A remote intelligent alarm method and system for intensive care

Abstract

The present application relates to the technical field of critical remote monitoring alarm, and particularly relates to a remote intelligent alarm method and system for intensive care. The method and system obtain patient vital sign data through a remote monitoring platform; generate a patient monitoring overview based on user configuration, which contains data signal connection relationships between various medical devices based on a treatment process and monitoring parameter ranges of the various medical devices; synchronize the patient monitoring overview with real-time data signals of the various medical devices; calculate a remaining duration from a current time to an end time of a preset monitoring period, determine an adaptive prediction strategy according to the remaining duration, generate a risk prediction value corresponding to the patient monitoring overview according to the adaptive prediction strategy based on real-time monitoring parameters of the various medical devices and the data signal connection relationships, compare the risk prediction value with a risk reference threshold set by a user, obtain a deviation index, and output an alarm signal, thereby providing more reasonable and efficient early warning support for an intensive care scene.

Description

Technical Field

[0001] This invention relates to the field of remote monitoring and alarm technology for critical care, specifically to a remote intelligent alarm method and system for critical care. Background Technology

[0002] In intensive care settings, patients' conditions are often complex and change rapidly, with vital signs constantly shifting, making real-time monitoring and timely warnings extremely urgent. Currently, intensive care largely relies on decentralized monitoring by various medical devices, such as electrocardiogram monitors, ventilators, and blood purification machines. These devices operate independently, and monitoring data is stored in different systems or terminals. Medical staff need to switch between various medical devices and interfaces to view data, which not only increases their workload but may also cause them to miss critical changes in the patient's condition due to untimely data integration. Studies have shown that the number of alarms in intensive care units is extremely high, and a significant proportion are non-action / false positive alarms. One study recorded over 2.5 million monitoring alarms in a month, averaging about 187 audible alarms per bed per day. Nearly 90% of these arrhythmia alarms were false positives, directly inducing "alarm fatigue" and reducing sensitivity to genuine danger signals. Furthermore, low-quality signals (such as body movement and lead loss) and conservative default threshold algorithms from manufacturers further amplify the false alarm rate, exacerbating information noise.

[0003] Most existing monitoring systems rely on fixed parameter thresholds for individual medical devices, triggering an alarm when a parameter exceeds a preset range. However, treating critically ill patients involves the coordinated action of multiple medical devices, and an abnormality in a single parameter may be closely related to the operational status of other devices. For example, adjustments to ventilator parameters can affect changes in blood oxygen saturation. Relying solely on isolated blood oxygen saturation values ​​for alarms can easily overlook the correlation between parameter changes, leading to false alarms or missed alarms. Previous research has shown that, compared to fixed thresholds, adaptive thresholds based on individual baselines and dynamic patterning, along with individual-specific algorithms, can significantly reduce invalid alarms. Machine learning / interpretable learning frameworks that integrate multi-parameter trends help identify key variables and threshold ranges upon which genuine danger signals depend. Furthermore, in real intensive care units, customizing alarm strategies and conducting strategy mining (such as batch optimization of alarm parameters and policy evaluation) can significantly reduce the number of alarms and improve response efficiency.

[0004] Current systems lack dynamic consideration of the monitoring cycle. Treatment for critically ill patients is typically divided into different stages, each with different monitoring focuses and risk points. For example, the fluctuation range of monitoring parameters and the need for risk warnings differ significantly between the initial postoperative period and the recovery period. Current prediction strategies often employ uniform algorithm models, failing to consider the impact of the remaining time within the monitoring cycle on risk assessment. This can lead to over-warnings at the beginning of the cycle and delayed warnings near the end. Clinical observational studies indicate that frontline healthcare workers expect monitoring systems to identify specific clinical situations and dynamic changes in treatment processes, and automatically adjust alarm strategies accordingly. For example, this could involve dynamically setting thresholds and response rules based on key milestones such as sedation depth, ventilation mode, and weaning / extubation. However, current tools lack the ability to quantitatively evaluate and optimize staged alarm strategies. Existing systems largely rely on static prediction methods, which struggle to accurately reflect the stage-specific characteristics of the disease progression, limiting the timeliness and accuracy of warnings.

[0005] With the development of telemedicine, remote monitoring platforms are increasingly being applied in the field of intensive care. However, existing platforms still have shortcomings in data synchronization and integration. Remotely acquired data often suffers from latency, and the data formats of different medical devices are inconsistent, making it difficult to quickly generate a comprehensive patient monitoring view. Medical staff at the remote end find it difficult to intuitively grasp the connections and collaborative effects between various medical devices, posing challenges to remote diagnostic and treatment decisions. These issues make it difficult for existing remote monitoring alarm mechanisms to provide accurate and timely risk warnings when dealing with the complex conditions of critically ill patients, and they fail to fully meet the actual needs of clinical monitoring. Summary of the Invention

[0006] The purpose of this invention is to provide a remote intelligent alarm method and system for intensive care, so as to solve the problems in the prior art.

[0007] To achieve the above objectives, the present invention provides a remote intelligent alarm method for intensive care, comprising the following:

[0008] The patient's vital signs data are acquired through a remote monitoring platform and a patient monitoring overview is generated. The patient monitoring overview includes the data signal connection relationship between various medical devices based on the treatment process, as well as the real-time monitoring parameter range of each medical device.

[0009] A medical device topology network is constructed based on preset medical device types, medical device connection rules, and monitoring parameter threshold ranges. Nodes in the medical device topology network represent medical devices, and edges represent data interaction relationships between medical devices. The monitoring parameter threshold ranges are used to assign parameter monitoring conditions to each node. The medical device topology network and parameter monitoring conditions are then integrated to generate the patient monitoring overview.

[0010] The patient monitoring overview is synchronized with the real-time data signals of each medical device; the remaining time between the current time and the end time of the preset monitoring cycle is calculated, and an adaptive prediction strategy is determined based on the remaining time. Based on the real-time monitoring parameters of each medical device and the data signal connection relationship, a risk prediction value corresponding to the patient monitoring overview is generated according to the adaptive prediction strategy; the risk prediction value is compared with a preset risk reference threshold to obtain a deviation index and output an alarm signal.

[0011] The remaining time determines the adaptive prediction strategy, including calling a preset first mapping table, inputting the remaining time into the first mapping table to match and obtain an initial prediction interval, obtaining the number of risk types corresponding to the risk reference threshold, querying a second mapping table based on the number of risk types to obtain an adjustment factor, multiplying the initial prediction interval with the adjustment factor to obtain an optimized prediction interval, and formulating the adaptive prediction strategy based on the optimized prediction interval.

[0012] The process of generating a risk prediction value corresponding to the patient monitoring overview based on the real-time monitoring parameters of each medical device and the data signal connection relationship, according to the adaptive prediction strategy, includes loading a pre-trained feature extraction module, which completes the training process using a historical dataset; controlling the feature extraction module to perform feature fusion on the real-time monitoring parameters of each medical device and the data signal connection relationship according to the optimized prediction interval, generating a fused feature vector; the fused feature vector is used to calculate the risk prediction value through a dynamic evaluation model, and the risk prediction value represents the comprehensive risk level of all medical devices in the patient monitoring overview within the preset monitoring period; wherein:

[0013] The final output of the feature extraction module is a fixed-dimensional fused feature vector, which is formally expressed as:

[0014] ;

[0015] in: represents the feature tensor input to the convolutional layer, which is a stack of node feature matrices and adjacency matrices; conv represents a composite convolution operation including edge convolution and regular spatial convolution; σ represents the activation function; h(x) is the output fused feature vector.

[0016] The present invention also includes a remote intelligent alarm system for intensive care, used to implement the aforementioned remote intelligent alarm method for intensive care. The system comprises a remote monitoring platform, a data acquisition module, an overview construction module, a prediction module, and an alarm output module, sequentially connected by data signals. The data acquisition module acquires patient vital sign data through the remote monitoring platform, including electrocardiogram (ECG) signals and blood oxygen saturation parameters. The overview construction module generates a patient monitoring overview based on user configuration. The patient monitoring overview includes data signal connections between various medical devices based on the treatment process, as well as the real-time monitoring parameter ranges of each medical device. The patient monitoring overview is synchronized with each medical device in real-time. The prediction module calculates the remaining time between the current moment and the end of a preset monitoring cycle, determines an adaptive prediction strategy based on the remaining time, and generates a risk prediction value corresponding to the patient monitoring overview according to the adaptive prediction strategy based on the real-time monitoring parameters of each medical device and the data signal connections. The alarm output module compares the risk prediction value with a user-defined risk reference threshold, obtains a deviation index, and outputs an alarm signal.

[0017] The significant advantages of this invention compared to existing technologies are as follows:

[0018] First, this invention provides more reasonable and efficient early warning support for intensive care scenarios through multi-stage optimized design. At the data integration level, the patient monitoring overview generated based on user configuration incorporates the connection relationships between various medical devices and the monitoring parameter ranges of each device, breaking the isolation of medical device data in traditional monitoring. Medical staff can intuitively grasp the collaborative relationships between different medical devices in the treatment process through the patient monitoring overview, making the scattered monitoring data form an organic whole, avoiding information omissions caused by data fragmentation, and making the judgment of the condition more holistic. In intensive care units, data fragmentation is often associated with a high proportion of invalid alarms. The integrated information helps to identify the causal relationships and trends between parameters, reducing false triggers from the source, such as "information noise" caused by lead detachment or overly conservative default thresholds.

[0019] Secondly, the real-time data signal synchronization mechanism ensures immediate updates of data between the overall patient monitoring system and various medical devices, reducing the impact of data delays. In intensive care, the timeliness of data directly affects the speed of capturing changes in the patient's condition. Real-time synchronization allows medical staff to conduct risk assessments based on the latest monitoring data, avoiding untimely warnings due to data lag and providing a foundation for rapid response to changes in the patient's condition. In the context of septic shock, this invention aligns with the "ROSE" four-stage fluid management concept (Resuscitation, Optimization, Stabilization, and Evacuation). The system can monitor or detect key indicators and automatically switch focus and warning thresholds in real time: During the resuscitation / optimization phase, it displays arterial pressure, arterial lactate, urine output, skin capillary refill time (CRT), dynamic fluid responsiveness indicators (such as central venous pressure, PPV / SVV, passive leg raise test, and ultrasound-indicated volume status), and changes in vasoactive drug infusion rates, emphasizing timely, adequate, and dynamically validated fluid resuscitation to restore perfusion; during the stabilization phase, it prioritizes maintaining perfusion and avoiding fluid overload; during the evacuation phase, it combines microcirculation, organ perfusion indicators, and fluid responsiveness to promote fluid drainage and control fluid intake; the system provides alerts at strategy transition points (such as when lactate returns to normal, CRT returns to normal, and fluid responsiveness is negative), reducing the risk of over-resuscitation and delayed evacuation. For example, in patients with cerebral hemorrhage, due to differences in blood pressure control during the acute and chronic phases and before and after surgery, the system can automatically switch the focus of monitoring and threshold settings according to the stage. For instance, the system can synchronize the drug pump rate with invasive / non-invasive blood pressure, heart rate, and intracranial pressure / cerebral perfusion pressure measured by transcranial Doppler (TCD), optic nerve sheath (ONS), TCCD, invasive pressure device, etc., on the same screen in real time. Once there are early signs of breakthrough intracranial hypertension or insufficient perfusion, the system will trigger a deviation prompt in the overview and link the relevant parameters to reduce the recognition delay caused by data asynchrony.

[0020] Third, the introduction of adaptive prediction strategies enables risk prediction to adapt to the dynamic changes of the monitoring cycle. By calculating the remaining time between the current moment and the preset end of the monitoring cycle, the prediction logic is adjusted accordingly, allowing the prediction model to exhibit adaptability at different stages of the cycle. For example, in the early stages of the cycle, the focus is on trend identification, such as the trajectory of slow improvement in oxygenation and lung mechanics in patients with acute respiratory distress syndrome (ARDS). At the end of the cycle or at key points (such as when weaning / extubation is expected or during sedation tapering), attention is strengthened to short-term drastic fluctuations. In the early stages of the cycle, the focus can be on capturing trend changes to provide clues for early identification of potential risks; as the end of the cycle approaches, attention can be strengthened to short-term fluctuations to ensure timely prediction of impending risks. This prediction approach, which adjusts with the progress of the cycle, improves the scenario adaptability of risk assessment.

[0021] Fourth, the generation of risk prediction values ​​combines real-time monitoring parameters of various medical devices with the interconnections between them, rather than relying on isolated judgments of a single parameter. In critical care, the synergistic effect between various medical devices has a significant impact on the patient's condition; an abnormality in the parameter of one medical device may be a chain reaction leading to changes in the operating status of other medical devices. Comprehensive prediction based on interconnections can consider the correlation of parameter changes, reduce misjudgments caused by abnormalities in a single parameter, and make risk assessment more comprehensive. For example, the system can combine transcranial color Doppler (such as pulsatility index PI, blood flow velocity, cervical compression test, optic nerve sheath width, etc.) with the presence of neurological pathological signs, pupil size, and pupillary light reflex to identify trends of high intracranial pressure or rule out the possibility of intracranial hypertension. It can combine EEG / quantitative EEG (reactivity, inhibition-burst, etc.) for early warning of adverse prognosis, and combine serum biomarkers (such as NSE) for stratified assessment. The above signals can then be correlated with blood pressure, ventilation and perfusion indicators on a unified time axis to form consistent evidence in the three dimensions of "structure-function-biochemistry", thereby indicating the risk of high ICP and adverse neurological outcomes earlier and more robustly.

[0022] Fifth, the predicted risk value is compared with the user-defined risk reference threshold to obtain a deviation index and output an alarm signal, making alarm triggering more targeted. Users can flexibly configure the threshold according to the individual patient's condition and treatment needs, allowing the alarm mechanism to adapt to the different patient's condition characteristics, avoiding the problem of excessively frequent alarms or insufficient warnings under a uniform threshold, and improving the practical value of the alarm signal. This invention can significantly reduce invalid alarms and limit alarm fatigue through individualized thresholds, alarm delay, and configuration optimization. Attached Figure Description

[0023] Figure 1 This is a timing diagram for a remote intelligent alarm method used in intensive care.

[0024] Figure 2 A flowchart for determining an adaptive prediction strategy for intensive care.

[0025] Figure 3 A multi-dimensional correlation analysis diagram for an adaptive prediction strategy used in intensive care.

[0026] Figure 4 This is a flowchart of a method for generating risk prediction values ​​for intensive care.

[0027] Figure 5 A multi-module analysis diagram of the risk assessment process used in intensive care.

[0028] Figure 6 A flowchart illustrating a method for determining the trigger trajectory of an alarm device used in intensive care. Detailed Implementation

[0029] The specific embodiments of the present invention will be further described in detail below with reference to the examples and accompanying drawings.

[0030] This invention provides a remote intelligent alarm method for intensive care and a remote intelligent alarm system for implementing this method (hereinafter referred to as the "system"), in conjunction with embodiments and appendices. Figure 1-6 A detailed description of the specific implementation method is provided.

[0031] Example 1. Refer to Figure 1 The present invention proposes a remote intelligent alarm method for intensive care, comprising the following:

[0032] The patient's vital signs data are acquired through a remote monitoring platform and a patient monitoring overview is generated. The patient monitoring overview includes the data signal connection relationship between various medical devices based on the treatment process, as well as the real-time monitoring parameter range of each medical device.

[0033] A medical device topology network is constructed based on preset medical device types, medical device connection rules, and monitoring parameter threshold ranges. Nodes in the medical device topology network represent medical devices, and edges represent data interaction relationships between medical devices. The monitoring parameter threshold ranges are used to assign parameter monitoring conditions to each node. The medical device topology network and parameter monitoring conditions are then integrated to generate the patient monitoring overview.

[0034] The patient monitoring overview is synchronized with the real-time data signals of each medical device; the remaining time between the current time and the end time of the preset monitoring cycle is calculated, and an adaptive prediction strategy is determined based on the remaining time. Based on the real-time monitoring parameters of each medical device and the data signal connection relationship, a risk prediction value corresponding to the patient monitoring overview is generated according to the adaptive prediction strategy; the risk prediction value is compared with a preset risk reference threshold to obtain a deviation index and output an alarm signal.

[0035] The remaining time determines the adaptive prediction strategy, including calling a preset first mapping table, inputting the remaining time into the first mapping table to match and obtain an initial prediction interval, obtaining the number of risk types corresponding to the risk reference threshold, querying a second mapping table based on the number of risk types to obtain an adjustment factor, multiplying the initial prediction interval with the adjustment factor to obtain an optimized prediction interval, and formulating the adaptive prediction strategy based on the optimized prediction interval.

[0036] The process of generating a risk prediction value corresponding to the patient monitoring overview based on the real-time monitoring parameters of each medical device and the data signal connection relationship, according to the adaptive prediction strategy, includes loading a pre-trained feature extraction module, which completes the training process using a historical dataset; controlling the feature extraction module to perform feature fusion on the real-time monitoring parameters of each medical device and the data signal connection relationship according to the optimized prediction interval, generating a fused feature vector; the fused feature vector is used to calculate the risk prediction value through a dynamic evaluation model, and the risk prediction value represents the comprehensive risk level of all medical devices in the patient monitoring overview within the preset monitoring period; wherein:

[0037] The final output of the feature extraction module is a fixed-dimensional fused feature vector, which is formally expressed as:

[0038] ;

[0039] in: represents the feature tensor input to the convolutional layer, which is a stack of node feature matrices and adjacency matrices; conv represents a composite convolution operation including edge convolution and regular spatial convolution; σ represents the activation function; h(x) is the output fused feature vector.

[0040] Preferably, the method further includes: calling a preset third mapping table; querying a second mapping table based on the number of risk types to obtain an adjustment factor, including extracting a first adjustment value from the second mapping table based on the number of risk types; counting the total number of medical devices in the patient monitoring overview, and querying the third mapping table based on the total number of medical devices to extract a second adjustment value; and performing a weighted summation of the first adjustment value and the second adjustment value as the adjustment factor.

[0041] Preferably, the risk reference threshold is obtained by: normalizing the reference values ​​of different risk types, and then fusing the converted values ​​to obtain the risk reference threshold.

[0042] Preferably, the method further includes: extracting key features from the patient's vital sign data to obtain a physiological feature sequence; generating a static risk response matrix from the physiological feature sequence through a pre-trained response learning module to obtain an interventional path; and controlling the interventional path of the alarm device based on each static risk response value in the static risk response matrix and the corresponding medical device location.

[0043] Preferably, the patient's vital signs data undergoes key feature extraction, including segmenting the patient's vital signs data into multiple time windows, extracting waveform features and trend features within each time window, and combining the waveform features and trend features to form the physiological feature sequence; wherein, the waveform features include electrocardiogram variability and ST segment, T wave dynamics, SpO2 short-term variability and rate of decline, and the count of asynchronous events between ventilator plateau pressure / driving pressure / tidal volume and end-tidal carbon dioxide; the trend features include the oxygenation index, ROX index, the fluctuation mean line and inflection point of alveolar-arterial oxygen partial pressure difference Aa, shock index, lactate level, ScvO2, (P(va)CO2), the gradient relationship between intake and output balance and norepinephrine equivalent dose.

[0044] Preferably, the static risk response values ​​and corresponding medical device locations in the static risk response matrix control the trigger trajectory of the alarm device on the intervened path, including parsing the intervened path into multiple path segments, calculating the sum of the static risk response values ​​of adjacent medical device locations corresponding to each path segment as the information concentration of that path segment, and using a path optimization algorithm to determine the trigger trajectory of the alarm device from the overall starting point.

[0045] The patient's vital signs data acquired through the remote monitoring platform described in this invention include, but are not limited to, electrocardiogram signals (heart rate / rhythm, ST segment), oxygen saturation (SpO2), non-invasive / invasive arterial blood pressure and MAP (NIBP / ABP / MAP), respiratory rate and end-expiratory carbon dioxide (EtCO2) waveform, body temperature, key ventilator parameters (FiO2, PEEP, tidal volume VT, ventilator frequency, peak pressure Ppeak, plateau pressure Pplat, driving pressure ΔP, compliance, ventilation mode and alarm status), arterial blood gas and metabolic indicators (pH, PaO2, PaCO2, HCO3-, lactate, blood glucose), advanced invasive hemodynamic parameters (CVP, cardiac output / cardiac index CO / CI, stroke volume SV, SVV / PPV), and circulatory perfusion indicators (urine output). The data includes capillary refill time (CRT), urine output, infusion pump parameters (vasoactive drugs such as norepinephrine / epinephrine, sedation and analgesia pump rate and cumulative volume), fluid intake and output and net balance, bladder pressure, and continuous renal replacement therapy (CRRT) and extracorporeal membrane oxygenation (ECMO) operating parameters, as well as medical device / process events (positional changes, suctioning, blood transfusion, and CRRT start / stop, etc.). The process of generating a patient monitoring overview based on user configuration involves transforming the user's abstract treatment intentions and nursing standards into a structured digital model that can be recognized and processed by a computer system. This model not only statically describes the layout and parameter requirements of medical devices, but also dynamically defines the data flow and logical relationships between medical devices, providing a complete contextual environment for subsequent risk prediction and alarm triggering.

[0046] The system of this invention first receives and parses configuration data provided by the user through a graphical interface or configuration file. This data serves as the blueprint and raw material for generating a monitoring overview. For example, if a doctor in an intensive care unit needs to monitor a post-cardiac surgery patient, he will select a series of medical device types connected to the patient in the system. These types may include electrocardiogram monitors, ventilators, central venous pressure / central venous oxygen saturation (CVP / ScvO2) monitoring devices, core body temperature monitoring devices (esophageal / bladder probes), transpulmonary thermodilution or pulse contour analysis cardiac output monitoring (PiCCO), transthoracic bedside echocardiography, and bedside coagulation function monitoring (ACT). The system includes TEG / ROTEM, blood gas analyzers (pH, PaO2, PaCO2, Pv-aCO2, BE, lactate), infusion pumps (sedatives / analgesics / vasoactive drugs), closed thoracic drainage devices (negative pressure / drainage volume), urine analyzers and fluid intake / output monitoring devices. Depending on the post-cardiac surgery management needs, the system may also include defibrillator monitoring and event marking devices, continuous blood purification (CRRT) instruments, temporary epicardial pacemakers / external pacemakers (intra-aortic balloon counterpulsation IABP), and extracorporeal membrane oxygenation (ECMO). Each type of medical device has its own digital definition in the system, including the set of physiological parameters it can monitor and the data communication protocol.

[0047] Users need to define the connection relationships between the various medical devices. This does not refer to physical wiring connections, but rather to data-level interactions and logical relationships. The system provides a visual drag-and-drop interface, allowing users to draw the data flow between medical devices. For example, users can define tidal volume data from a ventilator as a condition for triggering an alarm, while heart rate data from an ECG monitor is an important reference for assessing the safety of infusion pump drug delivery rates. Users can set a rule: when the ECG monitor detects that the patient's heart rate is continuously higher than a certain threshold, the system should automatically retrieve the rate information of the currently infused vasoactive drugs (such as dopamine) from the infusion pump and perform a comprehensive risk assessment. After these rules are parsed by the system, they are transformed into directed edges in the medical device topology network. In this network, each medical device becomes a node, and each directed edge represents the impact of one medical device's data output on the status or alarm logic of another medical device. The system automatically checks the logical consistency of this network to prevent errors such as circular dependencies.

[0048] Users need to set monitoring parameter ranges for each medical device node, i.e., parameter monitoring conditions, which form the threshold basis for alarm logic. For invasive blood pressure monitors, users need to set safe upper and lower limits for systolic blood pressure, diastolic blood pressure, and mean arterial pressure. For example, mean arterial pressure can be set to be maintained between 65 mmHg and 110 mmHg. For ventilators, users need to set normal ranges for respiratory rate and oxygen concentration, and pay special attention to whether tidal volume (e.g., 4–6 mL / kg predicted body weight), plateau pressure (≤30 cmH2O), and driving pressure (≤15 cmH2O) are within the range allowed by the lung-protective ventilation strategy (or individualized according to the disease). For pulse oximetry... For the device, blood oxygen saturation should typically be maintained above 94%; the end-tidal carbon dioxide (EtCO2) monitoring node is usually set within the target range of 35–45 mmHg, triggering a composite alarm when EtCO2 drops by ≥10 mmHg or the waveform disappears and is accompanied by abnormal airway pressure / tidal volume; the threshold for NIRSrSO2 monitoring is ≥60%, or a decrease of ≥20% from the individual baseline lasting ≥60 seconds triggers an alarm; all these thresholds are not isolated, but are tightly bound to their respective medical device nodes; the system assigns a unique identifier to each parameter monitoring condition and associates it with a specific node in the topology network.

[0049] After parsing and initially processing the configuration data, the system executes the construction process. Building the medical device topology network is a process of transforming abstract rules into concrete data structures. The system uses a graph structure to store this network. Each medical device instance is a node object, and node attributes include metadata such as medical device ID, type, model, and IP address. Nodes are connected by edge objects. Each edge contains a source node ID, a target node ID, and a relationship description field. This field details what data from the source node will affect the state assessment of the target node under what conditions. For example, an edge from the "ECG monitor" node to the "infusion pump" node might have the relationship description "IF heart rate > 130 beats / min THEN query current infusion rate and assess its safety." This graph structure completely depicts how the various medical devices work together and how data flows in the treatment process, thus forming a virtual monitoring environment similar to a digital twin.

[0050] While constructing the network topology, the system assigns parameter monitoring conditions to each node in parallel. This process essentially involves deeply integrating user-defined threshold data with the network nodes. Each node object contains one or more monitoring condition objects. Each monitoring condition object defines in detail a monitored parameter (such as "mean arterial pressure"), its lower and upper limits of safe range, acceptable short-term fluctuation range, and the initial alarm level to be triggered when the parameter exceeds the limits. These conditions become the direct basis for making preliminary judgments on the real-time data signal stream at that node.

[0051] Finally, the system integrates the constructed medical device topology network with all allocated parameter monitoring conditions to generate the final patient monitoring overview. This patient monitoring overview is a composite data structure containing a complete medical device relationship graph and detailed monitoring rules attached to each node in the graph. Internally, the patient monitoring overview object is continuously used to drive the engine for real-time data signal synchronization. The real-time data signal stream flows and aggregates between medical device nodes on demand according to the topological relationships defined in the patient monitoring overview. For example, when the real-time heart rate data of the ECG monitor node becomes abnormal, the system will immediately know, based on the topological relationships, that this abnormality may affect the infusion pump node and the ventilator node, thereby triggering more complex risk prediction calculations based on multi-parameter fusion for the latter two.

[0052] By parsing user configurations, constructing a medical device topology network, allocating parameter monitoring conditions, and performing deep integration, the system successfully generates a patient monitoring overview that can both statically reflect the layout and parameter requirements of medical devices and dynamically describe the data interaction logic. This patient monitoring overview provides indispensable, structured contextual information for subsequent adaptive risk prediction and accurate alarms, enabling the entire alarm system to intelligently understand the treatment process rather than judging individual parameters in isolation, greatly improving the accuracy and clinical applicability of alarms.

[0053] Example 2. Refer to Figure 2The process of determining the adaptive prediction strategy for intensive care is as follows: A preset first mapping table is invoked, the remaining time is input into the first mapping table, and an initial prediction is obtained through matching; the number of risk types corresponding to the risk reference threshold is obtained; an adjustment factor is obtained by querying a second mapping table based on the number of risk types; the initial prediction interval is multiplied by the adjustment factor to obtain an optimized prediction interval; and the adaptive prediction strategy is formulated based on the optimized prediction interval. Specifically, in determining the adaptive prediction strategy, the system uses clinical key nodes as time anchors (such as the early 24 hours postoperatively, sedation / reduction / discontinuation of sedation drugs, extubation trials, and various stages of septic shock ROSE, etc.). Dynamic strategy adjustment is mainly achieved through a comprehensive analysis of three key factors: remaining time, number of risk types (divided by organ system / treatment domain, such as respiratory, circulatory, neurological, renal, hepatic, coagulation, metabolic / endocrine, infection, gastrointestinal / nutrition, device / treatment safety), and the total number of medical devices involved in monitoring. The remaining time refers to the time difference between the current time and the end time of the preset monitoring period. There is a preset first mapping table that stores the correspondence between different remaining time ranges and the initial prediction interval. In specific operation, the calculated remaining time value is used as the input item to find the matching entry in the first mapping table, thereby obtaining an initial prediction interval value. This initial prediction interval represents the basic value of the system risk prediction and assessment frequency based on time urgency, without considering other factors.

[0054] The risk reference threshold is a warning line set by the user based on actual monitoring needs. Associated with this risk reference threshold is the number of different risk types it covers. The system needs to identify and count the total number of specific risk categories corresponding to the risk reference threshold. There is a preset second mapping table that defines the mapping rules between the number of different risk types and a value called the first adjustment value. Based on the identified number of risk types, the corresponding first adjustment value can be extracted by querying the second mapping table.

[0055] The system needs to count the number of all medical devices involved in the patient monitoring overview, i.e., the total number of medical devices; there is a preset third mapping table, which records the relationship between different total numbers of medical devices and another value called the second adjustment value; based on the counted total number of medical devices, the corresponding second adjustment value can be extracted by querying the third mapping table.

[0056] After obtaining the first adjustment value and the second adjustment value, they need to be integrated to form the final adjustment factor. This integration is performed using a weighted summation method. The system has fixed weight coefficients, which are assigned to the first adjustment value and the second adjustment value respectively. The first weight coefficient is applied to the first adjustment value, and the second weight coefficient is applied to the second adjustment value. The first adjustment value is multiplied by the first weight coefficient, and the second adjustment value is multiplied by the second weight coefficient. Then the two products are added together, and the result is the adjustment factor.

[0057] After the adjustment factor is generated, it is mathematically operated on with the initial prediction interval previously obtained from the first mapping table; specifically, the initial prediction interval value is multiplied by the adjustment factor value, and the result of this product operation is defined as the optimized prediction interval; the optimized prediction interval takes into account the influence of time factors, risk complexity and medical equipment scale, and reflects a more reasonable risk prediction and assessment interval time requirement under the current specific monitoring scenario.

[0058] Based on the calculated optimized prediction interval, the system formulates a final adaptive prediction strategy. The core of this strategy is to control the execution frequency of the risk prediction module. The system maintains a sequence of prediction execution time points. Whenever the current time reaches the next preset time point in the sequence, a new risk prediction assessment process is triggered. The adaptive prediction strategy dynamically sets and updates the interval between these prediction execution time points, setting it as the value of the optimized prediction interval. In this way, the frequency of risk prediction assessment is no longer fixed, but will be automatically and dynamically adjusted according to the real-time remaining monitoring time, the complexity of the risk type to be monitored, and the number of medical devices actually involved in the monitoring. When the remaining monitoring time is short, the prediction interval is shortened and the assessment frequency is increased. When the risk type is complex or involves many medical devices, the prediction interval may be adjusted accordingly to match the intensity of the monitoring needs.

[0059] In the initial prediction interval query phase, the design of the first mapping table typically reflects a correspondence where the shorter the remaining time, the smaller the initial prediction interval value is set. This design allows the system to automatically increase the frequency of risk prediction assessments as the monitoring period nears its end, in order to cope with more likely changes under time pressure. The construction of the second mapping table considers the impact of the complexity of risk categories on the prediction frequency. The more risk types there are, the more the first adjustment value is usually set, resulting in a compression effect on the original prediction interval. This means that higher risk complexity requires the system to perform more frequent assessments to capture potential cross-risks. The third mapping table focuses on the impact of the scale of medical equipment. The larger the total number of medical equipment, the more data sources the system needs to process and analyze, and the more complex the potential correlations. The design of the second adjustment value also usually promotes a shorter prediction interval to adapt to the increased timeliness requirements in large-scale medical equipment scenarios.

[0060] The weighting coefficients used in the weighted summation process are preset by the system, and their values ​​are configured based on experience or application scenarios. The first weighting coefficient represents the importance of the complexity of the risk type, and the second weighting coefficient represents the importance of the quantity and scale of medical equipment. The relative size of these two coefficients determines whether the calculation of the adjustment factor is more inclined towards risk complexity factors or medical equipment scale factors. The reasonable configuration of weights plays a role in whether the final optimized prediction interval can effectively balance the needs of multiple factors.

[0061] In handling extreme situations, the system includes corresponding logical judgments. For example, when the calculated remaining time is lower than a preset minimum threshold, the first mapping table can directly return a minimum prediction interval, forcing the system to enter a high-frequency prediction mode and ignoring subsequent adjustment factor calculations to cope with emergency situations. When the number of risk types is zero or the total number of medical devices is zero, the system has a default processing mechanism that can return the set baseline adjustment value to maintain basic prediction functions. After the optimized prediction interval is calculated, the system will compare it with the maximum and minimum allowed values. If the calculation result exceeds the upper limit, the upper limit value will be automatically taken; if the calculation result is lower than the lower limit, the lower limit value will be automatically taken to ensure that the prediction interval is within a controllable and effective operating range.

[0062] This implementation method organically combines the remaining duration of the time dimension, the number of types of risks, and the scale of medical equipment in the resource dimension through multi-level mapping table queries and weighted calculations. This is ultimately quantified into the key parameter of optimizing the prediction interval, driving the dynamic changes of the adaptive prediction strategy. Specifically, the time dimension connects to key clinical decision-making nodes (such as the first 24 hours post-surgery, sedation / reduction / discontinuation of sedation drugs, extubation trials, early post-operative period after intracranial hematoma evacuation, and various stages of septic shock ROSE). The risk dimension is counted and stratified according to organ system / treatment domain (respiratory, circulatory, neurological, renal, hepatic, coagulation, metabolic / endocrine, infectious, gastrointestinal / nutritional, device / treatment safety). The resource dimension is determined by the actual number and complexity of the types of medical equipment currently involved in monitoring. The mapping table combines these three elements and translates them into regulatory signals, enabling the strategy to adapt to changes in disease progression and scenario. The entire process enables the system to autonomously adjust the pace of risk prediction and assessment based on real-time key parameters of the monitoring environment (time, risk complexity, number of medical devices) without frequent user intervention. This makes the assessment more aligned with the actual situation of the current monitoring task. For example, after patients with acute respiratory distress syndrome (ARDS) enter the extubation and light sedation phase, the time dimension becomes tighter and the weight of the respiratory domain increases, so the system automatically shortens the prediction interval. When postoperative cardiac patients require combined IABP, transthoracic echocardiography (TTE) related assessment parameters, invasive arterial pressure, microcirculation attention indicators, and complex vasoactive drug microinfusions, the complexity of the devices and circulatory domain increases, and the assessment frequency increases. Conversely, when patients enter the transitional monitoring unit during the perioperative period and their vital signs tend to stabilize, the time and risk intensity decrease, and the system correspondingly lengthens the prediction interval to reduce the alarm load, thus finding a balance between resource efficiency and timely risk warning. The design of the mapping table and the setting of the weight coefficients are the core of the parameterized configuration of this method, allowing for adjustments and optimizations based on different application scenarios or user preferences.

[0063] Reference Figure 3 A multi-dimensional correlation analysis diagram of adaptive prediction strategies for intensive care, including: sub- Figure 3-1 The relationship between remaining duration and detection prediction interval; sub Figure 3-2 The relationship between the number of risk types and the first adjustment value; sub Figure 3-3 The relationship between the total number of medical devices and the second adjustment value; sub Figure 3-4 The relationship between risk type and optimized prediction interval. As a multi-dimensional quantitative analysis platform for adaptive prediction strategies in intensive care, it fully presents the dynamic generation logic of the strategy through three core dimensions—time, risk, and resources—and multi-dimensional coupling results; among them: in the time dimension, Figure 3 sub Figure 3-1Using the remaining duration (remaining time of the monitoring cycle) as input, a first mapping table establishes a correlation that "the shorter the remaining duration, the smaller the initial prediction interval," reflecting the driving effect of time urgency on prediction frequency. That is, as the monitoring cycle progresses, the system spontaneously increases the prediction frequency, providing basic time parameters for the strategy to adapt to changes at the end of the cycle and meet demands. In the risk dimension, Figure 3 sub Figure 3-2 Based on the number of risk types covered by the risk reference threshold, a relationship is constructed using the second mapping table: "the more risk types there are, the stronger the compression of the prediction interval by the first adjustment value." This reflects the need for high-frequency assessment in complex risk scenarios. By forcing the system to assess more frequently to capture cross-risks, it provides a basis for risk correction of the adjustment factor. In terms of resources... Figure 3 sub Figure 3-3 Based on the total number of medical devices involved in monitoring, a third mapping table is used to establish the logic that "the larger the total number of medical devices, the shorter the prediction interval due to the second adjustment value." This adapts to the more complex data sources and relationships in large-scale medical device scenarios, providing resource correction parameters for the adjustment factors to ensure timely assessment. Figure 3 sub Figure 3-4 As a presentation of multi-dimensional coupling, the system integrates the dimensions of time (remaining duration) and risk (number of types) with the implicit impact of the scale of medical equipment. A heatmap visually displays the optimized prediction interval. The multi-factor coupling effect is quantified through the product of the initial prediction interval and the adjustment factor (generated by a weighted sum of the first and second adjustment values). This clearly presents the dynamic adaptation rule that "the shorter the remaining duration and the more risk types, the shorter the optimized prediction interval," driving the system to autonomously adjust the prediction frequency to balance resource efficiency and timely warnings. The entire chart system, through its hierarchical design of "single-dimensional mapping - multi-dimensional coupling," transforms time pressure, risk complexity, and resource scale into quantifiable optimized prediction intervals. This effectively supports the dynamic adjustment of the adaptive prediction strategy of the intensive care system, achieving full-cycle, full-scenario monitoring needs matching without manual intervention.

[0064] Example 3. Refer to Figure 4The process for generating risk prediction values ​​for intensive care is as follows: A pre-trained feature extraction module is loaded, which completes the training process using historical datasets; based on the optimized prediction interval, the feature extraction module is controlled to perform feature fusion on the real-time monitoring parameters of each medical device and the connection relationship; a fused feature vector is generated; based on the fused feature vector, the risk prediction value is calculated through a dynamic evaluation model; the risk prediction value represents the comprehensive risk level of all medical devices in the patient monitoring overview within the preset monitoring period. Specifically, the generation process of risk reference thresholds involves the integration and processing of multi-source data. The system receives a set of original reference values ​​configured by the user for different risk types. These values ​​are distributed across different dimensions and numerical ranges, and normalization transformation is executed as a key step; the system identifies the type attribute of each original reference value and its corresponding risk category, extracting the pre-set upper and lower baseline values ​​for that category; a linear scaling operation is performed: the original reference value is subtracted from the lower baseline value, and the difference is divided by the difference between the upper and lower baseline values. This operation maps all original reference values ​​to a closed interval between 0 and 1, forming a standard. Standardization eliminates dimensional differences, making the warning thresholds for different risk types numerically comparable. After standardization, the system initiates the fusion processing phase, maintaining a dynamic weight configuration table that assigns weight coefficients based on the priority of risk types in clinical guidelines. The sum of these weight coefficients is strictly limited to 1. The fusion processing executes a weighted average algorithm: each standardized reference value is multiplied by its corresponding weight coefficient, and all products are summed to obtain the final risk reference threshold. Users can adjust the weight configuration through the management interface to adapt the threshold generation strategy to the diagnostic and treatment requirements of different diseases or medical institutions.

[0065] The calculation process for risk prediction values ​​relies on a pre-trained feature extraction module. This module was established through a historical data training phase, and the training dataset contains complete historical monitoring parameter sequences of medical devices and their clinical risk annotation results. The module architecture uses a multi-scale convolutional neural network as its main structure, including convolutional layers in the time dimension and connection relationship processors in the spatial dimension. When the real-time monitoring process starts, the system first loads the trained feature extraction module parameters into memory. Based on the optimized prediction interval timestamp output by the adaptive prediction strategy, the system sets the trigger time point sequence for feature extraction. Each time it is triggered, the module reads the latest monitoring parameter set of each medical device from the data buffer. The key operation lies in the modeling and processing of connection relationships: each medical device is abstracted as a node in the topological network, and the data flow or physical connection between medical devices is abstracted as an edge; the node feature vector is filled with real-time monitoring parameters, and the edge feature vector describes the connection type and strength; a dedicated edge convolutional layer is configured in the convolutional neural network, which extracts the correlation features between adjacent medical devices; each convolutional kernel performs feature cross-operation between the node features and the features of its neighboring medical devices to capture local relationship patterns; after multiple rounds of edge convolution, the high-level feature tensor simultaneously encodes the state of the medical device itself and its correlation influence in the network; the final output of the feature extraction module is a fixed-dimensional fused feature vector, which fully represents the real-time status of the patient monitoring overview and the dynamic interaction effects between medical devices.

[0066] The dynamic assessment model is implemented using a multilayer perceptron architecture. The input layer receives the fused feature vector generated by the feature extraction module. The hidden layer is designed as three fully connected layers, each equipped with a linear transformation unit and an activation function. The output layer generates a scalar value between 0 and 1 through a sigmoid activation function. This scalar value is labeled as a risk prediction value, which quantifies the overall risk probability level of all medical devices covered by the patient monitoring overview within a preset monitoring period. The system sets a timer that is bound to the execution of the dynamic assessment model. Whenever the time window for optimizing the prediction interval expires, the system automatically starts the feature extraction process, sends the generated fused feature vector into the dynamic assessment model for forward propagation calculation, and updates the calculation results to the risk prediction value cache in real time. The trend between the old and new prediction values ​​can be used to assess the rate of deterioration of the patient's condition and the risk prediction. The calculation process for the value is completely independent of the alarm judgment logic. The generated results are transmitted to the alarm analysis module via shared memory. Users can view historical prediction curves on the monitoring interface to help judge the trend of disease development. The parameter files of the feature extraction module and the dynamic evaluation model are stored independently in the system configuration library, supporting hot model replacement operations according to the characteristics of different intensive care units. When a new type of medical equipment is added, the model supports incremental learning mode to adapt to the new data feature pattern. The system log records the input parameters, feature vector summary and output prediction value of each prediction process in detail, forming a traceable calculation audit chain. The risk reference threshold and the risk prediction value are expressed in the same dimension, and the difference between the two values ​​directly reflects the system's early warning needs. The entire process does not require manual intervention in the parameter processing stage, realizing end-to-end automated conversion from raw data to decision indicators.

[0067] Reference Figure 5 A multi-module analysis diagram of the risk assessment process for intensive care, including: sub-modules Figure 5-1 Standardize the distribution of reference values ​​for risk types; sub- Figure 5-2 For the trend of risk prediction interval changes; sub Figure 5-3 For the distribution of characteristic importance; sub Figure 5-4 To dynamically evaluate model performance. Figure 5 As a key quantitative analysis tool for risk assessment processes in intensive care, this document comprehensively presents the dynamic logic of risk assessment from the dimensions of threshold generation, prediction calculation, feature correlation, and model performance: In the risk reference threshold generation stage, corresponding sub-... Figure 5-1The "Standardized Reference Value Distribution for Risk Types" is based on multi-source raw reference values. First, it identifies the upper and lower limits of risk category benchmarks and normalizes the raw values ​​by linear scaling to the [0,1] interval, eliminating dimensional differences and forming standardized reference values. Then, relying on a dynamic weighting table allocated according to clinical guidelines and summing to 1, it uses a weighted average to fuse the standardized values, generating risk reference thresholds that can be flexibly adjusted to adapt to disease types and institutional treatment needs. The distribution of standardized values ​​for different risk types is presented in charts to support this process. For risk prediction value calculation, it is deeply integrated with an adaptive prediction strategy; the system will trigger the entire process of feature extraction and prediction calculation in an orderly manner according to the optimized prediction interval. Figure 5-2 The "risk prediction interval change trend" presents the pattern of risk prediction value changes over time under this prediction interval, clearly reflecting the risk fluctuation at different time points and providing data support for medical staff to intuitively grasp the rhythm of changes in patients' conditions. In the specific feature processing stage, the system relies on a pre-trained multi-scale convolutional neural network to perform feature extraction operations. In this process, medical devices are abstracted as nodes in a topological network, and the connections between medical devices are abstracted as edges, thereby constructing a node-edge feature vector containing the state of the medical device itself and the relationship between medical devices. With the help of the edge convolutional layer in the network, the real-time monitoring parameter features of each medical device are cross-operated with the features of adjacent medical devices, and finally a fused feature vector is output. This vector fully encodes the real-time status of the patient monitoring overview and the dynamic interaction effects between various medical devices. Figure 5-3 The “feature importance distribution” clearly reflects the criticality of each type of feature in risk assessment, which can help understand the model’s dependence on different monitoring information and provide direction for subsequent model interpretation and optimization. The dynamic assessment model composed of multi-layer sensing mechanisms receives the above-mentioned fused feature vectors, and after information processing by multiple fully connected layers and nonlinear transformation by activation functions, it outputs a risk prediction value in the 0-1 range. This value is used to quantitatively represent the comprehensive risk probability associated with all medical devices in the patient monitoring overview within the preset monitoring period. Figure 5 sub Figure 5-4The "Dynamic Evaluation of Model Performance" verifies the model's reliability in the risk prediction process by showcasing its core performance indicators, ensuring that the output risk prediction values ​​accurately reflect the risk status in actual monitoring scenarios. Throughout the process, the system log records in detail the input parameters, generated fusion feature vectors, and final output risk prediction values ​​for each prediction process, forming a complete audit chain for easy traceability and verification. The system also supports hot model replacement and incremental learning; when new medical equipment types are added or monitoring scenarios change, the model can be updated to adapt to new application requirements. The chart system, through a hierarchical design of "threshold standardization - feature association modeling - dynamic prediction verification," transforms the processing of multi-source data, the relationships between medical devices, and the model's computational logic into quantifiable risk indicators. This effectively supports the automated operation, interpretable analysis, and dynamic adaptation capabilities of the intensive care system's risk assessment, ultimately achieving end-to-end process coverage from raw vital sign monitoring data to risk warning results.

[0068] Example 4. The feature extraction process of patient vital signs data described in this invention adopts a multi-stage processing architecture. The system receives raw vital signs data streams from monitoring medical devices. These data streams include various physiological parameters such as electrocardiogram signals, blood oxygen waveforms, respiratory curves, non-invasive / invasive arterial pressure and central venous pressure waveforms, end-tidal carbon dioxide and flow-pressure curves, ventilator pressure / flow / volume loops, electroencephalogram (EEG), near-infrared brain oxygenation (NIRSrSO2), transthoracic echocardiography (TTE) related indicators, and pupillary parameters. The data preprocessing unit first performs quality detection on the raw signals, automatically identifying and removing data caused by body movement and turning, suctioning and back percussion, lead / tube dislodgement or loosening, water accumulation or blockage in the sampling tube, and low perfusion. Abnormal signals caused by vasoconstriction, IABP / ECMO / CRRT operation, electrosurgical excision, and defibrillation interference are analyzed. Specifically, ECG leads are tested for lead contact and pacing / electromyography artifacts; SpO2 is analyzed based on perfusion index and waveform consistency to identify body motion / hypoperfusion artifacts; and ventilator waveforms are screened for leaks, self-triggers, airway obstruction, and secretion effects, which are deemed unacceptable or treated as missing values. Valid data segments are fed into a time window segmentation module, which automatically determines the window length based on signal type. ECG signals use a fixed 300-millisecond window, oxygen saturation data uses a 500-millisecond window, and respiratory waveforms use an 800-millisecond window. Data within each time window undergoes independent feature calculations.

[0069] The waveform feature extractor comprises three parallel processing channels: the first channel calculates the statistical characteristics of the signal within the window, including basic statistics such as peak-to-peak amplitude, zero-crossing rate, and interquartile range; the second channel performs morphological analysis, detecting feature points in the waveform using a sliding difference algorithm, identifying key positions such as R-wave peaks and P-wave initiation points, and calculating time history parameters between each feature point; the third channel performs frequency domain transformation, applying a fast Fourier transform to the window data and extracting the energy proportion of the 0.04-0.15Hz frequency band as a low-frequency oscillation indicator. Trend feature analysis employs a double sliding window mechanism: the outer window span is set to five times the duration of the current window, and the inner window is aligned with the current processing window. The ratio of the current window mean to the outer window mean is calculated as a short-term trend indicator, and the angle between the current window slope and the outer window slope is used as a trend consistency indicator.

[0070] In the feature combination stage, waveform features and trend features are structurally integrated. The feature set generated for each time window is encoded into a fixed-dimensional feature vector. The vector elements are arranged in a specific order: statistical features occupy the first 8 dimensions, morphological features occupy the middle 12 dimensions, frequency domain features occupy the last 4 dimensions, and trend indicators occupy the last 3 dimensions. The feature vectors of consecutive time windows are stacked in chronological order to form a physiological feature sequence matrix. The row dimension of this matrix corresponds to the number of time windows, and the column dimension corresponds to the length of the feature vector. A missing value handling strategy is implemented during matrix generation. When the calculation of a feature in a certain window fails, the moving average of the first three valid windows is used to fill in the missing value.

[0071] The response learning module employs a bidirectional long short-term memory network architecture to process physiological feature sequences. The network input layer receives the feature matrix, and the embedding layer maps the original features to a high-dimensional space. The forward propagation path contains four LSTM layers with 128, 64, 32, and 16 units per layer, respectively. The backpropagation path uses a symmetrical structure. An attention mechanism is introduced in the intermediate layers to calculate the relative importance weights of features in each time window. The formula for generating these attention weights is as follows:

[0072] ;

[0073] in: This represents the LSTM hidden state vector at time t. For trainable weight matrix, For bias vectors, For attention parameter vectors, This refers to the attention weight at time t. The network output layer connects to a fully connected layer, converting the final feature representation into a static risk response matrix. The rows of this matrix correspond to the location numbers of medical devices, and the columns represent different risk dimensions. The matrix element values ​​range from 0 to 1, quantifying the potential threat level of each location in each risk dimension.

[0074] The planning of interventionable paths is based on the actual physical layout of the medical environment. The passage cost of a path segment needs to comprehensively consider the physical distance (quantified by the distance coefficient) and the complexity of turning operations (quantified by the turning penalty). The distance coefficient is the converted weight of the actual length of the path segment, and the turning penalty is the additional cost corresponding to the turning action. The specific parameters are preset according to the layout of the monitoring area. The system stores a digital plan of the monitoring area, which marks the fixed location coordinates of all medical devices. The path network is modeled as a graph structure, where nodes represent the location of medical devices and critical path points, and edges represent walkable path segments. The passage cost of a path segment is jointly determined by the distance coefficient and the turning penalty.

[0075] The values ​​in the static risk response matrix are mapped to the corresponding medical device location nodes, forming a risk heatmap. The calculation of the alarm device's trigger trajectory employs an improved heuristic search algorithm. This algorithm maintains a priority queue, with queue elements ordered based on a balance function between the cumulative risk value of the path segment and the path length. During the search process, the weight parameters of the heuristic function are dynamically adjusted to ensure that the trajectory covers high-risk areas while maintaining reasonable movement efficiency. After trajectory generation, it is converted into a sequence of control commands and sent to the alarm device's actuator via a wireless communication protocol. As the device moves along the predetermined trajectory, it monitors the deviation between its own position and the planned trajectory in real time. When the deviation exceeds a threshold, a trajectory recalculation process is triggered.

[0076] During system implementation, the update frequency of physiological characteristic sequences is synchronized with the original data acquisition rate. The parameters of the response learning module are automatically updated incrementally every 24 hours, incorporating the latest acquired monitoring data for fine-tuning training. The refresh cycle of the static risk response matrix is ​​set to 5 minutes, matching the routine rounds of clinical nursing. The topology information of the interventionable path is manually triggered for updating when the location of medical equipment changes. The motion status of the alarm device is monitored in real time through the indoor positioning system, and the location data is fed back to the path planning module to form a closed-loop control. Throughout the entire processing flow, the computational load of feature extraction and risk modeling is distributed between edge computing nodes and the central server, achieving optimized allocation of computing resources. Data transmission between each processing stage uses an encryption protocol to ensure the protection of patient privacy information. The system operation log records detailed input and output snapshots of each processing step, supporting post-event auditing and fault diagnosis. The user interface provides a real-time visualization of the risk heatmap, assisting medical staff in quickly locating high-risk areas.

[0077] Example 5. Refer to Figure 6The process for determining the trigger trajectory of an alarm device used in intensive care is as follows: The interveneable path is analyzed into multiple path segments; the locations of adjacent medical devices corresponding to each path segment are determined; the sum of static risk response values ​​of adjacent medical devices corresponding to each path is calculated; the sum is used as the information concentration of the path segment, and a path optimization algorithm is used to determine the trigger trajectory of the alarm device from the starting point. Specifically, in the actual application of intensive care units, the planning of interveneable paths and the generation of alarm device trigger trajectories need to be combined with the specific medical environment layout and medical equipment distribution; assuming that the floor plan layout of an intensive care unit is as shown in Table 1, it includes 6 main medical equipment location points (numbered D1-D6) and 3 critical path nodes (numbered P1-P3); these location points are interconnected through corridors and passageways, forming a complete path network topology. The three key path nodes (numbered P1-P3) are core transit nodes set in the interventionable path network based on the actual physical layout of the intensive care unit and the distribution of medical equipment locations. They are used to connect different medical equipment locations and their function is to realize path connectivity and trajectory planning adaptation between the locations of various medical equipment, which matches the construction logic of the interventionable path and the control requirements of the alarm device triggering the trajectory. Among them, the critical path node P1 is a transit node near the ward entrance area. One end connects to the nurse station (the starting point S of the interventionable path) and the other end directly connects to the medical equipment location point D1. At the same time, it is indirectly connected to the medical equipment location point D2 through a corridor branch. It is an essential node on the path between the ward entrance area and the core monitoring area (the area where D1 and D2 are located). It can provide key transit support for the alarm device to move from the starting point to the medical equipment location near the entrance. It meets the requirement of path segment connection nodes when the interventionable path is analyzed as multiple path segments. The critical path node P2 is a core transit node in the middle of the ward. It is directly connected to the medical equipment location points D3 and D4 respectively, and connects the critical path nodes P1 and P3 through a horizontal corridor. It is the core hub for the interaction between the central monitoring area (the area where D3 and D4 are located) and other areas in the ward.The setup of this node ensures that the alarm device can quickly switch to different medical equipment locations when moving within the central monitoring area, meeting the requirements for path flexibility and efficiency when controlling the alarm device's trigger trajectory on the intervened path based on the static risk response matrix. Critical path node P3 is a transit node near the inner ward area, connecting medical equipment locations D5 and D6 at one end and to critical path node P2 via a longitudinal corridor at the other, while also extending to the auxiliary passageway inside the ward. This node primarily supports the movement of the alarm device to the inner monitoring area (the area where D5 and D6 are located). Especially when risk warnings occur at high-risk medical equipment locations such as D6, it provides a direct path connection for the alarm device to move from the core transit area (the area where P2 is located) to the inner area, adapting to the technical solution of calculating path segment information concentration and determining the trigger trajectory, ensuring the timeliness of risk response to medical equipment locations in the inner area. The static risk response values ​​of each medical equipment location in this intensive care unit are shown in Table 1.

[0078] Table 1: Distribution of Risk Response Values ​​for Medical Device Locations

[0079]

[0080] The path resolution process first divides the entire interventionable path into several path segments. Taking the path from the nurse station (starting point S) to the medical equipment location D6 as an example, the system identifies three possible paths: S-P1-D1-D6, S-P2-D3-D6, and S-P3-D5-D6. Each path is automatically divided into several segments; for example, the path S-P1-D1-D6 is decomposed into three path segments: S→P1, P1→D1, and D1→D6. The system calculates an information concentration value for each path segment, which is determined by the static risk response values ​​of the nodes at both ends of the segment. For example, the information concentration of the path segment P1→D1 is calculated as a weighted average of the combined risk values ​​of node P1 (default risk value 0.5) and node D1.

[0081] During the information concentration calculation, the system adopts a dynamic weight allocation strategy. For straight corridor segments, the distance weight accounts for 70% and the risk value weight accounts for 30%. For turning path segments, the distance weight is reduced to 50% and the risk value weight is increased to 50%. This allocation method takes into account that more attention needs to be allocated to turning areas during actual movement. Taking the path segment D1→D6 as an example, this segment is a straight corridor with a length of 8 meters. The comprehensive risk value of D1 is 0.68 and the comprehensive risk value of D6 is 0.88. Its information concentration is calculated as (0.68×0.3+0.88×0.3)+(8×0.7 / 10) standardized value.

[0082] The path optimization algorithm is implemented using a modified Dijkstra's algorithm. During initialization, the starting point S is added to the visited set, and its distance to itself is recorded as 0. Then, all path segments directly connected to S are traversed, and the comprehensive cost of these path segments is calculated. The comprehensive cost consists of two parts: the physical length (in meters) of the path segment and a risk adjustment coefficient. The risk adjustment coefficient is non-linearly mapped according to the information concentration of the path segment; path segments with high information concentration receive lower travel cost weights. The algorithm maintains a priority queue, and each time, the path segment with the lowest current comprehensive cost is taken for expansion.

[0083] During trajectory generation, the system introduces a backtracking mechanism. When the algorithm reaches the endpoint D6, it not only records the shortest path but also retains alternative, suboptimal paths. These alternative paths have a high cumulative information density but may slightly increase the physical path length. During actual operation, the alarm device can dynamically switch between these pre-calculated paths based on real-time risk monitoring. The trajectory output is a series of continuous spatial coordinate points, each containing three-dimensional position information and a recommended movement speed. The motion control system of the alarm device converts these trajectory points into specific motor control commands.

[0084] In practical applications, when the system detects a sudden increase in the risk value of medical device D6 to 0.88, it triggers a trajectory replanning process. The system starts from the current location of the alarm device (let's say P2) and calculates the optimal trajectory to D6. Based on the real-time updated risk data, it may generate a path P2-D3-D6 that passes through D3, with a total length of 15 meters and a cumulative information concentration of 2.47. Alternatively, it may generate a direct path P2-D6, with a length of 12 meters, but which requires passing through a high-risk area. The system selects the former as the execution trajectory based on preset balance parameters, as it achieves a better balance between path length and risk coverage.

[0085] During the trajectory execution phase, the alarm device is equipped with multiple sensors to achieve precise positioning and environmental perception; infrared ranging sensors detect obstacles ahead, inertial measurement units track their own motion status, and RFID readers identify position markers on the path; these sensor data are compared with the predetermined trajectory in real time, and when the deviation exceeds the threshold (such as a position error greater than 0.5 meters or a directional deviation greater than 15 degrees), a local trajectory adjustment is triggered; during the adjustment process, the system comprehensively considers the current position, the risk distribution of the remaining path segment, and physical constraints to generate a smooth transition trajectory.

[0086] During system operation, the system continuously monitors the risk value changes of each medical device location. When the risk distribution of a certain path segment changes significantly (e.g., the comprehensive risk value of D3 increases from 0.79 to 0.85), the information concentration of the relevant path segment is immediately recalculated. If the change exceeds a preset threshold, the system automatically initiates a background replanning task to pre-calculate possible alternative trajectories, but does not immediately switch to execution. Online trajectory switching is only implemented when the current trajectory cannot meet the risk coverage requirements, thus avoiding unnecessary movement interruptions.

[0087] The trajectory visualization interface provides medical staff with intuitive operational feedback. On the electronic floor plan of the ward, the current location of the alarm device and the planned trajectory are displayed in real time. Different colors represent the risk level of the path segment: red represents high risk, yellow represents medium risk, and green represents low risk. Medical staff can adjust the balance parameters through the touch screen and manually select different trajectory schemes that emphasize risk coverage or path length. The system records all trajectory execution logs, including detailed information such as planning time, execution duration, and risk points covered, supporting subsequent analysis and optimization.

[0088] A maintenance and update mechanism ensures the accuracy of the path network. Medical equipment locations are calibrated monthly, and laser rangefinders are used to verify the actual distances between nodes. When ward layouts are adjusted or medical equipment locations change, administrators update the digital floor plan information using configuration tools, and the system automatically reconstructs the path network topology. Temporary obstacles (such as mobile equipment carts) can be marked in the system using a temporary marking function; these markings remain valid until manually removed or a preset expiration time is reached. The entire system design considers the dynamic characteristics of the clinical environment, maintaining a balance between accuracy and flexibility in path planning.

[0089] Example 6. During the execution of the trigger trajectory by the alarm device, the system simultaneously activates the real-time environmental parameter monitoring function; multiple environmental sensors are deployed at the location of the medical equipment, including a temperature and humidity composite sensor, a noise level detector, and an air quality monitoring unit; the sensors establish a continuous communication link with the IoT gateway, collecting and uploading raw environmental parameter data every three seconds (the collection frequency is the basic frequency of environmental monitoring, used for obtaining stable parameters in routine monitoring scenarios, supporting the data foundation for daily risk assessment); the data packet structure is clearly defined, including four environmental dimension parameters: timestamp, medical equipment location code, temperature value (degrees Celsius), relative humidity value (percentage), noise decibel value, and PM2.5 concentration (μg / m³); the receiving end service performs validity verification on the raw data, filters out abnormal data points that suddenly change or exceed physical limits, and distributes the data stream to the dynamic risk calculation engine.

[0090] The dynamic risk response value is generated using a real-time modeling method. The system maintains an environmental baseline file for each medical device location, which records the average environmental parameter levels for that location over the past 7 days at the same time. The difference between the currently collected environmental parameter vector and the corresponding baseline vector is fed into the evaluation model. This evaluation model is configured with a multi-dimensional weight adjustment matrix, and different risk sensitivity coefficients are set for different environmental parameters. The temperature parameter uses a non-linear response function, and the sensitivity coefficient is automatically increased when the deviation exceeds ±2℃. The noise parameter is processed in frequency bands, with noise in the 1000-4000Hz band assigned a higher risk weight. The deviation values ​​of each parameter are multiplied by the corresponding sensitivity coefficients and then normalized and combined. The calculation results are mapped to a dynamic risk response value between 0 and 1 using an S-curve function. This value is updated every five seconds and labeled with a location tag.

[0091] The response difference calculation module receives two inputs: static risk response value and dynamic risk response value. The system performs precise numerical matching based on the location code of the medical device, and data pairs with inconsistent location codes are automatically discarded. For valid data pairs, mathematical operations are performed to calculate the absolute difference between the dynamic risk response value and the static risk response value as the response difference. This response difference reflects the degree of deviation between the actual environmental state and the expected model prediction. The system is configured with multi-level judgment thresholds: below 0.02 is marked as blue (ignorable fluctuation), 0.02-0.05 is marked as yellow (observation level fluctuation), 0.05-0.1 is set as orange (warning level), and above 0.1 is marked as red (intervention level). Each calculation result, along with the timestamp and the location of the medical device, is written into the alarm device control command stream.

[0092] The alarm device employs a tiered response strategy for directional adjustment. When the response difference is at the blue level, the device maintains its original trajectory, with only the control panel indicator light displaying the current monitoring value. Upon entering the yellow level, the device's movement speed is reduced to 60% of its original rate, and the rotating platform scans the abnormal position direction every ten seconds, continuously sending directional alert signals to the control terminal without altering the physical travel route. When the difference exceeds the orange threshold, the directional control system initiates a trajectory correction program. This program first pauses the current movement command and generates a new turning angle based on the alarm device's current position coordinates, the abnormal medical equipment's position coordinates, and real-time environmental parameters. The turning calculation combines the shortest path principle with environmental risk distribution: prioritizing low-risk channels and avoiding high-temperature and high-humidity areas; if it is necessary to pass through a high-risk area, the protective shield opening command is activated. The red level response triggers the highest priority control, interrupting all predetermined tasks and immediately moving along the risk gradient descent direction, while simultaneously activating the multimodal alarm signal.

[0093] The environmental anomaly handling and recording mechanism fully tracks all operational events. Each directional adjustment generates an operation log, including response difference values, device position change information, speed adjustment status, and subsequent handling codes. Medical personnel can use the control terminal to trace the device's movement trajectory and corresponding environmental parameter curves at any time period. When the device reaches the target adjustment position, it automatically extends the dwell time at that point to three times the basic patrol duration, performs a comprehensive environmental scan, and generates a scan report. This report displays peak parameters, duration, and fluctuation patterns. The system automatically compares with historical similar events to generate a brief analysis summary. After all anomalies are handled, the device automatically executes a return procedure, returning to the nearest task path point to reload the main operating trajectory, while environmental monitoring returns to the base frequency. Throughout the process, the motor drive module of the alarm device's motion system receives high-precision control signals, achieving a steering angle resolution of 0.5° and a straight-line positioning error controlled within ±3 cm. The system automatically executes an environmental sensor calibration procedure every 24 hours, correcting monitoring deviations using data from the baseline environmental chamber. After the night mode is activated, the device's lighting system and noise suppression device are activated in tandem, reducing light intensity while maintaining 50% of the steering sensitivity parameters.

[0094] Example 7. The present invention proposes a remote intelligent alarm system for intensive care to implement the above method, comprising a remote monitoring platform, a data acquisition module, an overview construction module, a prediction module, and an alarm output module connected sequentially by data signals. The data acquisition module acquires patient vital sign data through the remote monitoring platform, including electrocardiogram signals and blood oxygen saturation parameters. The overview construction module generates a patient monitoring overview based on user configuration. The patient monitoring overview includes the data signal connection relationships between various medical devices based on the treatment process, as well as the real-time monitoring parameter ranges of each medical device. The patient monitoring overview is synchronized with each medical device in real-time. The prediction module calculates the remaining time between the current moment and the end time of a preset monitoring cycle, determines an adaptive prediction strategy based on the remaining time, and generates a risk prediction value corresponding to the patient monitoring overview according to the adaptive prediction strategy based on the real-time monitoring parameters of each medical device and the data signal connection relationships. The alarm output module compares the risk prediction value with a user-set risk reference threshold, obtains a deviation index, and outputs an alarm signal.

[0095] Preferably, the data acquisition module establishes a stable data transmission connection with the electrocardiogram monitor and pulse oximeter configured for patients in the intensive care unit through a remote monitoring platform, supports network communication protocols that comply with medical industry standards, and receives patient vital sign data output by medical devices in real time.

[0096] Preferably, the electrocardiogram (ECG) monitor acquires the patient's ECG signal every 1.5 seconds. This ECG signal includes heart rate, heart rhythm waveform, and ST segment changes. The pulse oximeter acquires blood oxygen saturation parameters and pulse rate information every 3 seconds. After preliminary processing by the medical equipment, this raw data is uploaded to the remote monitoring platform via an encrypted network link. The data acquisition module extracts the patient's vital signs data from the designated data interface of the remote monitoring platform, parses the coded data output by the ECG monitor into intuitive heart rate values ​​(unit: beats / min), converts the blood oxygen saturation data into a percentage format of 0-100%, and removes abnormal values ​​caused by signal interference during data transmission (such as invalid data where blood oxygen saturation drops sharply from 96% to 80% in a short period of time without clinical symptom support). Only valid data with fluctuations within a reasonable range over two consecutive acquisition cycles is retained to ensure that the valid data reflects the patient's true physiological state.

[0097] Preferably, the overview construction module receives configuration data input from the system user interface. This configuration data includes the types of medical devices required for the patient's treatment process, such as electrocardiogram monitors, body temperature monitors, blood gas analyzers, intracranial pressure monitors, infusion pumps, ventilators, and other monitoring instruments for critical care. It also includes medical device connection rules, specifying that the pressure data from the intracranial pressure monitor needs to be linked and analyzed with the heart rate data from the electrocardiogram monitor, the respiratory rate parameters of the ventilator need to be adjusted in conjunction with blood oxygen saturation data, and the mannitol administration rate of the infusion pump needs to be referenced to the intracranial pressure data. It also includes the real-time monitoring parameter ranges for each medical device. For example, the heart rate of the electrocardiogram monitor is set to 55-105 beats / min, the pressure range of the intracranial pressure monitor is 0-20 mmHg, the mannitol infusion pump rate is 10-20 ml / h, and the respiratory rate of the ventilator is 12-20 breaths / min. After parsing the configuration data, the overview module constructs a medical device topology network according to the medical device connection rules. Each medical device in the network corresponds to a node, and the node is labeled with the medical device name, model, and current working status. The data interaction relationship between nodes is represented by lines with arrows. For example, the line from the "intracranial pressure monitor node" to the "infusion pump node" indicates "the mannitol administration rate needs to be adjusted when intracranial pressure > 15 mmHg". Then, parameters are assigned to each node. The monitoring conditions are defined by binding the monitoring parameter ranges of each medical device to the corresponding nodes. For example, the "intracranial pressure monitor node" is bound to "marked as abnormal state when pressure > 20 mmHg". Finally, the medical device topology network and parameter monitoring conditions are integrated to generate the patient monitoring overview, which is displayed in the form of a visual chart on the terminal interface of the medical staff's workstation. At the same time, a real-time data signal synchronization mechanism is activated, which obtains the latest monitoring parameters from each medical device every 4 seconds and updates them to the corresponding nodes in the monitoring overview. For example, when the ventilator's respiratory rate is adjusted from 16 breaths / min to 18 breaths / min, the parameter information of the "ventilator node" in the patient monitoring overview is updated within 4 seconds to keep it consistent with the actual operating parameters of the medical device.

[0098] The prediction module first calculates the remaining time between the current moment and the end of the preset monitoring cycle. The patient's preset monitoring cycle is 48 hours, and the current moment is the 12th hour since monitoring started, resulting in a calculated remaining time of 36 hours. The prediction module then calls a preset first mapping table, which stores the correspondence between different remaining times and initial prediction intervals. The initial prediction interval for a remaining time of 36 hours is 25 minutes, and the prediction module uses this to determine the initial prediction interval. Next, it obtains the number of risk types corresponding to the user-defined risk reference threshold. The current risk reference threshold covers four risk types: abnormal heart rate, high intracranial pressure, insufficient blood oxygen saturation, and abnormal respiratory rate. The prediction module queries the second mapping table and extracts the first adjustment value corresponding to a risk type count of 4. Simultaneously, it counts the total number of medical devices in the patient monitoring overview as 4, and queries the third mapping table to extract the second adjustment value corresponding to a total of 4 medical devices. The two are then weighted and summed according to preset weights to obtain an adjustment factor. The initial prediction interval of 25 minutes is multiplied by the adjustment factor to obtain the optimized prediction interval. Based on this, an adaptive prediction strategy is formulated, generating risk prediction values ​​according to the optimized interval frequency.

[0099] According to the adaptive prediction strategy, the prediction module loads a pre-trained feature extraction module, which is trained on a large dataset of historical monitoring data from similar traumatic brain injury patients. This module is capable of extracting feature information from the monitoring parameters of medical devices. The feature extraction module obtains real-time monitoring parameters and connectivity relationships of each medical device from the overview construction module at optimized prediction intervals. For example, in one instance, the extracted data might include: ECG monitor heart rate 78 bpm, intracranial pressure monitor pressure 12 mmHg, infusion pump mannitol rate 15 ml / h, ventilator respiratory rate 16 breaths / min, and blood oxygen saturation 97%. Combining this with the rule of "intracranial pressure and infusion pump linkage" in the medical device connectivity relationships, these parameters are converted into a fused feature vector. The prediction module then inputs this fused feature vector into a dynamic assessment model, which calculates a risk prediction value. This value reflects the overall risk level of all medical devices within the remaining monitoring period.

[0100] The alarm output module obtains the risk prediction value from the prediction module and retrieves the user-defined risk reference threshold, which is set by medical staff based on the severity of the patient's condition. The alarm output module compares the risk prediction value with the risk reference threshold and calculates the deviation index. If the risk prediction value is lower than the risk reference threshold, the deviation index is within a safe range, and the module only integrates the risk prediction value, real-time parameters of each medical device, and deviation index into a monitoring record, stores it in the system database, and displays it in regular font on the terminal interface. If the risk prediction value exceeds the risk reference threshold, the deviation index reaches the alarm trigger condition, and the alarm output module immediately generates an alarm signal, which is then alerted by the audible and visual alarm on the medical staff's workstation. At the same time, an alarm window pops up on the terminal interface, clearly displaying the risk prediction value, deviation index, and marking the specific parameters exceeding the threshold and the location of the corresponding medical device, facilitating medical staff to quickly locate the source of risk and take timely and targeted clinical intervention measures.

[0101] It should be noted that, in the above specific embodiments, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, 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 a process, method, article, or apparatus.

[0102] In the specific embodiments of the present invention, any descriptions not covered herein are known in the art and may be implemented with reference to such known techniques.

[0103] The above specific implementation methods and embodiments are specific support for the technical concept of a remote intelligent alarm method and system for intensive care proposed in this invention, and should not be used to limit the scope of protection of this invention. Any equivalent changes or modifications made based on the technical concept of this invention shall still fall within the scope of protection of this invention.

Claims

1. A remote intelligent alarm method for intensive care, characterized in that, The specific steps include the following: The patient's vital signs data are acquired through a remote monitoring platform and a patient monitoring overview is generated. The patient monitoring overview includes the data signal connection relationship between various medical devices based on the treatment process, as well as the real-time monitoring parameter range of each medical device. A medical device topology network is constructed based on preset medical device types, medical device connection rules, and monitoring parameter threshold ranges. Nodes in the medical device topology network represent medical devices, and edges represent data interaction relationships between medical devices. The monitoring parameter threshold ranges are used to assign parameter monitoring conditions to each node. The medical device topology network and parameter monitoring conditions are then integrated to generate the patient monitoring overview. The patient monitoring overview is synchronized with the real-time data signals of each medical device; the remaining time between the current time and the end time of the preset monitoring cycle is calculated, and an adaptive prediction strategy is determined based on the remaining time. Based on the real-time monitoring parameters of each medical device and the data signal connection relationship, a risk prediction value corresponding to the patient monitoring overview is generated according to the adaptive prediction strategy; the risk prediction value is compared with a preset risk reference threshold to obtain a deviation index and output an alarm signal. The remaining time determination adaptive prediction strategy includes: calling a preset first mapping table; inputting the remaining time into the first mapping table to match and obtain an initial prediction interval; obtaining the number of risk types corresponding to the risk reference threshold; extracting a first adjustment value from a second mapping table based on the number of risk types; counting the total number of medical devices in the patient monitoring overview; querying a third mapping table based on the total number of medical devices to extract a second adjustment value; performing a weighted sum of the first adjustment value and the second adjustment value as an adjustment factor; multiplying the initial prediction interval by the adjustment factor to obtain an optimized prediction interval; and formulating the adaptive prediction strategy based on the optimized prediction interval. The process of generating risk prediction values ​​corresponding to the patient monitoring overview based on the real-time monitoring parameters of each medical device and the data signal connection relationships, according to the adaptive prediction strategy, includes loading a pre-trained feature extraction module, which completes the training process using historical datasets; and controlling the feature extraction module to perform feature fusion on the real-time monitoring parameters of each medical device and the data signal connection relationships according to the optimized prediction interval, generating a fused feature vector. The fused feature vectors are used to calculate the risk prediction value through a dynamic evaluation model. This risk prediction value represents the overall risk level of all medical devices in the patient monitoring overview during the preset monitoring period; wherein: The final output of the feature extraction module is a fixed-dimensional fused feature vector, which is formally expressed as: ； in: represents the feature tensor input to the convolutional layer, which is a stack of node feature matrices and adjacency matrices; conv represents a composite convolution operation including edge convolution and regular spatial convolution; σ represents the activation function; h(x) is the output fused feature vector. The method further includes: extracting key features from the patient's vital signs data to obtain a physiological feature sequence; generating a static risk response matrix from the physiological feature sequence through a pre-trained response learning module to obtain an interventional path; and controlling the interventional path of the alarm device based on each static risk response value in the static risk response matrix and the corresponding medical device location.

2. The remote intelligent alarm method for intensive care as described in claim 1, characterized in that, The risk reference threshold is obtained by normalizing the reference values ​​of different risk types, and then fusing the converted values ​​to obtain the risk reference threshold.

3. The remote intelligent alarm method for intensive care as described in claim 1, characterized in that, The patient's vital signs data are subjected to key feature extraction, including segmenting the patient's vital signs data into multiple time windows, extracting waveform features and trend features within each time window, and combining the waveform features and trend features to form the physiological feature sequence.

4. The remote intelligent alarm method for intensive care as described in claim 1, characterized in that, The static risk response values ​​and corresponding medical device locations in the static risk response matrix are used to control the trigger trajectory of the alarm device on the intervened path. This includes parsing the intervened path into multiple path segments, calculating the sum of the static risk response values ​​of adjacent medical device locations corresponding to each path segment as the information concentration of that path segment, and using a path optimization algorithm to determine the trigger trajectory of the alarm device from the overall starting point.

5. A remote intelligent alarm system for intensive care, used to implement the remote intelligent alarm method for intensive care as described in any one of claims 1 to 4, characterized in that, The system comprises a remote monitoring platform, a data acquisition module, an overview construction module, a prediction module, and an alarm output module, all connected sequentially by data signals. The data acquisition module acquires patient vital sign data, including electrocardiogram (ECG) signals and blood oxygen saturation parameters, through the remote monitoring platform. The overview construction module generates a patient monitoring overview based on user configuration. This overview includes the data signal connection relationships between various medical devices based on the treatment process, as well as the real-time monitoring parameter ranges of each medical device. The patient monitoring overview is synchronized with each medical device in real-time. The prediction module calculates the remaining time between the current moment and the end of a preset monitoring cycle, determines an adaptive prediction strategy based on the remaining time, and generates a risk prediction value corresponding to the patient monitoring overview according to the adaptive prediction strategy, based on the real-time monitoring parameters of each medical device and the data signal connection relationships. The alarm output module compares the risk prediction value with a user-defined risk reference threshold, derives a deviation index, and outputs an alarm signal.

6. A remote intelligent alarm system for intensive care as described in claim 5, characterized in that, The data acquisition module establishes a stable data transmission connection with the electrocardiogram monitor and pulse oximeter configured for patients in the intensive care unit through a remote monitoring platform. It supports network communication protocols that comply with medical industry standards and receives patient vital sign data output by medical devices in real time.

7. A remote intelligent alarm system for intensive care as described in claim 6, characterized in that, The electrocardiogram monitor collects the patient's electrocardiogram signal every 1.5 seconds. The electrocardiogram signal includes heart rate value, heart rhythm waveform and ST segment changes. The pulse oximeter collects blood oxygen saturation parameters and pulse rate information every 3 seconds.

8. A remote intelligent alarm system for intensive care as described in claim 5, characterized in that, The overview module receives configuration data input from the system user interface. This configuration data includes the types of medical equipment required for the patient's treatment process, such as electrocardiogram monitors, body temperature monitors, blood gas analyzers, intracranial pressure monitors, infusion pumps, ventilators, and other monitoring instruments for critical care. It also includes medical equipment connection rules, specifying that the pressure data from the intracranial pressure monitor must be analyzed in conjunction with the heart rate data from the electrocardiogram monitor, the respiratory rate parameters of the ventilator must be adjusted in conjunction with blood oxygen saturation data, and the mannitol administration rate of the infusion pump must be referenced to the intracranial pressure data. It also includes the real-time monitoring parameter ranges for each medical device.

Images