Intelligent feedback method and system for lung function exercise

Abstract

The invention discloses an intelligent feedback method and system for lung function exercise, and the method comprises the steps: firstly obtaining real-time multi-dimensional synchronous data, carrying out the deviation evaluation based on an individualized standard mode, and generating a fine deviation time sequence; then, performing shared feature coding on the time series data by using a recurrent neural network so as to capture the dynamic change of the lung state; a task context vector containing rich spatio-temporal information is extracted from the coding feature sequence through double-channel conjoint analysis, so that the limitation of information dissociation and static fusion in the prior art can be effectively overcome; and then a multi-task decoder is used to identify the root cause of the deviation, the specific focus area and the severity thereof from the context vector, and a solid basis is provided for subsequent feedback. In this way, the accuracy and fineness of deviation attribution are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of intelligent feedback, and more specifically, to an intelligent feedback method and system for lung function training. Background Technology

[0002] Pulmonary function training plays a crucial role in the rehabilitation of respiratory diseases, effectively improving patients' lung function and quality of life. However, traditional pulmonary function training often lacks real-time, precise guidance, leading to poor results for patients due to improper form or force application, and potentially causing unnecessary compensation or injury. Therefore, developing a pulmonary function training program that provides intelligent feedback to ensure patients can correctly and effectively perform the training is of great significance for improving rehabilitation outcomes.

[0003] However, existing technologies have significant limitations in providing intelligent feedback for lung function training. Current systems typically employ a rule-based mapping mechanism, directly mapping detected symptoms to preset instructions. This approach fails to delve into the underlying physiological causal relationships behind the deviations. Specifically, a simple rule-based system cannot distinguish between these situations, resulting in vague and impractical instructions that fail to effectively guide users on how to correct their behavior. Furthermore, even more advanced machine learning-based classification models often only learn the correlation between deviation features and feedback instructions, rather than the true causal chain. Therefore, the fundamental bottleneck of existing technologies lies in their treatment of multidimensional respiratory time-series signals as flat feature vectors, failing to model the inherent physiological causal structure of the human respiratory system. Consequently, when deviations occur, the root cause cannot be inferred. This technological deficiency leads to insufficient accuracy, precision, and interpretability of feedback, severely restricting the efficiency and effectiveness of user rehabilitation training.

[0004] Therefore, we look forward to an optimized intelligent feedback method for lung function training. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide an intelligent feedback method and system for lung function training. First, it acquires real-time multidimensional synchronous data and performs deviation assessment based on an individualized standardized pattern, generating a refined deviation time series. Then, it uses a recurrent neural network to encode shared features of these time series data to capture dynamic changes in lung status. Next, through dual-channel joint analysis, it extracts a task context vector containing rich spatiotemporal information from the encoded feature sequence, effectively overcoming the limitations of information dissociation and static fusion in existing technologies. Subsequently, a multi-task decoder is used to identify the root cause, specific focal area, and severity of the deviation from this context vector, providing a solid basis for subsequent feedback. This approach significantly improves the accuracy and precision of deviation attribution.

[0006] According to one aspect of this application, a smart feedback method for lung function training is provided, comprising: Obtain real-time synchronized data; Based on the personalized specification model, deviation assessment is performed on real-time synchronized data to obtain the deviation time sequence; The deviation time series is encoded using shared features based on a recurrent neural network model to obtain the lung state deviation encoded feature sequence; A dual-channel joint analysis was performed on the lung state deviation encoded feature sequence to obtain the task context vector; Multi-task decoding of the task context vector is performed to obtain the cause probability, focus region probability, and severity score; Based on the probability of cause, the probability of focal area, and the severity score, actionable feedback is generated.

[0007] According to another aspect of this application, a smart feedback system for lung function training is provided, comprising: The data acquisition module is used to acquire real-time synchronized data; The deviation assessment module is used to assess the deviation of real-time synchronized data based on the personalized specification mode to obtain the deviation time sequence. The shared feature encoding module is used to encode the deviation time series based on a recurrent neural network model to obtain the lung state deviation encoded feature sequence; The dual-channel joint analysis module is used to perform dual-channel joint analysis on the lung state deviation coding feature sequence to obtain the task context vector; The multi-task decoding module is used to perform multi-task decoding on the task context vector to obtain the cause probability, focus region probability, and severity score; The actionability feedback module generates actionable feedback based on cause probability, focus area probability, and severity score.

[0008] Compared with existing technologies, the intelligent feedback method and system for lung function training provided in this application first acquires real-time multidimensional synchronous data and performs deviation assessment based on an individualized standardized pattern to generate a refined deviation time series. Then, a recurrent neural network is used to encode shared features of these time series data to capture dynamic changes in lung status. Next, through dual-channel joint analysis, a task context vector containing rich spatiotemporal information is extracted from the encoded feature sequence, effectively overcoming the limitations of information dissociation and static fusion in existing technologies. Subsequently, a multi-task decoder is used to identify the root cause, specific focal area, and severity of the deviation from this context vector, providing a solid basis for subsequent feedback. This approach significantly improves the accuracy and precision of deviation attribution. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of the intelligent feedback method for lung function training according to an embodiment of this application; Figure 2 This is a schematic diagram of the data flow in the intelligent feedback method for lung function training according to an embodiment of this application; Figure 3 This is a block diagram of an intelligent feedback system for lung function training according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0016] The technical solution of this application proposes an intelligent feedback method for lung function training. Figure 1 This is a flowchart of a lung function training intelligent feedback method according to an embodiment of this application. Figure 2 This is a system architecture diagram of the intelligent feedback method for lung function training according to an embodiment of this application. Figure 1 and Figure 2 As shown, the intelligent feedback method for lung function training according to an embodiment of this application includes the following steps: S1, acquiring real-time synchronized data; S2, performing deviation assessment on the real-time synchronized data based on a personalized standard mode to obtain a deviation time series; S3, performing shared feature encoding on the deviation time series based on a recurrent neural network model to obtain a lung state deviation encoding feature sequence; S4, performing dual-channel joint analysis on the lung state deviation encoding feature sequence to obtain a task context vector; S5, performing multi-task decoding on the task context vector to obtain causal probability, focal region probability, and severity score; S6, generating actionable feedback opinions based on the causal probability, focal region probability, and severity score.

[0017] Specifically, S1 involves acquiring real-time synchronized data. It should be understood that the evaluation of the effectiveness of lung function training and the diagnosis of deviations rely on the accurate perception and dynamic capture of the user's physiological state during exercise. Only by continuously and accurately acquiring the user's multi-dimensional physiological signals can the system establish a comprehensive understanding of the user's current breathing pattern, and then compare and analyze it with a personalized standard pattern, thereby providing a reliable data foundation for subsequent deviation assessment, cause identification, and feedback generation. In this process, "real-time" emphasizes the timeliness of data acquisition, meaning that data should be collected simultaneously with or within a very short delay of the user's exercise to reflect the current true physiological state, rather than historical data; "synchronization" refers to the consistency of different physiological signals on the time axis, ensuring that multiple related indicators, such as chest expansion and abdominal movement, and changes in inspiratory flow and diaphragmatic activity, can be accurately correlated. For example, in a scenario of judging chest-abdominal coordination disorder, if the data on chest movement and abdominal movement are not strictly synchronized, the system will find it difficult to accurately assess the relative temporal relationship between the two, thus failing to effectively diagnose coordination problems. This synchronization is crucial for subsequent time-series alignment based on multidimensional dynamic time warping and for calculating the signal deviation between the personalized specification pattern and the real-time synchronized data. It ensures the accuracy and reliability of deviation assessment. If there is a lag, inaccuracy, or asynchrony in the data acquisition process, it will directly affect the accuracy of the system's judgment of the user's status, rendering all subsequent intelligent analysis and feedback worthless and unable to effectively guide the user to perform correct lung function exercises.

[0018] In practice, this typically involves using various sensor devices to monitor and collect multiple physiological indicators of the user during lung function exercises in real time. These data may include, but are not limited to, respiratory flow, respiratory pressure, chest and abdominal movement amplitude, heart rate, and electromyography (EMG) signals. These signals are crucial for assessing lung activity, respiratory muscle coordination, and overall respiratory efficiency. During data acquisition, the various sensors need to operate in parallel at high sampling rates, ensuring that the different physiological signals collected are strictly synchronized in time. This means that all data points should be accompanied by precise timestamps to ensure accurate alignment in subsequent analysis, capturing the instantaneous correlations and dynamic evolution between different physiological phenomena. This real-time synchronized data acquisition mechanism is key to ensuring that the system can accurately assess the user's breathing patterns, identify any subtle deviations, and understand their temporal characteristics.

[0019] Specifically, S2, based on a personalized standard model, performs deviation assessment on real-time synchronized data to obtain the deviation time series. It should be understood that the effectiveness assessment of pulmonary function training is not simply a comparison with a general standard, but requires full consideration of individual differences. Each user's physiological condition, disease type, rehabilitation stage, and even body shape habits may lead to differences in their "ideal" breathing pattern. Therefore, by establishing a personalized standard model, the system can more accurately understand the correct training method, thereby ensuring that subsequent deviation assessments accurately reflect the user's actual performance and the gap between the target and the individual level. Specifically, the purpose of this step is to quantitatively compare the multi-dimensional physiological data collected by the user in real time with a pre-set ideal training model that conforms to their individual characteristics, thereby revealing the specific deviations the user exhibits during training and their changes over time, providing core data support for subsequent intelligent diagnosis and accurate feedback.

[0020] In practical implementation, the personalized training model and real-time synchronized data are first aligned temporally using multidimensional dynamic time warping (MDTW) to obtain a warped path. It is understandable that although users attempt to follow a personalized training model, there are often natural and unavoidable non-linear scaling and offsets between their real-time physiological data and the model on the time axis. For example, the duration of inhalation or exhalation may vary slightly, or the rhythm of the entire respiratory cycle may change subtly. Without precise alignment of these temporal differences, directly comparing the two sets of data may lead to incorrect deviation calculations, such as misjudging rhythmic lag as insufficient amplitude, thus severely affecting the accuracy of subsequent deviation attribution. Therefore, in the technical solution of this application, through multidimensional dynamic time warping (MDTW) technology, the system can intelligently align real-time synchronized data to achieve optimal temporal matching with the personalized training model. This ensures that subsequent signal deviation calculations are performed between the truly corresponding physiological stages, accurately revealing the user's actual physiological deviation, rather than a simple temporal misalignment.

[0021] In this process, firstly, a local distance matrix is ​​constructed. Assuming the personalized canonical pattern is a multi-dimensional time series and the real-time synchronized data is another multi-dimensional time series, the system calculates the local distance between any two corresponding elements in the two multi-dimensional time series. For multi-dimensional data, this distance is typically Euclidean distance to quantify the similarity or difference between the two multi-dimensional vectors. Secondly, a cumulative distance matrix is ​​constructed. After calculating all local distances, the system constructs a cumulative distance matrix where each element represents the minimum cumulative distance obtained by optimally aligning the first i time points of the personalized canonical pattern with the first j time points of the real-time synchronized data. This matrix is ​​calculated using dynamic programming, where each element records the minimum cumulative cost to reach that point. Finally, a backtracking and normalization path is performed. After the cumulative distance matrix is ​​fully filled, the system starts from the bottom right element and backtracks along the path with the minimum cumulative distance towards the top left element. This backtracking path is called the normalization path. This path satisfies constraints such as monotonicity, continuity, and boundary conditions. It precisely describes how to distort real-time data in time to achieve the best match with the normalization pattern, thereby eliminating nonlinear differences on the time axis. Next, based on the normalized path, signal deviation calculations are performed on the personalized standard pattern and real-time synchronized data to obtain the original deviation time series. It should be understood that while the normalized path solves the nonlinear matching problem on the time axis, it does not directly provide the degree of deviation. In the technical solution of this application, through signal deviation calculation, the system can transform the abstract alignment relationship into specific, quantifiable physiological signal differences, thereby clearly revealing that during exercise, a specific physiological indicator deviated from the standard pattern at a certain point in time, and the degree of deviation. This quantified original deviation time series is the direct basis for subsequent diagnosis of the cause of the deviation, location of the focal area, and assessment of its severity.

[0022] During this process, the system traverses every index pair along the regularization path. For each alignment point in the regularization path, the system extracts the corresponding multidimensional physiological signal vectors from the personalized standardization pattern and the real-time synchronized data. Both vectors contain multiple physiological indicators of the user at a specific moment, such as respiratory flow, chest movement amplitude, and abdominal movement amplitude. Subsequently, the system performs a dimension-by-dimensional difference calculation on these two aligned physiological signal vectors to quantify the degree of deviation of each physiological indicator. By repeating the above difference calculation for all alignment points along the regularization path, the system obtains a series of multidimensional deviation vectors. Connecting these multidimensional deviation vectors in the order of the regularization path forms the final raw deviation time series. This deviation time series records in detail the quantitative deviation of all physiological indicators relative to the personalized standardization pattern at each alignment point throughout the entire exercise process, thus providing direct and accurate input data for subsequent in-depth analysis. Furthermore, the original deviation time series is standardized and shaped to obtain the deviation time series. It should be understood that different physiological signal channels (e.g., respiratory flow deviation and chest-abdominal coordination deviation) may have different dimensions, numerical ranges, and statistical characteristics. Directly inputting them into the subsequent recurrent neural network model may cause the model to overemphasize signals with larger values ​​while ignoring smaller but equally important signals, thus affecting the fairness and accuracy of model learning. In addition, the length of the original deviation time series may vary with the user's training duration, while most deep learning models, especially recurrent neural networks, typically require fixed-length input sequences for efficient and stable training. Therefore, standardizing the deviation signal can eliminate the dimensional differences between different physiological signals, making all deviation signals numerically comparable. Windowing shaping can transform variable-length continuous time series data into fixed-length, structured input samples, thereby better adapting to the processing requirements of the recurrent neural network model and ensuring the effectiveness and robustness of subsequent shared feature encoding.

[0023] In this process, firstly, each physiological signal channel in the original deviation time series is independently standardized to obtain a standardized deviation time series. After obtaining the multidimensional original deviation time series, it contains the deviation values ​​of multiple physiological signal channels at different time points. To address the issue of inconsistencies in the dimensions and numerical ranges of different physiological signals, the system needs to independently standardize each physiological signal channel in the original deviation time series. This standardization typically employs the Z-score standardization method, aiming to transform the deviation data of each channel into a distribution with a mean of 0 and a standard deviation of 1, thereby eliminating dimensional differences and making the deviation data of all channels numerically comparable. For the original deviation value of a physiological signal channel c at time point t in the original deviation time series... Its standardized deviation value The calculation formula is as follows: , in, This represents the mean of the physiological signal channel c throughout the entire original bias time series. This represents the standard deviation of the physiological signal channel c throughout the entire original deviation time series. This independent channel-by-channel standardization ensures that the deviation information of each physiological signal channel is treated fairly in the subsequent model, preventing certain signals with larger values ​​from dominating the model's learning process. Furthermore, the standardized deviation time series is windowed to obtain the final deviation time series. It should be understood that although the standardized deviation time series obtained in the previous step eliminates dimensional differences, it is still a continuous time series with potentially variable lengths. To better serve as input to the shared feature encoding module of the recurrent neural network model, it is further windowed to divide the continuous standardized deviation time series data into a series of predefined-length time windows. Each time window serves as an independent input sample, containing complete information about the multidimensional standardized deviation signal within a specific time period. For example, the system can set a fixed-length time window of L and slide it over the standardized deviation time series with a certain step size (e.g., a fraction of the window length or a non-overlapping step size), and successively extract deviation time series of length L. These are then input into the subsequent recurrent neural network in a uniform format, enabling the model to effectively capture local time series patterns and features within each time window, thereby extracting stable and representative lung state deviation coding feature sequences.

[0024] Specifically, in S3, the deviation time series is encoded using a shared feature model based on a recurrent neural network (RNN) to obtain a lung state deviation encoded feature sequence. It should be understood that although the deviation time series has been standardized and shaped, it still represents a relatively low-level raw signal representation. These time series data contain rich temporal dependencies and potential patterns; however, traditional linear analysis methods or simple feature extraction techniques struggle to capture the complex nonlinear relationships and long-range dependencies. Recurrent neural network (RNN) models, especially their variants such as Long Short-Term Memory (LSTM) networks, have an inherent advantage in processing sequential data, effectively learning and memorizing long-term dependencies in time series, thereby transforming the raw, potentially noisy deviation signal into a more abstract, high-dimensional, and semantically rich lung state deviation encoded feature representation. This encoding process is crucial for achieving accurate deviation attribution and actionable feedback. It provides high-quality, high-information-density input for subsequent dual-channel joint analysis and multi-task decoding, enabling the system to identify deeper physiological causal relationships rather than merely superficial phenomena.

[0025] In practice, the deviation time sequence is input into a shared bidirectional long short-term memory (BiLSTM) network to obtain the lung state deviation encoding feature sequence. The BiLSTM network is a special recurrent neural network architecture that captures more comprehensive contextual information by processing the input sequence in two directions (forward and backward). For each input time point (or time window), BiLSTM considers not only its previous historical information but also its future information, thus enabling more accurate encoding of the features at the current moment. Specifically, when the deviation time sequence is input into BiLSTM, the network performs information transmission and feature extraction along the forward (from start to end) and backward (from end to start) directions of the time sequence, respectively. The hidden states in these two directions are combined at each time step to form the output features for that time step. This bidirectional processing mechanism allows BiLSTM to effectively understand that a deviation at a certain moment in a breathing pattern may be related to previous or subsequent physiological events. For example, an insufficient inhalation may affect subsequent exhalation, and vice versa; BiLSTM can capture these dependencies simultaneously. Finally, after processing by the BiLSTM network, for each input deviation time-series sample, a corresponding lung state deviation encoding feature sequence is output. This sequence consists of the hidden states (or combinations thereof) generated by BiLSTM at each time step. It is no longer the original deviation value, but a highly abstract and compressed feature representation. This feature sequence captures the key patterns, temporal dependencies, and potential physiological meanings in the original deviation time series, providing rich contextual information for subsequent diagnostic tasks. It is worth noting that "sharing" means that all input deviation time-series samples are encoded using the same BiLSTM network structure and parameters, ensuring consistency and generalization ability in feature extraction.

[0026] Specifically, in step S4, a dual-channel joint analysis is performed on the lung state deviation encoded feature sequence to obtain a task context vector. It should be understood that although the lung state deviation encoded feature sequence contains rich temporal and feature information, it is still a variable-length sequence. For subsequent multi-task decoding (such as cause identification, focus localization, and severity assessment), a fixed-length vector representation that comprehensively summarizes the core features of the entire respiratory deviation event is typically required. Simple global averaging or max pooling may lose crucial local information or important feature dimensions. Therefore, in the technical solution of this application, dual-channel joint analysis is used to intelligently identify and aggregate the most important time points and the most diagnostically valuable physiological feature information from the complex encoded feature sequence, thereby generating a task context vector that includes both temporal and feature context.

[0027] In practice, firstly, temporal channel attention is applied to the lung state deviation encoding feature sequence to obtain temporal attention weights. It should be understood that the lung state deviation encoding feature sequence is a multidimensional, temporal feature representation encompassing the user's entire breathing exercise process. However, not all information at all time points is equally important for identifying specific breathing deviations. For example, when diaphragmatic dysfunction, data from the early inspiratory phase may be more diagnostically valuable than data from the late expiratory phase; and when assessing chest-abdominal coordination, information from certain key inflection points may be particularly important. Simply treating all time steps equally without distinguishing the importance of these time points may introduce noise, dilute key information, and reduce the accuracy of subsequent deviation attribution. Therefore, applying temporal channel attention aims to enable the model to intelligently focus on the key time points in the sequence most relevant to the current diagnostic task, dynamically assigning an importance weight to each time step to more effectively extract meaningful temporal context information and improve the overall accuracy of the analysis.

[0028] In this process, the system analyzes each time step in the lung state deviation encoding feature sequence and assesses its importance to the overall deviation diagnosis task. Specifically, this process is typically implemented through a simple neural network layer. Each feature vector is fed into a feedforward network (e.g., a fully connected layer with an activation function) to generate a raw importance score. This feedforward network learns how to determine its contribution to the entire sequence based on the content of the lung state deviation encoding features. Finally, to convert these raw importance scores into interpretable and probabilistically meaningful weights, the system typically applies a softmax function for normalization. Through the softmax function, the attention weights at all time steps are compressed to the range (0,1], and their sum is 1. In this way, the system can dynamically focus on critical moments when the deviation occurs or evolves, avoiding interference from irrelevant or secondary time point information. Next, feature channel attention is applied to the lung state deviation encoding feature sequence to obtain feature attention weights. It should be understood that not all physiological feature dimensions are equally important when diagnosing different types of respiratory deviations. For example, when assessing diaphragmatic insufficiency, feature dimensions related to abdominal movement may be more diagnostically valuable than those related to chest movement; while when identifying limited chest wall movement, features related to chest expansion may be more critical. Treating all physiological feature dimensions equally may introduce feature noise irrelevant to the current task, diluting crucial diagnostic information and affecting the model's accuracy in identifying the cause of the deviation. Therefore, applying feature channel attention aims to enable the model to intelligently focus on the physiological feature dimensions in the sequence most relevant to the current diagnostic task, dynamically assigning an importance weight to each feature channel, thereby more effectively extracting the most diagnostically valuable feature context information and improving the overall accuracy of the analysis.

[0029] In this process, the system evaluates each feature channel (e.g., respiratory flow deviation feature, chest movement deviation feature, etc.) in the lung state deviation encoded feature sequence, and then extracts a representation from this time series that can represent the overall information of that feature channel across all time steps. This is typically achieved by performing a global pooling operation in the time dimension, such as global average pooling, which averages the values ​​of each feature channel across all time steps to obtain a feature description vector. Next, the system applies an attention mechanism to this feature description vector to calculate the importance score of each feature channel. This is usually implemented through a small neural network structure that learns how to determine the contribution of each feature dimension to the overall task based on the content of hˉhˉ. Specifically, the feature description vector is input into a feedforward network (e.g., a fully connected layer with an activation function) to generate a raw importance score that quantifies the importance of the feature channel. Finally, to convert these raw importance scores into interpretable and probabilistically meaningful weights, the system typically applies a Softmax function for normalization. Through the Softmax function, the attention weights of all feature channels are compressed to the range (0,1], and their sum is 1. Furthermore, based on temporal attention weights and feature attention weights, the lung state deviation encoding feature sequence is subjected to multi-channel collaborative aggregation to obtain the task context vector. It should be understood that although temporal attention weights and feature attention weights quantify important time points and important physiological feature dimensions in the sequence, respectively, they are decoupled and static. The complexity of respiratory physiology lies in the fact that the importance of physiological features often exhibits significant time-varying characteristics; that is, at different time points, different physiological features may play a dominant role or exhibit specific synergistic relationships. For example, the importance of abdominal movement features may peak during the initial stage of inspiration, while it becomes relatively secondary during the expiratory phase. Simply concatenating or processing temporal and feature context information independently will fail to effectively capture this dynamic and interactive physiological relationship, leading to the loss of key information and thus limiting the accuracy of deviation attribution and the granularity of feedback. Therefore, in the technical solution of this application, multi-channel collaborative aggregation is used to explicitly model the dynamic interaction between time and feature dimensions, and generate a task context vector that deeply integrates key spatiotemporal information, thereby providing a more comprehensive, accurate and physiologically meaningful abstract representation of the entire respiratory deviation event, laying a solid foundation for subsequent accurate cause identification, focus localization and severity assessment.

[0030] In this process, firstly, temporal-feature co-attention instance gating is performed on the temporal attention weights and feature attention weights to generate a co-attention instance gating matrix. It should be understood that in this preferred scheme, the interaction between temporal attention and feature attention needs to be explicitly modeled, rather than left independent. Through this interaction, a gating mechanism that dynamically reflects the joint importance of different features at different time points can be constructed. Specifically, the temporal attention weight vector and the feature attention weight vector are outer-producted to initially establish the correlation matrix between them. Then, a learnable parameter matrix of the same dimension is introduced and multiplied element-wise with the outer-product result to allow the model to adaptively learn and adjust the complex interaction patterns between different time-feature combinations. Finally, the final temporal-feature co-attention instance gating matrix is ​​generated using the tanh activation function. Specifically, this process is expressed by the following formula: , in, Representative example gated matrix; It is a learnable parameter matrix with the same dimension as the gating matrix; It represents the Hadamardi (or Hadama) stack; It is the time attention weight; Indicates the outer product; It refers to feature attention weights. In other words, this step creates a dynamic two-dimensional filter or template. Each element in its matrix They all precisely quantified at time. Features Joint contribution to the final diagnostic task. The introduction of this technology endows the process with the ability to learn complex physiological interactions, such as the antagonistic or synergistic relationships between muscle groups during respiration. This generates a high-information-content gating matrix, which itself precisely captures the core information of key features at critical moments, providing a foundation for subsequent feature modulation. Next, the lung state deviation coding feature sequence is gated and modulated based on the coherent instance gating matrix to obtain the lung state modulated feature sequence. This step aims to apply the dynamically generated spatiotemporal joint importance information to the original lung state deviation coding feature sequence to achieve fine-grained filtering and focusing of feature information. Specifically, the system uses the coherent instance gating matrix... The product is applied to the original encoded feature sequence matrix H through element-wise multiplication (Hadamard product). Specifically, this process is expressed by the formula: , in, H represents the modulated, new feature sequence matrix, where H is the original encoded feature sequence matrix. Similarly, it represents the Hadamardi (or Hadamardi). This is the coherence instance gating matrix generated in the previous step. This step can be based on... The size of the elements in the middle selectively preserves or enhances key feature information that has a high joint contribution at critical moments, while effectively suppressing information that is less important in time and space or belongs to background noise. For example, interference signals generated by slight body swaying during the stable exhalation phase, if they are in If the corresponding weight is low, it will be effectively suppressed. Through gating modulation, a lung state modulation feature sequence is generated, which has a much higher information density and signal-to-noise ratio than the original sequence, and is more focused on the core bias patterns related to the diagnostic task; Furthermore, context aggregation is performed on the modulated feature sequence of the lung state to obtain a task context vector. It should be understood that although the modulated feature sequence has high information purity, it is still variable-length time-series data and needs to be aggregated into a fixed-length vector representation to adapt to the input requirements of the subsequent task decoding head. Specifically, the modulated feature sequence... Summation or global average pooling is performed along the time dimension to compress and distill the sequence information into a final context vector. Specifically, this process is expressed by the formula: , in, It is the final generated task context vector. It is the modulation feature sequence matrix at time 10:00. The feature vector is generated by this aggregation operation, which synthesizes the essential information after spatiotemporal joint attention, producing a single vector that comprehensively and accurately summarizes the core features of the entire respiratory deviation event. This task context vector is no longer a simple concatenation of two independent summaries, but a deep fusion representation that inherently contains dynamic interaction information of time and features, providing higher-quality and more relevant input for downstream tasks.

[0031] Through the aforementioned mechanism, the accuracy, precision, and interpretability of the deviation attribution and feedback generation stages in the intelligent feedback system for lung function training are significantly improved. Specifically, by introducing time-feature collaborative instance gating, a task context vector with extremely high information density and deep integration of key spatiotemporal information is generated. This enables the system to more accurately distinguish similar deviation patterns with different underlying causes, such as precisely identifying whether the deviation stems from insufficient diaphragmatic strength or simply an error in chest-abdominal coordination technique. Furthermore, the collaborative instance gating matrix G itself has high visualization potential. By analyzing its weight distribution, the timing of the model's judgments and the physiological characteristics upon which they are based can be clearly revealed when diagnosing specific problems. This transforms the model's decision-making process from a black box into a verifiable analytical framework, greatly enhancing the system's credibility.

[0032] Specifically, in step S5, the task context vector is multi-task decoded to obtain the cause probability, focal region probability, and severity score. It should be understood that the task context vector itself is still an abstract numerical representation; it has not yet been directly converted into a user-understandable, diagnostically significant conclusion. To provide users with accurate and targeted feedback, the technical solution of this application performs multi-task decoding on the task context vector. Through multi-task decoding, the system can extract diagnostic information from the same high-quality context vector in parallel, thereby laying the final foundation for constructing comprehensive, personalized, and clinically guiding feedback, significantly improving the practicality and effectiveness of intelligent feedback.

[0033] In practice, the first step is to input the task context vector into a cause identification head to obtain the cause probability. This step aims to identify the root causes of deviations in the user's lung function training. During this process, the system inputs the task context vector into a specially trained cause identification head. This head typically consists of one or more fully connected layers (i.e., a feedforward neural network), which maps the input context vector to a predefined cause category space. For example, potential causes might include insufficient diaphragmatic strength, incorrect breathing techniques, or limited chest wall movement. The last layer of the network typically applies a Softmax activation function, outputting a probability distribution vector where each element represents the likelihood that the currently detected deviation belongs to a specific cause. A higher probability value indicates a greater likelihood of that cause. Next, the task context vector is input into the focus localization head to obtain the focus region probability. This step aims to accurately locate the physiological regions or movement segments primarily affected by the user's lung function training deviations. During this process, the system inputs the task context vector into a specially trained focus localization head. Similar to the cause identification head, this localization head typically consists of one or more fully connected layers. It learns to decode the context vector into probability distributions for different physiological focus regions. These focus regions may encompass chest movements, abdominal movements, the airway, diaphragmatic activity, etc. The final layer of the network can use a Softmax or Sigmoid activation function, outputting a probability vector indicating the likelihood that each focus region is affected by the deviation. For example, if the deviation is mainly reflected in insufficient abdominal movement, the probability of the abdominal movement focus region will be higher. Next, the task context vector is input into the severity assessment head to obtain a severity score. This step quantifies the severity of the current lung function training deviation. During this process, the system inputs the task context vector into the severity assessment head, which also consists of one or more fully connected layers. Its goal is to map the context vector to a numerical value representing the severity level. The severity score can be a discrete level (e.g., mild, moderate, severe), in which case the last layer typically uses a Softmax activation function to classify these levels; or it can be a continuous numerical value, in which case the last layer is typically a linear output layer (without an activation function or using Sigmoid / ReLU to restrict the output to a specific range). This score can intuitively reflect the degree of the user's deviation and is of significant reference value in guiding users to adjust training intensity or seek further medical assistance.

[0034] Specifically, in step S6, actionable feedback is generated based on causal probability, focal area probability, and severity score. It should be understood that simply providing causal probability, focal area probability, and severity score will not allow users to directly grasp how these abstract diagnostic results translate into practical actions. Traditional feedback mechanisms often lack specificity and cannot provide personalized corrective strategies based on the user's specific problems, making it difficult for users to effectively improve their training results. Therefore, in the technical solution of this application, this highly diagnostic information is integrated and transformed into actionable feedback that is easy to understand, directly addresses the core of the problem, and clearly guides users on how to operate, thereby fundamentally improving the efficiency and effectiveness of the user's rehabilitation training and truly achieving the goal of intelligent assisted training.

[0035] In practice, the system receives three key diagnostic results from the multi-task decoding in the previous step: causal probability, focal area probability, and severity score. The system then performs a comprehensive judgment and prioritization based on these results. It identifies the one or more most likely root causes from the causal probability distribution (e.g., selecting the cause with the highest probability). Similarly, it identifies the most affected physiological area or motor segment from the focal area probability distribution. The severity score is used to assess the urgency of the current problem and the intensity of the feedback tone. For example, if the severity score is high, the feedback tone may emphasize immediate correction and potential risks. Next, the system enters the feedback generation phase. This phase typically relies on a pre-defined feedback rule base or a template-based natural language generation module, combined with a deep learning model. The feedback rule base stores standardized feedback templates and corrective suggestions for different combinations of causes, focus areas, and severity. For example: Diagnostic information integration: The system first integrates the most likely cause, the most prominent focal area, and the severity level. For example, if the most likely cause is "insufficient diaphragmatic strength," the focal area is "abdominal movement," and the severity is "moderate." Selecting a Feedback Template: Based on the integrated diagnostic information, the system selects the most suitable template from a pre-set feedback template library. These templates are designed to provide specific guidance directly addressing specific physiological problems. Personalized content filling and refinement: The selected template will be filled and refined with content based on the user's personalized norm pattern and specific deviation details. For example, if the feedback is about "strengthening the diaphragm," it may be further combined with the user's specific deviation value in inhalation depth to give a quantitative suggestion such as "Please try to expand your abdomen outward by X centimeters during inhalation." If the user's norm pattern requires deep breathing, this will also be emphasized in the feedback; Actionable guidance: Feedback focuses on providing specific, actionable corrective actions or techniques. It clearly tells users "how to do it," rather than simply pointing out "what's wrong." This might include guiding users to adjust their breathing rhythm, change the point of force exertion, correct body posture, or provide specific auxiliary exercises. For example, for "insufficient diaphragmatic strength" and "insufficient abdominal movement," feedback might suggest: "When inhaling, actively feel your abdomen expand and try to use your hands to help push your abdomen outward, ensuring your chest remains relatively stable." Tone and modality adjustment: The tone of feedback will be adjusted accordingly based on the severity score. For minor deviations, the tone may be more encouraging and suggestive; for serious deviations, the tone may emphasize risk and timely correction. Ultimately, the system will output one or more clear, specific, and instructive actionable feedback statements directly to the user.

[0036] In summary, the intelligent feedback method for lung function training according to the embodiments of this application is explained. It first acquires real-time multidimensional synchronous data and performs deviation assessment based on an individualized standardized pattern to generate a refined deviation time series. Subsequently, a recurrent neural network is used to encode shared features of these time series data to capture dynamic changes in lung status. Then, through dual-channel joint analysis, a task context vector containing rich spatiotemporal information is extracted from the encoded feature sequence, effectively overcoming the limitations of information dissociation and static fusion in existing technologies. Finally, a multi-task decoder is used to identify the root cause, specific focal area, and severity of the deviation from this context vector, providing a solid basis for subsequent feedback. In this way, the accuracy and precision of deviation attribution are significantly improved.

[0037] Furthermore, an intelligent feedback system for lung function training is also provided.

[0038] Figure 3 This is a block diagram of an intelligent feedback system for lung function training according to an embodiment of this application. Figure 3As shown, the intelligent feedback system 300 for lung function training according to an embodiment of this application includes: a data acquisition module 310 for acquiring real-time synchronized data; a deviation assessment module 320 for performing deviation assessment on the real-time synchronized data based on a personalized standard mode to obtain a deviation time sequence; a shared feature encoding module 330 for performing shared feature encoding on the deviation time sequence based on a recurrent neural network model to obtain a lung state deviation encoding feature sequence; a dual-channel joint analysis module 340 for performing dual-channel joint analysis on the lung state deviation encoding feature sequence to obtain a task context vector; a multi-task decoding module 350 for performing multi-task decoding on the task context vector to obtain causal probability, focal region probability, and severity score; and an operability feedback module 360 ​​for generating operability feedback opinions based on causal probability, focal region probability, and severity score.

[0039] As described above, the intelligent feedback system 300 for lung function training according to the embodiments of this application can be implemented in various wireless terminals, such as servers with intelligent feedback algorithms for lung function training. In one possible implementation, the intelligent feedback system 300 for lung function training according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or a hardware module. For example, the intelligent feedback system 300 for lung function training can be a software module in the operating system of the wireless terminal, or it can be an application developed for the wireless terminal; of course, the intelligent feedback system 300 for lung function training can also be one of many hardware modules of the wireless terminal.

[0040] Alternatively, in another example, the intelligent feedback system 300 for lung function training and the wireless terminal can also be separate devices, and the intelligent feedback system 300 for lung function training can be connected to the wireless terminal via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0041] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method of intelligent feedback for lung function exercise, characterized in that, The method comprises the following steps: obtaining real-time synchronization data; performing deviation evaluation on the real-time synchronization data based on a personalized standard mode to obtain a deviation time sequence; performing shared feature coding on the deviation time sequence based on a recurrent neural network model to obtain a lung state deviation coding feature sequence; performing double-channel joint analysis on the lung state deviation coding feature sequence to obtain a task context vector; performing multi-task decoding on the task context vector to obtain a reason probability, a focus area probability and a severity score; generating operability feedback based on the reason probability, the focus area probability and the severity score.

2. The intelligent feedback method for lung function exercise according to claim 1, wherein, The method for performing deviation evaluation on the real-time synchronization data based on the personalized standard mode to obtain the deviation time sequence comprises the following steps: performing time sequence alignment on the personalized standard mode and the real-time synchronization data based on multi-dimensional dynamic time warping to obtain a warping path; performing signal deviation calculation on the personalized standard mode and the real-time synchronization data based on the warping path to obtain an original deviation time sequence; performing deviation signal standardization and shaping on the original deviation time sequence to obtain the deviation time sequence.

3. The intelligent feedback method for lung function exercise of claim 2, wherein, The method for performing deviation signal standardization and shaping on the original deviation time sequence to obtain the deviation time sequence comprises the following steps: performing independent standardization processing on each physiological signal channel of the original deviation time sequence to obtain a standardized deviation time sequence; performing window shaping on the standardized deviation time sequence to obtain the deviation time sequence.

4. The intelligent feedback method for lung function exercise of claim 1, wherein, The method for performing shared feature coding on the deviation time sequence based on the recurrent neural network model to obtain the lung state deviation coding feature sequence comprises the following step:

5. The intelligent feedback method for lung function exercise of claim 1, wherein, inputting the deviation time sequence into a shared bidirectional long short-term memory network to obtain the lung state deviation coding feature sequence. The method for performing double-channel joint analysis on the lung state deviation coding feature sequence to obtain the task context vector comprises the following steps: applying time channel attention to the lung state deviation coding feature sequence to obtain time attention weights; applying feature channel attention to the lung state deviation coding feature sequence to obtain feature attention weights; 6. The intelligent feedback method for lung function exercise of claim 5, wherein, performing multi-channel collaborative aggregation on the lung state deviation coding feature sequence based on the time attention weights and the feature attention weights to obtain the task context vector. The method for performing multi-channel collaborative aggregation on the lung state deviation coding feature sequence based on the time attention weights and the feature attention weights to obtain the task context vector comprises the following steps: generating time-feature collaborative example gating based on the time attention weights and the feature attention weights to obtain a collaborative example gating matrix; performing gated feature modulation on the lung state deviation coding feature sequence based on the collaborative example gating matrix to obtain a lung state modulation feature sequence; 7. The intelligent feedback method for lung function exercise of claim 6, wherein, performing context aggregation on the lung state modulation feature sequence to obtain the task context vector. where, represents a gating matrix for the collaborative example; is a learnable parameter matrix of the same dimension as the gating matrix; denotes a Hadamard product; is a temporal attention weight; denotes an outer product; is a feature attention weight.

8. The intelligent feedback method for lung function exercise of claim 1, wherein, The method for generating time-feature collaborative example gating based on the time attention weights and the feature attention weights to obtain the collaborative example gating matrix comprises the following step: generating time-feature collaborative example gating based on the time attention weights and the feature attention weights according to the following formula: The method for performing multi-task decoding on the task context vector to obtain the reason probability, the focus area probability and the severity score comprises the following steps: inputting the task context vector into a reason recognition head to obtain the reason probability; inputting the task context vector into a focus positioning head to obtain the focus area probability; The task context vector is input to the severity assessment head to obtain a severity score.

9. A lung function exercise intelligent feedback system, characterized in that, The method comprises the following steps: a data acquisition module configured to acquire real-time synchronization data; a bias evaluation module configured to perform bias evaluation on the real-time synchronization data based on a personalized normative pattern to obtain a bias time sequence; a shared feature encoding module configured to perform shared feature encoding on the bias time sequence based on a recurrent neural network model to obtain a lung state bias encoding feature sequence; a dual-channel joint analysis module configured to perform dual-channel joint analysis on the lung state bias encoding feature sequence to obtain a task context vector; a multi-task decoding module configured to perform multi-task decoding on the task context vector to obtain a reason probability, a focus area probability, and a severity score; an operability feedback module configured to generate operability feedback opinions based on the reason probability, the focus area probability, and the severity score.

Images