Method and device for identifying geriatric occult emergency, and storage medium

Abstract

The application discloses a method and device for identifying senile occult emergency and a storage medium, and relates to the technical field of medical information processing. The application generates and dynamically updates individualized dynamic baseline through multi-scene multi-source data acquisition and preprocessing, construction of senile exclusive physiological decline indicators, and improved LSTM-Attention neural network combined with the introduction of a decline weight factor. Then, the application identifies occult abnormalities through a multi-level occult signal enhancement mechanism, and outputs risk and suspected emergency types by combining GBDT-GNN double model fusion with clinical rule constraints. In addition, the application matches multi-terminal differentiated early warning and graded response, and finally optimizes the model and rule base through clinical feedback closed loop, effectively adapts to the physiological decline characteristics of the elderly, realizes the precise early identification of senile occult emergency, improves the clinical interpretability and medical acceptance of the identification results, realizes the efficient linkage of early warning and intervention, and reduces the incidence risk and medical burden of senile occult emergency.

Description

Technical Field

[0001] This application relates to the field of medical information processing technology, specifically to a method, device, and storage medium for identifying latent acute illnesses in the elderly. Background Technology

[0002] Elderly individuals are highly susceptible to insidious acute illnesses such as infections, cardiovascular events, hypoglycemia, and pulmonary embolism due to natural physiological decline, coexisting underlying diseases, and weakened immunity. These insidious acute illnesses are characterized by atypical symptoms, insidious onset, and rapid progression. Initially, they often present with nonspecific symptoms such as lethargy, decreased appetite, and reduced activity, lacking the typical signs of traditional acute illnesses, significantly increasing the risk of death and the medical burden on elderly patients.

[0003] In the field of medical monitoring, accurate prediction and timely early warning of physiological signals are of great significance for early disease intervention and patient safety. However, traditional technologies still have significant shortcomings in data collection, predictive modeling, and early warning response. Traditional elderly health monitoring systems often focus on monitoring single or a few physiological indicators, neglecting individual differences and comprehensive assessment of overall health status, resulting in limited sensitivity and specificity of the early warning system. Furthermore, traditional models lack clinical evidence, leading to low acceptance by medical staff and difficulties in clinical implementation. In addition, traditional intelligent monitoring systems cannot quickly provide emergency solutions when they judge or predict that a patient may experience a health condition, hindering rapid rescue efforts and failing to adequately protect the health status of elderly patients. Summary of the Invention

[0004] The purpose of this application is to provide a method, device, and storage medium for identifying latent acute diseases in the elderly, in order to solve the problems of low accuracy and lack of clinical relevance in traditional monitoring of latent acute diseases in the elderly.

[0005] To achieve the above objectives, the first aspect of this application provides a method for identifying latent acute illnesses in the elderly, comprising: Multi-source data of elderly users in multiple scenarios were collected, the multi-source data were preprocessed, and physiological decline indicators of the elderly users were constructed. The basic functional features are fused with the physiological decline indicators and input into the improved LSTM-Attention neural network to train and generate an individualized dynamic baseline. The baseline is dynamically updated based on the decline curve. The improved LSTM-Attention neural network introduces a decline weight factor into the attention mechanism. Based on the individualized dynamic baseline, the hidden abnormalities of the elderly users are identified through a multi-level hidden signal enhancement mechanism. The GBDT-GNN dual-model fusion strategy is adopted, and the node edge weights of the GNN model are constrained by clinical rules to output the risk level and suspected acute condition type. The system outputs warning information matching the risk level and suspected acute illness type to multiple terminal devices, and triggers a graded response based on the risk level. The warning information includes clinical rule basis and interpretable warning information based on decline impact analysis. Based on the clinical feedback data following the graded response, the node edge weights and associated clinical rule base of the GNN model are updated. A second aspect of this application provides a device for identifying latent acute illnesses in the elderly, comprising: The data acquisition module is used to collect multi-source data of elderly users in multiple scenarios, preprocess the multi-source data, and construct physiological decline indicators of the elderly users. The training module is used to fuse basic functional features with the physiological decline indicators and input them into the improved LSTM-Attention neural network to train and generate an individualized dynamic baseline and dynamically update the baseline based on the decline curve. The improved LSTM-Attention neural network introduces a decline weight factor in the attention mechanism. The output module is used to identify the hidden abnormalities of the elderly user based on the individualized dynamic baseline through a multi-level hidden signal enhancement mechanism, and to output the risk level and suspected acute condition type by adopting the GBDT-GNN dual-model fusion strategy and combining clinical rules to constrain the node edge weights of the GNN model. The early warning module is used to output early warning information matching the risk level and suspected acute illness type to multiple terminal devices respectively, and trigger a graded response according to the risk level. The early warning information includes clinical rule basis and interpretable early warning information based on decline impact analysis. An update module is used to update the node edge weights and associated clinical rule base of the GNN model based on the clinical feedback data after the graded response.

[0006] A third aspect of this application provides a computer-readable storage medium storing a program that can be loaded by a processor and executed as described above regarding the method for identifying latent acute conditions in the elderly.

[0007] The beneficial effects of this application are: This application utilizes multi-scenario, multi-source data collection and preprocessing, constructs age-specific physiological decline indicators, combines an improved LSTM-Attention neural network with a decline weight factor to generate and dynamically update an individualized dynamic baseline, identifies hidden abnormalities through a multi-level hidden signal enhancement mechanism, and outputs risk and suspected acute disease types through GBDT-GNN dual-model fusion and clinical rule constraints. It also incorporates multi-terminal differentiated early warning and graded response, and finally optimizes the model and rule base through clinical feedback closed-loop optimization. This effectively adapts to the characteristics of age-related physiological decline, achieving accurate early identification of hidden acute diseases in the elderly, improving the clinical interpretability and healthcare acceptance of the identification results, reducing missed diagnoses and misjudgments, and simultaneously achieving efficient linkage between early warning and intervention. This reduces the incidence risk and medical burden of hidden acute diseases in the elderly and is suitable for various elderly health management scenarios, including home, community, and hospital settings.

[0008] Other features and advantages of this application will be described in detail in the following detailed description section. Attached Figure Description

[0009] Figure 1 This is a schematic diagram illustrating an application scenario of a method for identifying latent acute illnesses in the elderly, as provided in this application embodiment. Figure 2 This is a flowchart illustrating a method for identifying latent acute illnesses in the elderly, as provided in an embodiment of this application. Figure 3 This is a schematic diagram of the structure of a device for identifying latent acute diseases in the elderly, as provided in an embodiment of this application. Detailed Implementation

[0010] 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, and 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.

[0011] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "a plurality of" means two or more, unless otherwise explicitly specified. Details are set forth in the following description for illustrative purposes. It should be understood that those skilled in the art will recognize that this application can be implemented without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but rather to be consistent with the broadest scope of the principles and features disclosed herein.

[0012] The method for identifying latent acute illnesses in the elderly in this application embodiment is applied to a device for identifying latent acute illnesses in the elderly, and the device for identifying latent acute illnesses in the elderly is installed in an electronic device. For example... Figure 1 As shown, Figure 1 This is a schematic diagram illustrating an application scenario of the method for identifying latent acute illnesses in the elderly in this application embodiment. The application scenario of the method for identifying latent acute illnesses in the elderly in this application embodiment includes an electronic device 110 for the method of identifying latent acute illnesses in the elderly. The electronic device 110 integrates a device for identifying latent acute illnesses in the elderly to run a computer-readable storage medium corresponding to the method for identifying latent acute illnesses in the elderly, so as to execute the steps of the method for identifying latent acute illnesses in the elderly.

[0013] Understandable Figure 1 The electronic devices in the application scenario of the method for identifying latent acute diseases in the elderly, or the devices contained in the electronic devices, do not constitute a limitation on the embodiments of this application. That is, the number or type of devices in the application scenario of the method for identifying latent acute diseases in the elderly, or the number or type of devices contained in each device, do not affect the overall implementation of the technical solution in the embodiments of this application, and can all be considered as equivalent substitutions or derivatives of the technical solutions claimed in the embodiments of this application.

[0014] In this application embodiment, the electronic device 110 can be an independent device, or a device network or device cluster composed of devices. For example, the electronic device 110 described in this application embodiment includes, but is not limited to, a computer, a network host, a single network device, a set of multiple network devices, or a cloud device composed of multiple devices. Among them, the cloud device is composed of a large number of computers or network devices based on cloud computing.

[0015] Those skilled in the art will understand that Figure 1The application scenarios shown are merely one application scenario corresponding to the technical solution of this application, and do not constitute a limitation on the application scenarios of the technical solution of this application. Other application scenarios may include more than one application scenario. Figure 1 The number of more or fewer electronic devices shown, or the network connections of electronic devices, for example Figure 1 Only one electronic device is shown in the image. It is understood that the scenario of this method for identifying latent acute diseases in the elderly may also include one or more other electronic devices, which are not specifically limited here. The electronic device 110 may also include a memory and a processor. The memory is used to store information related to the method for identifying latent acute diseases in the elderly.

[0016] Furthermore, in the application scenario of the method for identifying latent acute illnesses in the elderly according to this application embodiment, the electronic device 110 may be equipped with a display device, or the electronic device 110 may not have a display device but may be communicatively connected to an external display device 120. The display device 120 is used to output the results of the method for identifying latent acute illnesses in the elderly executed by the electronic device. The electronic device 110 can access the background database 130. The background database 130 may be the local storage of the electronic device 110 or a cloud database located in the cloud. The background database 130 stores information related to the method for identifying latent acute illnesses in the elderly.

[0017] It should be noted that, Figure 1 The application scenario of the method for identifying latent acute diseases in the elderly shown is merely an example. The application scenario of the method for identifying latent acute diseases in the elderly described in this application embodiment is to more clearly illustrate the technical solution of this application embodiment and does not constitute a limitation on the technical solution provided in this application embodiment.

[0018] Based on the application scenarios of the above-mentioned methods for identifying latent acute illnesses in the elderly, embodiments of these methods are proposed. A detailed description is provided below with reference to the accompanying drawings.

[0019] Figure 2 This is a flowchart illustrating a method for identifying latent acute illnesses in the elderly, as provided in an embodiment of this application. Figure 2 As shown, this identification method can be executed by the processor in the above-mentioned electronic device 110, and steps 201-205 are described in detail below.

[0020] Step 201: Collect multi-source data of elderly users in multiple scenarios, preprocess the multi-source data, and construct physiological decline indicators for elderly users.

[0021] The multi-source data in this application's embodiments across multiple scenarios can include, but is not limited to, home millimeter-wave radar data, wearable device data, medical examination and treatment data, and data reported by family members' terminals. For example, physiological data, behavioral data, environmental data, family member-reported data, and medical indicator data of elderly users are collected in multiple scenarios such as home, outings, and medical treatment. The data is then preprocessed, including noise removal, missing value completion, data standardization, and time-series alignment. This achieves full-scenario health data coverage and improves data quality and usability. Quantitative indicators reflecting the gradual decline in cardiopulmonary, muscular, and cognitive functions in the elderly are selected to construct physiological decline indicators for elderly users. These indicators are calculated based on physical examination data and real-time multi-source data and are dynamically corrected periodically. Elderly-specific decline indicators are established to provide core features for subsequent individualized modeling and address the issue that general indicators are unsuitable for the elderly.

[0022] Step 202: Integrate basic functional features with physiological decline indicators and input them into an improved LSTM-Attention neural network to train and generate an individualized dynamic baseline and dynamically update the baseline based on the decline curve.

[0023] Basic functional characteristics refer to routine health features such as blood pressure, heart rate, mental state, and activity level. These basic functional characteristics are fused with physiological decline indicators into a multi-dimensional input vector, which is then fed into an improved LSTM-Attention network. The improved LSTM-Attention (Improved Long Short-Term Memory-Attention) is an enhanced neural network based on a long short-term memory-attention mechanism. Specifically, the improved LSTM-Attention neural network introduces a decline weight factor into the attention mechanism. This decline weight factor is obtained by weighting physiological decline indicators and is used to strengthen the model's focus on decline features. Training the model using a loss function with a decline penalty generates an individualized dynamic baseline that fits the individual. This individualized dynamic baseline is based on the individual's own decline trajectory and represents the normal range for that individual, rather than a universal threshold. This allows for greater focus on age-related decline features, a baseline that more closely reflects the individual's true state, and the identification of baseline-drifting hidden anomalies, significantly reducing the missed diagnosis rate.

[0024] Step 203: Based on the individualized dynamic baseline, the hidden abnormalities of elderly users are identified through a multi-level hidden signal enhancement mechanism. The GBDT-GNN dual-model fusion strategy is adopted, and the node edge weights of the GNN model are constrained by clinical rules to output the risk level and suspected acute condition type.

[0025] Multi-level hidden signal enhancement mechanisms can include time-series trend analysis, multi-indicator correlation mining, and cross-scenario cross-validation. Using an individualized dynamic baseline as a reference, a three-level mechanism—single-indicator time-series detection, multi-indicator clinical rule correlation, and cross-scenario signal verification—identifies true hidden anomalies, accurately capturing early, weak, and atypical hidden anomalies. The GBDT-GNN (Gradient Boosting Decision Tree-Graph Neural Network) dual model is a gradient boosting decision tree-graph neural network. GBDT excels at feature importance analysis and risk grading, and in this scheme, it is used to output multi-level risk levels and core risk contribution features. GNN excels at mining inter-node relationships; in this scheme, non-specific symptoms-specific indicators-emergency types are used as nodes, combined with clinical rule constraints on edge weights, to mine the correlation between features and emergency types. A multi-dimensional feature fusion model, incorporating decay indicators, baseline bias, and other features, is constructed and input into GBDT to obtain risk levels and key features. Then, clinical rules constrain the GNN node edge weights to mine emergency types for high-risk samples. Clinical rule constraints use prior clinical knowledge to limit the GNN node edge weights. The GNN results are used to adjust the GBDT weights, forming a closed-loop optimization, and finally outputting structured results. The two models have clear division of labor, are more efficient, and clinical constraints improve interpretability and healthcare acceptance.

[0026] Step 204: Output early warning information matching the risk level and suspected acute illness type to multiple terminal devices respectively, and trigger a graded response according to the risk level.

[0027] Multiple terminal devices can include medical staff terminals, family member terminals, and elderly user terminals. Differentiated early warning information matching the risk level is output to each of these terminals, such as pushing information with different levels of detail according to the three target groups: medical staff, family members, and the elderly. The early warning information includes clinical rule basis and interpretable early warning information based on the analysis of the impact of decline. Interpretable early warning information refers to explanations with judgment criteria and the impact of decline, rather than simply alarms. Tiered response can implement intervention measures of different intensities according to high risk, medium risk, and low risk. For example, high risk triggers emergency response, medium risk triggers outpatient follow-up, and low risk triggers home health monitoring. The early warning information in this embodiment is easy to understand, reliable, and executable, enabling the linkage between early warning and intervention, improving emergency response efficiency, and reducing the risk of acute condition deterioration.

[0028] Step 205: Based on the clinical feedback data after the graded response, update the node edge weights and associated clinical rule base of the GNN model.

[0029] Clinical feedback data can include confirmation results from medical staff, family feedback, medical diagnoses, and prognostic information. Feedback information such as clinical diagnoses, corrective opinions, and prognostic data are collected after a tiered response. Then, the edge weights of the GNN model nodes are adjusted based on the feedback, and the clinical rule base is updated and expanded, allowing for continuous iterative optimization of the model and rules. Iterative optimization is a complete cycle of identification → early warning → response → feedback → model optimization. The model continuously evolves with individual decline and updates in clinical knowledge, maintaining high accuracy over the long term, making the system more closely aligned with clinical practice and more readily applicable.

[0030] This application's embodiments utilize multi-scenario, multi-source data acquisition and preprocessing, construction of age-specific physiological decline indicators, and the generation and dynamic updating of individualized dynamic baselines through an improved LSTM-Attention neural network incorporating a decline weight factor. Furthermore, it identifies hidden abnormalities through a multi-level hidden signal enhancement mechanism, and outputs risk and suspected acute disease types using GBDT-GNN dual-model fusion and clinical rule constraints. Combined with multi-terminal differentiated early warning and graded response, and finally optimizing the model and rule base through clinical feedback closed-loop optimization, this approach effectively adapts to the characteristics of age-related physiological decline, achieving accurate early identification of hidden acute diseases in the elderly. This improves the clinical interpretability and healthcare acceptance of the identification results, reduces missed diagnoses and misjudgments, and simultaneously achieves efficient linkage between early warning and intervention, reducing the incidence risk and medical burden of hidden acute diseases in the elderly. It is suitable for various elderly health management scenarios, including home, community, and hospital settings.

[0031] In this embodiment, the multi-source data may include home scenario data, outing scenario data, and medical scenario data. The home scenario refers to the indoor environment where elderly users live and engage in daily activities. The outing scenario refers to the outdoor environment where elderly users go out. The medical scenario refers to scenarios such as elderly users undergoing examinations, tests, diagnoses, and physical examinations at medical institutions. As shown in Table 1, Table 1 illustrates an example of multi-scenario data collection and deployment.

[0032] Table 1 In step 201, the first physiological data, first scene data, and first reported data of the elderly user can be collected through millimeter-wave radar, desktop temperature and humidity sensors, and family member terminals deployed in the home environment. Millimeter-wave radar is a non-contact sensing device used to collect physiological behavioral signals such as micro-movements, respiration, and body movements of the elderly. The first physiological data, collected by the millimeter-wave radar, reflects the physiological and behavioral data of the elderly user at night and at home, and may include the frequency of turning over, breathing rhythm, and number of times the user gets up at night. The first scene data is environmental data such as temperature and humidity in the home environment, and may include ambient temperature and humidity. The first reported data is text data uploaded by family members through the terminal regarding the elderly user's subjective state, such as mental state, appetite, and sleep, and may include text data of the subjective state reported by family members.

[0033] Secondary physiological data of elderly users is collected through contact-based devices adapted for use in outdoor scenarios. These contact-based devices are wearable physiological monitoring devices suitable for elderly users, such as smart bracelets, smartwatches, and blood pressure monitors. The secondary physiological data consists of real-time physiological signals and limb activity data collected in outdoor scenarios, which can include both physiological signal data and limb activity data.

[0034] Medical care indicator data for elderly users is collected through medical information systems or medical devices in healthcare settings. This data comes from structured or textual data such as test results, examinations, diagnoses, and physical examinations conducted by medical institutions, and may include textual data representing medical care indicators for elderly users.

[0035] The wearable device in this application embodiment can be made of medical-grade flexible material. Before sampling, voice prompts (speech rate ≤ 120 words / minute) reduce tension and interference, and the signal-to-noise ratio is ≥ 40dB. The non-contact sensing device can use a radiation-free millimeter-wave radar deployed on the bedroom ceiling to capture concealed behavioral signals, reducing the constraint of wearing it. Through multi-sensor cross-validation (such as radar breathing signals to fill gaps in blood oxygen data), the accuracy of data completion is ≥ 95%.

[0036] This application employs a home-based, contactless data collection method, reducing discomfort and low compliance associated with wearable devices, making it suitable for long-term continuous monitoring. Through multi-scenario, multi-source data collection, it achieves full data coverage across home, outings, and medical visits, providing more comprehensive data dimensions and capturing health changes in elderly users under different conditions. Combining objective physiological data, environmental data, subjective descriptions from family members, and clinical medical data provides rich, accurate, and multi-dimensional data support for subsequent construction of physiological decline indicators and individualized baseline modeling. The cross-verification of multi-source data effectively reduces data bias caused by single-device errors or environmental interference, improving the reliability of subsequent identification.

[0037] In step 201, the acquired multi-source data can also be preprocessed. For physiological data, a combined algorithm of wavelet transform and adaptive filtering can be used. Wavelet transform and adaptive filtering are combined algorithms for denoising physiological signals, which can effectively remove electromyographic noise and motion artifacts. Electromyographic noise and motion artifacts are interference signals generated by muscle activity and body movement during physiological signal acquisition. For text data, invalid characters and repetitive information can be filtered out to achieve noise removal from multi-source data points.

[0038] For missing values ​​in multi-source data, the missing value mask and historical monitoring data are input into a multi-source completion model based on an attention mechanism. The model outputs the completed missing values, and the completion results are cross-validated using sensors from multiple scenarios to ensure that the accuracy of the completion results is greater than or equal to a set accuracy. The missing mask is used to mark which locations in the data contain missing information. The multi-source completion model uses historical data and an attention mechanism to learn data patterns and achieve high-precision completion of missing values.

[0039] The Z-score standardization algorithm was used to standardize the physiological data, and the Likert 5-level rating method was used to quantify the text data, mapping it to the [0,1] interval. The Z-score standardization algorithm converts the physiological data into standardized data with a mean of 0 and a variance of 1, eliminating the influence of units. The Likert 5-level rating method quantifies the subjective text state into a 1-5 level rating and maps it to the [0,1] interval.

[0040] Based on microsecond-level timestamps, a dynamic time warping algorithm is employed to correct non-uniformly sampled multi-source data, aligning the time sequence of multi-source data from multiple scenarios and unifying the format of physiological data with the text data. The dynamic time warping algorithm is used to align data with different sampling frequencies and time axes. Time alignment and format unification unify data collected from multiple scenarios and devices to the same time axis and data format.

[0041] Data quality is improved by effectively removing noise from physiological signals and text data. High-precision missing values ​​are filled in, and multi-scenario cross-validation ensures reliability, reducing the impact of missing data on modeling. Standardization and quantization eliminate dimensional differences, allowing different types of data to be input into the model. Temporal alignment and format unification resolve the issue of asynchronous sampling from multiple devices, providing high-quality, directly usable, and well-organized data for subsequent decay indicator construction and dynamic baseline training.

[0042] In step 201, constructing physiological decline indicators may include the following steps. First, the annual cardiopulmonary function decline rate, muscle mass loss rate, and cognitive function decline coefficient are selected as physiological decline indicators for elderly users. These physiological decline indicators are quantitative indicators specifically reflecting the gradual and irreversible decline in cardiopulmonary, muscle, and cognitive functions with age in the elderly. The initial values ​​of these indicators are calibrated based on the elderly users' annual physical examination data. The annual cardiopulmonary function decline rate is an indicator that characterizes the degree of cardiopulmonary function decline in elderly users each year, with vital capacity as its core. The muscle mass loss rate is an indicator that characterizes the rate of muscle mass reduction year by year, based on the muscle mass measured at the initial physical examination and combined with limb activity data. The cognitive function decline coefficient is a quantitative coefficient that comprehensively assesses the degree of cognitive decline by combining subjective reports from family members and physiological status.

[0043] The lung capacity test value is determined by combining the annual physical examination data of elderly users with respiratory data from multiple sources, and a first value corresponding to the annual cardiopulmonary function decline rate is determined based on the lung capacity test value. A baseline value for muscle mass is determined by combining the elderly users' first annual physical examination data with limb activity data from multiple sources, and a second value corresponding to the muscle mass loss rate is determined based on the muscle mass baseline value. A third value corresponding to the cognitive function decline coefficient is obtained by combining text data filled in by family members with the elderly users' physiological data through weighted calculation. The first, second, and third values ​​are all initial values ​​for physiological decline indicators, that is, the starting point value of the decline indicators is determined based on the annual physical examination data.

[0044] In one example, the annual rate of decline in cardiopulmonary function can be calculated using the following formula: η cardio = (vital capacity 当年 -vital capacity 上年 ) / vital capacity 上年 * 100. The rate of muscle loss can be calculated using the following formula: η muscle =(muscle mass) 基准 -Muscle mass 当前 / Monitoring duration (months). The cognitive decline coefficient can be calculated using the following formula: Family member reporting (5-level rating) + weighted calculation based on device behavior analysis.

[0045] By combining multi-source data collected from multiple scenarios at set intervals, the first, second, and third values ​​of physiological decline indicators can be recalculated and corrected, dynamically adjusting the multi-source data to maintain consistency with the individual's real-time status. For example, the initial values ​​are calibrated by annual physical examinations, and the decline rate is corrected every 3 months based on monitoring data.

[0046] Specifically designed for the physiological characteristics of the elderly, this system features a unique set of decline indicators, distinct from general health indicators, and better suited to identifying latent acute conditions in the elderly. The integration of multi-source data with physical examination data ensures the indicators are more objective, comprehensive, and accurate. A dynamic update and correction mechanism allows the indicators to change in real time with the individual's aging process, providing reliable core characteristics for subsequent individualized dynamic baselines. Three categories of decline indicators cover the three key systems of respiration, movement, and cognition, providing crucial evidence for the early identification of latent abnormalities and significantly reducing the rate of missed diagnoses.

[0047] In step 202, the basic functional features can be fused with the physiological decline indicators to form a multidimensional input vector. Basic functional features are core characteristics reflecting the basic health status of elderly users, including routine health-related features such as blood pressure, heart rate, blood oxygen saturation, mental state, and daily activity levels, serving as the basis for assessing an individual's basic health level. The multidimensional input vector is a feature set conforming to the neural network input format, formed by fusing the basic functional features with the physiological decline indicators, containing information on both basic health and physiological decline.

[0048] For example, the input features can be 20 basic features (12 functional baselines + 8 physiological indicators) + 4 degradation indicators, forming a 24-dimensional input vector with a time step of T=90 days (covering 3 months of historical data). Then, feature weights are initialized, and the degradation indicators are mapped to [0,1] using Min-Max normalization and assigned an initial weight of 0.3 (differentiated from the basic features).

[0049] Then, based on a three-layer LSTM network (128 / 64 / 32 neurons), a decay weight factor was introduced into the Bahdanau attention mechanism to construct an improved LSTM-Attention neural network. The attention formula was optimized to: e t =v T tanh(W h ht+W s st-1+W d ωt), where ωt is the decay weight factor, which strengthens the focus on decay-related nodes. The three-layer LSTM network excels at capturing the changing patterns of time-series data and can effectively handle the temporal characteristics of multi-source data (such as the temporal correlation of physiological and behavioral data). The Bahdanau attention mechanism is a classic attention mechanism whose core function is to allow the model to automatically focus on key features in the input vector during training and inference. In this embodiment, it is used to allow the model to focus on features related to physiological decline in the elderly. The decay weight factor is obtained by weighted summation of physiological decline indicators, used to quantify the overall physiological decline degree of elderly users, guiding the attention mechanism to focus on decay-related features, and ultimately constructing an improved LSTM-Attention neural network adapted to the physiological characteristics of the elderly.

[0050] Next, an improved LSTM-Attention neural network is trained using the Huber loss function with a decay penalty, coupled with the AdamW optimizer (initial learning rate 0.001, decaying by 10% every 15 epochs). This outputs the individual baseline ranges for each indicator for elderly users, generating individualized dynamic baselines. The Huber loss function with a decay penalty is a loss function that adds a decay penalty term to the traditional Huber loss function (which has strong resistance to outliers). The decay penalty term is introduced as follows: ζ=0.2. Its core function is to penalize the model for ignoring the physiological decline characteristics of aging, ensuring that the trained baseline closely matches the individual's decline characteristics. The AdamW optimizer is a commonly used neural network training optimizer that can effectively reduce model overfitting, accelerate training convergence, and adapt to the incremental training requirements of this application embodiment. The individualized dynamic baseline is a range of normal indicators specific to each elderly user, trained based on their own data, rather than a fixed threshold used in the industry. It can be dynamically updated as the individual's physiological decline progresses and is a core reference for identifying hidden abnormalities in the elderly.

[0051] At each predetermined time interval, preprocessed multi-source data is fed into the system to incrementally train the improved LSTM-Attention neural network, updating the baseline threshold of the individualized dynamic baseline. Incremental training eliminates the need to retrain the entire neural network; it only fine-tunes the model parameters using the newly added multi-source data, enabling rapid model updates, saving computational resources, and improving efficiency. Furthermore, at each predetermined time interval, the baseline range of indicators for the next predetermined time interval is predicted based on the decline curve. The decline curve is a curve plotted based on the physiological decline indicators constructed in step 201, combined with historical data, used to visually represent the gradual and irreversible decline trend of elderly users' physiological functions. Through predictive baseline updates, individual decline trends can be adapted in advance. For example, the baseline threshold is updated every 7 days through incremental training, and the normal range for the next year is predicted every 3 months based on the decline curve. For instance, if cardiopulmonary function declines by 10%, the normal threshold for blood oxygen saturation is lowered by 2% in advance. The individual normal range (mean ± 2 standard deviations) for each indicator is output, along with an explanation of the impact of the decline trend (e.g., "Bottom blood pressure is lowered by 1% due to muscle loss").

[0052] This application's embodiments construct an individualized dynamic baseline adapted to the physiological decline characteristics of aging through a process of "feature fusion → improved model construction → baseline training → dynamic update." This addresses the core pain points of traditional techniques, such as general and static baselines, which deviate significantly from the individual state of the elderly and fail to adapt to the decline trend. Simultaneously, improvements to the neural network and the design of the decline penalty loss function further enhance the model's focus on the characteristics of aging, providing an accurate and real-time reference standard for identifying hidden anomalies and laying the foundation for subsequent accurate identification of hidden acute conditions in the elderly.

[0053] In step 203, a single-indicator time-series in-depth detection is first performed. Using an individualized dynamic baseline as a reference standard, a sliding window is used to analyze real-time data segment by segment for physiological data. By calculating the trend slope and fluctuation frequency of the indicators within the sliding window, it is determined whether there is a continuous worsening trend of the critical abnormality. The sliding window divides continuous real-time physiological data into fixed-length windows, analyzing the change characteristics of the data within each window segment by segment, which facilitates the capture of gradual changes in indicators (adapting to the slow onset characteristics of latent abnormalities in the elderly). The trend slope is a parameter that quantifies the rate of change of physiological indicators within the sliding window; the larger the slope, the faster the indicator deteriorates / becomes abnormal. The fluctuation frequency is a parameter that quantifies the stability of changes in physiological indicators within the sliding window; the lower the frequency, the more stable the abnormal state, ruling out the possibility of random fluctuations.

[0054] If real-time data falls within the critical range for occult acute conditions in the elderly, with a trend slope greater than or equal to a set slope and a fluctuation frequency less than or equal to a set frequency, the real-time data is marked as a primary occult abnormality, initially screening for abnormal signals suspected of being precursors to acute conditions. Abnormal signals initially screened that meet the critical range and show a worsening trend may contain occasional fluctuations or interference signals, requiring further verification. For example, calculating the trend slope plus fluctuation frequency within a 30-minute window, such as blood oxygen saturation remaining between 90% and 93% for 30 minutes and decreasing by 0.5% every 5 minutes, triggers a primary warning. Critical abnormality determination can be based on clinically defined thresholds, such as a body temperature of 37.5-38℃ with a slope ≥0.1℃ / hour, which is marked as a primary occult abnormality.

[0055] Then, multi-indicator time-series correlation detection is performed. A clinical rule base is invoked, which is a set of rules built based on the "Guidelines for the Diagnosis and Treatment of Occult Acute Diseases in the Elderly." This rule base can contain multi-indicator correlation logic, covering typical signal combinations of various occult acute diseases in the elderly. For example, the clinical rule base can contain multi-indicator correlation logic, covering typical signal combinations of common occult acute diseases in the elderly such as infections and cardiovascular events. Each rule has been clinically validated. The Apriori algorithm is used to perform time-series correlation mining on the target signal combination composed of labeled primary occult anomalies and concurrent indicator data to find the inherent correlation patterns between indicators. The Apriori algorithm is a commonly used association rule mining algorithm used to mine the correlation relationships between indicators from multiple sets of data, adapting to the correlation mining requirements of multi-indicator combinations and acute disease types in this scheme. The target signal combination is a signal set composed of labeled primary occult anomalies and other concurrently collected indicator data, used for subsequent correlation mining.

[0056] If a target signal combination matching a predefined signal combination is detected, the predefined signal combination is a set of typical indicators corresponding to a certain type of latent acute illness in the elderly (e.g., "increased respiratory rate + decreased urine output + lethargy" corresponds to infectious acute illnesses). In other words, if it matches a predefined signal combination in the clinical rule base corresponding to a certain type of latent acute illness, the primary latent abnormality in the target signal combination is upgraded to an intermediate latent abnormality, further narrowing the scope of abnormalities and improving the specificity of abnormality detection. Intermediate latent abnormalities are primary latent abnormalities that match clinical acute illness signal combinations after multi-indicator association mining verification, significantly improving the reliability of the abnormality.

[0057] Multi-source data from multiple scenarios are extracted and cross-validated. If an intermediate-level hidden anomaly matches signals from two or more different scenarios simultaneously, it is determined to be a true hidden anomaly. A true hidden anomaly is an abnormal signal that accurately reflects the prodromal symptoms of acute illness in elderly users, confirmed after cross-validation of multi-scenario, multi-source data and elimination of interference. It serves as the core input for subsequent risk grading and acute illness prediction. If an anomaly occurs only in a single scenario, further verification using historical data is required to eliminate interference.

[0058] For example, the association rule base matching can call a library containing 120 clinical rules (such as respiratory rate >22 breaths / min + decreased urine output >30% + lethargy → infection risk), and use the Apriori algorithm to mine temporal association patterns. Cross-scenario signal verification is achieved by using signals such as nighttime restlessness on home radar + a 20% increase in heart rate on wearable devices + a sudden drop in appetite on family members' devices; if two or more of these three signals are met, the warning level is upgraded.

[0059] Using individualized dynamic baselines as a reference instead of universal thresholds addresses the problem of large discrepancies between traditional technical reference standards and the individual's declining condition, as well as the inability to identify baseline-drifting hidden anomalies, thus improving the specificity of anomaly identification. A combined analysis of sliding window, trend slope, and fluctuation frequency accurately distinguishes between occasional fluctuations and critical anomalies of continuous deterioration, effectively reducing misjudgments and lowering the missed diagnosis rate of hidden anomalies in the elderly. The introduction of a clinical rule base and the Apriori algorithm enables multi-indicator time-series correlation mining, overcoming the limitation of single-indicator anomalies in determining acute risk, ensuring that anomaly identification aligns with clinical diagnosis and treatment logic, and improving the reliability of anomaly judgment. A multi-scenario cross-validation mechanism eliminates interference from single devices and environments, further improving the accuracy of identifying true hidden anomalies, ensuring the accuracy of subsequent risk stratification and acute condition prediction, and providing high-quality input for dual-model fusion.

[0060] In step 203, identification can also be performed using GBDT-GNN dual-model fusion. Based on true hidden anomalies, baseline deviation features between the current indicators and the individualized dynamic baseline, basic health status features of elderly users, ABCDE framework features, and physiological decline indicators, a multi-dimensional fusion feature vector adapted for dual-model analysis is constructed. Baseline deviation features are the deviation values ​​between the current physiological indicators and the individualized dynamic baseline generated in step 202, quantifying the degree to which the indicators deviate from the individual's normal range. The ABCDE framework features are the core framework features used for elderly health assessment, covering mental state (A), activity level (B), cognitive function (C), nutritional status (D), and excretory function (E), comprehensively reflecting the overall health status of elderly users. The multi-dimensional fusion feature vector is a feature set integrating five categories of features: true hidden anomalies, baseline deviation, basic health, ABCDE framework, and physiological decline, providing comprehensive and accurate input for the dual model.

[0061] Multi-dimensional fused feature vectors are input into the GBDT model, outputting multi-level risk grades (e.g., high, medium, and low risk) and core risk contribution features. The multi-level risk grades are based on the severity and deterioration trend of the true hidden anomalies, guiding subsequent tiered responses. Core risk contribution features are the key features that cause the GBDT model to output the corresponding risk level, clarifying the core source of risk and improving the interpretability of the results. Simultaneously, feature data corresponding to high-level risks is extracted based on the risk grades. Only feature data corresponding to high-level risks is extracted as input to the subsequent GNN model, reducing the computational resource consumption of low-risk data.

[0062] Using non-specific symptoms and specific indicators from multi-source data, as well as types of latent acute conditions in the elderly (such as infection and cardiovascular events), as nodes, a Generative Neural Network (GNN) model is constructed, with initial values ​​for edge weights between nodes constrained by a clinical rule base. The GNN model is then trained by constraining the edge weights based on clinical logic. Non-specific symptoms are early, atypical symptoms of latent acute conditions in the elderly (such as lethargy and decreased appetite), lacking clear directional indications and requiring correlation with other indicators for accurate diagnosis. Specific indicators are physiological and laboratory indicators that clearly point to a particular type of acute condition (e.g., elevated C-reactive protein (CRP) indicates infection, and elevated B-type natriuretic peptide (BNP) indicates cardiovascular events). In the graph structure of the GNN model, the initial values ​​of the correlation strength parameters between different nodes (symptoms, indicators, and acute conditions) are constrained by the clinical rule base to ensure that the correlations align with clinical logic.

[0063] The high-risk feature data output by the GBDT model is input into the GNN model to mine the association between the high-risk feature data and the hidden acute disease type, and output the suspected acute disease type and the corresponding probability information.

[0064] The emergency prediction of the GNN model is triggered based on the risk level output by the GBDT model, and the suspected emergency type output by the GNN model is fed back to the GBDT model to dynamically adjust the weight allocation of physiological decline indicators and baseline deviation features, forming a closed-loop optimization of risk grading and emergency prediction. A two-way collaborative mechanism between the GBDT model and the GNN model is established. Closed-loop optimization is a mechanism of two-way feedback and mutual adjustment between GBDT and GNN, enabling risk grading and emergency prediction to dynamically adapt to the health status of elderly users and continuously improve accuracy. On the one hand, based on the risk level output by GBDT, the emergency prediction of the GNN model is selectively triggered (only high-risk levels are triggered, low and medium risks are not activated, saving computational resources). On the other hand, the suspected emergency type output by GNN is fed back to the GBDT model to dynamically adjust the weight allocation of physiological decline indicators and baseline deviation features (e.g., if the prediction is an infectious emergency, the weight of respiratory and temperature-related deviation features is increased), forming a closed loop of "risk grading → emergency prediction → weight optimization".

[0065] By integrating the risk level of the GBDT model with the suspected emergency type and probability information of the GNN model, and associating the core risk contribution characteristics with the corresponding clinical rules, the system outputs a structured result containing the risk level, suspected emergency name, core risk contribution characteristics, and clinical rules, ensuring that the judgment logic is traceable and interpretable.

[0066] For example, in GBDT risk grading, the input features are 42 dimensions (18 baseline biases + 12 basic states + 12 ABCDE framework features). The model parameters are 200 decision trees, a maximum depth of 5, a learning rate of 0.01, and the output is high / medium / low risk levels and the top 3 risk features. Clinical rule constraints on GNN emergency prediction can include graph structure constraints, with node edge weights initially limited by clinical rules (e.g., the edge weights for "elevated CRP" and "infection" are ≥0.7), and adjusted only within ±0.1 during training. The output results are suspected emergency types (12 categories) and probability values ​​(accuracy ≥88%). The fusion strategy is serial fusion (high / medium risk enters GNN prediction + backfeedback (GNN results adjust GBDT feature weights)).

[0067] By employing a multi-level hidden signal enhancement mechanism and a GBDT-GNN dual-model fusion design, a complete process is formed, encompassing anomaly screening, correlation verification, risk grading, emergency prediction, and closed-loop optimization. This addresses the pain points of traditional techniques, such as low accuracy in identifying hidden anomalies in the elderly, lack of clinical basis for emergency prediction, and lack of collaborative logic in dual-model fusion. On one hand, the multi-level progressive anomaly identification mechanism accurately captures slow, atypical abnormal signals of hidden emergencies in the elderly, reducing missed diagnoses and misjudgments. On the other hand, the clinically rule-constrained dual-model closed-loop fusion improves the accuracy of risk grading and emergency prediction while enhancing the clinical interpretability of the results. This provides high-quality, actionable core evidence for early warning responses, aligning with the physiological decline characteristics of the elderly and the needs of clinical diagnosis and treatment, demonstrating significant innovation and practicality.

[0068] In this embodiment, multiple terminal devices may include medical care terminals, family member terminals, and elderly user terminals. Risk levels may include high-risk, medium-risk, and low-risk levels. Medical care terminals are professional terminal devices (such as hospital workstations or mobile nursing tablets) used by medical personnel (doctors, nurses, etc.) to receive professional early warning information, view complete data, and support clinical decision-making. Family member terminals are personal terminals (such as mobile phones or tablets) used by family members of elderly users to receive early warning reminders, view response suggestions, and achieve real-time monitoring of the elderly user. Elderly user terminals are simple terminals adapted for elderly users (such as age-friendly mobile phones or smart bracelets), adhering to age-friendly design specifications, with simple operation and clear prompts.

[0069] In step 204, the first early warning information is output to the medical terminal. This first early warning information is a professional-grade warning for the medical terminal, characterized by its comprehensiveness, professionalism, and traceability. It includes complete judgment logic and data support, adapting to clinical diagnosis and treatment needs. It may include risk level, suspected emergency type and probability, core risk contribution characteristics, corresponding clinical rules, and decline impact analysis, along with multi-source data collected from multiple scenarios and preprocessed feature curves. The feature curves are physiological and behavioral indicator change curves (such as respiratory rate time-series curves and blood pressure change curves) plotted after multi-source data preprocessing, which can intuitively present the trend and abnormal nodes of the indicators.

[0070] The second warning message is delivered via pop-up windows and voice push notifications. This second warning message is a user-friendly warning message for family members' devices, characterized by its simplicity, ease of understanding, and operability. It avoids technical jargon, focuses on clearly defining response measures, and is tailored to the cognitive level of non-professional family members. It includes the risk level, the name of the suspected emergency, key symptom alerts, an interpretation of the effects of decline, and emergency treatment suggestions.

[0071] The third warning message is delivered in a manner adapted to the cognitive abilities of the elderly. This cognitively adapted approach is designed to address the cognitive decline and reduced visual and auditory functions of older adults. Key features include large fonts, high-contrast graphics and text, slow-paced speech, and concise, straightforward language. It reduces complex terminology and cumbersome operations, aligning with the cognitive habits and usage needs of seniors, and supports access to assistive devices such as screen readers. The third warning message is a simplified warning message for elderly users' terminals. Its core characteristics are simplicity, directness, and ease of execution. It avoids complex content, focusing on clearly defining the individual's condition and simple action requirements, including risk level, symptom description, and action guidance.

[0072] The multi-terminal differentiated early warning design accurately matches the core needs of three user groups: medical staff, family members, and elderly users, solving the problems of "one-size-fits-all" and poor adaptability of existing technology's early warning information. Medical staff receive professional and comprehensive decision support, family members receive easy-to-understand and practical coping guidelines, and elderly users receive simple and easy-to-understand status prompts, ensuring that early warning information can be efficiently understood and rationally utilized by different groups.

[0073] If the risk level is high, it means that the elderly's hidden acute illness is in a stage of rapid deterioration and is about to have an acute attack. If no immediate intervention is made, it may endanger their lives. This is the highest priority in the graded response. In this case, an emergency call response is triggered, an emergency priority indicator is pushed to the medical terminal, the elderly user's real-time location, basic medical history and current abnormal indicators are shared at the same time, an emergency contact pop-up is pushed to the family's terminal, and the phone number of the nearest emergency center and guidance on the green channel for medical treatment are provided.

[0074] If the risk level is medium risk, it means that the latent acute disease in the elderly is in a continuous development stage with no risk of acute onset, but medical intervention and regular check-ups are required. Otherwise, it may progress to a high risk level. The priority is between high risk and low risk. In this case, an outpatient check-up response is triggered, and a check-up reminder and a list of items to be evaluated are pushed to the medical staff terminal. An appointment registration link is pushed to the family terminal, and a monitoring plan is generated.

[0075] If the risk level is low, it means that the elderly user has only mild and hidden abnormalities with no worsening trend, which is in line with the normal fluctuations of physiological decline in old age. This risk level can be alleviated through home care and daily monitoring, and does not require immediate medical intervention. This is the lowest priority in the tiered response, which triggers the home attention response, pushes regular monitoring reminders to family members' terminals, and pushes healthy living suggestions to the elderly user's terminal.

[0076] Through a three-tiered differentiated response design (high, medium, and low), this system precisely matches the needs of elderly individuals with hidden acute conditions at different risk levels. It addresses the core pain points of traditional technologies, such as lack of targeted intervention after early warning, limited response methods, and poor adaptability. High-risk cases focus on rapid emergency care and seamless medical access to secure the golden treatment window. Medium-risk cases focus on convenient follow-up examinations and continuous monitoring to slow disease progression. Low-risk cases focus on home care and health guidance to reduce the probability of abnormal worsening. The entire response process achieves collaboration among medical staff, family members, and elderly users, aligning with the physiological characteristics and cognitive abilities of the elderly, conforming to clinical treatment guidelines, and prioritizing convenience and practicality. This seamless integration of early warning, response, and nursing / treatment further enhances the implementation and clinical applicability of the elderly hidden acute condition identification program, effectively reducing the risk of onset, deterioration, and medical burden associated with these conditions.

[0077] In step 205, closed-loop iterative optimization and data security assurance can be achieved.

[0078] For model iteration, feedback can be collected first, such as medical staff marking the accuracy of warnings (0-5 points) and family members reporting the results of abnormal handling. Then, the model is updated, with the GNN side weights updated every 100 pieces of feedback data, and the association rule base expanded every 3 months based on new clinical data. Next, the decline indicators are calibrated, and the decline rate calculation model is revised annually based on physical examination data.

[0079] To ensure data security, encrypted transmission can be used, such as the national standard SM4 algorithm (128-bit key). Patient privacy information is protected using differential privacy (dynamically adjusting Gaussian noise intensity). The cloud server records data access logs via blockchain, enabling full traceability.

[0080] Figure 3 This is a schematic diagram of the structure of a device 300 for identifying latent acute illnesses in the elderly, as provided in an embodiment of this application. Figure 3 As shown, the identification device 300 for latent acute diseases in the elderly may include a data acquisition module 301, a training module 302, an output module 303, an early warning module 304, and an update module 305.

[0081] The data acquisition module 301 is used to collect multi-source data of elderly users in multiple scenarios, preprocess the multi-source data, and construct physiological decline indicators for elderly users.

[0082] The training module 302 is used to fuse basic functional features with physiological decline indicators and input them into the improved LSTM-Attention neural network, train to generate an individualized dynamic baseline and dynamically update the baseline based on the decline curve. The improved LSTM-Attention neural network introduces a decline weight factor in the attention mechanism.

[0083] The output module 303 is used to identify the hidden abnormalities of the elderly user based on an individualized dynamic baseline through a multi-level hidden signal enhancement mechanism, and to output the risk level and suspected emergency type by adopting a GBDT-GNN dual-model fusion strategy and combining clinical rules to constrain the node edge weights of the GNN model.

[0084] The early warning module 304 is used to output early warning information matching the risk level and suspected acute illness type to multiple terminal devices respectively, and trigger a graded response according to the risk level. The early warning information includes clinical rule basis and interpretable early warning information based on the analysis of the impact of decline.

[0085] The update module 305 is used to update the node edge weights of the GNN model and the associated clinical rule base based on the clinical feedback data after the graded response.

[0086] The acquisition module 301, training module 302, output module 303, early warning module 304, and update module 305 can be used to execute steps 201-205 in the embodiments of the above-mentioned method for identifying latent acute diseases in the elderly. For the specific implementation of these modules and more details, please refer to the corresponding method section, which will not be elaborated here.

[0087] This application also provides a computer-readable storage medium storing a program that can be loaded by a processor and executed by any of the methods for identifying latent acute illnesses in this application.

[0088] Those skilled in the art will understand that all or part of the functions of the various methods in the above embodiments can be implemented by hardware or by computer programs. When all or part of the functions in the above embodiments are implemented by computer programs, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer programs, the program can also be stored in a server, another computer, disk, optical disk, flash drive, or external hard drive, etc., and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0089] The above examples illustrate this application only to aid understanding and are not intended to limit its scope. Those skilled in the art to which this application pertains can make various simple deductions, modifications, or substitutions based on the ideas presented.

Claims

1. A method for identifying latent acute illnesses in the elderly, characterized in that, include: Multi-source data of elderly users in multiple scenarios were collected, the multi-source data were preprocessed, and physiological decline indicators of the elderly users were constructed. The basic functional features are fused with the physiological decline indicators and input into the improved LSTM-Attention neural network to train and generate an individualized dynamic baseline. The baseline is dynamically updated based on the decline curve. The improved LSTM-Attention neural network introduces a decline weight factor into the attention mechanism. Based on the individualized dynamic baseline, the hidden abnormalities of the elderly users are identified through a multi-level hidden signal enhancement mechanism. The GBDT-GNN dual-model fusion strategy is adopted, and the node edge weights of the GNN model are constrained by clinical rules to output the risk level and suspected acute condition type. The system outputs warning information matching the risk level and suspected acute illness type to multiple terminal devices, and triggers a graded response based on the risk level. The warning information includes clinical rule basis and interpretable warning information based on decline impact analysis. Based on the clinical feedback data after the graded response, the node edge weights and associated clinical rule base of the GNN model are updated.

2. The method for identifying latent acute illnesses in the elderly according to claim 1, characterized in that, The multi-source data includes data from home scenarios, outing scenarios, and medical treatment scenarios. The collection of multi-source data from elderly users in multiple scenarios includes: By deploying millimeter-wave radar, desktop temperature and humidity sensors, and family member terminals in home settings, the system collects the elderly user's first physiological data, first scene data, and first reporting data. The first physiological data includes turning frequency, breathing rhythm, and number of times the user gets up at night. The first scene data includes ambient temperature and humidity. The first reporting data includes text data of the subjective state reported by family members. The elderly user’s second physiological data is collected through a contact device adapted to the elderly user in the outdoor scenario. The second physiological data includes physiological signal data and limb activity data. The medical information system or medical equipment in the medical treatment scenario are used to collect medical treatment indicator data of the elderly users. The medical treatment indicator data includes text data that represents the medical treatment indicators of the elderly users.

3. The method for identifying latent acute illnesses in the elderly according to claim 1, characterized in that, The preprocessing of the multi-source data includes: For physiological data, a combination algorithm of wavelet transform and adaptive filtering is used to remove electromyographic noise and motion artifacts. For text data, invalid characters and duplicate information in the text data are filtered to achieve the noise removal operation of the multi-source data points. For missing values ​​in the multi-source data, the missing mask corresponding to the missing value and historical monitoring data are input into a multi-source completion model based on an attention mechanism, and the completion result of the missing value is output. The completion result is cross-validated by sensors in multiple scenarios so that the accuracy of the completion result is greater than or equal to the set accuracy. The physiological data were standardized using the Z-score normalization algorithm, and the text data was quantified using the Likert 5-level scoring method and mapped to the [0,1] interval. Based on microsecond-level timestamps, a dynamic time warping algorithm is used to correct the non-uniformly sampled multi-source data, align the multi-source data from multiple scenarios in time sequence, and unify the format of the physiological data and the text data.

4. The method for identifying latent acute illnesses in the elderly according to claim 1, characterized in that, The construction of the physiological decline indicators for the elderly users includes: The annual rate of decline in cardiopulmonary function, the rate of muscle loss, and the cognitive function decline coefficient were selected as physiological decline indicators for elderly users. The initial values ​​of the physiological decline indicators were calibrated based on the annual physical examination data of the elderly users. The lung capacity test value is determined by combining the annual physical examination data of the elderly users with the respiratory data of the multi-source data, and the first value corresponding to the annual cardiopulmonary function decline rate is determined based on the lung capacity test value. A baseline value for muscle mass is determined by combining the elderly user's first annual physical examination data with the limb activity data from the multi-source data, and a second value corresponding to the rate of muscle loss is determined based on the baseline value for muscle mass. By combining the text data filled in by family members with the physiological data of the elderly users, a third value corresponding to the cognitive function decline coefficient is obtained through weighted calculation. At set intervals, the first, second, and third values ​​of the physiological decline index are recalculated and corrected by combining the multi-source data collected from multiple scenarios.

5. The method for identifying latent acute illnesses in the elderly according to claim 1, characterized in that, The process of fusing basic functional features with the physiological decline indicators and inputting them into an improved LSTM-Attention neural network to train and generate an individualized dynamic baseline, and then dynamically updating the baseline based on the decline curve, includes: The basic functional characteristics are fused with the physiological decline indicators to form a multidimensional input vector; Based on a three-layer LSTM network, a decay weight factor is introduced into the Bahdanau attention mechanism to construct an improved LSTM-Attention neural network. The decay weight factor is obtained by weighted summation of the physiological decay indicators. The improved LSTM-Attention neural network is trained using the Huber loss function with a decay penalty, and paired with the AdamW optimizer to output the individual baseline range of each indicator for the elderly user, generating an individualized dynamic baseline. At each first set time interval, preprocessed multi-source data is accessed, and the improved LSTM-Attention neural network is incrementally trained to update the baseline threshold of the individualized dynamic baseline. At each second set time interval, the indicator benchmark range for the next third set time interval is predicted based on the decay curve.

6. The method for identifying latent acute illnesses in the elderly according to claim 1, characterized in that, The method of identifying hidden anomalies in elderly users based on the individualized dynamic baseline through a multi-level occult signal enhancement mechanism includes: Using the individualized dynamic baseline as a reference standard, real-time data is analyzed segment by segment using a sliding window for physiological data. By calculating the trend slope and fluctuation frequency of the indicators within the sliding window, it is determined whether there is a continuous deterioration trend in the critical abnormality. If the real-time data is within the critical range of clinical latent acute diseases in the elderly, the trend slope is greater than or equal to a set slope, and the fluctuation frequency is less than or equal to a set frequency, then the real-time data is marked as a primary latent anomaly. The clinical rule base is invoked, and the Apriori algorithm is used to perform time-series correlation mining on the target signal combination composed of the labeled primary occult anomalies and the concurrent indicator data. The clinical rule base includes multi-indicator correlation logic, covering multiple set signal combinations of occult emergencies. If a target signal combination that matches the set signal combination is detected, the primary concealment anomaly in the target signal combination is upgraded to an intermediate concealment anomaly. Multi-source data from multiple scenarios are extracted for cross-validation, and if the intermediate-level hidden anomaly matches signals from two or more different scenarios, it is determined to be a true hidden anomaly.

7. The method for identifying latent acute illnesses in the elderly according to claim 6, characterized in that, The method employs a GBDT-GNN dual-model fusion strategy, combining clinical rules to constrain the node edge weights of the GNN model, and outputs risk levels and suspected emergency types, including: Based on the real hidden anomalies, the baseline deviation characteristics between the current indicators and the individualized dynamic baseline, the basic health status characteristics of the elderly users, the ABCDE framework characteristics, and the physiological decline indicators, a multi-dimensional fusion feature vector is constructed. The multi-dimensional fused feature vector is input into the GBDT model to output multi-level risk levels and core risk contribution features. At the same time, feature data corresponding to high-level risks are extracted based on the risk levels. Using non-specific symptoms and specific indicators from the multi-source data, as well as the types of latent acute diseases in the elderly, as nodes, and using the clinical rule base to limit the initial values ​​of the edge weights between nodes, a GNN model is constructed, and the edge weights are constrained based on clinical logic to train the GNN model. The high-risk feature data output by the GBDT model is input into the GNN model to mine the association between the high-risk feature data and the hidden acute illness type, and output the suspected acute illness type and the corresponding probability information. The emergency prediction of the GNN model is triggered based on the level output by the GBDT model, and the suspected emergency type output by the GNN model is fed back to the GBDT model to dynamically adjust the weight allocation of the physiological decline index and baseline deviation feature, forming a closed-loop optimization of risk classification and emergency prediction. The risk level of the GBDT model is integrated with the suspected emergency type and probability information of the GNN model, and the core risk contribution features are associated with the corresponding clinical rules. The output is a structured result containing risk level, suspected emergency name, core risk contribution features and clinical rules.

8. The method for identifying latent acute illnesses in the elderly according to claim 1, characterized in that, The multiple terminal devices include medical care terminals, family member terminals, and elderly user terminals, and the risk levels include high risk level, medium risk level, and low risk level; The step of outputting warning information matching the risk level and suspected acute illness type to multiple terminal devices, and triggering a tiered response based on the risk level, includes: The first warning information is output to the medical terminal. The first warning information includes risk level, suspected emergency type and probability, core risk contribution characteristics, corresponding clinical rule basis, decline impact analysis, and multi-source data collected from multiple scenarios and preprocessed feature curves. The second warning information is output through pop-up windows and voice push. The second warning information includes the risk level, the name of the suspected acute illness, the core symptom prompts, the interpretation of the impact of decline, and emergency treatment suggestions. The third warning information is output in a prompting manner adapted to the cognitive abilities of the elderly. The third warning information includes risk level, symptom description and action guidance. If the risk level is high, an emergency call response is triggered, an emergency priority indicator is pushed to the medical terminal, the real-time location, basic medical history and current abnormal indicators of the elderly user are shared simultaneously, an emergency contact pop-up is pushed to the family terminal, and the telephone number of the nearby emergency center and guidance on the green channel for medical treatment are provided. If the risk level is medium risk, an outpatient follow-up response will be triggered, pushing a follow-up reminder and a list of items to be evaluated to the medical staff terminal, pushing an appointment registration link to the family member terminal, and generating a monitoring plan. If the risk level is low, a home monitoring response will be triggered, sending regular monitoring reminders to family members' terminals and healthy living suggestions to elderly users' terminals.

9. A device for identifying latent acute illnesses in the elderly, characterized in that, include: The data acquisition module is used to collect multi-source data of elderly users in multiple scenarios, preprocess the multi-source data, and construct physiological decline indicators of the elderly users. The training module is used to fuse basic functional features with the physiological decline indicators and input them into the improved LSTM-Attention neural network to train and generate an individualized dynamic baseline and dynamically update the baseline based on the decline curve. The improved LSTM-Attention neural network introduces a decline weight factor in the attention mechanism. The output module is used to identify the hidden abnormalities of the elderly user based on the individualized dynamic baseline through a multi-level hidden signal enhancement mechanism, and to output the risk level and suspected acute condition type by adopting the GBDT-GNN dual-model fusion strategy and combining clinical rules to constrain the node edge weights of the GNN model. The early warning module is used to output early warning information matching the risk level and suspected acute illness type to multiple terminal devices respectively, and trigger a graded response according to the risk level. The early warning information includes clinical rule basis and interpretable early warning information based on decline impact analysis. An update module is used to update the node edge weights and associated clinical rule base of the GNN model based on the clinical feedback data after the graded response.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that can be loaded by a processor and executed as described in any one of claims 1 to 8 for the identification of latent acute conditions in the elderly.

Images