AICU critical patient risk early warning system based on medical data fusion

Abstract

The application discloses an AICU critical patient risk early warning system based on medical data fusion and relates to the technical field of medical information science; the application constructs a candidate directed graph according to extracorporeal membrane oxygenation and continuous renal replacement therapy combined operation data, generates an extracorporeal life support topology version by using a support adjusted pressure flow response, traces test data along the topology version to form a path, separates extracorporeal treatment contribution, generates patient side equivalent observation, accepts or reconstructs a patient state baseline according to the patient side equivalent observation, combines device side oxygen transfer changes, patient oxygenation response and continuous renal replacement therapy lactic acid net removal contribution, and outputs an extracorporeal support implementation attenuation risk early warning, thereby improving early warning reliability.

Description

Technical Field

[0001] This invention relates to the field of medical informatics technology, specifically to a risk warning system for critically ill patients in the AICU based on medical data fusion. Background Technology

[0002] As intelligent intensive care units gradually integrate extracorporeal membrane oxygenation (ECMO) equipment, continuous renal replacement therapy (CRRT) equipment, bedside testing systems, and vital sign monitoring systems, the equipment operation data and clinical test data generated during the treatment of critically ill patients can be used for patient status assessment. Taking a patient with severe respiratory failure who is simultaneously receiving intravenous extracorporeal membrane oxygenation (ECMO) and CRRT as an example, the CRRT circuit can operate through an independent vascular access or be connected to the ECMO circuit. The access position may be adjusted during treatment due to changes in access pressure, filter replacement, or vascular conditions. Before and after the adjustment, the sampling port name and test item name in the equipment interface may remain the same, but the oxygenator, filter, replacement fluid inlet node, and branch confluence relationship that the blood passes through before reaching the sampling port have changed. For example, after the CRRT reflux position is adjusted, the sampling port at the front end of the oxygenator may collect a mixed sample of blood that has been processed by the filter. Although the test value of this sample is within a reasonable range, the corresponding blood formation path is different from the same sample before the adjustment.

[0003] Existing risk warning schemes for critically ill patients can associate equipment operation records, vital signs, and test results with patient identification and collection time, and generate risk information based on threshold comparisons, trend changes, or predictive models. Existing extracorporeal life support monitoring can also analyze blood gas differences before and after oxygenation, extracorporeal blood flow status, and local anticoagulation status of filters, providing a data foundation for equipment operation management and patient treatment. When the extracorporeal circuit connection remains stable, continuous data from the same sampling port have relatively consistent source conditions. However, when the location of continuous renal replacement therapy is changed, the oxygenator is replaced, or the extracorporeal circuit is reconnected, data with different formation paths may still be classified into the same risk sequence based solely on the sampling port name and timestamp. If the system continues to use the patient risk baseline formed before the adjustment, changes in test values ​​caused by circuit reconstruction may be included in changes in the patient's condition, and local circuit processing results may also participate in the assessment of the patient's overall condition. Therefore, in the case of dynamic changes in the extracorporeal circulation topology, there is still a need for further improvement in how to confirm the actual formation path of test data and ensure that test data under different topological conditions correspond to consistent patient status objects. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: The AICU critical patient risk early warning system based on medical data fusion includes: The topology identification module acquires data on the combined operation of extracorporeal membrane oxygenation (ECMO) and continuous renal replacement therapy (CRRT), constructs candidate directed graphs based on port connection relationships, extracts the pressure-flow response that supports the adjustment, filters candidate directed graphs based on pressure-flow response constraints, and generates an extracorporeal life support topology version. The data correction module traces back from the test sampling node to the patient status reference node based on the in vitro life support topology version, generating a test data formation path; it establishes a material transfer model according to the test data formation path, performs an identifiability test on the contribution of the observed and the path to be estimated, separates the in vitro processing contribution, and generates equivalent observations on the patient side. The state reconstruction module inputs the equivalent observations from the patient side into the state space model and recursively estimates the patient state based on the observation relationship corresponding to the version of the extracorporeal life support topology. When the version of the extracorporeal life support topology changes, it performs a patient state baseline inheritance determination based on the innovative residuals of the new observations relative to the previous state, and reconstructs the patient state baseline if the inheritance conditions are not met. The risk warning module acquires the patient status sequence after baseline consolidation and establishes an individual conditional response interval based on changes in oxygen transfer on the device side and changes in the patient's oxygenation status. It generates a support implementation deviation based on the interval deviation where the patient's oxygenation response rate interval is lower than the lower boundary of the individual conditional response interval. It corrects the patient's lactate clearance rate interval based on the net lactate removal component of continuous renal replacement therapy corresponding to the path formed by test data. When the lower boundary of the device side oxygen transfer change interval is greater than zero, the support implementation deviation is greater than zero, and the upper boundary of the corrected lactate clearance rate interval is not greater than zero, it outputs an in vitro support implementation attenuation risk warning.

[0005] Furthermore, the specific operations for constructing candidate directed graphs are as follows: Establish a unified set of nodes corresponding to the patient-side interface, extracorporeal membrane oxygenation interface, continuous renal replacement therapy interface, and laboratory sampling interface; generate candidate directed edges based on port status, built-in blood flow direction of the device, and pipeline access records; for candidate directed edges with actual flow feedback, exclude candidate directed edges with inconsistent blood flow directions based on actual flow. Candidate directed graphs are generated by combining candidate directed edges based on the unidirectional transport relationship of the equipment, the continuity of the patient's drainage to return path, and the closure relationship of the continuous renal replacement therapy branch.

[0006] Furthermore, the specific operations for filtering candidate directed graphs are as follows: The system extracts support adjustment events that occur when the actual execution volume of the equipment is adjusted, and establishes stability check windows and response check windows based on the common sampling period, adjustment duration, and communication latency; it performs time-scale correction on pressure and flow data to generate response time intervals and local response sensitivities for monitoring nodes; it filters candidate directed graphs based on node reachability, allowed response direction, distinguishable response propagation order, and flow balance of measurable branch merging nodes; it generates an extracorporeal life support topology version when only one candidate directed graph is retained; and it generates topology-unidentifiable markers when no unique candidate directed graph is retained.

[0007] Furthermore, the specific steps for generating the test data formation path are as follows: Obtain a unique blood sample identifier, sampling time period, test sampling node, sample type, and test item identifier; determine the sampling validity based on the dead space of the sampling tubing, blood disposal records, and local fluid input status; trace back from the test sampling node along the directed graph corresponding to the extracorporeal life support topology version to the patient status reference node to generate a candidate test data formation path; The transport delay and blood sample source interval are determined based on the effective blood volume of the tube segment in the candidate test data formation path and the cumulative actual flow before sampling. When the blood sample source interval is located in the effective interval of the same in vitro life support topology version and the test sampling node has a unique node mapping, the candidate test data formation path is determined as the test data formation path. If the determination conditions are not met, the test data formation path is not generated.

[0008] Furthermore, the specific steps for generating patient-side equivalent observations are as follows: The accompanying detection quantities of the same blood sample, as well as the cross-node observations within the same protocol sampling event and the blood sample source interval located in the same stable topological time period, are combined to form a matching observation vector; the node transfer relationships corresponding to the test items are combined according to the blood flow sequence in the test data formation path to generate a path transfer operator; the feasible region of path contribution is established based on the observation error boundary, equipment operating range, and test item parameter range, and the feasible boundary of patient-side transport volume is obtained; When the feasible region of path contribution is non-empty and bounded, and the half-width of the feasible interval on the patient side is not greater than the corresponding direct measurement allowable error, an equivalent observation on the patient side is generated, and the risk access flag is set to valid; when the feasible region of path contribution is empty and unbounded, the risk access flag is set to invalid; when the half-width of the feasible interval on the patient side is greater than the corresponding direct measurement allowable error, the risk access flag is set to invalid.

[0009] Furthermore, the recursive processing of the state-space model is as follows: Establish fixed patient status coordinates according to the test item identifier and patient status reference node; group equivalent observations of patients with common time intersection in the blood sample source interval into the same status update group, and form observation intervals according to status components; form the status change rate interval based on the patient direct sampling observations formed before the current status update group, corresponding to different unique blood sample identifiers and with non-overlapping blood sample source intervals. When no support adjustments, device interlocks, or changes in the external life support topology occur between adjacent state update groups, the feasible interval of the preceding state is propagated, the propagation result is determined as the predicted state interval, and the intersection of the predicted state interval and the current observation interval is determined as the current feasible state interval; when the state change rate interval does not have finite boundaries, a temporary state interval is formed based on the current observation interval.

[0010] Furthermore, the specific procedures for baseline patient status processing are as follows: The shortest interval between the current observation interval and the predicted state interval is determined as the innovative residual. When the version of the extracorporeal life support topology changes, the last valid state interval of the previous extracorporeal life support topology is propagated to the first valid observation time of the current extracorporeal life support topology according to the state components, generating a cross-topology prediction interval. When the first valid observation interval and the cross-topology prediction interval have an intersection, the patient state baseline is inherited, and an initial state interval is formed based on the intersection. When the first valid observation interval and the cross-topology prediction interval are separated, baseline reconstruction is initiated. When a finite state interval is formed by two valid observations that correspond to different unique blood sample identifiers and whose blood sample source intervals do not overlap, baseline reconstruction is completed, and a patient state sequence after baseline inheritance is generated.

[0011] Furthermore, the specific procedures for determining the relationship between changes in oxygen transfer on the equipment side and changes in the patient's oxygenation status are as follows: The blood transport time interval is determined based on the effective blood volume and actual blood flow between the sampling nodes before and after the oxygenator; the blood sample source interval after the oxygenator is reverse-calculated to the sampling node before the oxygenator according to the blood transport time interval; when the reverse-calculated blood sample source interval intersects with the blood sample source interval before the oxygenator, the oxygen transfer interval on the device side is generated based on the local blood oxygen content interval before and after the oxygenator and the actual blood flow during the intersection period. The equipment-side oxygen transfer change range is determined based on the equipment-side oxygen transfer range before and after adjustment. When the lower boundary of the equipment-side oxygen transfer change range is greater than zero, an effective oxygen transfer adjustment event is determined. The patient oxygenation response rate range is generated based on the oxygenation status component of the same patient before and after the effective oxygen transfer adjustment event.

[0012] Furthermore, the specific steps for establishing an individual conditional response interval are as follows: From the historical effective oxygen transfer adjustment events of the same patient, select historical events that match the oxygenator instance, node mapping relationship, continuous renal replacement therapy device instance and operating status, and the patient's oxygenation status interval before adjustment with the current effective oxygen transfer adjustment event to form historical matching events; establish the feasible domain of individual response parameters based on the device-side oxygen transfer change interval and the patient's oxygenation response rate interval of the historical matching events. The current oxygen transfer change range on the device side is input into the feasible domain of the individual response parameters to generate the individual conditional response range. The individual conditional response range has finite boundaries and the lower boundary is greater than zero. When the upper boundary of the current patient oxygenation response rate range is lower than the lower boundary of the individual conditional response range, a deviation that supports realization is generated based on the interval and the minimum discernible response rate determined by the patient's oxygenation status measurement error and the response time interval. When the feasible domain of the individual response parameters is empty or unbounded, no deviation that supports realization is generated.

[0013] Furthermore, the specific procedures for correcting the effects of continuous renal replacement therapy on lactate changes and providing extracorporeal support to achieve attenuation risk warning are as follows: Four patient lactate status intervals were selected from the patient status sequence after baseline, and paired according to physiological time to form two lactate confirmation intervals and corresponding patient lactate clearance rate intervals. The four patient lactate status intervals correspond to different unique blood sample identifiers, and the blood sample source intervals do not overlap. The feasible region of path contribution established by the data correction module for lactate test data was projected onto the continuous renal replacement therapy lactate net removal component to generate the lactate net removal projection interval, and the lactate net removal intensity interval was generated according to the coverage duration within each lactate confirmation interval. A feasible region for lactate impact parameters is established based on the historical lactate evaluation intervals of the same patient under the same continuous renal replacement therapy device instance and operating conditions; the patient's lactate clearance rate interval is corrected based on the feasible region for lactate impact parameters and the interval of net lactate removal intensity, and a corrected lactate clearance rate interval is generated. When the deviation is greater than zero, both lactate confirmation intervals are completely covered by the net lactate removal projection interval, and the upper boundaries of both corrected lactate clearance rate intervals are not greater than zero, an in vitro support for attenuation risk warning is output; when the net lactate removal projection interval does not completely cover the lactate confirmation interval, an in vitro support for attenuation risk warning is not output; when the feasible domain of the lactate influence parameter is empty or unbounded, an in vitro support for attenuation risk warning is not output.

[0014] This invention provides a risk warning system for critically ill patients in the AICU based on medical data fusion, which has the following beneficial effects: 1. This invention uses the pressure and flow response after adjustments to the actual execution volume of the device as evidence for topology identification. Combining node reachability, response direction, response propagation order, and branch flow balance, multiple candidate directed graphs are excluded. An extracorporeal life support topology version is generated only when the connection relationship is uniquely identifiable. Thus, even if the interface name remains unchanged after the continuous renal replacement therapy branch is reconnected or the oxygenator or filter is replaced, the actual effective blood flow path can still be identified. If the connection relationship cannot be uniquely determined, subsequent risk calculations are blocked to avoid the entire patient status assessment chain being offset due to errors in static connection records.

[0015] 2. This invention further uses the extracorporeal life support topology version as the structural boundary, traces backward from the test sampling node to the patient status reference node, and uses the effective blood volume of the tube segment and the cumulative actual flow before sampling to restore the blood sample source interval; then, according to the blood flow sequence, the node transfer relationship is combined to jointly constrain the patient-side transport volume, filter net removal, branch dilution, and local mixing contribution. Only when the patient-side transport volume is identifiable relative to the extracorporeal processing contribution is the equivalent patient-side observation generated. Thus, the original detection values ​​of the blood sample after oxygenator, blood sample after filter, or branch mixed blood sample are no longer directly used as the patient's overall status. Instead, the data formed by different topologies and different sampling locations are converted into comparable observation results corresponding to the same patient status object, eliminating the mixing of local loop processing results into the patient status judgment from the path formation level.

[0016] 3. This invention propagates the last effective state of the previous topology version to the first effective observation time of the new topology version according to the patient state component. Based on the innovative residuals of the cross-topology prediction interval and the new observation interval, baseline continuation or baseline reconstruction is performed respectively. On this basis, the determined changes in oxygen transfer on the device side are compared with the historical conditional responses of the same patient, and the contribution of continuous renal replacement therapy to net lactate removal is separated from lactate changes. Thus, it is possible to distinguish the support realization attenuation of the device that has achieved oxygen transfer adjustment but the patient has not achieved corresponding oxygenation and lactate clearance responses from the apparent changes caused by loop reconnection, sampling path changes, and local clearance of continuous renal replacement therapy, so that the risk warning directly corresponds to the device support realization status and the patient's actual response under the current topology. Attached Figure Description

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

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This embodiment provides a risk warning system for critically ill patients in the AICU based on medical data fusion, including the following description: This implementation method is applied to critically ill patients in the AICU (Adult Intensive Care Unit) who receive combined extracorporeal membrane oxygenation (ECMO) and continuous renal replacement therapy (CRRT). Patient P017 received intravenous-venous ECMO support due to severe pulmonary infection with refractory hypoxia, and simultaneously received CRRT due to acute kidney injury and metabolic acidosis. The ECMO device example is ECMO-E03, and the oxygenator example is OX-0714A; the CRRT device example is CRRT-C02, and the blood filter example after replacement at 08:40 is HF-0714B.

[0020] After the blood filter was replaced, the inlet of the continuous renal replacement therapy (CRRT) was connected to the interface after the oxygenator, and the return end was connected to the interface before the extracorporeal membrane oxygenation (ECMO) pump. The treatment operation record only recorded that the CRRT was connected to the ECMO circuit, without recording the branch port number. At 11:20, an adjustment was made to the actual amount of oxygen transfer. After the adjustment, the lower boundary of the oxygen transfer change range on the device side was greater than 0, but the patient's oxygenation response rate range was lower than the individual condition response range. At the same time, the CRRT continued to remove lactate, and the uncorrected lactate detection value still showed a downward trend.

[0021] The joint operation data is processed in the order of topology identification, data correction, state reconstruction, and risk warning. The topology identification module outputs the in vitro life support topology version; the data correction module generates the test data formation path and equivalent observations on the patient side within this version; the state reconstruction module forms the patient state sequence after baseline consolidation; and the risk warning module outputs warning results based on changes in oxygen transfer on the device side, patient oxygenation response, and the contribution of net lactate removal from continuous renal replacement therapy.

[0022] Through hospital-authorized equipment interfaces, bedside data platforms, laboratory information systems, and treatment operation record interfaces, data such as port status, actual equipment execution volume, pressure and flow rate, equipment interlock records, laboratory sampling records, and laboratory results are collected. Each data record is associated with the patient identification replacement code P017 and bound to the equipment instance identifier, sample identifier, interface version, and parameter version. The patient's real identity mapping table is stored separately from the processed data, and data reading and result publication are controlled according to job permissions.

[0023] The object boundary in this implementation is jointly defined by the patient identification replacement code, device instance, logical port, unique blood sample identifier, and extracorporeal life support topology version. The time boundary includes the effective interval of the topology version, the stability check window and response check window that support adjustment events, the blood sample source interval of the test data, the common time interval of the status update group, the sampling interval before and after oxygen transfer adjustment, and the lactate confirmation interval. Data that crosses the corresponding time boundary is not merged with data outside the boundary.

[0024] The in vitro life support topology version consists of the currently valid directed graph, node mapping relationships, device instance identifiers, and valid time intervals; the test data formation path is the valid blood flow path within the same topology version, from the test sampling node back to the patient status reference node; the feasible domain of path contribution is the combination of patient-side transport volume and in vitro processing contribution that simultaneously satisfy local rank conditions, device operating range, test item parameter range, physical constraints, and observation residual conditions.

[0025] The equivalent observables on the patient side include the midpoint of the feasible interval on the patient side, the half-width of the interval, the blood sample source interval, the patient status reference node, and the risk access marker; the individual conditional response interval is jointly determined by the feasible domain of the individual response parameters formed by the same patient's historical matching events and the current equipment-side oxygen transfer change interval; the support realization deviation is the result of the interval interval between the current patient's oxygenation response rate interval and the individual conditional response interval after uncertainty scaling.

[0026] Each processing stage employs objective gating conditions. When a candidate directed graph cannot be uniquely confirmed, a topologically unidentifiable marker is generated. When the path contribution cannot be identified, the risk admission marker is set to invalid. When there are insufficient state observations, a temporary state interval is formed. When there is an observation conflict, the corresponding state component is set to unidentifiable. When there are insufficient historical events, invalid parameter feasible regions, or incomplete coverage of lactate removal records, the current event is kept in a pending confirmation state, and no external support is output to achieve attenuation risk warning.

[0027] Table 1 Key Time Nodes for Successive Implementation Methods:

[0028] The topology recognition module acquires data on the combined operation of extracorporeal membrane oxygenation (ECMO) and continuous renal replacement therapy (CRRT), constructs candidate directed graphs based on port connection relationships, extracts the pressure-flow response to support the adjustment, filters candidate directed graphs based on pressure-flow response constraints, and generates an extracorporeal life support topology version.

[0029] When constructing the candidate directed graph, a unified set of nodes is established corresponding to the patient-side interface, extracorporeal membrane oxygenation interface, continuous renal replacement therapy interface, and test sampling interface. Candidate directed edges are generated based on port status, built-in blood flow direction of the equipment, and pipeline access records. For candidate directed edges with actual flow feedback, candidate directed edges with inconsistent blood flow directions are excluded based on actual flow. Candidate directed edges are combined according to the unidirectional delivery relationship of the equipment, the continuity of the path from patient drainage to return flow, and the closure relationship of the continuous renal replacement therapy branch to generate the candidate directed graph.

[0030] After the replacement of the HF-0714B blood filter is completed at 08:40, the new filter, branch tubing, and sampling port are established as new node instances. The unified node set includes the patient-side drainage interface, patient-side return interface, ECMO blood pump pre-port interface, ECMO blood pump post-port interface, oxygenator pre-port interface, oxygenator post-port interface, continuous renal replacement therapy blood inlet interface, continuous renal replacement therapy blood return interface, blood filter pre-port interface, blood filter post-port interface, branch junction interface, and test sampling interface. Each node is bound to a device instance identifier and a logical port number. Test sampling interfaces that cannot be uniquely bound are set as nodes to be resolved.

[0031] Based on port status, built-in blood flow direction, and reconnection records, three candidate directed graphs are formed. Candidate directed graph G1 represents the continuous renal replacement therapy (CRRT) blood inlet interface connected to the oxygenator post-interface and the blood return interface connected to the extracorporeal membrane oxygenation (ECMO) pump pre-interface; candidate directed graph G2 represents the blood inlet interface connected to the ECMO pump post-interface and the blood return interface connected to the ECMO pump pre-interface; candidate directed graph G3 represents the blood inlet interface connected to the oxygenator post-interface and the blood return interface connected to the oxygenator pre-interface. The three candidate directed graphs are respectively tested for the unidirectional delivery relationship of the device, the continuity of the path from patient-side drainage to return, and the closure relationship of the continuous renal replacement therapy branch.

[0032] The oxygenator post-port, continuous renal replacement therapy (CRRT) blood inlet, and CRRT blood return ports have pressure or actual flow feedback, which can be used to verify the direction of candidate directed edges. The local tubing section from the ECMO blood pump post-port to the oxygenator pre-port does not have an independent flow sensor, and the main circuit flow is not used as the independent measured flow of this tubing section.

[0033] When screening candidate directed graphs, support adjustment events that occur when the actual execution quantity of the equipment is adjusted are extracted. Stability check windows and response check windows are established based on common sampling period, adjustment duration and communication latency. Time-scale correction is performed on pressure and flow data to generate response time intervals and local response sensitivities of monitoring nodes. Candidate directed graphs are screened based on node reachability, allowed response direction, distinguishable response propagation order and flow balance of measurable branch merging nodes. When only one candidate directed graph is retained, an extracorporeal life support topology version is generated. When no unique candidate directed graph is retained, an unidentifiable topology marker is generated.

[0034] The effective upload period of the devices participating in the response analysis is 1 to 2 seconds, and the preferred range of the common sampling period is 1 to 5 seconds, preferably 2 seconds, and the maximum effective upload period among the signals participating in the analysis is taken; when it exceeds 5 seconds, the current event is not used for propagation order judgment, the stability check window is preferably 20 to 60 common sampling points, preferably 30 common sampling points; the upper limit of the data missing ratio is preferably 10% to 20%, preferably 15%. In this case, the stability check window is 60 seconds, containing 29 effective sampling points, and the missing ratio is 3.3%.

[0035] When the change in the median value of the actual execution quantity of the device reaches the minimum resolution of the device and is maintained for 3 to 5 consecutive device sampling cycles, the time when adjustment is supported is determined. It is preferred to maintain 3 consecutive device sampling cycles. The relative amplitude of the supported adjustment should also be no less than the minimum recognizable relative adjustment amplitude formed by the minimum resolution of the device and the upper bound of the stable fluctuation. In this case, the minimum recognizable relative adjustment amplitude is 0.02. At 08:52, the actual blood flow of continuous renal replacement therapy was adjusted from 180 mL / min to 200 mL / min, with a relative adjustment amplitude of 0.111, which meets the event admission condition. The response check window is extended by 4 to 8 common sampling cycles after the 8-second adjustment duration, preferably by 6 common sampling cycles. The upper limit of the window is preferably 60 seconds. In this case, the response check window is 20 seconds. Events that do not reach the minimum recognizable relative adjustment amplitude do not form local response sensitivity and are not used for candidate directed graph screening.

[0036] The device timescale calibration uses two uniquely paired events.

[0037] The first item is a clamp opening event, which occurred at 08:42.

[0038] The second item is the blood pump restart event, which occurred at 08:47.

[0039] The timescale offsets of the extracorporeal membrane oxygenation (ECMO) device and the continuous renal replacement therapy (CRRT) device relative to the bedside server are 3.2 seconds and -1.1 seconds, respectively. The median timescale scaling ratio is 1.0004, and the correction residual is 0.4 seconds. The device communication delay is formed by the difference between the device's sending time and the platform's receiving time within the stability check window. The 10th percentile to the 90th percentile is taken as the effective delay interval, which in this implementation is 0.6 to 1.0 seconds. The corresponding communication delay fluctuation half-width is 0.2 seconds, and the time quantization half-width corresponding to the common sampling period is 1.0 seconds. The time quantization half-width, the timescale correction residual, and the communication delay fluctuation half-width are added together to obtain a response time uncertainty of 1.6 seconds. This method allows the response time uncertainty to be adjusted synchronously with the device upload cycle, timescale correction quality, and actual communication status.

[0040] Each monitoring node uses the median value of the original signal within the stability check window as the reference value, the median value of the slope of the sampling point as the reference drift rate, and the normalization scale is the larger of the difference between the 90th percentile value and the 10th percentile value within the stability check window, and twice the minimum resolution of the equipment. The response non-compliance is formed by adding the absolute value of the reference correction signal and the absolute value of the change in adjacent sampling.

[0041] The tolerance for false recognition of the response is preferably 0.05 to 0.10, and preferably 0.05; the number of consecutive out-of-bounds points is preferably 3 to 5 common sampling points, and preferably 3 common sampling points. When the rounded-up value of twice the uncertainty of the response time divided by the common sampling period is greater than 3, the rounded-up value is adopted.

[0042] The observable boundary of local response sensitivity is formed by dividing the response non-compliance boundary of the current monitoring node by the relative adjustment magnitude of the current support adjustment event, so that the node noise boundary and the event adjustment magnitude are in the same sensitivity dimension. The response non-compliance boundary formed by the oxygenator front interface within the stability check window is converted to a relative adjustment magnitude of 0.111, and the observable boundary is 0.21.

[0043] The response time intervals of the continuous renal replacement therapy (CRRT) blood inlet interface, blood return interface, and ECMO blood pump pre-port interface were 08:52:10.6 to 08:52:13.8, 08:52:12.4 to 08:52:15.6, and 08:52:14.2 to 08:52:17.4, respectively, with corresponding local response sensitivities of 1.18, 0.74, and -0.46. The local response sensitivity of the oxygenator pre-port interface was 0.05, which is below the observable boundary of 0.21.

[0044] Within the current constant pressure control mode and stable pipeline resistance range, the lower boundary of the normalized local response sensitivity of the oxygenator inlet interface corresponding to candidate directed graph G2 is 0.28, and the lower boundary of candidate directed graph G3 is 0.31, both higher than the observable boundary of 0.21. In the path equivalent resistance range formed by the 30 stable sampling points before the event, the path corresponding to G2 is 20.6 to 25.2 mmHg / (L / min), and the path corresponding to G3 is 22.9 to 28.0 mmHg / (L / min). The actual flow rate variation range in this event is 0.019 to 0.021 L / min, and the normalized scale of the oxygenator inlet interface is 1.40 mmHg. Multiplying the path equivalent resistance range by the actual flow rate variation range... The predicted pressure change range was obtained and divided by the normalized scale to obtain normalized responses of G2 and G3 of no less than 0.28 and 0.31, respectively. Both candidate directed graphs G2 and G3 require observable responses from the oxygenator pre-port, but the actual local response sensitivity is 0.05, which does not reach 0.21. Therefore, the corresponding candidate directed graphs conflict with the actual response evidence and are excluded. Candidate directed graph G1 does not require observable responses from the oxygenator pre-port and is consistent with the node reachability, allowed response direction, and the response propagation order of the continuous renal replacement therapy blood inlet interface before the extracorporeal membrane oxygenation pump pre-port. The actual response data of the current event is only used to screen candidate directed graphs and is not used to update the path equivalent resistance range and observable boundary in reverse.

[0045] Within the adjusted quasi-steady-state range, the inflow and outflow rates for continuous renal replacement therapy were 0.200 L / min and 0.199 L / min, respectively. The net fluid removal was converted to 0.0013 L / min and included as the net volume term of the outflow circuit. The flow imbalance was the absolute value of the result obtained by subtracting 0.199 from 0.200 and then subtracting 0.0013, which is 0.0003 L / min. The total allowable error for the calculation was 0.0027 L / min, and the normalized flow balance residual was approximately 0.11, which did not exceed 1.

[0046] At 08:57, the actual blood flow in the extracorporeal membrane oxygenation (ECMO) facility was adjusted from 4.20 L / min to 4.45 L / min, forming a second piece of evidence that is different from the adjustment source of the 08:52 event and whose response examination window does not overlap, still pointing to candidate directed graph G1.

[0047] At 08:58:06, candidate directed graph G1 was confirmed, and in vitro life support topology version T2 was generated.

[0048] From 08:40 to 08:58:06, the system remains in a topologically unidentifiable state, and data generated during this period will not be included in subsequent patient status processing.

[0049] The in vitro life support topology version T2 records the currently valid directed graph, node mapping relationships, and device instance identifiers.

[0050] T2 also records the online effective time, version termination time, and topology-unidentifiable markers, which serve as the structural basis for the data correction module to track and verify the data formation path.

[0051] The data correction module traces backward from the test sampling node to the patient status reference node based on the in vitro life support topology version, generating a test data formation path; it establishes a material transfer model according to the test data formation path, performs an identifiability test on the contribution of the observed and the path to be estimated, separates the in vitro processing contribution, and generates equivalent observations on the patient side.

[0052] When generating a test data formation path, the unique blood sample identifier, sampling time period, test sampling node, sample type, and test item identifier are obtained. The sampling validity is determined based on the dead space of the sampling tubing, blood disposal records, and local fluid input status. Candidate test data formation paths are generated by tracing backward from the test sampling node along the directed graph corresponding to the extracorporeal life support topology version to the patient status reference node. The transport delay and blood sample source interval are determined based on the effective blood volume of the tubing segment in the candidate test data formation path and the cumulative actual flow rate before sampling. If the blood sample source interval is located in the effective interval of the same extracorporeal life support topology version and the test sampling node has a unique node mapping, the candidate test data formation path is determined as the test data formation path. If the determination conditions are not met, no test data formation path is generated.

[0053] After T2 takes effect, the protocol sampling data generated from 09:02 to 10:30 is processed. Each test data record includes a unique blood sample identifier, protocol sampling event identifier, sampling start time, sampling end time, test sampling node, sample type, test item identifier, original test value, and the time when the test result is available online. For patient arterial blood, the arterial sampling interface A-ART is used as the patient status reference node; for blood samples before and after continuous renal replacement therapy filters, the blood flow position before entering the filter is used as the lactate transport reference node.

[0054] At 09:02, arterial blood sample B0902-A was collected from the patient. The sampling node and the patient status reference node were both A-ART. The blood sample source range was completely within the T2 effective range. The blood sample formed the patient's arterial oxygen content observation range of 132.0 to 135.0 mL / L and the patient's lactate observation range of 6.80 to 7.04 mmol / L. The half-width of the oxygen content and lactate ranges did not exceed the corresponding direct measurement allowable error. Therefore, the risk admission markers for both observations were set to effective and were used as the first effective patient status observation under T2.

[0055] Blood samples B0930-FOUT were collected after the blood filter at 09:30. The sampling period was from 09:30:02 to 09:30:10. The effective blood discard volume was determined to be 2 to 3 times the dead space volume of the sampling tubing, preferably 3 times. The dead space of the sampling tubing was 2.4 mL. The effective blood discard volume corresponding to the current parameter version was 7.2 mL. The actual blood discard volume in this case was 7.5 mL. The local flushing fluid was turned off during the sampling period, so the sampling was valid. If the blood discard volume was insufficient, the blood discard record was missing, or there was local fluid input during the sampling period, the current test data would not be eligible for patient status inversion.

[0056] Tracing backward from the post-filter sampling node along T2 to the upstream reference node of the continuous renal replacement therapy blood inlet, the candidate test data formation path sequentially passes through the post-filter tube section, the blood side of the filter, and the pre-filter tube section, with effective blood volumes of 18 mL, 72 mL, and 28 mL, respectively. Based on the actual flow rate of 196 to 202 mL / min before sampling, the volume is accumulated at each sampling interval. The time required for the cumulative volume to reach 118 mL for the first time is 35.6 seconds. Combining the 1-second time recording accuracy, the blood sample source interval is 09:29:25.4 to 09:30:11.0.

[0057] The blood sample source interval is completely located within the T2 effective interval, and the post-filter sampling node has a unique node mapping. Therefore, the candidate test data formation path is determined as the test data formation path. If the actual flow is missing, the flow direction changes, the flow stops continuously, the cumulative volume cannot replace the effective blood volume of the tube segment, or the blood sample source interval crosses the topology version boundary, no test data formation path will be generated.

[0058] The pre-oxygenator blood sample B0916-OI and post-oxygenator blood sample B0916-OO, formed at 09:16, belong to the same protocol sampling event. The effective blood volume between the two sampling nodes is 365 mL, corresponding to an actual blood flow of 4.42 to 4.48 L / min and a transport time interval of 4.9 to 5.0 seconds. The minimum duration of the common feasible time period required for cross-node pairing is preferably 2 to 5 seconds, preferably the larger value between twice the comprehensive time uncertainty and one common sampling cycle. In this embodiment, 3.2 seconds is used. The source interval of the post-oxygenator blood sample after reverse conversion and the source interval of the pre-oxygenator blood sample have a common intersection of 3.4 seconds, which meets the pairing condition. The two blood samples enter the same cross-node pairing observation group.

[0059] When generating equivalent observations on the patient side, the accompanying detection quantities of the same blood sample, as well as cross-node observations within the same protocol sampling event and blood sample source intervals located in the same stable topological time period, are combined to form a matching observation vector; the node transfer relationships corresponding to the test items are combined according to the blood flow sequence in the test data formation path to generate a path transfer operator; the feasible region of path contribution is established based on the observation error boundary, equipment operating range, and test item parameter range, and the feasible boundary of patient-side transport volume is obtained.

[0060] Oxygenation-related observations used hemoglobin, blood oxygen saturation, and oxygen partial pressure under the same unique blood sample identifier to calculate blood oxygen content, without using accompanying test quantities corresponding to other unique blood sample identifiers. The lactate matching observation vector included lactate entering the blood during continuous renal replacement therapy, lactate returning to the blood, lactate in the filtrate, actual blood flow, and filtrate flow rate. The inputs to the material transfer model were the matching observation vector, actual time-weighted carrier flow rate, equipment operating status, and test parameters. The variables to be estimated were the patient-side lactate transport, the net lactate removal component of continuous renal replacement therapy, and the local dilution component of the sampling branch. The output was the predicted lactate observation corresponding to each test sampling node.

[0061] The path transfer operator combines node transfer relationships according to the blood flow sequence of the inlet, hemofilter, and return ends of continuous renal replacement therapy.

[0062] The model parameters read the equipment specifications of the blood filter HF-0714B, the current testing method, and the stable parameter version confirmed before 09:30.

[0063] The current blood sample is not used to update the parameter version of the current path transfer operator.

[0064] Before solving the interval, a local rank test is performed to confirm that the patient-side transport volume has an independent observation direction relative to the net lactate removal component and the local dilution component of continuous renal replacement therapy.

[0065] Five observations correspond to three variables to be estimated. The preferred range for the local sensitivity perturbation amplitude is 1% to 5% of the allowable range of the variables to be estimated, with 2% being preferred. The observations are normalized according to the allowable error, and the variables to be estimated are normalized according to the allowable range, and a 2% perturbation is applied to each. The lower limit of the perturbation is used to ensure that the predicted change in the observation is higher than the numerical rounding error, and the upper limit of the perturbation is used to maintain the effectiveness of the local linear approximation.

[0066] The preferred rank determination threshold for the normalized sensitivity matrix is ​​between 0.08 and 0.20. First, based on the repeated test error, flow measurement error, and predicted observation fluctuations caused by 2% parameter perturbation under the same stable parameter version, a maximum normalized invalid fluctuation of 0.096 is formed. Then, multiplied by a separation coefficient of 1.25, an optimal rank determination threshold of 0.12 is obtained. The current three singular values ​​are 2.31, 0.88, and 0.24, all higher than 0.12, and the effective rank is 3. After projecting the patient-side lactate sensitivity column to the subspace spanned by the lactate net removal sensitivity column and the local dilution sensitivity column, the normalized residual is 0.24, which is higher than 0.12. The patient-side lactate transport has an independent observation direction. When the rank condition is not met, the feasible region of path contribution is not forced through residual search.

[0067] After passing the rank test, feasible combinations are searched within the equipment operating range, the range of test item parameters, and the observation error boundary. The maximum value of the difference between the predicted and actual observations divided by the corresponding allowable error is 0.68, which satisfies the condition that the maximum normalized residual is not greater than 1. All feasible combinations constitute the feasible region of path contribution.

[0068] The feasible range for patient-side lactate, given by the feasible domain of the path contribution, is 6.63 to 6.87 mmol / L, with a half-width of 0.12 mmol / L. The preferred range for the allowable error of direct lactate measurement is 0.15 to 0.25 mmol / L, preferably 0.20 mmol / L, determined based on the repeatability validation and quality control records of the current lactate analyzer. The feasible domain is non-empty and bounded, and the half-width of the feasible patient-side range is no greater than 0.20 mmol / L. Therefore, the equivalent patient-side observation of 6.75 mmol / L is generated, and the risk access marker is set to valid. Simultaneously, the feasible range for patient arterial oxygen content is 132.0 to 136.0 mL / L, with a half-width of 2.0 mL / L. The preferred range for the allowable error of direct oxygen content measurement is 2.0 to 3.0 mL / L, preferably 2.5 mL / L, determined based on the interval propagation results of the allowable errors for hemoglobin, oxygen saturation, and oxygen partial pressure. Therefore, the equivalent patient-side observation with valid risk access is generated.

[0069] If any of the following conditions are met: the feasible domain of path contribution is empty, the patient-side transport volume has no finite boundary, or the half-width of the patient-side feasible interval is greater than the corresponding direct measurement allowable error, the risk access flag will be set to invalid, and the current test data will only be retained as the original test record and the local loop analysis record.

[0070] The feasible region of the path contribution formed at 09:30 retains the projection range of net lactate removal of continuous renal replacement therapy from 0.25 to 0.31 mmol / min. This projection range is used by the risk warning module to correct the time variation of lactate status between adjacent patients and is not repeatedly deducted during the data correction stage.

[0071] In the 09:30 protocol sampling event, the pre-filter blood sample B0930-FIN, post-filter blood sample B0930-FOUT, and patient arterial blood sample B0930-A formed equivalent observation intervals of 6.60 to 6.90 mmol / L, 6.63 to 6.87 mmol / L, and 6.58 to 6.92 mmol / L for patient-side lactate, respectively. The three observations correspond to different unique blood sample identifiers, and the blood sample source intervals have a common temporal intersection. The common intersection of the observation intervals is 6.63 to 6.87 mmol / L, which is used as the lactate observation interval for the 09:30 status update group.

[0072] The state reconstruction module inputs the equivalent observations from the patient side into the state space model and recursively estimates the patient's state based on the observation relationships corresponding to the version of the extracorporeal life support topology. When the version of the extracorporeal life support topology changes, it performs a patient state baseline inheritance determination based on the innovative residuals of the new observations relative to the previous state, and reconstructs the patient state baseline if the inheritance conditions are not met.

[0073] The recursive processing of the state-space model is as follows: Establish fixed patient state coordinates according to the test item identifier and patient state reference node; group equivalent patient-side observations with common temporal intersections in the blood sample source intervals into the same state update group, and form observation intervals according to state components; form state change rate intervals based on patient-side direct sampling observations formed before the current state update group, corresponding to different unique blood sample identifiers and with non-overlapping blood sample source intervals; when no support adjustments, device interlocks, or changes in the external life support topology version occur between adjacent state update groups, propagate the preceding state feasible intervals, determine the propagation result as the predicted state interval, and determine the intersection of the predicted state interval and the current observation interval as the current state feasible interval; when the state change rate interval does not have finite boundaries, form a temporary state interval based on the current observation interval.

[0074] The fixed patient status coordinates include at least the patient's arterial oxygen content and the patient's lactate concentration. The status components are jointly determined by the test item identifier, the patient status reference node, the transport object, and the unit of measurement. Changes in the extracorporeal life support topology version only adjust the observation source and observation availability, without changing the medical object and unit of measurement corresponding to the status components.

[0075] The observation relationship corresponding to the extracorporeal life support topology version takes the node mapping relationship and the test item reference rule as input, maps the equivalent observation on the patient side to the fixed patient state coordinates, and generates the observation interval corresponding to the state component with the equivalent observation value on the patient side as the midpoint of the interval and the half-width of the feasible interval on the patient side as the base half-width. When the blood sample source interval crosses the support adjustment or equipment interlock, the observation mapping based on a single physiological time is not executed, and the current observation is only used as a temporary state constraint within the blood sample source interval.

[0076] The equivalent observations on the patient side were established into state update groups according to the blood sample source intervals. The source intervals of the three lactate observations around 09:30 were from 09:29:28 to 09:29:34. The three observations were assigned to the same state update group. The physiological assignment time was taken as the midpoint of the common intersection, 09:29:31. Data with only the first and last overlapping and no common intersection of all source intervals were not assigned to the same state update group.

[0077] The patient-side direct sampling observations used to establish the steady-state rate of change intervals were all taken from independent blood samples taken during the effective T1 period earlier than 08:40. The patient's arterial oxygen content at 02:20, 04:20, 06:20, and 08:20 were used to establish direct observation intervals of 130 to 134, 131 to 135, 132 to 136, and 132 to 136 mL / L, respectively. The lower boundary of the rate was formed by subtracting the upper boundary of the previous observation interval from the lower boundary of the subsequent observation interval, and the upper boundary of the rate was formed by subtracting the lower boundary of the previous observation interval from the upper boundary of the subsequent observation interval. These were then divided by a 120-minute time difference. The three effective observation pairs established steady-state rate of change intervals ranging from -0.033 to 0.042 mL / L / min.

[0078] The patient's lactate levels formed direct observation intervals of 6.90 to 7.10, 6.78 to 6.98, 6.66 to 6.86, and 6.54 to 6.74 mmol / L at the same time points. Three effective observation pairs formed a steady-state rate of change interval of -0.0027 to 0.0007 mmol / L / min. The number of effective observation pairs used to form the steady-state rate of change interval is preferably 3 to 6, with 3 being the preferred option. If there are fewer than 3 pairs, only a temporary state interval is formed. If there are more than 6 pairs, the 6 most recent effective observation pairs before the current state update group are used to limit the crossing of excessively long disease stages.

[0079] When there are no changes in support adjustment, device interlocking, or extracorporeal life support topology between adjacent state update groups, the previous state feasible interval is propagated according to the steady state change rate interval and the physiological attribution time difference to form the predicted state interval; the intersection of the predicted state interval and the current observation interval is determined as the current state feasible interval. During the support adjustment period, the steady state change rate interval is no longer used to test state continuity alone, and the last baseline valid state interval before adjustment is retained as the patient state reference before adjustment.

[0080] When there is a support adjustment or topology-indistinguishable interval between the last effective state of the previous version of extracorporeal life support and the first effective observation of the current version, a basic prediction interval is first formed according to the steady state change rate interval, and then the state transition extension is superimposed to generate a cross-topology prediction interval. The state transition extension is formed based on the same type of support adjustment event in the current treatment cycle of the same patient that is earlier than the current topology change. The number of historical events is preferably 3 to 6, preferably 3. The offset of the first patient-side direct observation interval after adjustment relative to the basic prediction interval is calculated, and the closed envelope of each offset is taken.

[0081] This implementation uses three similar diagnostic support adjustment events earlier than 08:20 to form a transitional expansion of the patient's arterial oxygen content status from -0.4 to 0.6 mL / L and a transitional expansion of the patient's lactate status from -0.01 to 0.01 mmol / L. When there are fewer than three historical events and the offset cannot form a finite boundary, the corresponding state component does not obtain the qualification to accept the cross-topological baseline and directly enters the baseline reconstruction.

[0082] When processing patient status baselines, the shortest interval between the current observation interval and the predicted status interval is determined as the innovative residual.

[0083] When the in vitro life support topology version changes, the last valid state interval of the previous in vitro life support topology version is propagated to the first valid observation time of the current in vitro life support topology version according to the state components, generating a cross-topology prediction interval.

[0084] When the first effective observation interval intersects with the cross-topology prediction interval, the patient status baseline is adopted, and an initial status interval is formed based on the intersection; when the first effective observation interval and the cross-topology prediction interval are separated, baseline reconstruction is initiated.

[0085] The baseline range of oxygen content at the last T1 test, formed at 08:20, from 132 to 136 mL / L, was used as the starting point for propagation.

[0086] The first effective oxygen content observation for T2 was formed at 09:02, with a propagation time of 42 minutes. Based on the steady-state change rate range of -0.033 to 0.042 mL / L / min, the basic prediction range of 130.6 to 137.8 mL / L was first obtained. Then, the state transition extension of -0.4 to 0.6 mL / L was superimposed to form the cross-topology prediction range of 130.2 to 138.4 mL / L.

[0087] The first effective oxygen content observation interval for T2 is 132.0 to 135.0 mL / L. This observation interval intersects with the cross-topology prediction interval, and the innovative residual is 0. The intersection of 132.0 to 135.0 mL / L forms the initial oxygen content interval under T2.

[0088] Starting from the T1 final lactate baseline interval of 6.54 to 6.74 mmol / L formed at 08:20, propagation proceeded for 42 minutes at a steady-state change rate interval of -0.0027 to 0.0007 mmol / L / min, first obtaining the basic prediction interval of 6.43 to 6.77 mmol / L. Then, a state transition extension of -0.01 to 0.01 mmol / L was superimposed to form a cross-topological prediction interval of 6.42 to 6.78 mmol / L. The first effective lactate observation for T2 was formed at 09:02, with an observation interval of 6.80 to 7.04 mmol / L and an innovation residual of 0.02 mmol / L. The two intervals were separate and did not inherit the lactate baseline. The inheritance threshold for the innovation residual was set to 0, that is, the existence of an intersection between the cross-topological prediction interval and the first effective observation interval was used as the baseline inheritance condition. Validation error, path inversion error, state transition uncertainty caused by support adjustment, and time attribution uncertainty were all included in the corresponding interval boundaries.

[0089] At 09:30, a second independent lactate status update group was formed, with lactate observation intervals ranging from 6.63 to 6.87 mmol / L. The preferred number of independent status observations for baseline reconstruction is 2 to 4, with 2 being the most preferred. Each status observation must be associated with at least one distinct blood sample identifier, and the corresponding blood sample source intervals must not overlap. The physiological attribution time of the first valid observation at 09:02 and the status update group at 09:30 differs by 28 minutes. The temporary status interval of 6.80 to 7.04 mmol / L formed at 09:02 is first divided into intervals of -0.0027 to 0.0007 mmol / L / min. The steady-state rate of change interval is propagated to 09:30, resulting in a predicted interval of 6.72 to 7.06 mmol / L. This interval is then intersected with the lactate observation interval at 09:30 to form a finite state interval of 6.72 to 6.87 mmol / L. Based on this, the lactate baseline reconstruction under T2 is completed. The steady-state rate of change interval is formed by direct sampling and observation on the patient side. Its state objects and measurement units do not change with the topology of the extracorporeal circuit. The temporary state interval is retained when only one valid state observation is formed. When the propagated predicted interval does not intersect with the observation interval of the second state update group, the baseline of the corresponding state component is kept invalid.

[0090] The midpoint of each feasible interval is used as the estimated value of the patient status, and half of the interval width is used as the uncertain half-width of the status. The patient status record simultaneously saves the physiological attribution time, the online effective time, the external life support topology version, and the baseline effective status. The late test result only forms a new patient status version from the actual online available time, without retroactively overwriting the published historical patient status records.

[0091] When the common intersection of observation intervals mapped to the same state component within the same state update group is empty, the state component is set to unidentifiable; other state components continue to be processed according to their respective observation intervals, and the risk access mark formed by the data correction module remains unchanged.

[0092] The risk warning module acquires the patient status sequence after baseline consolidation and establishes an individual conditional response interval based on changes in oxygen transfer on the device side and changes in the patient's oxygenation status. It generates a support realization deviation based on the interval deviation where the patient's oxygenation response rate interval is lower than the lower boundary of the individual conditional response interval. It corrects the patient's lactate clearance rate interval based on the net lactate removal component of continuous renal replacement therapy corresponding to the path formed by the test data. When the lower boundary of the device side oxygen transfer change interval is greater than zero, the support realization deviation is greater than zero, and the upper boundary of the corrected lactate clearance rate interval is not greater than zero, it outputs a risk warning of in vitro support realization attenuation.

[0093] When determining the relationship between changes in oxygen transfer on the device side and changes in the patient's oxygenation status, the blood transport time interval is determined based on the effective blood volume and actual blood flow between the sampling nodes before and after the oxygenator. The blood sample source interval after the oxygenator is reverse-calculated to the sampling node before the oxygenator according to the blood transport time interval. When there is an intersection between the reverse-calculated blood sample source interval and the blood sample source interval before the oxygenator, the device-side oxygen transfer interval is generated based on the local blood oxygen content interval before and after the oxygenator and the actual blood flow during the intersection period. The device-side oxygen transfer change interval is determined based on the device-side oxygen transfer interval before and after adjustment. When the lower boundary of the device-side oxygen transfer change interval is greater than zero, an effective oxygen transfer adjustment event is determined. The patient oxygenation response rate interval is generated based on the same patient's oxygenation status component before and after the effective oxygen transfer adjustment event.

[0094] At 11:20, the actual value of the gas source oxygen concentration was adjusted from 0.75 to 0.90, and the actual blood flow of the extracorporeal membrane oxygenation (ECMO) was adjusted from 4.45 L / min to 4.65 L / min. The preferred range of the quasi-steady-state control band after adjustment relative to the median value of the actual execution volume after adjustment is ±1% to ±3%, preferably ±2%. The number of continuous holding points is preferably 3 to 5 common sampling cycles, preferably 5 common sampling cycles. In this embodiment, the actual execution volume is continuously held within the ±2% control band for 10 seconds to confirm entry into the quasi-steady-state after adjustment. The equipment setting value is only used to locate candidate adjustment events. The effective oxygen transfer adjustment event is confirmed based on the oxygen transfer interval on the equipment side before and after adjustment.

[0095] Before and after adjustment, sampling was performed according to the protocol before and after the oxygenator. The effective blood volume between the sampling nodes before and after the oxygenator was 365 mL. Before adjustment, the actual blood flow was 4.42 to 4.48 L / min, corresponding to a blood transport time interval of 4.89 to 4.96 seconds. After adjustment, the actual blood flow was 4.62 to 4.68 L / min, corresponding to a blood transport time interval of 4.68 to 4.74 seconds. The minimum duration of the common feasible period was formed using the same method as the data correction module, with a preferred range of 2 to 5 seconds, and a preferred value of 3.2 seconds. The source interval of the blood sample after reverse conversion after the oxygenator and the source interval of the blood sample before the oxygenator formed a common feasible period of not less than 3.2 seconds. No flow rate adjustment, equipment interlocking, changes in the extracorporeal life support topology version, or oxygenator replacement occurred during the common feasible period.

[0096] Before the adjustment, during the common feasible period, the local oxygen content range before the oxygenator was 66.2 to 68.2 mL / L, and the local oxygen content range after the oxygenator was 135.0 to 137.2 mL / L. The lower boundary of the equipment-side oxygen transfer range was calculated by multiplying 4.42 L / min by the difference between 135.0 mL / L and 68.2 mL / L, and the upper boundary was calculated by multiplying 4.48 L / min by the difference between 137.2 mL / L and 66.2 mL / L, resulting in an equipment-side oxygen transfer range of 295 to 318 mL / min before the adjustment.

[0097] During the adjusted common feasible period, the local oxygen content range before the oxygenator was 65.8 to 68.0 mL / L, and the local oxygen content range after the oxygenator was 147.0 to 149.1 mL / L. The adjusted equipment-side oxygen transfer range of 365 to 390 mL / min was formed by multiplying the same ranges. The local oxygen content before and after the adjustment was calculated from the hemoglobin, blood oxygen saturation and oxygen partial pressure under the same unique blood sample identifier.

[0098] The oxygen transfer rate was adjusted by subtracting the oxygen transfer rate before adjustment from the adjusted oxygen transfer rate on the device side. The lower boundary of the change range was 47 mL / min (365 minus 318), and the upper boundary was 95 mL / min (390 minus 295). The oxygen transfer rate on the device side ranged from 47 to 95 mL / min. Since the lower boundary was greater than 0, an effective oxygen transfer adjustment event was determined. If the lower boundary was not greater than 0, the patient's oxygenation response deviation was not considered.

[0099] Before adjustment, the patient's oxygenation status was determined using the effective range of the last baseline before 11:15, which was 132.0 to 136.0 mL / L. After adjustment, the first eligible patient's oxygenation status was determined at 11:55, with a range of 132.4 to 135.8 mL / L. The physiological time difference was 35 minutes. Based on the boundary between the pre- and post-adjustment status ranges, the patient's oxygenation response rate range was determined to be -0.103 to 0.109 mL / L / min.

[0100] When establishing an individual conditional response interval, from the historical effective oxygen transfer adjustment events of the same patient, select historical events that match the oxygenator instance, node mapping relationship, continuous renal replacement therapy device instance and operating status, and the patient's oxygenation status interval before adjustment with the current effective oxygen transfer adjustment event to form historical matching events; establish a feasible domain for individual response parameters based on the device-side oxygen transfer change interval and the patient's oxygenation response rate interval of the historical matching events; input the current device-side oxygen transfer change interval into the feasible domain for individual response parameters to generate the individual conditional response interval.

[0101] The number of historical effective oxygen transfer adjustment events used to establish the feasible domain of individual response parameters is preferably 3 to 6, with 4 being the most preferred. The changes in oxygen transfer on the device side should cover at least 2 mutually separate levels, preferably 4 levels. In this study, 4 historical events were selected, with device-side oxygen transfer change ranges of 32 to 44, 46 to 58, 64 to 76, and 80 to 92 mL / min, respectively. The corresponding patient oxygenation response rate ranges were 0.13 to 0.19, 0.17 to 0.23, 0.22 to 0.29, and 0.27 to 0.34 mL / L / min, respectively. Each event used the same oxygenator instance, node mapping relationship, continuous renal replacement therapy operation status, and patient oxygenation status component.

[0102] Using the historical equipment-side oxygen transfer change range as input and the patient oxygenation response rate range as output constraint, for each historical event, the midpoint and half-width of the equipment-side oxygen transfer change range and the patient oxygenation response rate range are calculated respectively. The absolute difference between the midpoint of the patient oxygenation response rate and the sum of the baseline response term, the individual transfer coefficient multiplied by the midpoint of the equipment-side oxygen transfer change is calculated. When the absolute difference is not greater than the sum of the patient response half-width and the equipment-side input half-width converted by the individual transfer coefficient, the current combination of baseline response term and individual transfer coefficient is retained. The parameter combination that can simultaneously satisfy the four historical events constitutes the feasible region of individual response parameters.

[0103] The feasible range for the baseline response is 0.050 to 0.060 mL / L / min, and the feasible range for the individual transmission coefficient is 0.00255 to 0.00284 L. - ¹, The current event is not involved in the solution of this parameter version; Input the current range of oxygen transfer variation on the device side from 47 to 95 mL / min into the feasible domain, and calculate the minimum and maximum outputs under all parameters and input combinations respectively, to obtain the individual condition response range from 0.170 to 0.330 mL / L / min.

[0104] The individual conditional response interval has finite boundaries with a lower boundary greater than 0. The upper boundary of the current patient's oxygenation response rate interval (0.109 mL / L / min) is lower than the lower boundary of the individual conditional response interval (0.170 mL / L / min). Based on the interval interval and the minimum identifiable response rate, a deviation is generated to support the achievement. The allowable error for direct measurement of the patient's oxygenation status is consistent with that of the data correction module and is set to 2.5 mL / L. The minimum identifiable response rate is calculated by dividing twice the allowable error by 35 minutes, which is approximately 0.143 mL / L / min. The width of the individual conditional response interval is 0.160 mL / L / min, and the width of the current patient's oxygenation response rate interval is 0.212 mL / L / min. Therefore, the normalization scale is set to 0.212 mL / L / min, and the deviation to support the achievement is approximately 0.288.

[0105] If any of the following conditions are met: fewer than 3 historical matching events, empty feasible domain of individual response parameters, or no finite boundary of output in the current input direction, no support realization deviation is generated, and the current event is set to support response pending accumulation state; if the support realization deviation is greater than 0, the continuous renal replacement therapy lactate net removal contribution correction continues to be performed.

[0106] When correcting for the effect of continuous renal replacement therapy on lactate changes, four patient lactate status intervals were selected from the patient status sequence after baseline. These intervals were paired according to physiological time assignment to form two lactate confirmation intervals and corresponding patient lactate clearance rate intervals. Each of the four patient lactate status intervals corresponds to a unique blood sample identifier, and the blood sample source intervals do not overlap. The feasible region of path contribution established by the data correction module for lactate test data was projected onto the continuous renal replacement therapy lactate net removal component to generate a lactate net removal projection interval. Furthermore, a lactate net removal intensity interval was generated according to the coverage duration within each lactate confirmation interval.

[0107] To establish two lactate confirmation intervals that do not share blood samples, at least four blood samples are required to assess the patient's lactate status. The preferred range is four to six samples, with four samples being the most desirable to ensure independent confirmation and limit the warning waiting time. All four blood samples must originate after the 11:20 effective oxygen transfer adjustment event and before the next support adjustment, device interlock, or extracorporeal life support topology version change. Any additional blood samples beyond four are used for subsequent rolling confirmation and do not replace the four samples currently used for online evaluation. In this embodiment, the four patient lactate status records are as follows: B1125 corresponds to 6.58 to 6.82 mmol / L, B1155 corresponds to... The lactate clearance ranges from 6.38 to 6.62 mmol / L for B1200, 6.34 to 6.58 mmol / L for B1200, and 6.16 to 6.40 mmol / L for B1230. B1125 and B1155 form the first lactate clearance interval, and B1200 and B1230 form the second lactate clearance interval. Both lactate clearance intervals are 30 minutes long and do not share intermediate blood samples. Patient lactate clearance is positively correlated with a decrease in lactate concentration. The uncorrected lactate clearance rate intervals are -0.0013 to 0.0147 mmol / L / min and -0.0020 to 0.0140 mmol / L / min, respectively.

[0108] The patient's lactate status record is generated from blood samples taken from the patient's arterial sampling interface; the continuous renal replacement therapy (CRRT) net lactate removal projection record is generated from CRRT blood inflow, blood return, and filtrate samples taken within the same time period. The two are associated through the extracorporeal life support topology version, device instance, and coverage period. The patient's arterial lactate status blood sample is not directly used as the net lactate removal observation. Instead, the feasible domain of the path contribution corresponding to each protocol sampling is projected onto the CRRT net lactate removal component.

[0109] The first lactate confirmation interval is fully covered by three projection intervals: 0.26 to 0.30 mmol / min from 11:25 to 11:35, 0.27 to 0.31 mmol / min from 11:35 to 11:45, and 0.25 to 0.31 mmol / min from 11:45 to 11:55. These intervals are weighted according to their 10-minute coverage duration to form a lactate removal intensity interval of 0.26 to 0.31 mmol / min. The second lactate confirmation interval is fully covered by three projection intervals: 0.24 to 0.29 mmol / min from 12:00 to 12:10, 0.26 to 0.30 mmol / min from 12:10 to 12:20, and 0.25 to 0.31 mmol / min from 12:20 to 12:30, forming a lactate removal intensity interval of 0.25 to 0.30 mmol / min.

[0110] The continuous renal replacement therapy device instances, operating modes, hemofiltration instances, and extracorporeal life support topology versions remain consistent across the two confirmed intervals.

[0111] Feasible domains for lactate impact parameters are established based on historical lactate evaluation intervals of the same patient under the same continuous renal replacement therapy device instance and operating conditions. The number of historical lactate evaluation intervals is preferably 3 to 6, with 4 being the most preferred, and should cover at least 2 mutually separate net lactate removal intensity levels. The net lactate removal intensity intervals of the 4 historical intervals are 0.12 to 0.17, 0.18 to 0.23, 0.24 to 0.28, and 0.29 to 0.34 mmol / min, respectively, corresponding to patient lactate clearance rate intervals of 0.003 to 0.016, 0.007 to 0.020, 0.010 to 0.024, and 0.013 to 0.028 mmol / L / min, respectively. All historical intervals are earlier than the current risk event. The current 2 lactate confirmation intervals are not included in the establishment of feasible domains for lactate impact parameters to explain their own.

[0112] Using the net lactate removal intensity range as input and the patient's lactate clearance rate range as output constraint, the feasible range of basal lactate variation (-0.004 to 0.003 mmol / L / min) is first determined from three historical time periods of continuous renal replacement therapy for the same patient where the net lactate removal intensity range includes 0 and no extracorporeal membrane oxygenation (ECMO) adjustment occurs. Then, for four historical lactate evaluation intervals with different net lactate removal intensity levels, a conversion factor for net lactate removal is selected based on the interval midpoint and half-width constraints. The parameter combination that predicts lactate clearance rate compatibility with the historical lactate clearance rate range and simultaneously satisfies the equivalent conversion volume check constitutes the feasible domain of lactate influence parameters. Currently, the feasible range of the net lactate removal conversion factor is 0.061 to 0.072 L. - ¹, The parameters are all formed by historical observations and patient body fluid spatial constraints, and are not used as preset parameters.

[0113] The reciprocal of the lactate removal conversion factor forms the equivalent conversion volume, used to verify the consistency between the conversion factor and the patient's lactate-exchangeable fluid space. The preferred verification range for the equivalent conversion volume is 0.18 to 0.30 L / kg of the patient's baseline weight, with a preferred reference value of 0.22 L / kg. This range is adjusted based on the cumulative fluid balance of the current treatment cycle. For patient P017, whose weight is 72 kg, this corresponds to 12.96 to 21.60 L; the current conversion factor is 0.061 to 0.072 L. - ¹Corresponding to 13.9 to 16.4L, within this range, when it exceeds the verification range, the feasible domain of the lactate influence parameter does not obtain the correction qualification for the current event, and the current event enters the lactate correction pending confirmation state.

[0114] Multiplying the lactate removal intensity range by the lactate removal conversion factor yields the lactate removal contribution of the first lactate confirmation range (0.0159 to 0.0223 mmol / L / min) and the lactate removal contribution of the second lactate confirmation range (0.0153 to 0.0216 mmol / L / min). After separating the corresponding contributions from the uncorrected patient lactate clearance rate ranges, the corrected lactate clearance rate ranges are -0.0236 to -0.0012 mmol / L / min and -0.0236 to -0.0013 mmol / L / min, respectively. The upper boundaries of both ranges are not greater than 0.

[0115] If any of the following conditions are met: the net lactate removal projection interval does not completely cover any lactate confirmation interval, the feasible region of the lactate influence parameter is empty, or the net lactate removal conversion coefficient has no finite boundary, the current event is set to the lactate correction pending confirmation state, and no in vitro support for attenuation risk warning is output.

[0116] At 12:42, the test results of blood sample B1230 are available online. The lower boundary of the oxygen transfer change interval on the device side is greater than 0, which supports the deviation amount being greater than 0. Both lactate confirmation intervals are completely covered by the net lactate removal projection interval, and the upper boundaries of both corrected lactate clearance rate intervals are not greater than 0. Therefore, the output supports the realization of attenuation risk warning in vitro.

[0117] The warning system for the attenuation risk of extracorporeal support is linked to the extracorporeal life support topology version T2, the oxygenator instance OX-0714A, the continuous renal replacement therapy device instance CRRT-C02, the device-side oxygen transfer change interval, the patient oxygenation response rate interval, the individual condition response interval, the support implementation deviation, two lactate confirmation intervals, and the corresponding corrected lactate clearance rate intervals. The warning record also saves the protocol sampling event identifier and unique blood sample identifier that form each interval, so that the warning result can be traced back to the topology version, the test data formation path, the patient status record, and the net lactate removal projection record. This result is used to characterize the actual implementation level of extracorporeal support and is output in the form of device operation risk information, without directly forming disease diagnosis, treatment plan, or medical device control instructions.

[0118] After the above processing, the pressure and flow response constrains the incomplete static take-off record to a unique extracorporeal life support topology version. The test data formation path and path contribution feasible region separate the extracorporeal treatment contribution from the patient-side state. The cross-topology prediction interval incorporates the state transition uncertainty caused by the support adjustment into the baseline. The individual condition response interval and the lactate removal correction together limit the triggering conditions for risk warning. When any chain of evidence cannot form a finite and consistent interval, the corresponding unidentifiable, temporary, or pending confirmation result is output.

[0119] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0120] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0121] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A risk early warning system for critically ill patients in the AICU based on medical data fusion, characterized in that, include: The topology recognition module acquires data on the combined operation of extracorporeal membrane oxygenation and continuous renal replacement therapy, and constructs candidate directed graphs based on port connection relationships. Extract the pressure-flow response that supports the adjustment, filter candidate directed graphs based on the pressure-flow response constraints, and generate an in vitro life support topology version. The data correction module traces back from the test sampling node to the patient status reference node based on the in vitro life support topology version, generating a test data formation path; it establishes a material transfer model according to the test data formation path, performs an identifiability test on the contribution of the observed and the path to be estimated, separates the in vitro processing contribution, and generates equivalent observations on the patient side. The state reconstruction module inputs the equivalent observations from the patient side into the state space model and recursively estimates the patient's state based on the observation relationships corresponding to the external life support topology version. When the version of the extracorporeal life support topology changes, the patient status baseline is determined based on the innovative residuals of the new observations relative to the previous state, and the patient status baseline is reconstructed if the acceptance conditions are not met. The risk warning module acquires the patient status sequence after baseline, establishes individual conditional response intervals based on changes in oxygen transfer on the device side and changes in patient oxygenation status; generates support for deviation based on the interval deviation where the patient's oxygenation response rate interval is lower than the lower boundary of the individual conditional response interval; and corrects the patient's lactate clearance rate interval based on the continuous renal replacement therapy lactate net removal component corresponding to the path formed by the test data. When the lower boundary of the oxygen transfer change range on the device side is greater than zero, the deviation is greater than zero, and the upper boundary of the corrected lactate clearance rate range is not greater than zero, the output supports the attenuation risk warning.

2. The AICU critical patient risk early warning system based on medical data fusion according to claim 1, characterized in that, The specific operations for constructing a candidate directed graph are as follows: Establish a unified set of nodes corresponding to the patient-side interface, extracorporeal membrane oxygenation interface, continuous renal replacement therapy interface, and laboratory sampling interface; generate candidate directed edges based on port status, built-in blood flow direction of the device, and pipeline access records; for candidate directed edges with actual flow feedback, exclude candidate directed edges with inconsistent blood flow directions based on actual flow. Candidate directed graphs are generated by combining candidate directed edges based on the unidirectional transport relationship of the equipment, the continuity of the patient's drainage to return path, and the closure relationship of the continuous renal replacement therapy branch.

3. The AICU critical patient risk early warning system based on medical data fusion according to claim 2, characterized in that, The specific steps for filtering candidate directed graphs are as follows: The system extracts support adjustment events for actual equipment execution adjustments, establishes stability check windows and response check windows based on common sampling period, adjustment duration and communication latency; performs time-scale correction on pressure and flow data, generates response time intervals and local response sensitivities for monitoring nodes; and filters candidate directed graphs based on node reachability, allowed response direction, distinguishable response propagation order and flow balance of measurable branch merging nodes. Generate an in vitro life-supporting topology version when only one candidate directed graph is retained; When a unique candidate directed graph is not retained, topologically unidentifiable labels are generated.

4. The AICU critical patient risk early warning system based on medical data fusion according to claim 1, characterized in that, The specific steps for generating the test data formation path are as follows: Obtain a unique blood sample identifier, sampling time period, test sampling node, sample type, and test item identifier; determine the sampling validity based on the dead space of the sampling tubing, blood disposal records, and local fluid input status; trace back from the test sampling node along the directed graph corresponding to the extracorporeal life support topology version to the patient status reference node to generate a candidate test data formation path; Based on the effective blood volume of the tube segment in the candidate test data formation path and the cumulative actual flow before sampling, the transport delay and blood sample source range are determined; If the blood sample source range is within the valid range of the same in vitro life support topology version, and the test sampling node has a unique node mapping, the candidate test data formation path will be determined as the test data formation path; if the determination conditions are not met, no test data formation path will be generated.

5. The AICU critical patient risk early warning system based on medical data fusion according to claim 4, characterized in that, The specific steps for generating patient-side equivalent observations are as follows: The accompanying detection quantities of the same blood sample, as well as the cross-node observations within the same protocol sampling event and blood sample source intervals located in the same stable topological time period, are combined to form a matching observation vector; Based on the blood flow sequence in the test data formation path, combine the node transfer relationships corresponding to the test items to generate path transfer operators; The feasible region of path contribution is established based on the observation error boundary, equipment operating range, and test item parameter range, and the feasible boundary of patient-side transport volume is obtained. When the feasible region of path contribution is non-empty and bounded, and the half-width of the feasible interval on the patient side is not greater than the corresponding direct measurement allowable error, the equivalent observation on the patient side is generated, and the risk access flag is set to valid; when the feasible region of path contribution is empty and unbounded, the risk access flag is set to invalid. When the half-width of the feasible interval on the patient side is greater than the corresponding direct measurement allowable error, the risk access marker is set to invalid.

6. The AICU critical patient risk early warning system based on medical data fusion according to claim 1, characterized in that, The recursive processing of the state-space model is as follows: Establish fixed patient status coordinates according to the test item identifier and patient status reference node; group equivalent observations on the patient side with common time intersection in the blood sample source interval into the same status update group, and form observation intervals according to status components; Based on the patient-side direct sampling observations that were previously formed in the current status update group, corresponding to different unique blood sample identifiers and with non-overlapping blood sample source intervals, a status change rate interval is formed. When no support adjustments, device interlocks, or changes in the external life support topology occur between adjacent state update groups, the feasible interval of the preceding state is propagated, the propagation result is determined as the predicted state interval, and the intersection of the predicted state interval and the current observation interval is determined as the current feasible state interval; when the state change rate interval does not have finite boundaries, a temporary state interval is formed based on the current observation interval.

7. The AICU critical patient risk early warning system based on medical data fusion according to claim 6, characterized in that, The specific procedures for baseline patient status processing are as follows: The shortest interval between the current observation interval and the predicted state interval is determined as the innovative residual. When the version of the extracorporeal life support topology changes, the last valid state interval of the previous extracorporeal life support topology is propagated to the first valid observation time of the current extracorporeal life support topology according to the state components, generating a cross-topology prediction interval. When the first valid observation interval and the cross-topology prediction interval have an intersection, the patient state baseline is inherited, and an initial state interval is formed based on the intersection. When the first valid observation interval and the cross-topology prediction interval are separated, baseline reconstruction is initiated. When a finite state interval is formed by two valid observations that correspond to different unique blood sample identifiers and whose blood sample source intervals do not overlap, baseline reconstruction is completed, and a patient state sequence after baseline inheritance is generated.

8. The AICU critical patient risk early warning system based on medical data fusion according to claim 1, characterized in that, The specific procedures for determining changes in oxygen transfer on the device side and changes in the patient's oxygenation status are as follows: The blood transport time interval is determined based on the effective blood volume and actual blood flow between the sampling nodes before and after the oxygenator; the blood sample source interval after the oxygenator is reverse-calculated to the sampling node before the oxygenator according to the blood transport time interval; when the reverse-calculated blood sample source interval intersects with the blood sample source interval before the oxygenator, the oxygen transfer interval on the device side is generated based on the local blood oxygen content interval before and after the oxygenator and the actual blood flow during the intersection period. The equipment-side oxygen transfer change range is determined based on the equipment-side oxygen transfer range before and after adjustment. When the lower boundary of the equipment-side oxygen transfer change range is greater than zero, an effective oxygen transfer adjustment event is determined. The patient oxygenation response rate range is generated based on the oxygenation status component of the same patient before and after the effective oxygen transfer adjustment event.

9. The AICU critical patient risk early warning system based on medical data fusion according to claim 8, characterized in that, The specific steps for establishing an individual conditional response interval are as follows: From the historical effective oxygen transfer adjustment events of the same patient, select historical events that match the oxygenator instance, node mapping relationship, continuous renal replacement therapy device instance and operating status, and the patient's oxygenation status interval before adjustment with the current effective oxygen transfer adjustment event to form historical matching events; establish the feasible domain of individual response parameters based on the device-side oxygen transfer change interval and the patient's oxygenation response rate interval of the historical matching events. Input the current equipment-side oxygen transfer change range into the individual response parameter feasible region to generate the individual conditional response range; The individual conditional response interval has finite boundaries and the lower boundary is greater than zero. When the upper boundary of the current patient oxygenation response rate interval is lower than the lower boundary of the individual conditional response interval, a deviation supporting realization is generated based on the interval interval and the minimum discernible response rate determined by the patient's oxygenation status measurement error and the response time interval. When the feasible domain of the individual response parameter is empty or unbounded, no deviation supporting realization is generated.

10. The AICU critical patient risk early warning system based on medical data fusion according to claim 9, characterized in that, The specific procedures for correcting the effects of continuous renal replacement therapy on lactate changes and providing extracorporeal support to achieve attenuation risk warning are as follows: Four patient lactate status intervals were selected from the patient status sequence after baseline, and paired according to physiological time to form two lactate confirmation intervals and corresponding patient lactate clearance rate intervals. The four patient lactate status intervals correspond to different unique blood sample identifiers, and the blood sample source intervals do not overlap. The feasible region of path contribution established by the data correction module for lactate test data was projected onto the continuous renal replacement therapy lactate net removal component to generate the lactate net removal projection interval, and the lactate net removal intensity interval was generated according to the coverage duration within each lactate confirmation interval. A feasible region for lactate impact parameters is established based on the historical lactate evaluation intervals of the same patient under the same continuous renal replacement therapy device instance and operating conditions; the patient's lactate clearance rate interval is corrected based on the feasible region for lactate impact parameters and the interval of net lactate removal intensity, and a corrected lactate clearance rate interval is generated. When the deviation is greater than zero, both lactate confirmation intervals are completely covered by the net lactate removal projection interval, and the upper boundaries of both corrected lactate clearance rate intervals are not greater than zero, an in vitro support for attenuation risk warning is output; when the net lactate removal projection interval does not completely cover the lactate confirmation interval, an in vitro support for attenuation risk warning is not output; when the feasible domain of the lactate influence parameter is empty or unbounded, an in vitro support for attenuation risk warning is not output.

Images