that, A cloud server that communicates with the wearable physiological parameter monitoring device includes: Step 1: Acquire the time series or characteristic sequence of physiological signals collected by the physiological sensors of the monitoring device, preprocessed, and uploaded; Step 2: Input the physiological signal time series or its feature sequence together with the calibration memory data into the trained encoder-decoder type prediction network. The decoder of the prediction network performs cross-attention calculation on the calibration memory data and outputs the predicted value of the target physiological parameter. Step 3: Extract the attention weight distribution of the target query vector to the calibration memory sequence from at least one cross-attention sublayer of the decoder described in Step 2; Step 4: Based on the attention weight distribution obtained in Step 3, calculate a scalar metric that characterizes its degree of dispersion to obtain the degree of dispersion metric; Step 5: The dispersion measure obtained in Step 4, the predicted value of the target physiological parameter obtained in Step 2, and the prior anchor value are used as three features and input together into the confidence classification network that does not share parameters with the prediction network. Step Six: The confidence classification network outputs the reliability level corresponding to the predicted value of the target physiological parameter; Step 7: Based on the reliability level output in Step 6, the cloud server generates control commands and sends them to the monitoring device to control the output management control module of the monitoring device to perform at least one device-side action on the predicted value of the target physiological parameter, including output marking, display suppression, resampling prompt, recalibration prompt, or data transmission strategy adjustment.

2. The method according to claim 1, characterized in that, The dispersion measure mentioned in step four uses at least one of the following calculation methods: Shannon's information entropy: ; Normalized entropy: or ; Impurity of the gin: ; Valid number of entries: ; Top-k attention quality: ; The KL divergence between the attention weight distribution and the uniform distribution: ; Coefficient of variation: ; in, The length of the calibration memory sequence, The number of valid memory entries that were not masked in the calibration memory sequence. For the attention weight distribution corresponding to the first The weight value of each memory location, This is a numerical stability constant used to avoid zero values ​​in logarithmic operations. Represents the natural logarithm function. It is a preset positive integer. This indicates the top ranking in the attention weight distribution. The set of position indices of the maximum weight value and These are the standard deviation function and the mean function of the attention weight distribution, respectively.

3. The method according to claim 1, characterized in that, In step five, the step of inputting the discreteness measure obtained in step four, the predicted value of the target physiological parameter obtained in step two, and the prior anchor value as three features into the confidence classification network includes: after the prior anchor value, the predicted value of the target physiological parameter, and the discreteness measure are respectively normalized, they are concatenated into a three-dimensional feature vector. The three-dimensional feature vector is projected through two fully connected layers and connected by a nonlinear activation function to map it into a high-dimensional embedding vector. The high-dimensional embedding vector is added to the learnable token embedding to initialize the confidence classification query vector of the confidence classification network decoder. The prior anchor value is obtained in one of the following ways: the reference value obtained from the most recent offline physical calibration, the sliding weighted average of the predicted values ​​at the most recent preset number of times, the low-frequency physiological trend estimate extracted by the low-frequency physiological trend estimation module, the weighted average of the aforementioned preset number of reference values ​​with the highest similarity to the current input in the calibration memory sequence, or the baseline estimate output by an independent baseline estimation network.

4. The method according to claim 1, characterized in that, In step seven, when the reliability level is low confidence category, the device-side action controlled by the generated control command to control the monitoring device to perform includes at least one of the following: reducing the data upload priority flag; Report the abnormal status indicator to the cloud server; Switch to the backup prediction model within the cloud server; The predicted values ​​of the target physiological parameters belonging to the low-confidence category are restricted from entering the downstream closed-loop control module of the monitoring device.

5. The method according to claim 1, characterized in that, In step two, each record in the calibration memory data consists of the following four features: a historical physiological signal waveform segment, the corresponding reference value, the similarity between the current target physiological signal and the historical physiological signal in the low-dimensional feature space, and the relative acquisition time interval between the acquisition time of the record and the current inference time.

6. A serial training method for a confidence classification network in a cloud server, characterized in that, include: Phase 1: Using training samples containing time series of physiological signals and target reference values, train the encoder-decoder type prediction network through regression loss until convergence, and then freeze the network parameters of the prediction network. Second stage: With the network parameters frozen and not participating in gradient updates, the training samples are input into the prediction network to perform a single forward inference, obtain fixed target physiological parameter prediction values, and extract the attention weight distribution from at least one cross-attention sublayer of the prediction network decoder to calculate the degree of dispersion measure. Based on the absolute or relative error between the predicted value of the fixed target physiological parameter and the target reference value, and at least two preset error thresholds... The numerical comparison relationship is used to generate stable discrete classification labels for the training samples offline across training rounds; The fixed target physiological parameter prediction value, the dispersion measure and the prior anchor value are jointly input into the confidence classification network as three features. The parameters of the confidence classification network are updated with weighted cross-entropy loss, wherein the discrete classification labels of the low-confidence categories are assigned higher class weights than those of the high-confidence categories.

7. The serial training method according to claim 6, characterized in that, The weighted cross-entropy loss takes the form of: in, For a small batch of samples, The size of the mini-batch sample set. The number of categories for reliability levels. For the first The discrete classification labels of each sample For the confidence classification network, the first The output of the nth sample Class probability value, For the first Class weights This is an indicator function that takes the value 1 if the condition within the parentheses is true, and 0 otherwise. The function is the natural logarithm; the category weights satisfy... This allows the low-confidence category to receive the highest category weight.

8. A cloud server, characterized in that, include: processor; The processor also includes a memory that is communicatively connected to the processor, the memory storing computer program instructions that, when executed by the processor, implement the steps of the predictive reliability assessment method as described in any one of claims 1 to 5, or the steps of the serial training method as described in claim 6 or 7.

9. A wearable physiological parameter monitoring device, characterized in that, Communication connection with a cloud server includes: a physiological signal sensor configured to acquire physiological signals; The signal preprocessing circuit is configured to preprocess the physiological signal and convert it into a physiological signal time series or its characteristic sequence. The communication module is configured to send the physiological signal time series or its characteristic sequence to the cloud server and receive control instructions issued by the cloud server, wherein the control instructions are generated by the cloud server after executing the evaluation method as described in any one of claims 1 to 5 based on the received data; The output management control module is configured to perform at least one device-side action on the predicted value of the target physiological parameter, including output marking, display suppression, resampling prompt, recalibration prompt, or data transmission strategy adjustment, based on the received control command.

10. The wearable physiological parameter monitoring device according to claim 9, characterized in that, The physiological signal sensor is a photoplethysmography (PPG) sensor, and the physiological signal time series is a PPG signal time series; the predicted value of the target physiological parameter is one of the following: non-invasive blood glucose estimate, continuous blood pressure estimate based on pulse wave conduction time or multi-channel pulse wave, or blood oxygen saturation estimate based on dual-wavelength PPG.

Images