Intensive care multi-dimensional heterogeneous data processing and adaptive early warning system

Abstract

The application discloses a kind of intensive care multidimensional heterogeneous data processing and adaptive early warning system, including data acquisition and preprocessing layer, multiscale time sequence feature extraction and fusion module, interpretable dynamic rule-neural hybrid decision engine, adaptive alarm grading and intelligent push module and feedback closed-loop learning module connected in turn;Wherein: the data acquisition and preprocessing layer is used to automatically collect discrete nursing operation event data packet in hospital information system by hospital information system interface, simultaneously parallelly collects the multimodal physiological parameters and waveform data of multiple bedside monitoring equipment, and carries out time stamp alignment based on sliding time window to the nursing operation event data packet and the physiological parameters and waveform data, and triggers historical state recalculation when detecting that the nursing operation event data packet exists entry delay.

Description

Technical Field

[0001] This invention relates to a multidimensional heterogeneous data processing and adaptive early warning system for intensive care, belonging to the field of medical monitoring. Background Technology

[0002] The Intensive Care Unit (ICU) is the core facility for treating critically ill patients, where medical staff need to monitor various physiological indicators in real time. Existing monitoring systems primarily rely on bedside monitors with fixed thresholds to trigger alarms. When parameters exceed these thresholds, alerts are sent to the nurses' station or mobile devices via the network. With the development of artificial intelligence technology, some systems are beginning to incorporate deep learning models to fuse and analyze multiple parameters, thereby improving the lead time for early warnings.

[0003] However, existing monitoring and early warning technologies still face significant challenges in clinical applications. First, alarm fatigue is a major problem. A high percentage of false alarms are caused by technical issues such as sensor detachment, patient movement, or equipment interference, leading to desensitization among medical staff and potentially masking genuine emergencies. Second, while pure deep learning models possess high nonlinear fitting capabilities, their "black box" nature makes early warning results lack medical interpretability, hindering rapid decision-making by medical staff based on a single probability output. Furthermore, existing alarm push mechanisms are relatively simplistic, failing to adequately consider the real-time workload and geographical location of medical staff, resulting in uneven distribution of push information or delayed response. Therefore, improvements to existing technologies are urgently needed to address the shortcomings in accuracy, interpretability, and effectiveness of early warnings in ICU environments. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings and deficiencies of the existing technology and to provide a multidimensional heterogeneous data processing and adaptive early warning system for intensive care, aiming to solve the above problems.

[0005] A multidimensional heterogeneous data processing and adaptive early warning system for intensive care includes a data acquisition and preprocessing layer, a multi-scale temporal feature extraction and fusion module, an interpretable dynamic rule-neural hybrid decision engine, an adaptive alarm grading and intelligent push module, and a feedback closed-loop learning module, all connected in sequence. The data acquisition and preprocessing layer is used to automatically acquire discrete nursing operation event data packets from the hospital information system via a hospital information system interface, while simultaneously acquiring multimodal physiological parameters and waveform data from multiple bedside monitoring devices. It performs timestamp alignment of the nursing operation event data packets with the physiological parameters and waveform data based on a sliding time window, and triggers historical state overwrite and recalculation when an input delay is detected in the nursing operation event data packets. The multi-scale temporal feature extraction and fusion module is used to extract physiological features from different time windows and generate multi-dimensional feature vectors. The interpretable dynamic rule-neural hybrid decision engine receives the multidimensional feature vector, outputs a first risk probability and feature importance weight distribution through the neural network submodule, and outputs a second risk assessment based on clinical guideline baselines through the symbol rule library submodule; and merges the first risk probability and the second risk assessment according to the set conflict arbitration strategy to obtain the final risk level, and matches medical knowledge graph nodes based on the feature importance weight distribution to generate natural language explanation text. The adaptive alarm grading and intelligent push module receives the final risk level and explanation text, combines the real-time acquired quantitative value of medical staff workload and spatial positioning data, dynamically calculates the optimal receiving terminal and delivery delay time, and performs targeted grading push. The feedback closed-loop learning module is used to receive labeled feedback on the early warning results from the medical staff and to incrementally update the neural network submodule.

[0006] Preferably, the data acquisition and preprocessing layer includes a morphology-based artifact recognition submodule. Before feature extraction, this submodule is used to identify and remove physiological waveform artifact signals caused by patient movement or equipment detachment. The artifact recognition submodule is located in the upstream data link of the multi-scale temporal feature extraction and fusion module to block technical false alarms at the source.

[0007] Furthermore, the multi-scale temporal feature extraction and fusion module adopts a three-layer edge-end-cloud architecture, including: an edge computing unit deployed at the bedside monitoring device or a dedicated edge box, running a lightweight temporal model to extract short-term abnormal features at the second to minute level; a ward-level server used to fuse 12 to 24-hour medium- and long-term trend features of multiple patients in the ward; and a cloud / hospital-level server used for cross-ward data aggregation, knowledge distillation, and global model updates. The three-layer structure interacts with data through independent data synchronization interface units to achieve hierarchical isolation.

[0008] Preferably, the conflict arbitration strategy specifically includes: when the deviation between the first risk probability and the second risk assessment exceeds a preset threshold, if the feature importance weight distribution is concentrated on a new feature dimension not defined by the rule base, then the first risk probability is used as the main factor, and the result is marked as an exploratory warning; if the feature importance weight distribution triggers the mandatory safety red line set by the symbol rule base, then the symbol rule base will perform a mandatory overwrite, and the second risk assessment will be used as the final risk level.

[0009] Furthermore, the adaptive alarm grading and intelligent push module also includes an expected alarm silencing unit; when the data acquisition and preprocessing layer receives a specific nursing operation event input, the expected alarm silencing unit automatically sets a dynamic time window for the associated physiological parameters directly affected by the nursing operation, and within the time window, it shields or downgrades non-fatal expected physiological parameter fluctuation alarms.

[0010] Preferably, the adaptive alarm grading and intelligent push module further includes a timeout escalation unit; when the assigned preferred receiving terminal fails to generate confirmation feedback within the set time window, or when the real-time workload quantification value of the preferred medical staff is higher than the full load threshold, the timeout escalation unit automatically transfers the warning information horizontally to the nearest secondary idle medical staff terminal based on spatial positioning data, or vertically escalates it to the superior management terminal.

[0011] Furthermore, the intelligent push module categorizes alarms into three types—extremely high risk, medium risk, and low risk—based on the final risk level, and executes different push channel strategies: extremely high risk events are pushed concurrently via bedside audio-visual alarms, nurse station main screen display, strong prompts on mobile terminals, and voice broadcasts; medium risk events are pushed via mobile terminals and bedside audio-visual alarms with delayed triggering; low risk events are only recorded in the background log and archived with explanatory text, without actively triggering push notifications.

[0012] More specifically, the process of generating natural language explanation text includes: extracting features whose absolute value is greater than a preset threshold as core control features; using the core control features and the corresponding positive and negative signs of SHAP values ​​as search conditions; mapping them to the causal relationship chain triple nodes of the control knowledge graph; and using the built-in template engine combined with historical running feature data to automatically assemble and generate structured multidimensional anomaly prompt information.

[0013] Preferably, the feedback closed-loop learning module includes a feedback confidence assessment unit; the feedback confidence assessment unit is used to calculate the frequency and response latency of the annotation operation on the medical staff end; when the frequency of ignored or false alarm annotations per unit time is higher than the preset fatigue blind spot threshold, it is determined that the confidence of the annotation feedback is low, and the weight of the batch of feedback data in the incremental update is reduced, so as to prevent the erroneous data caused by fatigue operation from polluting the model.

[0014] Furthermore, the feedback closed-loop learning module also includes an objective backtracking verification unit; the objective backtracking verification unit is used to perform secondary verification and alignment of the historical subjective annotation feedback after obtaining the patient's final electronic medical record outcome data; if there is a conflict between the electronic medical record outcome data and the subjective annotation, the electronic medical record outcome data is used as the objective gold standard to generate error correction samples and feed them back to the edge computing unit for lightweight federated learning updates.

[0015] The core of this invention, a multidimensional heterogeneous data processing and adaptive early warning system for intensive care, lies in constructing a hybrid decision-making and perception-based push closed loop that deeply integrates clinical logic. Compared to existing technologies, this invention offers significant advantages: By employing edge artifact recognition and a rule-based neural arbitration strategy, false alarms caused by equipment interference are significantly reduced, alleviating alarm fatigue. Simultaneously, the system's output natural language explanation transforms abstract algorithmic probabilities into concrete medical descriptions, enabling medical staff to quickly assess the authenticity of alerts and enhancing clinical trust. Furthermore, a dynamic push mechanism based on spatial location and workload optimizes resource flow efficiency within the ICU. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.

[0017] Figure 1 This is a system module diagram of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0019] It should be noted that all uses of "first" and "second" in the embodiments of the present invention are for the purpose of distinguishing two entities or parameters with the same name but different names. It is clear that "first" and "second" are only for the convenience of expression and should not be construed as limiting the embodiments of the present invention. Subsequent embodiments will not explain this in detail.

[0020] The directional and positional terms used in this invention, such as "up," "down," "front," "back," "left," "right," "inner," "outer," "top," "bottom," and "side," are merely for reference to the accompanying drawings. Therefore, the directional and positional terms used are for illustrating and understanding this invention, and not for limiting the scope of protection of this invention.

[0021] This application proposes a multidimensional heterogeneous data processing and adaptive early warning system for intensive care, see reference [link to relevant documentation]. Figure 1 The system module diagram of this invention includes, in sequence, a data acquisition and preprocessing layer, a multi-scale temporal feature extraction and fusion module, an interpretable dynamic rule-neural hybrid decision engine, an adaptive alarm grading and intelligent push module, and a feedback closed-loop learning module; wherein: The data acquisition and preprocessing layer is used to automatically acquire discrete nursing operation event data packets from the hospital information system through the hospital information system interface, and simultaneously acquire multimodal physiological parameters and waveform data from multiple bedside monitoring devices, and perform timestamp alignment and preliminary cleaning based on a sliding time window on the nursing operation event data packets and the physiological parameters and waveform data. The multi-scale temporal feature extraction and fusion module is used to extract physiological features from different time windows and generate multi-dimensional feature vectors. The interpretable dynamic rule-neural hybrid decision engine receives the multi-dimensional feature vector, outputs a first risk probability and feature importance weight distribution through the neural network submodule, and outputs a second risk assessment based on the operation guide baseline through the symbol rule library submodule; and merges the first risk probability and the second risk assessment according to the set conflict arbitration strategy to obtain the final risk level, and extracts core control features based on the feature importance weight distribution, matches the causal relationship chain triple nodes in the control knowledge graph, and generates natural language anomaly explanation prompt information through template filling; The adaptive alarm grading and intelligent push module receives the final risk level and anomaly explanation prompt information, and combines the real-time acquired quantitative value of the workload of the staff at the system receiving end and spatial positioning data to dynamically calculate the optimal receiving terminal and delivery delay time, and execute targeted grading push; The feedback closed-loop learning module is used to receive the labeled feedback of the warning results from the receiving end and to incrementally update the neural network submodule.

[0022] In practical applications, timestamp alignment of the data acquisition and preprocessing layer refers to unifying the time base of data from different bedside monitoring devices. This can be achieved by using a network time protocol to synchronize the global clock, or by adjusting the position of discrete sampling points through a linear interpolation algorithm. Specifically, the timestamp alignment includes: using the standard execution timestamp in the acquired nursing operation event data packet as the center, constructing a sliding time window by extending a preset span in both the historical and future directions, and extracting timing parameters falling within the sliding time window for alignment; when the gateway receives a timestamp later than the standard execution timestamp by more than a preset delay threshold, triggering a historical state overlay recalculation, the system automatically backtracks the calculation clock to the historical moment corresponding to the standard execution timestamp, and re-calls the parameters at that moment for overlay calculation; preliminary cleaning can use digital filters to remove high-frequency noise interference, or remove outlier data points based on statistical distribution thresholds, thereby blocking false alarms caused by device asynchrony or signal distortion from the data source.

[0023] Furthermore, the multi-scale time-series feature extraction and fusion module can extract physiological features from different time windows by using wavelet transform to decompose the signal to capture multi-scale dynamic changes, or by calculating statistics such as mean and variance through a sliding window; the generation of multi-dimensional feature vectors can integrate short-term and long-term representations using feature splicing methods. Specifically, this process is implemented as parallel extraction of second-level instantaneous fluctuations and hour-level trend evolution, avoiding the limitations of feature representation under a single time scale.

[0024] This application reduces technical false alarms through timestamp alignment and cleaning mechanisms in the data acquisition and preprocessing layers; enhances anomaly correlation representation through multi-time window feature capture in the multi-scale temporal feature extraction and fusion module; provides transparent decision-making basis through dual-path decision-making and natural language anomaly explanation prompts in the interpretable dynamic rule-neural hybrid decision engine; achieves accurate information delivery by combining workload and spatial positioning in the adaptive alarm grading and intelligent push module; and forms a continuous optimization closed loop through annotation feedback and model updates in the feedback closed-loop learning module. In this way, it systematically solves the problems of alarm fatigue, lack of data interpretability of early warning results, and push mechanism not adapting to the staff status caused by sensor detachment or equipment interference in the intensive care unit.

[0025] Specifically, the architecture is designed to block false alarms at the data source, balance data-driven and operational norms through hybrid decision-making, dynamically adjust the push strategy based on real-time context, and continuously iterate the model using closed-loop feedback, thereby improving personnel decision-making efficiency while ensuring the accuracy of early warnings. As a preferred implementation, the collaborative work of the above modules is realized as an end-to-end intelligent early warning ecosystem, where feature extraction and decision-making processes are strictly confined within an operational knowledge framework, and the dynamic adjustment of the push strategy is constrained to personnel availability boundaries, thus ensuring the system's practicality in complex ICU environments.

[0026] This application further proposes that the data acquisition and preprocessing layer includes a morphology-based artifact recognition submodule; before feature extraction, this submodule is used to identify and remove physiological waveform artifact signals caused by non-physiological sudden movements or equipment detachment, and the artifact recognition submodule is set in the upstream data link of the multi-scale temporal feature extraction and fusion module to block technical false alarms from the source.

[0027] In practical applications, the morphology-based artifact recognition submodule refers to a component that uses waveform shape feature analysis technology to identify artifact signals. It can be implemented using morphological filters, wavelet transforms, or feature extractors based on convolutional neural networks. The purpose is to accurately distinguish between non-physiological artifacts caused by external physical motion or equipment detachment and real physiological fluctuations, avoiding misjudgments by traditional thresholding methods. Identifying and removing physiological waveform artifact signals can be understood as actively removing noise signals during the data preprocessing stage. This can include signal mutation detection, amplitude anomaly analysis, or rhythm irregularity assessment. The purpose is to ensure that subsequent feature extraction is based on clean signal data and prevent artifacts from interfering with the generation of feature vectors. Specifically, the artifact recognition submodule is located in the upstream data link, which can be understood as this submodule being located within the data acquisition and preprocessing layer and in the data flow path before the multi-scale temporal feature extraction and fusion module. Its purpose is to block the propagation of technical false alarms at the source and prevent noise signals from spreading to subsequent decision-making stages.

[0028] Specifically, the solution in this application deploys a morphology-based artifact recognition submodule within the data acquisition and preprocessing layer, placing it upstream of the multi-scale temporal feature extraction and fusion module in the data flow path. This allows for real-time artifact recognition and removal before physiological waveform data enters the feature extraction stage. This submodule analyzes waveform structural anomalies (such as abrupt changes, irregular amplitudes, or non-physiological rhythms) rather than relying on a single threshold, effectively distinguishing between technical artifacts and genuine physiological fluctuations. When artifact signals caused by external motion or equipment detachment are detected, they are immediately removed from the data stream, ensuring that the signal input to the multi-scale temporal feature extraction and fusion module is clean data. This architecture embeds artifact processing into the initial stage of data preprocessing, blocking the generation path of technical false alarms at the source and preventing noisy data from contaminating subsequent multi-dimensional feature vectors. This proactive prevention of false alarms rather than post-event correction is achieved in the overall early warning process.

[0029] This application further proposes that the multi-scale temporal feature extraction and fusion module adopts a three-layer edge-end-cloud architecture, including: an edge computing unit deployed at the bedside monitoring device or a dedicated edge box, running a lightweight temporal model to extract short-term abnormal features at the second to minute level; a ward-level server used to fuse 12 to 24-hour medium- and long-term trend features from multiple device channels; and a cloud / hospital-level server used for cross-channel data aggregation, knowledge distillation, and global model updates; and the three-layer structure interacts with data through an independent data synchronization interface unit to achieve hierarchical isolation.

[0030] Specifically, the solution in this application directly processes raw physiological waveforms at the bedside device via edge computing units, quickly identifying short-term abnormal features at the second to minute level and eliminating artifact signals, thereby reducing the amount of invalid data transmitted to the central node; the ward-level server receives feature data from multiple edge nodes, and integrates the correlation patterns between multiple channels based on a 12 to 24-hour time window to identify systemic risk trends, avoiding interference from short-term fluctuations of a single node with the mid-term assessment; the cloud / hospital-level server aggregates cross-channel feature data and performs knowledge distillation, feeding back global model update parameters to lower-level nodes to adapt to different operating environment characteristics; the three-layer structure conducts controllable data interaction through independent data synchronization interface units, ensuring seamless integration of short-term abnormal features and medium- to long-term trend features under hierarchical isolation conditions, forming a complete multi-scale feature support system.

[0031] This application further proposes that the conflict arbitration strategy specifically includes: when the deviation between the first risk probability and the second risk assessment exceeds a preset threshold, if the feature importance weight distribution is concentrated on a new feature dimension not defined by the rule base, then the first risk probability is used as the main factor, and the result is marked as an exploratory warning; if the feature importance weight distribution triggers the mandatory safety red line set by the symbol rule base, then the symbol rule base will perform a mandatory overwrite, and the second risk assessment will be used as the final risk level.

[0032] Specifically, the proposed solution constructs conditional branching logic within the conflict arbitration strategy. When the deviation between the first risk probability and the second risk assessment exceeds a preset threshold, the system automatically triggers the feature importance weight distribution analysis process. Based on the determination result of the distribution concentration location, if the weights are highly concentrated on novel feature dimensions not defined by the rule base, the first risk probability output by the neural network submodule is adopted as the final risk level, and an exploratory warning mark is added to prompt personnel for manual verification. If the weight distribution activates the mandatory safety red line set by the symbolic rule base, the neural network output path is immediately interrupted, and the risk assessment result is forcibly overwritten by the symbolic rule base. This mechanism enables the system to adaptively adjust the decision weights according to the essential source of the deviation, forming a dynamic balance between novel risk exploration and safety norms. This avoids the underreporting problem caused by the rigidity of the rule base and ensures strict adherence to operational safety principles in critical scenarios that endanger safety.

[0033] This application further proposes that the adaptive alarm grading and intelligent push module also includes an expected alarm silencing unit; when the data acquisition and preprocessing layer listens to the input of a specific nursing operation event through the hospital information system interface, the expected alarm silencing unit automatically sets a dynamic time window for the associated physiological parameters directly affected by the nursing operation, and within the time window, it shields or downgrades non-fatal expected physiological parameter fluctuation alarms.

[0034] Specifically, the solution in this application uses this interface to capture discrete nursing operation event inputs in real time as trigger sources. The expected alarm silencing unit automatically identifies the set of physiological parameters directly affected by the operation based on a preset event-parameter association rule base, and generates an appropriate dynamic time window length by calling the historical fluctuation database in combination with the operation type. During the duration of this window, the system performs masking or downgrading processing on non-fatal alarms of the associated parameters, while fatal alarms are still pushed at the original level, thus forming a closed-loop logic chain of event triggering, window generation and alarm filtering, ensuring that the silencing response is strictly synchronized with the operation process, and avoiding the defect that fixed time strategies cannot match dynamic scenarios.

[0035] This application further proposes that the adaptive alarm grading and intelligent push module is also connected to a timeout escalation transfer unit; when the assigned preferred receiving terminal does not generate confirmation feedback within the set time window, or when the real-time workload quantification value of the preferred personnel is higher than the full load threshold, the timeout escalation transfer unit automatically transfers the warning information horizontally to the nearest secondary idle receiving terminal based on spatial positioning data, or vertically escalates it to the superior management terminal.

[0036] Specifically, the timeout escalation unit continuously monitors the confirmation feedback status of the primary receiving terminal and the workload quantification value of relevant personnel. When no confirmation feedback is generated within a set time window, or the workload quantification value exceeds the full load threshold, the unit automatically activates. Based on real-time acquired spatial positioning data, the unit calculates the nearest available secondary receiving terminal. If a suitable secondary terminal exists, a lateral transfer is performed. If the lateral transfer conditions are not met, such as no available terminal or a high risk level, a vertical escalation to the superior management terminal is performed. This dual-transfer strategy eliminates the lag in location information through the real-time nature of spatial positioning data, ensuring that early warning information can be quickly redirected to available resources when the primary path fails, reducing response delay.

[0037] This application further proposes that the intelligent push module classifies alarms into three categories—extremely high risk, medium risk, and low risk—based on the final risk level, and executes different push channel strategies: extremely high risk uses concurrent pushes of bedside audio-visual alarms, main screen display, strong mobile terminal prompts, and voice broadcasts; medium risk uses mobile terminal pushes and delayed bedside audio-visual alarms; low risk events only record background logs and archive abnormal explanation prompts, without actively triggering pushes.

[0038] Specifically, the proposed solution receives the final risk level and anomaly explanation prompts from the interpretable dynamic rule-neural hybrid decision engine via an intelligent push module. Combining this with quantitative values ​​of personnel workload and spatial positioning data, the alarm is first dynamically categorized into three levels: extremely high risk, medium risk, and low risk based on preset thresholds. When an alarm is determined to be extremely high risk, the bedside audible and visual alarm unit, main screen display unit, mobile terminal strong prompt unit, and voice broadcast unit are immediately activated for multi-channel concurrent push notifications, forming redundant notification paths. When an alarm is determined to be medium risk, the warning information is prioritized to be sent via the mobile terminal push unit, and a delay timer is started. The bedside audible and visual alarm unit is only triggered if no confirmation feedback is received within a set time window. When an alarm is determined to be low risk, only the background log recording unit is used to store archived anomaly explanation prompt data, completely suppressing proactive push behavior. This hierarchical strategy and dynamic binding mechanism between risk levels ensures that push resource allocation is strictly matched to the urgency of the event, preventing high-risk events from being missed due to a single push channel, while simultaneously eliminating the ineffective occupation of personnel attention by low-risk events.

[0039] This application further proposes that the process of generating natural language anomaly explanation prompt information includes: extracting features whose absolute value is greater than a preset threshold as core control features, using the core control features and the corresponding positive and negative signs of SHAP values ​​as retrieval conditions, mapping them to the corresponding entity nodes of the control knowledge graph, and using a built-in template engine combined with historical running feature data to automatically assemble and generate structured multidimensional anomaly prompt information.

[0040] This application's solution maps high-weight feature points and SHAP values ​​output by a neural network submodule to corresponding entity nodes in the control knowledge graph, thereby transforming technical indicators into data semantic concepts. Simultaneously, it utilizes a built-in template engine combined with historical operational feature data to construct a personalized interpretation framework, ultimately generating clearly structured, multi-dimensional anomaly alerts. In this process, the quantification of SHAP values ​​ensures accurate expression of feature contributions, the entity nodes in the control knowledge graph provide standardized data concept anchors, the structured constraints of the template engine prevent arbitrary alert text, and the integration of historical operational feature data deeply couples the interpretation content with the specific scenario context.

[0041] Specifically, the mapping and text assembly process is as follows: the system iterates through the SHAP values ​​of all features and compares them with a preset threshold, filters out features exceeding the threshold, and combines their positive and negative contribution signs. It then retrieves the corresponding abnormal association path triples in the graph using a Cypher query. Subsequently, the retrieved triple data is filled into a preset semantic slot template, assembling it into an intermediate attribution prompt that does not contain a confirmed disease conclusion. The above mechanisms work together to ensure that the generated prompt text maintains both technical rigor and readability, effectively bridging the cognitive gap between AI decision-making and operational practice.

[0042] This application further proposes that the feedback closed-loop learning module includes a feedback confidence evaluation unit; the feedback confidence evaluation unit is used to calculate the frequency and response latency of the receiving end annotation operation; when the frequency of ignored or false alarm annotations per unit time is higher than the preset fatigue blind spot threshold, it is determined that the confidence of the annotation feedback is low, and the weight of the batch of feedback data in the incremental update is reduced, so as to prevent the model from being polluted by erroneous data caused by fatigue operation.

[0043] In practical applications, the feedback confidence assessment unit refers to a computational entity used to dynamically quantify the reliability of human feedback behavior. It can be implemented using time-series analysis-based software services or dedicated embedded logic circuits. Its purpose is to extract objective quality indicators from behavioral patterns rather than relying on original annotation content. The frequency and response latency of annotation operations can be understood as behavioral characteristic parameters reflecting workload status. Their calculation can be achieved by statistically counting operation events and the event response time difference sequence within a unit of time, aiming to accurately identify phenomena such as distraction or decision-making delay. The preset fatigue blind spot threshold is a critical value set based on operational experience. It can be implemented using a dynamic threshold model trained on historical fatigue data or a fixed empirical threshold, aiming to distinguish between systematic mislabeling behavior and random errors. The mechanism for determining low confidence in annotation feedback can be understood as the decision-making process for generating confidence labels, aiming to provide a reliable basis for subsequent data processing. Reducing the weight of this batch of feedback data in incremental updates is specifically achieved by adjusting the sample contribution ratio or the learning rate decay factor, aiming to suppress noise influence while retaining some historical information and avoiding the loss of learning efficiency caused by completely discarding feedback.

[0044] This application further proposes that the feedback closed-loop learning module also includes an objective backtracking verification unit; the objective backtracking verification unit is used to perform secondary verification and alignment of the historical subjective annotation feedback after obtaining the final system discharge outcome data; if there is a conflict between the system discharge outcome data and the subjective annotation, the system discharge outcome data is used as the objective gold standard to generate error correction samples and feed them back to the edge computing unit for lightweight federated learning updates.

[0045] Specifically, the proposed solution automatically acquires the system's discharge outcome data after the process is completed, triggering an objective backtracking verification unit to systematically verify historical subjective annotations. When a conflict is detected between subjective annotations and objective outcomes, the system forces the generation of error correction samples based on the system's discharge outcome and distributes update instructions to edge computing units through a lightweight federated learning mechanism. This process ensures that model iteration is always based on real-world operational experience rather than error feedback, effectively maintaining the rationality of the early warning logic while avoiding privacy risks associated with centralized data transmission, and achieving efficient iterative optimization of model updates.

[0046] The above description discloses only preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

[0047] While the invention has been described with reference to several specific embodiments, it should be understood that the invention is not limited to the disclosed specific embodiments. The invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A multidimensional heterogeneous data processing and adaptive early warning system for intensive care, characterized in that, It includes a data acquisition and preprocessing layer with sequential communication connections, a multi-scale temporal feature extraction and fusion module, an interpretable dynamic rule-neural hybrid decision engine, an adaptive alarm grading and intelligent push module, and a feedback closed-loop learning module; among which: The data acquisition and preprocessing layer is used to automatically acquire discrete nursing operation event data packets in the hospital information system through the hospital information system interface, and simultaneously acquire multimodal physiological parameters and waveform data from multiple bedside monitoring devices in parallel. It performs timestamp alignment of the nursing operation event data packets with the physiological parameters and waveform data based on a sliding time window, and triggers historical state overwrite recalculation when an input delay is detected in the nursing operation event data packets. The multi-scale temporal feature extraction and fusion module is used to extract physiological features from different time windows and generate multi-dimensional feature vectors. The interpretable dynamic rule-neural hybrid decision engine receives the multidimensional feature vector, outputs a first risk probability and feature importance weight distribution through a neural network submodule, and outputs a second risk assessment based on clinical guideline baselines through a symbolic rule library submodule; it also fuses the first risk probability and the second risk assessment according to a set conflict arbitration strategy to obtain the final risk level, extracts core control features based on the feature importance weight distribution, matches causal relationship chain triple nodes in the control knowledge graph, and generates natural language abnormality explanation prompts through template filling; The adaptive alarm grading and intelligent push module receives the final risk level and explanation text, combines the real-time acquired quantitative value of medical staff workload and spatial positioning data, dynamically calculates the optimal receiving terminal and delivery delay time, and performs targeted grading push. The feedback closed-loop learning module is used to receive labeled feedback on the early warning results from the medical staff and to incrementally update the neural network submodule.

2. The critical care multidimensional heterogeneous data processing and adaptive early warning system according to claim 1, characterized in that, The data acquisition and preprocessing layer includes a morphology-based artifact recognition submodule. Before feature extraction, this submodule is used to identify and remove physiological waveform artifact signals caused by patient movement or equipment detachment. The artifact recognition submodule is located in the upstream data link of the multi-scale temporal feature extraction and fusion module to block technical false alarms at the source.

3. The critical care multidimensional heterogeneous data processing and adaptive early warning system according to claim 1, characterized in that, The multi-scale temporal feature extraction and fusion module adopts an edge-device-cloud three-layer architecture, including: Edge computing units, deployed at bedside monitoring devices or dedicated edge boxes, run lightweight time-series models to extract short-term abnormal features at the second to minute level; A ward-level server is used to integrate the 12- to 24-hour medium- to long-term trend characteristics of multiple patients in a ward. Cloud / hospital-grade servers are used for cross-ward data aggregation, knowledge distillation, and global model updates. Furthermore, the three-layer structure interacts with data through independent data synchronization interface units to achieve hierarchical isolation.

4. The critical care multidimensional heterogeneous data processing and adaptive early warning system according to claim 1, characterized in that, The conflict arbitration strategy specifically includes: When the deviation between the first risk probability and the second risk assessment exceeds a preset threshold, if the feature importance weight distribution is concentrated in a new feature dimension not defined by the rule base, then the first risk probability is used as the main factor, and the result is marked as an exploratory warning. If the feature importance weight distribution triggers the mandatory safety red line set by the symbol rule base, then the symbol rule base will forcibly overwrite it, and the second risk assessment will be used as the final risk level.

5. The critical care multidimensional heterogeneous data processing and adaptive early warning system according to claim 1, characterized in that, The adaptive alarm grading and intelligent push module also includes an expected alarm silencing unit; When the data acquisition and preprocessing layer receives a specific nursing operation event input, the expected alarm silencing unit automatically sets a dynamic time window for the associated physiological parameters directly affected by the nursing operation, and within the time window, it shields or downgrades non-fatal expected physiological parameter fluctuation alarms.

6. The critical care multidimensional heterogeneous data processing and adaptive early warning system according to claim 1, characterized in that, The adaptive alarm grading and intelligent push module also includes a timeout upgrade transfer unit; When the assigned primary receiving terminal fails to generate confirmation feedback within the set time window, or when the real-time workload quantification value of the primary medical staff exceeds the full load threshold, the timeout upgrade transfer unit automatically transfers the warning information horizontally to the nearest secondary idle medical staff terminal based on spatial positioning data, or upgrades it vertically to the superior management terminal.

7. The critical care multidimensional heterogeneous data processing and adaptive early warning system according to claim 1 or 6, characterized in that, The intelligent push module categorizes alarms into three levels—extremely high risk, medium risk, and low risk—based on the final risk level, and executes different push channel strategies accordingly. For extremely high-risk cases, a combination of bedside audible and visual alarms, nurse station main screen display, strong prompts on mobile terminals, and voice broadcasts are used for simultaneous push notifications. For medium-risk cases, a bedside audible and visual alarm is provided via mobile terminal push and delayed triggering. Low-risk events are only logged in the background and the explanatory text is archived; no push notifications are actively triggered.

8. The critical care multidimensional heterogeneous data processing and adaptive early warning system according to claim 1, characterized in that, The process of generating natural language explanation text includes: extracting features whose absolute value is greater than a preset threshold as core control features; using the core control features and the corresponding positive and negative signs of SHAP values ​​as search conditions; mapping them to the causal relationship chain triple nodes of the control knowledge graph; and using the built-in template engine combined with historical running feature data to automatically assemble and generate structured multidimensional anomaly prompt information.

9. The critical care multidimensional heterogeneous data processing and adaptive early warning system according to claim 1, characterized in that, The feedback closed-loop learning module includes a feedback confidence evaluation unit; The feedback confidence assessment unit is used to calculate the frequency and response latency of annotation operations on the medical staff side. When the frequency of ignored or false alarm annotations per unit time is higher than the preset fatigue blind spot threshold, the confidence of the annotation feedback is determined to be low, and the weight of the batch of feedback data in the incremental update is reduced to prevent erroneous data caused by fatigue operations from polluting the model.

10. The critical care multidimensional heterogeneous data processing and adaptive early warning system according to claim 9, characterized in that, The feedback closed-loop learning module also includes an objective backtracking verification unit; The objective backtracking verification unit is used to perform secondary verification and alignment of historical subjective annotation feedback after obtaining the patient's final electronic medical record outcome data. If there is a conflict between the electronic medical record outcome data and the subjective annotation, the electronic medical record outcome data is used as the objective gold standard to generate error correction samples and feed them back to the edge computing unit for lightweight federated learning updates.

Images