A multi-parameter vital sign AI fusion analysis and early deterioration risk early warning system

Abstract

The application relates to the technical field of medical intelligent monitoring, and discloses a multi-parameter vital sign AI fusion analysis and early deterioration risk early warning system, which comprises a multi-modal sign collection module, a data preprocessing module, a feature extraction module, an AI fusion analysis module, a risk early warning module, a data storage module, a terminal interaction module, a medical knowledge graph module, a model self-optimization module and a multi-terminal linkage module; the multi-modal collection mode combining the contact type and the non-contact type is adopted, multi-dimensional vital sign data and auxiliary data of a human body are comprehensively collected, and the problem that the existing system has a single collection mode and incomplete data is solved; through an improved TFT-multi time sequence fusion transformer algorithm, end-to-end joint fusion analysis of multi-parameters is realized, physiological correlations between different signs are captured, the defects that the existing system has single-parameter independent analysis and weak fusion capability are solved, and the precision of health state evaluation and risk prediction is improved.

Description

Technical Field

[0001] This invention belongs to the field of medical intelligent monitoring technology, specifically a multi-parameter vital signs AI fusion analysis and early warning system for deterioration risk. Background Technology

[0002] Vital signs are core indicators reflecting the human body's health status, including heart rate, blood pressure, blood oxygen saturation, and respiratory rate. Changes in these parameters often indicate changes in the body's health status, especially for critically ill patients, postoperative patients, and patients with chronic diseases. Subtle abnormalities in vital signs may be early signs of disease deterioration. Timely capture and analysis of these signals are of great significance for reducing mortality and improving treatment outcomes.

[0003] Currently, most existing vital sign monitoring systems employ independent monitoring of single parameters or simple overlay methods, which have several shortcomings: First, they lack the ability to deeply integrate multimodal data, failing to capture the physiological correlations between different vital signs, resulting in biased analysis results and low accuracy in early warnings; second, they have weak capabilities in handling missing values ​​and interference data in the collected data, especially for sparsely sampled vital sign parameters, leading to large prediction errors and affecting the reliability of early warnings; third, the early warning mechanisms are simplistic, often using fixed threshold alarms without considering individual differences, easily resulting in false alarms and missed alarms, and the early warnings are delayed, failing to predict the risk of early deterioration; fourth, the models lack self-optimization capabilities, making it difficult to adapt to the monitoring needs of different populations and scenarios, and unable to achieve accurate analysis by combining clinical knowledge; fifth, the multi-terminal linkage is poor, and early warning information cannot be synchronized to medical staff, family members, and the emergency medical system in a timely manner, affecting the efficiency of emergency response.

[0004] Furthermore, some existing monitoring systems rely on contact-based data collection, which imposes a burden on patients and is unsuitable for long-term monitoring; while some non-contact monitoring systems suffer from low data accuracy and weak anti-interference capabilities. Meanwhile, existing AI analysis models mostly employ single time-series prediction algorithms, failing to achieve simultaneous fusion analysis of multiple parameters, thus struggling to meet the clinical demands for high-precision and timely early warning of deterioration risks. Therefore, developing an intelligent system capable of deep multi-parameter fusion, accurate analysis, early warning, multi-terminal linkage, and self-optimization capabilities has become an urgent need in the field of intelligent medical monitoring. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, this invention provides a multi-parameter vital signs AI fusion analysis and early deterioration risk warning system, which effectively solves the problems raised in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-parameter vital signs AI fusion analysis and early deterioration risk warning system, comprising a multimodal vital signs acquisition module, a data preprocessing module, a feature extraction module, an AI fusion analysis module, a risk warning module, a data storage module, a terminal interaction module, a medical knowledge graph module, a model self-optimization module, and a multi-terminal linkage module; The output of the multimodal vital sign acquisition module is electrically connected to the input of the data preprocessing module. The output of the data preprocessing module is electrically connected to the input of the feature extraction module. The output of the feature extraction module is electrically connected to the input of the AI ​​fusion analysis module. The AI ​​fusion analysis module is bidirectionally electrically connected to the risk warning module, the medical knowledge graph module, and the model self-optimization module. The output of the risk warning module is electrically connected to the input of the multi-terminal linkage module. The data storage module is bidirectionally electrically connected to the data preprocessing module, the AI ​​fusion analysis module, and the medical knowledge graph module. The terminal interaction module is bidirectionally electrically connected to the data storage module, the risk warning module, and the model self-optimization module. All modules work together to complete multi-parameter vital sign fusion analysis and early deterioration risk warning.

[0007] Preferably, the multimodal vital sign acquisition module includes a contact acquisition unit, a non-contact acquisition unit, and an auxiliary acquisition unit; The contact-type acquisition unit uses medical-grade biosensors to collect data on heart rate, blood pressure, blood oxygen saturation, body surface temperature, respiratory rate, and heart rate variability. The non-contact acquisition unit uses a high-sensitivity PVDF piezoelectric thin film sensor and millimeter-wave radar to acquire human cardiac impact signals, respiratory body movement signals and bed exit status data based on BCG technology, without the need for human body wear; The auxiliary data acquisition unit is used to collect data on human activity levels, body position, and ambient temperature and humidity. The sampling frequency of each acquisition unit is uniformly set to 15 minutes / time, and automatically switches to 1 minute / time in emergency situations. The acquired data is converted into digital signals by A / D converters and then transmitted to the data preprocessing module.

[0008] Preferably, the data preprocessing module includes a standardization processing unit, a missing value repair unit, and an outlier removal unit; The standardization processing unit uses the Z-score standardization algorithm to convert vital sign data of different dimensions into normalized data with a unified standard, and the normalization range is [0,1]. The missing value repair unit uses an improved interpolation algorithm based on masking technology to fill in missing values ​​by combining adjacent time-series data and baseline data of the same type of population, with a filling error of no more than 5%. The outlier removal unit uses the 3σ principle combined with Granger causality test to identify and remove interfering data generated during the data collection process, while retaining valid vital sign data.

[0009] Preferably, the feature extraction module includes a temporal feature extraction unit and a correlation feature extraction unit; The time-series feature extraction unit uses wavelet transform algorithm to extract the time-domain and frequency-domain features of each vital sign data. The time-domain features include mean, variance, peak value, valley value and rate of change, while the frequency-domain features include power spectral density and characteristic frequency. The correlation feature extraction unit uses an attention mechanism to extract correlation features between different vital signs parameters, quantifies the influence weight of each vital sign parameter on the risk of disease deterioration, and retains three decimal places in the weight calculation.

[0010] Preferably, the AI ​​fusion analysis module adopts an improved TFT-multi temporal fusion transformer algorithm. This algorithm optimizes the output layer and loss function based on the traditional TFT algorithm, introduces a multivariate joint training mechanism, and simultaneously receives temporal features and correlation features output by the feature extraction module and clinical knowledge data output by the medical knowledge graph module. It completes end-to-end joint fusion analysis of multi-parameter vital signs, outputs the evolution trend of vital sign parameters and the correlation of deterioration risk, and the fusion analysis delay does not exceed 30 seconds.

[0011] Preferably, the risk warning module includes a risk classification unit, a warning threshold setting unit, and a warning information generation unit; The risk grading unit divides the risk of early deterioration into four levels: normal level (risk value < 0.2), attention level (0.2 ≤ risk value < 0.4), warning level (0.4 ≤ risk value < 0.7), and critical level (risk value ≥ 0.7). The warning threshold setting unit supports personalized threshold adjustment based on individual medical history, age, and physical characteristics, with an adjustment step size of 0.05. The early warning information generation unit generates corresponding early warning information based on the risk level, including the risk level, abnormal vital signs parameters, changing trends, and preliminary intervention suggestions.

[0012] Preferably, the data storage module adopts a dual storage architecture of "local + cloud"; the local storage uses a solid-state drive to store real-time vital sign data and early warning records for the past 30 days, with a storage capacity of not less than 1TB; the cloud storage uses a distributed database to store historical vital sign data, AI fusion analysis results, medical knowledge graph data and model parameters, supports encrypted data transmission and backup, with a backup frequency of 24 hours / time, and a data storage period of not less than 5 years, meeting the medical data retention standards.

[0013] Preferably, the medical knowledge graph module includes a clinical rule base, a vital sign association base, and an intervention suggestion base; The clinical rule base stores clinical diagnosis and treatment guidelines and criteria for judging abnormal signs related to intensive care and chronic disease management; The vital sign association database stores the physiological correlations and pathological influence patterns among different vital sign parameters; The intervention suggestion library stores personalized intervention plans, medication references, and medical guidance corresponding to different risk levels and abnormal signs; The medical knowledge graph module supports regular updates, with an update frequency of no less than once a month, to ensure the timeliness of the knowledge.

[0014] Preferably, the model self-optimization module includes a data incremental training unit, a model performance evaluation unit, and a parameter adaptive adjustment unit; The incremental data training unit regularly uses newly added vital sign data and clinical case data to incrementally train the AI ​​fusion analysis model, with a training cycle of 7 days. The model performance evaluation unit uses MAE error, accuracy and early warning sensitivity as evaluation indicators. When any indicator fails to meet the preset threshold, parameter adjustment is triggered. The parameter adaptive adjustment unit uses the Adam optimizer with a learning rate set to 1e-3 to automatically adjust the model weight parameters, ensuring continuous optimization of model analysis accuracy and early warning performance.

[0015] Preferably, the multi-terminal linkage module includes a medical terminal linkage unit, a family terminal linkage unit, and an emergency terminal linkage unit; The medical terminal linkage unit is used to push early warning information to the workstations of medical staff and simultaneously transmit abnormal vital signs data and fusion analysis results; The family member terminal linkage unit pushes early warning information to the patient's family members via a mobile APP, including the risk level and contact suggestions; When a critical-level warning is triggered, the emergency terminal linkage unit automatically links with the 120 emergency medical system and the nearest medical institution to push the patient's location, vital signs data and medical history information, so as to achieve rapid emergency response. The information push delay of each linkage unit shall not exceed 10 seconds to ensure that the early warning information is delivered in a timely manner.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention utilizes multimodal acquisition and deep fusion to improve analysis accuracy. It employs a multimodal acquisition method combining contact and non-contact methods to comprehensively collect multidimensional vital sign data and auxiliary data of the human body, solving the problems of single acquisition methods and incomplete data in existing systems. Through an improved TFT-multi time-series fusion transformer algorithm, it achieves end-to-end joint fusion analysis of multiple parameters, capturing the physiological correlation between different vital signs, overcoming the shortcomings of existing systems in independent analysis of single parameters and weak fusion capabilities, and improving the accuracy of health status assessment and risk prediction. 2. This invention provides efficient data processing and improves data reliability. It adopts a data preprocessing process that combines standardization, missing value repair, and outlier removal. Combined with an improved interpolation algorithm and Granger causality test, it effectively repairs missing data and removes interfering data, solving the problems of low data processing accuracy and missing data affecting analysis results in existing systems. This provides high-quality data support for subsequent analysis and improves the reliability of system operation. 3. This invention provides personalized graded early warning, which improves the timeliness and specificity of early warning. It establishes a four-level risk grading system, supports personalized early warning threshold adjustment based on individual differences, and generates targeted early warning information and intervention suggestions by combining clinical knowledge provided by medical knowledge graph. This solves the problems of existing systems having a single early warning mechanism, high false alarm and false alarm rates, and delayed early warning, and achieves accurate early warning of early deterioration risk, thus buying time for clinical intervention and emergency treatment. 4. The model of this invention is self-optimizing, which improves the system's adaptability. The model self-optimization module is set up to achieve continuous optimization of the AI ​​fusion analysis model through regular incremental training, performance evaluation and adaptive parameter adjustment. This solves the problem that the existing system model has no self-optimization ability and cannot adapt to different groups of people and scenarios, ensuring that the model is always in the optimal operating state and improving the long-term adaptability and stability of the system. 5. This invention features multi-terminal linkage to improve emergency response efficiency. It establishes a multi-terminal linkage mechanism among medical terminals, family terminals, emergency terminals (and community terminals), enabling rapid and synchronous push of early warning information. This solves the problems of insufficient multi-terminal linkage and untimely transmission of early warning information in existing systems, ensuring that emergency response can be quickly initiated in critical situations, improving emergency response efficiency, and reducing the risk of worsening condition. 6. This invention features convenient human-computer interaction, adapts to multiple application scenarios, designs a clear and intuitive terminal interface, supports multiple input methods, simplifies operation processes, and adopts a "local + cloud" dual storage architecture to balance data security and scalability. It is suitable for various medical and health scenarios such as ICU monitoring, postoperative rehabilitation, home-based elderly care, and chronic disease management, thereby improving the system's versatility and practicality. 7. This invention has sufficient knowledge support, enhances the professionalism of analysis, sets up a medical knowledge graph module to store authoritative clinical knowledge data, provides professional support for AI fusion analysis and risk warning, solves the problems of existing systems lacking clinical knowledge support and insufficient professionalism of analysis results, and improves the scientificity and rationality of system analysis and warning. 8. This invention combines non-contact monitoring with precise monitoring, improving patient comfort. It adopts a non-contact data collection method, eliminating the need for patients to wear any devices, reducing the burden on patients and improving the comfort of long-term monitoring. At the same time, the contact data collection unit uses medical-grade sensors to ensure the accuracy of the collected data, balancing comfort and accuracy, and is suitable for long-term monitoring scenarios. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0018] In the attached diagram: Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the system workflow of the present invention; Figure 3 This is a logic block diagram of the AI ​​fusion analysis module of the present invention; Figure 4 This is a logic block diagram of the risk warning module of the present invention; Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] Depend on Figures 1-4 This invention relates to a multi-parameter vital sign AI fusion analysis and early deterioration risk warning system, comprising a multimodal vital sign acquisition module, a data preprocessing module, a feature extraction module, an AI fusion analysis module, a risk warning module, a data storage module, a terminal interaction module, a medical knowledge graph module, a model self-optimization module, and a multi-terminal linkage module. These modules are connected via wired or wireless communication to collaboratively complete the multi-parameter vital sign fusion analysis and early deterioration risk warning. The specific structure and operation are as follows: 1. Multimodal vital sign acquisition module This module is the system's data input source, used to comprehensively collect multi-dimensional vital sign data and related auxiliary data from the human body. It employs a combination of contact and non-contact acquisition methods, balancing data accuracy and patient comfort. Specifically, it includes a contact acquisition unit, a non-contact acquisition unit, and an auxiliary acquisition unit. The contact-type data acquisition unit uses medical-grade biosensors that have been clinically validated and meet medical standards. It can directly contact human skin to collect data on heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, blood oxygen saturation (SpO2), body surface temperature, respiratory rate, and heart rate variability (HRV). The acquisition accuracy meets clinical monitoring requirements, with blood pressure measurement error not exceeding ±2 mmHg, blood oxygen saturation measurement error not exceeding ±1%, and heart rate measurement error not exceeding ±1 beat / minute.

[0021] The non-contact acquisition unit uses a high-sensitivity PVDF piezoelectric film sensor and millimeter-wave radar. Based on BCG (cardiac impaction mapping) technology, it does not require any wearable devices on the human body. Patients only need to lie in bed or sit still to acquire cardiac impact signals, respiratory body movement signals, and data on their status after getting out of bed. The PVDF piezoelectric film sensor captures the weak body movements caused by the heartbeat, converts them into analog voltage signals, and then converts them into digital signals via an A / D converter. The millimeter-wave radar is used to monitor respiratory rate and changes in body position. It has strong anti-interference capabilities and can effectively avoid the influence of environmental noise on the acquired data.

[0022] The auxiliary acquisition unit uses motion sensors and environmental sensors to collect data on human activity, body position (supine, lateral, prone, sitting, standing) and environmental temperature and humidity. This data is used for environmental interference correction and personalized analysis of vital signs during subsequent data preprocessing. The environmental temperature and humidity acquisition accuracy is ±0.1℃ and ±1%RH, respectively.

[0023] The sampling frequency of each acquisition unit is uniformly set to 15 minutes / time to meet the needs of routine monitoring. When the system detects that a certain vital sign parameter is close to the warning threshold, it automatically switches to a high-frequency sampling mode of 1 minute / time to ensure that subtle changes in vital sign parameters are captured. The acquired data is converted into digital signals by an A / D converter and then transmitted to the data preprocessing module in real time.

[0024] 2. Data Preprocessing Module This module receives digital signals transmitted from the multimodal vital sign acquisition module and is used to standardize and purify the raw vital sign data, eliminate interference factors, and repair data defects, providing high-quality data support for subsequent feature extraction and AI fusion analysis. Specifically, it includes a standardization processing unit, a missing value repair unit, and an outlier removal unit. The standardization processing unit uses the Z-score standardization algorithm to normalize vital sign data of different dimensions and ranges, converting all vital sign data into normalized data with a unified standard. The normalization range is set to [0,1] to eliminate the influence of dimensional differences on the fusion analysis. The standardization calculation formula is: x'=(x-μ) / σ, where x is the original vital sign data, μ is the mean of the vital sign parameter, and σ is the standard deviation of the vital sign parameter.

[0025] The missing value repair unit employs an improved interpolation algorithm based on masking technology. For missing data generated during the acquisition process due to sensor detachment, signal interference, or other reasons, it constructs a temporal mask matrix to mark the location of missing data. Combined with valid vital sign data from adjacent time series and baseline vital sign data from individuals of the same age, physical condition, and medical history, it completes the missing value filling. The filling error is strictly controlled within 5% to avoid the impact of missing data on the analysis results. Simultaneously, this unit marks data that are missing more than three times consecutively and transmits it to the terminal interaction module to remind staff to check the acquisition equipment and patient status.

[0026] The outlier removal unit uses the 3σ principle combined with Granger causality test to identify and remove anomalies in the standardized vital sign data: First, the 3σ principle is used to identify outlier data that exceeds the normal range (i.e., data that deviates from the mean by more than 3 standard deviations). Then, the Granger causality test is used to determine whether the outlier data is due to genuine vital sign abnormalities or data acquisition interference. If it is due to data acquisition interference, it is removed; if it is due to genuine vital sign abnormalities, the data is retained and marked as key monitoring data, and transmitted to subsequent modules for in-depth analysis.

[0027] The preprocessed valid vital signs data are transmitted to the feature extraction module in real time and backed up to the data storage module at the same time.

[0028] 3. Feature Extraction Module This module receives valid vital sign data transmitted from the data preprocessing module. It extracts key features from the preprocessed vital sign data that reflect the human body's health status and disease trends, providing feature input to the AI ​​fusion analysis module. Specifically, it includes a time-series feature extraction unit and a correlation feature extraction unit. The time-series feature extraction unit uses wavelet transform algorithm to extract features from the time-series data of each individual vital sign parameter. The extracted time-series features include time-domain features and frequency-domain features: time-domain features include the mean, variance, peak value, trough value, rate of change, and fluctuation amplitude of the vital sign parameter within a set time period, reflecting the overall level and short-term trend of the vital sign parameter; frequency-domain features include power spectral density and characteristic frequency, reflecting the periodic change law of the vital sign parameter. The set time period can be customized through the terminal interaction module, and the default setting is 24 hours.

[0029] The correlation feature extraction unit employs an attention mechanism to analyze the correlation between different vital sign parameters, extract correlation features, and quantify the impact weight of each vital sign parameter on the risk of disease deterioration, with the weight calculation precision retained to three decimal places. For example, it analyzes the correlation between blood pressure fluctuations and changes in blood oxygen saturation, and the correlation between heart rate variability and respiratory rate, identifies the synergistic change patterns among different vital sign parameters, and uncovers hidden health risk signals. Simultaneously, this unit matches the extracted correlation features with the vital sign association rules in the medical knowledge graph module to verify the effectiveness of the features.

[0030] After feature extraction is completed, the temporal features and related features are integrated into a feature vector, which is transmitted to the AI ​​fusion analysis module in real time and stored in the data storage module for subsequent model training and optimization.

[0031] 4. AI Fusion Analysis Module This module is the core analysis unit of the system, used to perform deep fusion analysis on the feature vectors transmitted by the feature extraction module. Combined with the clinical knowledge provided by the medical knowledge graph module, it achieves accurate assessment of human health status and prediction of disease deterioration trends. It adopts an improved TFT-multi temporal fusion transformer algorithm, and its specific working method is as follows: The improved TFT-multi algorithm optimizes the output layer and loss function based on the traditional Temporal Fusion Transformer (TFT) algorithm, and introduces a multivariate joint training mechanism to solve the shortcomings of the traditional algorithm in modeling single parameters independently and failing to capture the physiological correlation between vital signs, thus realizing end-to-end joint fusion analysis of multi-parameter vital signs.

[0032] The algorithm receives temporal feature vectors and associated feature vectors output by the feature extraction module, and simultaneously calls upon data from the clinical rule base and vital sign association database in the medical knowledge graph module as the basis for analysis. It aligns and fuses the temporal features of multiple vital signs through a temporal fusion layer, and deeply mines the associated features between vital signs through an association fusion layer. Combined with an attention weight allocation mechanism, it highlights the influence of key vital sign parameters and key features. Through an optimized loss function, it reduces the impact of missing and outlier data on the analysis results, improving the accuracy of the fusion analysis.

[0033] The improved TFT-multi algorithm, based on the traditional TFT algorithm, changes the output layer to a dual-output structure (outputting the health status assessment result and the deterioration risk value respectively), optimizes the loss function to a weighted sum of the cross-entropy loss function and the MAE loss function, and introduces a multivariate joint training mechanism to solve the defect of the traditional algorithm's independent modeling of a single parameter. The output results of the AI ​​fusion analysis module include two parts: first, the current health status assessment result of the human body, which is divided into four levels: normal, basically normal, abnormal, and severely abnormal; second, the early deterioration risk assessment result of the disease, including the deterioration risk value (between 0 and 1) and the prediction of the evolution trend of vital signs parameters, predicting the change trend of each vital sign parameter in the next 24-72 hours, providing data support for risk warning.

[0034] The module's fusion analysis latency is strictly controlled within 30 seconds to ensure the timeliness of the analysis results. At the same time, the module transmits the fusion analysis results to the risk warning module, data storage module, and model self-optimization module in real time for subsequent warning, storage, and model optimization.

[0035] 5. Risk Warning Module This module receives health status assessment results and deterioration risk assessment results from the AI ​​fusion analysis module. It is used to complete the graded early warning of early deterioration risk, generate warning information, and transmit it to relevant modules. Specifically, it includes a risk grading unit, a warning threshold setting unit, and a warning information generation unit. The risk grading unit classifies early deterioration risk into four levels based on the deterioration risk value output by the AI ​​fusion analysis module. The grading standards are clear and fixed: Normal level (risk value < 0.2), indicating that the human body's health status is stable and there is no obvious risk of deterioration; Attention level (0.2 ≤ risk value < 0.4), indicating that the human body's health status is basically stable, but some vital signs and parameters are slightly abnormal, requiring enhanced monitoring; Warning level (0.4 ≤ risk value < 0.7), indicating that there is an early deterioration risk and vital signs and parameters are significantly abnormal, requiring timely intervention; Critical level (risk value ≥ 0.7), indicating that there is a serious risk of deterioration and the possibility of acute illness, requiring immediate emergency measures.

[0036] The warning threshold setting unit supports personalized threshold adjustment based on individual differences. Medical staff or workers can adjust the warning threshold for each risk level through the terminal interaction module according to factors such as the patient's age, physical condition, medical history, and treatment plan. The adjustment step size is set to 0.05 to ensure the targeted nature of the warning. At the same time, the unit presets a set of standard warning thresholds, which are applicable to the general population without special medical history, as the default threshold.

[0037] The early warning information generation unit generates corresponding early warning information based on the risk classification results. The early warning information includes the risk level, the name and specific value of abnormal vital signs parameters, the trend of changes in vital signs parameters, a description of the risk of deterioration, and preliminary intervention suggestions (such as medical guidance, medication adjustment suggestions, lifestyle adjustment suggestions, etc.). Different risk levels correspond to different early warning methods: the normal level does not generate obvious early warning information, but only displays the health status on the terminal interaction module; the attention level generates a yellow prompt message with no sound alarm; the early warning level generates an orange early warning message accompanied by a low-volume sound alarm; and the critical level generates a red emergency early warning message accompanied by a high-volume sound alarm, while also triggering a light alarm.

[0038] After the warning information is generated, it is transmitted in real time to the multi-terminal linkage module and the terminal interaction module, and backed up to the data storage module to ensure that the warning information is traceable.

[0039] 6. Data storage module This module stores all data generated during system operation, including raw vital sign data, preprocessed valid data, extracted feature data, AI fusion analysis results, risk warning records, medical knowledge graph data, model parameters, and terminal operation records. It employs a dual-storage architecture of "local + cloud" to balance data security, timeliness, and scalability. Local storage uses solid-state drives (SSDs) to store real-time vital signs data, early warning records, and commonly used medical knowledge graph data for the past 30 days. The storage capacity is no less than 1TB, and the read speed is no less than 500MB / s to ensure efficient data reading during terminal interaction. The local storage device has data encryption to prevent data leakage and also supports power failure protection to avoid data loss.

[0040] The cloud storage employs a distributed database deployed on a dedicated medical cloud server to store historical vital sign data, AI fusion analysis results, complete medical knowledge graph data, model parameters, and all operation records, supporting massive data storage and efficient retrieval. The cloud storage uses the AES encryption algorithm to encrypt data throughout the transmission and storage process, and also supports regular backups at a frequency of 24 hours per backup. Backup data is stored on servers in different regions to prevent data loss. The data storage period is no less than 5 years, complying with relevant medical data retention regulations, facilitating subsequent clinical traceability, research analysis, and dispute resolution.

[0041] This module supports bidirectional data transmission. It can read and transmit relevant data according to the instructions of the terminal interaction module. At the same time, it regularly organizes the stored data, deletes invalid data, frees up storage space, and ensures smooth system operation.

[0042] 7. Terminal Interaction Module This module serves as the system's human-computer interaction interface, enabling staff and medical personnel to interact with the system. Functions include data viewing, parameter setting, command input, and alert confirmation. Specifically, it comprises a display unit, an input unit, and a feedback unit. The display unit uses a high-definition touch screen with a resolution of no less than 1920×1080. It can display the patient's basic information, current vital signs parameters, preprocessing results, feature extraction results, AI fusion analysis results, risk warning information, and historical data trend charts in real time. The display interface adopts a partitioned design, which is clear and intuitive. It supports the simultaneous display of information from multiple patients (up to 16 patients), which is convenient for medical staff to monitor in batches. At the same time, it supports data export, which can export the required data to Excel, PDF and other formats for clinical records and scientific research analysis.

[0043] The input unit includes three input methods: touch input, keyboard input, and voice input. Staff can use the input unit to set system parameters (such as sampling frequency, early warning threshold, storage period, etc.), input basic patient information (such as age, gender, medical history, treatment plan, etc.), and issue commands (such as model training commands, data backup commands, early warning confirmation commands, etc.). Voice input supports Mandarin Chinese recognition with an accuracy rate of no less than 95%, facilitating quick operation by staff.

[0044] The feedback unit is used to receive the operating status information of each module of the system. When the system malfunctions (such as abnormal acquisition equipment, data transmission failure, storage failure, etc.), it generates fault prompt information in a timely manner, displays it through the display unit and is accompanied by an audible alarm to remind staff to handle it in time. At the same time, it receives the early warning confirmation instruction, transmits the confirmation result to the risk early warning module and the multi-terminal linkage module, and stops the relevant early warning operation.

[0045] 8. Medical Knowledge Graph Module This module provides clinical knowledge support for the system's AI fusion analysis and risk warning, and is used to store and manage clinical diagnosis and treatment-related knowledge data, specifically including a clinical rule base, a vital sign association base, and an intervention suggestion base. The clinical rule base stores clinical diagnosis and treatment guidelines, criteria for judging abnormal signs, and rules for judging the deterioration of the condition related to intensive care, postoperative rehabilitation, and chronic disease management (such as hypertension, diabetes, and coronary heart disease). The data comes from clinical guidelines of authoritative medical institutions and a large number of clinical cases, ensuring the authority and accuracy of the knowledge.

[0046] The vital sign association database stores the physiological correlations and pathological influence patterns between different vital sign parameters, including the correlation patterns of vital signs under normal physiological conditions and the correlation patterns of abnormal vital signs under pathological conditions, such as the correlation between elevated blood pressure and increased heart rate, and the correlation between decreased blood oxygen saturation and abnormal respiratory rate, providing support for the correlation feature verification and fusion analysis of the AI ​​fusion analysis module.

[0047] The intervention suggestion database stores personalized intervention plans, medication references, and medical guidance corresponding to different risk levels and abnormal signs, including intervention measures corresponding to the attention level, warning level, and critical level, as well as personalized intervention suggestions for patients with different chronic diseases and postoperative conditions, to ensure that intervention measures after risk warning are scientific and reasonable.

[0048] This module supports regular updates of knowledge data, with an update frequency of no less than once a month. Staff can add, modify, and delete knowledge data through the terminal interaction module to ensure the timeliness of the knowledge data and adapt to the development of clinical diagnosis and treatment technologies. At the same time, this module supports knowledge retrieval. The AI ​​fusion analysis module can retrieve relevant knowledge data in real time according to analysis needs, as a basis for analysis.

[0049] 9. Model self-optimization module This module is used to achieve adaptive optimization of the AI ​​fusion analysis model, ensuring that the model's analytical accuracy and early warning performance continuously meet the needs of clinical monitoring. Specifically, it includes a data incremental training unit, a model performance evaluation unit, and a parameter adaptive adjustment unit. The incremental data training unit regularly uses newly added vital sign data, clinical case data, and early warning feedback data from the system to incrementally train the AI ​​fusion analysis model, with a training cycle of 7 days. During incremental training, the original effective parameters of the model are retained, and only the weight parameters of the model are adjusted to avoid overfitting and improve the model's adaptability to different populations and scenarios. The training data uses preprocessed effective data and clinical case data stored in the data storage module to ensure the authenticity and effectiveness of the training data.

[0050] The model performance evaluation unit uses MAE error (mean absolute error), accuracy, and early warning sensitivity as evaluation metrics for model performance. MAE error is used to evaluate the accuracy of model fusion analysis, accuracy is used to evaluate the accuracy of health status assessment, and early warning sensitivity is used to evaluate the timeliness and accuracy of risk warnings. Preset evaluation thresholds are: MAE error ≤ 0.05, accuracy ≥ 95%, and early warning sensitivity ≥ 90%. After each incremental training iteration, model performance is evaluated. If all evaluation metrics meet the preset thresholds, model optimization is complete. If any metric fails to meet the preset threshold, the parameter adaptive adjustment unit is triggered.

[0051] The parameter adaptive adjustment unit uses the Adam optimizer with a learning rate set to 1e-3. Based on the model performance evaluation results, it automatically adjusts the weight parameters, loss function parameters, and fusion layer parameters of the AI ​​fusion analysis model. After adjustment, the model performance is evaluated again until all evaluation indicators meet the preset thresholds. At the same time, the unit stores the optimized parameters to the data storage module, replacing the original model parameters, to ensure that the model is always in the optimal operating state. Staff can view the model performance evaluation results and parameter adjustment records through the terminal interaction module, and can also manually trigger model optimization operations.

[0052] 10. Multi-terminal linkage module This module is used to realize the synchronous push and multi-terminal linkage of early warning information, ensuring that early warning information is delivered to relevant personnel in a timely manner and improving emergency response efficiency. Specifically, it includes a medical terminal linkage unit, a family terminal linkage unit, and an emergency terminal linkage unit. The medical terminal linkage unit is used to push risk warning information and related data to medical staff workstations, nurse station terminals, and doctors' mobile terminals (such as mobile phones and tablets). The pushed information includes basic patient information, risk level, abnormal vital signs parameters, AI fusion analysis results, vital sign change trends, and preliminary intervention suggestions. At the same time, it supports medical staff to provide feedback on intervention measures through the medical terminal. The feedback information is transmitted to the terminal interaction module and data storage module in real time to realize the closed loop of diagnosis and treatment.

[0053] The family member terminal linkage unit pushes risk warning information to the patient's family members' mobile terminals via a dedicated mobile APP. The information pushed includes the risk level, warning prompts, and contact suggestions (such as contacting the attending physician, going to the hospital for treatment, etc.). At the same time, the family members can also view the patient's real-time vital signs data and health status through the APP to understand the changes in the patient's condition. The APP supports message push reminders to ensure that family members receive warning information in a timely manner.

[0054] When the risk warning module triggers a critical level warning, the emergency terminal linkage unit automatically links with the 120 emergency medical system and the emergency terminals of the nearest medical institutions. The information pushed includes the patient's basic information, current vital signs data, AI fusion analysis results, the patient's precise location, and medical history information. At the same time, the local storage module of the linkage system quickly transmits the patient's vital signs data and analysis results from the past 72 hours to the emergency terminal, providing emergency personnel with accurate diagnostic and treatment references, shortening emergency preparation time, and improving emergency efficiency.

[0055] The information push delay of each linkage unit is strictly controlled within 10 seconds to ensure that the early warning information is delivered in a timely manner. At the same time, the module supports linkage status monitoring. When the linkage of a certain end fails, a linkage failure prompt is generated in a timely manner and transmitted to the terminal interaction module to remind the staff to manually push the early warning information.

[0056] 11. Overall System Workflow The overall workflow of this system is as follows: 1) The multimodal vital sign acquisition module collects multidimensional vital sign data and auxiliary data of the human body and transmits them to the data preprocessing module; 2) The data preprocessing module standardizes the raw data, repairs missing values, and removes outliers, outputting valid vital sign data and backing it up to the data storage module; 3) The feature extraction module extracts temporal and correlation features from the effective vital sign data, generates feature vectors, and transmits them to the AI ​​fusion analysis module; 4) The AI ​​fusion analysis module calls the knowledge data of the medical knowledge graph module, uses the improved TFT-multi algorithm to perform fusion analysis on the feature vectors, outputs the health status assessment results and deterioration risk assessment results, and transmits them to the risk warning module, data storage module and model self-optimization module; 5) The risk warning module completes risk classification based on the deterioration risk assessment results, generates warning information, transmits it to the multi-terminal linkage module and terminal interaction module, and backs it up to the data storage module; 6) The multi-terminal linkage module will simultaneously push the early warning information to the medical terminal, family terminal and emergency terminal (in case of critical level early warning). 7) The terminal interaction module displays relevant data and warning information, receives operation instructions from staff, and provides feedback on operation results; 8) The model self-optimization module regularly uses new data to incrementally train, evaluate performance, and adjust parameters of the AI ​​fusion analysis model to ensure optimal model performance; 9) The data storage module stores all data during the entire system operation process and supports data reading, backup, and export.

[0057] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-parameter vital sign AI fusion analysis and early warning system for deterioration risk, characterized in that, It includes a multimodal vital sign acquisition module, a data preprocessing module, a feature extraction module, an AI fusion analysis module, a risk warning module, a data storage module, a terminal interaction module, a medical knowledge graph module, a model self-optimization module, and a multi-terminal linkage module; The output of the multimodal vital sign acquisition module is electrically connected to the input of the data preprocessing module. The output of the data preprocessing module is electrically connected to the input of the feature extraction module. The output of the feature extraction module is electrically connected to the input of the AI ​​fusion analysis module. The AI ​​fusion analysis module is bidirectionally electrically connected to the risk warning module, the medical knowledge graph module, and the model self-optimization module. The output of the risk warning module is electrically connected to the input of the multi-terminal linkage module. The data storage module is bidirectionally electrically connected to the data preprocessing module, the AI ​​fusion analysis module, and the medical knowledge graph module. The terminal interaction module is bidirectionally electrically connected to the data storage module, the risk warning module, and the model self-optimization module. All modules work together to complete multi-parameter vital sign fusion analysis and early deterioration risk warning.

2. The multi-parameter vital signs AI fusion analysis and early deterioration risk warning system according to claim 1, characterized in that: The multimodal vital sign acquisition module includes a contact acquisition unit, a non-contact acquisition unit, and an auxiliary acquisition unit; The contact-type acquisition unit uses medical-grade biosensors to collect data on heart rate, blood pressure, blood oxygen saturation, body surface temperature, respiratory rate, and heart rate variability. The non-contact acquisition unit uses a high-sensitivity PVDF piezoelectric thin film sensor and millimeter-wave radar to acquire human cardiac impact signals, respiratory body movement signals and bed exit status data based on BCG technology, without the need for human body wear; The auxiliary data acquisition unit is used to collect data on human activity levels, body position, and ambient temperature and humidity. The sampling frequency of each acquisition unit is uniformly set to 15 minutes / time, and automatically switches to 1 minute / time in emergency situations. The acquired data is converted into digital signals by A / D converters and then transmitted to the data preprocessing module.

3. The multi-parameter vital sign AI fusion analysis and early deterioration risk warning system according to claim 1, characterized in that: The data preprocessing module includes a standardization processing unit, a missing value repair unit, and an outlier removal unit. The standardization processing unit uses the Z-score standardization algorithm to convert vital sign data of different dimensions into normalized data with a unified standard, and the normalization range is [0,1]. The missing value repair unit uses an improved interpolation algorithm based on masking technology to fill in missing values ​​by combining adjacent time-series data and baseline data of the same type of population, with a filling error of no more than 5%. The outlier removal unit uses the 3σ principle combined with Granger causality test to identify and remove interfering data generated during the data collection process, while retaining valid vital sign data.

4. The multi-parameter vital signs AI fusion analysis and early deterioration risk warning system according to claim 1, characterized in that: The feature extraction module includes a temporal feature extraction unit and a correlation feature extraction unit; The time-series feature extraction unit uses wavelet transform algorithm to extract the time-domain and frequency-domain features of each vital sign data. The time-domain features include mean, variance, peak value, valley value and rate of change, while the frequency-domain features include power spectral density and characteristic frequency. The correlation feature extraction unit uses an attention mechanism to extract correlation features between different vital signs parameters, quantifies the influence weight of each vital sign parameter on the risk of disease deterioration, and retains three decimal places in the weight calculation.

5. The multi-parameter vital signs AI fusion analysis and early deterioration risk warning system according to claim 1, characterized in that: The AI ​​fusion analysis module adopts an improved TFT-multi temporal fusion transformer algorithm. This algorithm optimizes the output layer and loss function based on the traditional TFT algorithm, introduces a multivariate joint training mechanism, and simultaneously receives temporal features and correlation features output by the feature extraction module and clinical knowledge data output by the medical knowledge graph module. It completes end-to-end joint fusion analysis of multi-parameter vital signs, outputs the evolution trend of vital sign parameters and the correlation of deterioration risk, and the fusion analysis delay does not exceed 30 seconds.

6. The multi-parameter vital signs AI fusion analysis and early deterioration risk warning system according to claim 1, characterized in that: The risk warning module includes a risk classification unit, a warning threshold setting unit, and a warning information generation unit; The risk grading unit divides the risk of early deterioration into four levels: normal, attention, warning, and critical. The warning threshold setting unit supports personalized threshold adjustment based on individual medical history, age, and physical characteristics, with an adjustment step size of 0.

05. The early warning information generation unit generates corresponding early warning information based on the risk level, including the risk level, abnormal vital signs parameters, changing trends, and preliminary intervention suggestions.

7. The multi-parameter vital signs AI fusion analysis and early deterioration risk warning system according to claim 1, characterized in that: The data storage module adopts a "local + cloud" dual storage architecture; the local storage uses a solid-state drive to store real-time vital signs data and early warning records for the past 30 days, with a storage capacity of no less than 1TB; The cloud storage uses a distributed database to store historical vital sign data, AI fusion analysis results, medical knowledge graph data and model parameters. It supports encrypted data transmission and backup, with a backup frequency of 24 hours per backup and a data storage period of no less than 5 years, meeting medical data retention standards.

8. The multi-parameter vital signs AI fusion analysis and early deterioration risk warning system according to claim 1, characterized in that: The medical knowledge graph module includes a clinical rule base, a vital sign association base, and an intervention suggestion base. The clinical rule base stores clinical diagnosis and treatment guidelines and criteria for judging abnormal signs related to intensive care and chronic disease management; The vital sign association database stores the physiological correlations and pathological influence patterns among different vital sign parameters; The intervention suggestion library stores personalized intervention plans, medication references, and medical guidance corresponding to different risk levels and abnormal signs; The medical knowledge graph module supports regular updates, with an update frequency of no less than once a month, to ensure the timeliness of the knowledge.

9. The multi-parameter vital signs AI fusion analysis and early deterioration risk warning system according to claim 1, characterized in that: The model self-optimization module includes a data incremental training unit, a model performance evaluation unit, and a parameter adaptive adjustment unit. The incremental data training unit regularly uses newly added vital sign data and clinical case data to incrementally train the AI ​​fusion analysis model, with a training cycle of 7 days. The model performance evaluation unit uses MAE error, accuracy and early warning sensitivity as evaluation indicators. When any indicator fails to meet the preset threshold, parameter adjustment is triggered. The parameter adaptive adjustment unit uses the Adam optimizer with a learning rate set to 1e-3 to automatically adjust the model weight parameters, ensuring continuous optimization of model analysis accuracy and early warning performance.

10. The multi-parameter vital signs AI fusion analysis and early deterioration risk warning system according to claim 1, characterized in that: The multi-terminal linkage module includes a medical terminal linkage unit, a family terminal linkage unit, and an emergency terminal linkage unit; The medical terminal linkage unit is used to push early warning information to the workstations of medical staff and simultaneously transmit abnormal vital signs data and fusion analysis results; The family member terminal linkage unit pushes early warning information to the patient's family members via a mobile APP, including the risk level and contact suggestions; When a critical-level warning is triggered, the emergency terminal linkage unit automatically links with the 120 emergency medical system and the nearest medical institution to push the patient's location, vital signs data and medical history information, so as to achieve rapid emergency response. The information push delay of each linkage unit shall not exceed 10 seconds to ensure that the early warning information is delivered in a timely manner.

Images