A method and system for unsupervised domain adaptive index calibration in multiple scenarios for pulmonary rehabilitation

Abstract

This application relates to the field of smart healthcare technology and discloses a method and system for calibrating unsupervised domain adaptive indicators for multi-scenario pulmonary rehabilitation. The method includes real-time acquisition of user-defined scenario data, including data from hospitals, communities, and homes; constructing and training a domain adaptive neural network model to obtain a trained unsupervised domain adaptive model; adding a scenario-conditional embedding module to the unsupervised domain adaptive model, inputting the scenario data to be calibrated into the unsupervised domain adaptive model to obtain prediction results; determining a scenario deviation vector; and performing online calibration of the original evaluation indicators in the prediction results based on the scenario deviation vector to obtain standardized indicators that match the user and the current scenario. This method eliminates the need for cross-scenario paired data annotation, achieving consistency of pulmonary rehabilitation evaluation indicators in continuous management across scenarios, and improving the reliability and effectiveness of remote rehabilitation management.

Description

Technical Field

[0001] This application relates to the field of smart healthcare technology, and in particular to a method and system for unsupervised domain adaptive index calibration in multiple scenarios for lung rehabilitation. Background Technology

[0002] Pulmonary rehabilitation is a core component of chronic respiratory disease management, and its effectiveness depends on continuous and accurate assessment of the patient's physiological state. With the development of wearable devices and telemedicine technologies, pulmonary rehabilitation management has expanded from a single hospital setting to the community and home, forming an integrated continuous management model of "hospital-community-home." However, inherent differences in different rehabilitation scenarios significantly affect the consistency and comparability of assessment indicators: First, monitoring devices (such as different models of pulse oximeters and ECG patches) have different measurement principles, accuracy, and noise characteristics; second, environmental conditions (such as electromagnetic interference in the home and noise from community activities) vary; and finally, there are systematic differences in the activity patterns and intensity distribution of rehabilitation training for patients in different rehabilitation scenarios. These factors lead to significant baseline shifts or scale differences in key quantitative indicators such as the "respiratory muscle fatigue index" and "training load index" calculated from data collected from the same patient under the same physiological state in different rehabilitation scenarios.

[0003] In related technologies, solving the problem of indicator consistency often employs simple linear calibration or supervised calibration methods. Linear calibration requires pre-conducting paired measurements on the same group of patients in multiple scenarios to establish a conversion formula, which is costly and difficult to implement on a large scale in clinical practice. Supervised calibration methods require a large amount of cross-scenario data with consistent clinical annotations for model training; however, "consistent clinical annotations" are themselves difficult to obtain due to scenario differences. Therefore, how to automatically learn and eliminate the impact of scenario differences on quantitative indicators in the absence of cross-scenario paired-annotated data (i.e., unsupervised) has become a technical bottleneck for achieving truly continuous and reliable pulmonary rehabilitation management. Therefore, there is an urgent need to develop a multi-scenario unsupervised domain adaptive indicator calibration method and system for pulmonary rehabilitation. Summary of the Invention

[0004] In view of this, this application provides a method and system for unsupervised domain adaptive index calibration in multiple scenarios of pulmonary rehabilitation. It does not require cross-scenario paired labeled data, and can achieve consistency of pulmonary rehabilitation assessment indicators in continuous management across scenarios, thereby improving the reliability and effectiveness of remote rehabilitation management.

[0005] Specifically, the following technical solutions are included:

[0006] This application provides a method for unsupervised domain adaptive index calibration in multiple scenarios for pulmonary rehabilitation, including the following steps:

[0007] Real-time acquisition of users’ data on scenarios to be calibrated, including data on hospitals, communities, and homes to be calibrated;

[0008] Construct and train a domain adaptive neural network model to obtain a trained unsupervised domain adaptive model;

[0009] Add a scene-conditional embedding module to the unsupervised domain adaptive model, input the scene data to be calibrated into the unsupervised domain adaptive model, and obtain the prediction results;

[0010] Determine the scene deviation vector;

[0011] The original evaluation metrics in the prediction results are calibrated online based on the scene deviation vector to obtain standardized metrics that match the user and the current scene.

[0012] In some embodiments, the domain adaptive neural network model includes a feature extractor, a scene discriminator, and an index regressor; the feature extractor is used to extract features from the input data, and its output is connected to the scene discriminator and the index regressor, respectively; the scene discriminator is used to classify the features into scenes, and the index regressor is used to output the original evaluation index.

[0013] In some embodiments, the training domain adaptive neural network model includes:

[0014] Collect time-series multimodal physiological-behavioral data of the target user in at least two different rehabilitation scenarios. The time-series multimodal physiological-behavioral data includes, but is not limited to, respiratory signals, electrocardiogram signals, and motion signals.

[0015] Partial annotation of temporal multimodal physiological-behavioral data;

[0016] Temporal multimodal physiological-behavioral data labeled with evaluation metrics are used as source scene training data, and temporal multimodal physiological-behavioral data without evaluation metrics are used as target scene training data. The source scene is the hospital scene, and the target scene is the home scene and / or community scene.

[0017] The source scene training data and the target scene training data are input into the domain adaptive neural network model. The feature extractor obtains the deep feature representations corresponding to each scene. The deep feature representations of each scene are used as a unified input for the training and feature alignment of the domain adaptive neural network model. The domain adaptive neural network model is jointly optimized and trained by minimizing the index regression loss of the index regressor on the source scene training data and minimizing the adversarial loss between the feature extractor and the scene discriminator. The index regression loss is used to constrain the prediction accuracy of the evaluation index of the domain adaptive neural network model on the source scene training data, and the adversarial loss is used to constrain the feature extractor to learn a consistent feature distribution across scenes, so that the feature extractor extracts the scene-independent domain-invariant feature representations.

[0018] In some embodiments, the domain adaptive neural network model adopts a domain adversarial neural network structure, and its overall loss function is:

[0019]

[0020] in, For feature extractor, For scene discriminator, For index regressors, , These are the evaluation metrics for the source scene training data and their corresponding annotations. , These are the mixed input data consisting of source scene training data and target scene training data, and their corresponding scene labels. For the regression loss of the indicator, To combat the losses, This is a tradeoff parameter used to balance the regression loss and adversarial loss of the aforementioned index.

[0021] In some embodiments, adding the scene-conditional embedding module to the unsupervised adaptive model includes:

[0022] The scene conditional embedding module is connected to the input of the index regressor to encode the embedding vectors corresponding to different scenes; the scene data to be calibrated is input into the feature extractor to obtain the domain invariant feature representation, and the domain invariant feature representation is fused with the embedding vector of the corresponding scene and then input into the index regressor to obtain the original evaluation index.

[0023] In some embodiments, determining the scene deviation vector includes:

[0024] The difference between the target scene output of the domain-invariant features in the index regressor and the conditional virtual output of the source scene obtained by fusing the domain-invariant features with the source scene embedding vector and inputting it into the index regressor is used to obtain the scene bias estimate of the current data segment. The specific calculation formula is as follows:

[0025]

[0026] in, Input data for the target scenario. For feature extractor, For index regressors, Embed vectors for the target scene. Embed vectors for the source scene. This represents the scene deviation estimate for the corresponding data segment;

[0027] Statistical analysis is performed on the scene deviation estimates corresponding to multiple data segments in the target scene to obtain the scene deviation vector of the target scene relative to the source scene. The specific calculation formula is as follows:

[0028]

[0029] in, This is the scene deviation vector of the target scene relative to the source scene. Training data for the target scene.

[0030] In some embodiments, the method further includes:

[0031] Establish user profiles and record users' historical deviation data in various scenarios; when the data accumulation of the same user in the same scenario exceeds the threshold, use Bayesian update or online learning to adjust the user's steady-state deviation vector in that scenario for subsequent online calibration.

[0032] In some embodiments, the different rehabilitation scenarios include hospital scenarios, home scenarios, and community scenarios. Differences between these scenarios include differences in monitoring equipment models, environmental electromagnetic and motion noise levels, and the intensity distribution of typical patient activity patterns. These differences are discriminated and modeled by a scenario discriminator during model training, used for learning constraint-domain-invariant features during adversarial training of the feature extractor, and encoded as scenario embedding vectors in the scenario conditional embedding module for subsequent index regression and scenario bias vector calculation.

[0033] This application provides a multi-scenario unsupervised domain adaptive index calibration system for pulmonary rehabilitation, used to implement the aforementioned multi-scenario unsupervised domain adaptive index calibration method for pulmonary rehabilitation. The system includes:

[0034] The acquisition module is used to acquire the user's calibration scenario data in real time. The calibration scenario data includes calibration hospital data, calibration community data, and calibration home data.

[0035] The construction and training module is used to build and train the domain adaptive neural network model to obtain a trained unsupervised domain adaptive model.

[0036] The prediction module is used to add the scene conditional embedding module to the unsupervised domain adaptive model. The scene data to be calibrated is input into the unsupervised domain adaptive model to obtain the prediction result.

[0037] The determination module is used to determine the scene deviation vector;

[0038] The calibration module is used to perform online calibration of the original evaluation indicators obtained by real-time calculation of the scene data to be calibrated based on the scene deviation vector, so as to obtain the calibrated standardized indicators.

[0039] The beneficial effects of the technical solutions provided in this application include at least the following:

[0040] (1) No need for cross-scenario data annotation: The innovative introduction of unsupervised domain adaptive technology into the multi-scenario calibration problem of lung rehabilitation breaks through the bottleneck of data annotation and makes large-scale application possible.

[0041] (2) High calibration accuracy: By learning complex nonlinear scene differences through deep networks, it can capture the composite deviations caused by the coupling of equipment, environment and behavior better than traditional linear methods.

[0042] (3) Support for personalization and online learning: It can adaptively adjust calibration parameters based on the patient's personal historical data and support online model updates to adapt to equipment replacement or slow environmental changes.

[0043] (4) Ensure the continuity of assessment: It fundamentally solves the problem of inconsistent indicator readings in different rehabilitation scenarios, enabling doctors to track the entire rehabilitation process of patients based on a unified and continuous indicator curve and make more accurate judgments. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the structure of the domain adaptive neural network model provided in the embodiments of this application.

[0046] Figure 2 The diagram shows the effect of the scenario-conditional index prediction mechanism provided in the embodiments of this application.

[0047] Figure 3 The diagram shows the effect of determining the scene deviation vector provided in the embodiment of this application. Detailed Implementation

[0048] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0049] To make the technical solutions and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0050] This application provides a method for unsupervised domain adaptive index calibration in multiple scenarios for pulmonary rehabilitation, including the following steps:

[0051] It acquires users' data on scenarios to be calibrated in real time, including data from hospitals, communities, and homes. By uniformly collecting raw physiological and behavioral data from multiple scenarios (hospitals, communities, and homes), it provides a complete cross-scenario data foundation for subsequent unsupervised adaptive calibration, avoids model bias caused by missing scenarios, and supports continuous management of the entire "hospital-community-home" chain of lung rehabilitation.

[0052] A domain-adaptive neural network model is constructed and trained to obtain a well-trained unsupervised domain-adaptive model. An adversarial training mechanism is then used to enable the feature extractor to learn scene-invariant features. This process eliminates feature distribution shifts caused by differences in device, environment, and activity patterns without requiring any cross-scene pairwise labeled data, significantly improving the model's generalization ability in target scenarios (such as home and community scenarios).

[0053] A scene-conditional embedding module is added to an unsupervised adaptive model. The scene data to be calibrated is input into the unsupervised adaptive model to obtain prediction results. A scene-conditional index prediction mechanism is designed to enable the model to separate general physiological information from scene-specific interference in the features. By introducing the scene-conditional embedding module into the pre-trained model, the model can explicitly separate general physiological states from scene-specific interference. On one hand, the index regressor can generate a virtual output in the style of the source scene based on "domain-invariant features + source scene encoding"; on the other hand, the original output is obtained based on the target scene encoding. This mechanism provides an interpretable, end-to-end differentiable computational path for subsequent accurate estimation of scene bias.

[0054] The system determines a scene deviation vector. By comparing the differences between the conditional output of the source scene and the original output of the target scene using the same domain-invariant features, or by statistically analyzing the deviation distribution of a batch of historical data, the systematic deviation of each target scene relative to the source scene is quantitatively estimated. This deviation vector can capture nonlinear, multi-factor coupled scene interference, is more accurate than traditional linear calibration methods, and can adapt to personalized offsets for different users and different device types.

[0055] In the real-time application phase, the unsupervised adaptive model first infers from the data of the scenario to be calibrated, outputting raw evaluation indicators. Then, based on the scenario deviation vector, the raw evaluation indicators are calibrated online to obtain standardized indicators comparable to the source scenario (e.g., a hospital). This calibration eliminates the need for repeated collection of cross-scenario paired data and changes to patients' daily usage habits, thus achieving low-cost, high-real-time standardized output of pulmonary rehabilitation indicators, effectively supporting remote rehabilitation early warning, longitudinal comparison of treatment efficacy, and multi-center clinical studies.

[0056] In some embodiments, the domain adaptive neural network model includes a feature extractor, a scene discriminator, and an index regressor; the feature extractor is used to extract features from the input data, and its output is connected to the scene discriminator and the index regressor, respectively; the scene discriminator is used to classify the features into scenes, and the index regressor is used to output the original evaluation index.

[0057] In some embodiments, the training domain adaptive neural network model includes:

[0058] Collect time-series multimodal physiological-behavioral data of the target user in at least two different rehabilitation scenarios. The time-series multimodal physiological-behavioral data includes, but is not limited to, respiratory signals, electrocardiogram signals, and motion signals.

[0059] Partial annotation of temporal multimodal physiological-behavioral data;

[0060] Temporal multimodal physiological-behavioral data labeled with evaluation metrics are used as source scene training data, and temporal multimodal physiological-behavioral data without evaluation metrics are used as target scene training data. The source scene is the hospital scene, and the target scene is the home scene and / or community scene.

[0061] The source scene training data and the target scene training data are input into the domain adaptive neural network model. The feature extractor obtains the deep feature representations corresponding to each scene. The deep feature representations of each scene are used as a unified input for the training and feature alignment of the domain adaptive neural network model. The domain adaptive neural network model is jointly optimized and trained by minimizing the index regression loss of the index regressor on the source scene training data and minimizing the adversarial loss between the feature extractor and the scene discriminator. The index regression loss is used to constrain the prediction accuracy of the evaluation index of the domain adaptive neural network model on the source scene training data, and the adversarial loss is used to constrain the feature extractor to learn a consistent feature distribution across scenes, so that the feature extractor extracts the scene-independent domain-invariant feature representations.

[0062] In some embodiments, the domain adaptive neural network model adopts a domain adversarial neural network structure, and its overall loss function is:

[0063]

[0064] in, For feature extractor, For scene discriminator, For index regressors, , These are the evaluation metrics for the source scene training data and their corresponding annotations. , These are the mixed input data consisting of source scene training data and target scene training data, and their corresponding scene labels. For the regression loss of the indicator, To combat the losses, This is a tradeoff parameter used to balance the regression loss and adversarial loss of the aforementioned index.

[0065] In some embodiments, adding the scene-conditional embedding module to the unsupervised adaptive model includes:

[0066] The scene conditional embedding module is connected to the input of the index regressor to encode the embedding vectors corresponding to different scenes; the scene data to be calibrated is input into the feature extractor to obtain the domain invariant feature representation, and the domain invariant feature representation is fused with the embedding vector of the corresponding scene and then input into the index regressor to obtain the original evaluation index.

[0067] In some embodiments, determining the scene deviation vector includes:

[0068] The difference between the target scene output of the domain-invariant features in the index regressor and the conditional virtual output of the source scene obtained by fusing the domain-invariant features with the source scene embedding vector and inputting it into the index regressor is used to obtain the scene bias estimate of the current data segment. The specific calculation formula is as follows:

[0069]

[0070] in, Input data for the target scenario. For feature extractor, For index regressors, Embed vectors for the target scene. Embed vectors for the source scene. This represents the scene deviation estimate for the corresponding data segment;

[0071] Statistical analysis is performed on the scene deviation estimates corresponding to multiple data segments in the target scene to obtain the scene deviation vector of the target scene relative to the source scene. The specific calculation formula is as follows:

[0072]

[0073] in, This is the scene deviation vector of the target scene relative to the source scene. Training data for the target scene.

[0074] In some embodiments, the method further includes:

[0075] Establish user profiles and record users' historical deviation data in various scenarios; when the data accumulation of the same user in the same scenario exceeds the threshold, use Bayesian update or online learning to adjust the user's steady-state deviation vector in that scenario for subsequent online calibration.

[0076] In some embodiments, the different rehabilitation scenarios include hospital scenarios, home scenarios, and community scenarios. Differences between these scenarios include differences in monitoring equipment models, environmental electromagnetic and motion noise levels, and the intensity distribution of typical patient activity patterns. These differences are discriminated and modeled by a scenario discriminator during model training, used for learning constraint-domain-invariant features during adversarial training of the feature extractor, and encoded as scenario embedding vectors in the scenario conditional embedding module for subsequent index regression and scenario bias vector calculation.

[0077] In some embodiments, the method may be specifically implemented as follows:

[0078] (1) In the hospital setting, multimodal temporal physiological-behavioral data of 50 target users under standard pulmonary rehabilitation training were collected. The data included, but was not limited to, respiratory signals, electrocardiogram signals, and motion signals. Training load assessment indicators (such as low, medium, and high training load levels) were labeled for each time segment according to clinical assessment criteria. In the home setting, unlabeled multimodal temporal physiological-behavioral data of the same group or another group of users during daily activities or rehabilitation training at home were collected. All data were standardized and preprocessed.

[0079] (2) For example Figure 1 As shown, a domain adversarial neural network (i.e., a domain adaptive neural network model) is constructed. Feature extractor A one-dimensional convolutional neural network is used to extract high-level features from raw time-series data. Scene discriminator. It's a binary classifier that attempts to determine whether the input features originate from a hospital or home setting. (Indicator regressor) It is a fully connected network used to predict training load evaluation metrics.

[0080] (3) During training, hospital data (with labeled load assessment indicators) and household data (without labeled load assessment indicators) are simultaneously input into the domain adversarial neural network. Through backpropagation, the index regressor is optimized. and scene discriminator At the same time, for feature extractors Employing a gradient inversion layer enables the feature extractor to... The extracted features should be used to obfuscate the scene discriminator as much as possible. Thus, it becomes a domain-invariant feature independent of the scene.

[0081] (4) such as Figure 2 As shown, a scene-conditional embedding module is added to the trained unsupervised adaptive model. This module can provide a low-dimensional encoding vector for each scene learning.

[0082] (5) Input the scene data to be calibrated into the unsupervised adaptive model to obtain the prediction results.

[0083] (6) Determine the scene bias vector: Calculate the difference between the target scene output of the domain-invariant features in the index regressor and the conditional virtual output of the source scene obtained by fusing the domain-invariant features with the source scene embedding vector and inputting it into the index regressor, to obtain the scene bias estimate of the current data segment. Specifically, on hospital data, the model learns the source scene embedding vector. (i.e., hospital code). For the entered family data... After feature extractor Obtaining domain-invariant features Domain-invariant features Embedded vectors of home scenes respectively (i.e., family coding, initialized as a zero vector or learnable) and hospital scene embedding vectors To perform fusion, input index regressor This yields two predicted values: the original output based on the family scene embedding vector. Virtual output based on hospital scene embedding vectors The difference between the two is defined as the scene bias estimate for the current data segment. Statistical analysis (e.g., averaging) is performed on the scene deviation estimates corresponding to multiple data segments in the home scenario to obtain the scene deviation vector of the home scenario relative to the hospital scenario. .

[0084] (7) Online calibration: When a user is undergoing rehabilitation in a target scenario (such as a home scenario), the original evaluation index is first obtained through an unsupervised domain adaptive model; then the steady-state deviation vector of the corresponding scenario is called to perform calibration calculation on the original evaluation index: standardized index = original evaluation index - scenario deviation vector (or corrected by a mapping function), thereby obtaining a unified index that can be directly compared with the hospital scenario.

[0085] (8) Personalized Adaptation: Establish user profiles and record users' historical deviation data in various scenarios; when a user accumulates valid data in the home scenario for a certain period of time (such as 2 weeks), use Bayesian update or online learning to update the user's global home scenario deviation vector. Fine-tuning is performed to obtain personalized deviations for this user. This is used for subsequent, more precise calibration.

[0086] Through the above embodiments, this application realizes automatic, high-precision, and personalized calibration of pulmonary rehabilitation indicators without the need for cross-scenario labeling, providing key technical support for building a continuous, unified, and reliable digital management system for pulmonary rehabilitation.

[0087] In summary, this application can achieve consistency of pulmonary rehabilitation assessment indicators in continuous management across different scenarios without requiring paired annotation data across scenarios, thereby improving the reliability and effectiveness of remote rehabilitation management.

[0088] This application provides a multi-scenario unsupervised domain adaptive index calibration system for pulmonary rehabilitation, used to implement the aforementioned multi-scenario unsupervised domain adaptive index calibration method for pulmonary rehabilitation. The system includes:

[0089] The acquisition module is used to acquire the user's calibration scenario data in real time. The calibration scenario data includes calibration hospital data, calibration community data, and calibration home data.

[0090] The construction and training module is used to build and train the domain adaptive neural network model to obtain a trained unsupervised domain adaptive model.

[0091] The prediction module is used to add the scene conditional embedding module to the unsupervised domain adaptive model. The scene data to be calibrated is input into the unsupervised domain adaptive model to obtain the prediction result.

[0092] The determination module is used to determine the scene deviation vector;

[0093] The calibration module is used to perform online calibration of the original evaluation metrics in the prediction results based on the scene deviation vector, so as to obtain standardized metrics that match the user and the current scene.

[0094] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only.

[0095] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for unsupervised domain adaptive index calibration in multi-scenario pulmonary rehabilitation, characterized in that, Includes the following steps: Real-time acquisition of users' data on scenarios to be calibrated, including data from hospitals, communities, and homes. Construct and train a domain adaptive neural network model to obtain a trained unsupervised domain adaptive model; Add a scene-conditional embedding module to the unsupervised domain adaptive model, input the scene data to be calibrated into the unsupervised domain adaptive model, and obtain the prediction results; Determine the scene deviation vector; The original evaluation metrics in the prediction results are calibrated online based on the scene deviation vector to obtain standardized metrics that match the user and the current scene.

2. The method for unsupervised domain adaptive index calibration in multi-scenario lung rehabilitation according to claim 1, characterized in that, The domain adaptive neural network model includes a feature extractor, a scene discriminator, and an indicator regressor; the feature extractor is used to extract features from the input data, and its output is connected to the scene discriminator and the indicator regressor respectively; the scene discriminator is used to classify the features into scenes, and the indicator regressor is used to output the original evaluation indicator.

3. The method for unsupervised domain adaptive index calibration in multi-scenario lung rehabilitation according to claim 2, characterized in that, The training domain adaptive neural network model includes: Collect time-series multimodal physiological-behavioral data of the target user in at least two different rehabilitation scenarios. The time-series multimodal physiological-behavioral data includes, but is not limited to, respiratory signals, electrocardiogram signals, and motion signals. Partial annotation of temporal multimodal physiological-behavioral data; Temporal multimodal physiological-behavioral data labeled with evaluation metrics are used as source scene training data, and temporal multimodal physiological-behavioral data without evaluation metrics are used as target scene training data. The source scene is the hospital scene, and the target scene is the home scene and / or community scene. The source scene training data and the target scene training data are input into the domain adaptive neural network model. The feature extractor obtains the deep feature representations corresponding to each scene. The deep feature representations of each scene are used as a unified input for the training and feature alignment of the domain adaptive neural network model. The domain adaptive neural network model is jointly optimized and trained by minimizing the index regression loss of the index regressor on the source scene training data and minimizing the adversarial loss between the feature extractor and the scene discriminator. The index regression loss is used to constrain the prediction accuracy of the evaluation index of the domain adaptive neural network model on the source scene training data, and the adversarial loss is used to constrain the feature extractor to learn a consistent feature distribution across scenes, so that the feature extractor extracts the scene-independent domain-invariant feature representations.

4. The method for multi-scenario unsupervised adaptive index calibration in pulmonary rehabilitation according to claim 3, characterized in that, The domain adaptive neural network model adopts a domain adversarial neural network structure, and its overall loss function is: in, For feature extractors, For scene discriminator, For index regressors, , These are the evaluation metrics for the source scene training data and their corresponding annotations. , These are the mixed input data consisting of source scene training data and target scene training data, and their corresponding scene labels. For the regression loss of the indicator, To combat the losses, This is a tradeoff parameter used to balance the regression loss and adversarial loss of the aforementioned index.

5. The method for unsupervised domain adaptive index calibration in multi-scenario pulmonary rehabilitation according to claim 1, characterized in that, The step of adding the scene-conditional embedding module to the unsupervised domain adaptive model includes: connecting the scene-conditional embedding module to the input of the index regressor to encode the embedding vectors corresponding to different scenes; inputting the scene data to be calibrated into the feature extractor to obtain the domain-invariant feature representation, and then fusing the domain-invariant feature representation with the embedding vector of the corresponding scene and inputting it into the index regressor to obtain the original evaluation index.

6. The method for unsupervised domain adaptive index calibration in multi-scenario pulmonary rehabilitation according to claim 3, characterized in that, The determination of the scene deviation vector includes: calculating the target scene output of the domain-invariant features in the index regressor, and the difference between the conditional virtual output of the source scene obtained by fusing the domain-invariant features with the source scene embedding vector and inputting it into the index regressor, to obtain the scene deviation estimate of the current data segment. The specific calculation formula is as follows: in, Input data for the target scenario. For feature extractors, For index regressors, Embed vectors for the target scene. Embed vectors for the source scene. This represents the scene deviation estimate for the corresponding data segment; Statistical analysis is performed on the scene deviation estimates corresponding to multiple data segments in the target scene to obtain the scene deviation vector of the target scene relative to the source scene. The specific calculation formula is as follows: in, This is the scene deviation vector of the target scene relative to the source scene. Training data for the target scene.

7. The method for unsupervised domain adaptive index calibration in multi-scenario pulmonary rehabilitation according to claim 1, characterized in that, The method further includes: Establish user profiles and record users' historical deviation data in various scenarios; when the data accumulation of the same user in the same scenario exceeds the threshold, use Bayesian update or online learning to adjust the user's steady-state deviation vector in that scenario for subsequent online calibration.

8. The method for multi-scenario unsupervised adaptive index calibration in pulmonary rehabilitation according to claim 3, characterized in that, The different rehabilitation scenarios include hospital scenarios, home scenarios, and community scenarios.

9. A multi-scenario unsupervised domain adaptive index calibration system for pulmonary rehabilitation, characterized in that, For implementing the multi-scenario unsupervised domain adaptive index calibration method for lung rehabilitation as described in any one of claims 1-8, the system comprises: The acquisition module is used to acquire the user's calibration scenario data in real time. The calibration scenario data includes calibration hospital data, calibration community data, and calibration home data. The construction and training module is used to build and train the domain adaptive neural network model to obtain a trained unsupervised domain adaptive model. The prediction module is used to add the scene conditional embedding module to the unsupervised domain adaptive model. The scene data to be calibrated is input into the unsupervised domain adaptive model to obtain the prediction result. The determination module is used to determine the scene deviation vector; The calibration module is used to perform online calibration of the original evaluation metrics in the prediction results based on the scene deviation vector, so as to obtain standardized metrics that match the user and the current scene.

Images