Needle pull-out risk anticipation method and system based on data credibility constraints

Abstract

The application discloses a needle pulling risk expectation method and system based on data credibility constraint, relates to the technical field of clinical nursing risk assessment, and proposes the following scheme, which comprises the following steps: extracting diagnosis and treatment data sequences and intervention event sequences in a backtracking time window and unifying the sequences to a time axis with reference to a benchmark time; fusing data quality and time decay to calculate basic credibility; marking an action object, correcting a penalty and repairing a neighborhood according to intervention action rules; forming a final credibility and gating aggregation to construct a risk feature vector; and inputting the risk feature vector into a risk prediction model to obtain needle pulling risk expectation. The application can effectively inhibit the interference of data disturbance caused by arterial pipeline maintenance operation and clinical treatment on risk assessment, avoid misjudgment of operation artifacts as real risk signals, and thus reduce risk warning false alarms caused by changes in nursing operation rhythm.

Description

Technical Field

[0001] This invention relates to the field of clinical nursing risk assessment technology, and more specifically, this application relates to a method and system for predicting needle removal risk based on data credibility constraints. Background Technology

[0002] In the nursing care of invasive arterial access devices such as arterial pressure monitoring catheters and arterial sheaths after interventional procedures, complications such as persistent bleeding and hematoma progression at the puncture site are prone to occur shortly after removal (catheter / needle removal), requiring nursing interventions such as pressure application, pressure bandaging, and graded observation. Nurses typically combine multi-source information, including vital signs and arterial waveform monitoring information, anticoagulation / antiplatelet medication administration, coagulation-related test results, puncture site dressing and bleeding observation records, to determine whether to prolong compression, increase the frequency of rounds, or trigger escalation measures. Existing auxiliary methods mostly rely on fixed procedures, threshold alarms, or simple rule scoring. In some scenarios, multi-source data is aligned and input into a model to output a risk score for alerts.

[0003] However, in actual practice, the aforementioned multi-source information commonly suffers from inconsistent collection and recording rhythms, missing and supplementary recordings, timestamp deviations, and differences in caliber. This results in data within the same observation window being usable but not necessarily reliable. On the other hand, routine maintenance operations of arterial lines (such as flushing, transducer zeroing / calibration) and clinical procedures (such as postural adjustments and sedation) can cause short-term fluctuations in monitoring data or changes in recording rhythm, making it difficult for existing methods to distinguish between operational artifacts and true risk signals. Without data reliability metrics and mechanisms for handling contradictory information specific to this scenario, risk assessment results are prone to fluctuations, increasing false positives / false negatives, and it is difficult to explain the basis and reliability of risk outputs. This, in turn, affects the consistency of nursing staff's adoption and handling of warning prompts. Therefore, a needle removal risk prediction method and system based on data reliability constraints are proposed to address this problem. Summary of the Invention

[0004] To address the aforementioned technical problems, this technical solution provides a method and system for predicting needle removal risk based on data credibility constraints, thus resolving the issues raised in the background section.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: Firstly, this application provides a method for predicting needle removal risk based on data credibility constraints, the method comprising: Obtain the baseline time of needle removal and nursing scenario for the target patient, and extract the preset retrospective time window, diagnosis and treatment data set and intervention event type set and their effect rules based on the nursing scenario; Extract the diagnosis and treatment data sequence and intervention event sequence within the retrospective time window, and align them to a unified timeline with the baseline time as the reference point; For each piece of diagnostic and treatment data in the diagnostic and treatment data sequence, the basic reliability is obtained by integrating the quality of the diagnostic and treatment data with the time-sensitivity decay relative to the baseline time. Based on the action rules, determine the influence window, applicable diagnostic and therapeutic data types, and associated penalty intensity of each intervention event in the intervention event sequence, and mark the diagnostic and therapeutic data that are earlier than the intervention event and match the applicable diagnostic and therapeutic data types within the influence window as the action objects; Based on the completeness of the treatment data type of the affected object relative to the treatment rule, and the trend consistency of all treatment data within the affected window, feedback correction is performed on the penalty intensity; Based on the revised penalty intensity and the relative time distance between the corresponding target and the intervention event, penalty weights or hard failure markers are applied to each target to update the basic reliability of the corresponding diagnosis and treatment data; For diagnostic and treatment data that are not marked as targets and have abnormal continuity of basic credibility, neighborhood repair is performed based on adjacent updated basic credibility. The updated or neighborhood-repaired basic credibility is recorded as the final credibility. Based on this, the diagnostic and treatment data sequence is gated or weighted to construct a risk feature vector, which is then input into the risk prediction model to obtain the expected risk of needle removal.

[0006] Secondly, this application provides a needle removal risk prediction system based on data credibility constraints, used to implement the aforementioned needle removal risk prediction method based on data credibility constraints, including: The rule preset module is used to obtain the baseline time of needle removal and nursing scenario of the target patient, and extract the preset retrospective time window, diagnosis and treatment data type set and intervention event type set and their effect rules according to the nursing scenario; The timeline alignment module is used to extract the diagnosis and treatment data sequence and intervention event sequence within the backtracking time window and align them to a unified timeline with the reference time as the reference point; The credibility calculation module is used to calculate the basic credibility of each diagnosis and treatment data in the diagnosis and treatment data sequence by combining the quality of the diagnosis and treatment data with the time-related decay relative to the baseline time. The object marking module is used to determine the influence window, applicable diagnostic and treatment data types and their associated penalty intensity of each intervention event in the intervention event sequence according to the action rules, and to mark the diagnostic and treatment data that are earlier than the intervention event and match the applicable diagnostic and treatment data types within the influence window as the action objects; The feedback correction module is used to perform feedback correction on the penalty intensity based on the completeness of the data type of the affected object relative to the action rule, and the trend consistency of all data types affecting the window. The update execution module is used to apply penalty weights or hard failure flags to each target based on the corrected penalty intensity and the relative time distance between the target and the intervention event, so as to update the basic reliability of the corresponding diagnosis and treatment data. The aggregation prediction module is used to perform neighborhood repair on diagnostic and treatment data that are not marked as targets and have abnormal continuity of basic credibility. The updated basic credibility is recorded as the final credibility, and the diagnostic and treatment data sequence is gated or weighted and aggregated according to it to construct a risk feature vector. The vector is then input into the risk prediction model to obtain the expected risk of needle removal.

[0007] Thirdly, this application provides a computer device including a memory and a processor, the memory storing code, and the processor being configured to acquire the code and execute the above-described method for predicting needle removal risk based on data credibility constraints.

[0008] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for anticipating needle removal risk based on data credibility constraints.

[0009] Compared with the prior art, the beneficial effects of the present invention are as follows: This application can effectively suppress the interference of data disturbances introduced by arterial pipeline maintenance operations and clinical procedures on risk assessment by identifying the impact window of intervention events and imposing credibility penalties on the affected diagnosis and treatment data within the window, thereby avoiding misjudging operation artifacts as real risk signals and reducing false alarms of risk warnings caused by changes in the rhythm of nursing operations. This application establishes basic credibility by integrating the quality and timeliness decay of medical data, and performs feedback correction on the intensity of punishment so that it adaptively adjusts with the level of evidence support. This avoids over-punishing valid information when data support is insufficient or ignoring potential risks when data is contradictory, thus maintaining the consistency and stability of the evaluation results even in the presence of missing data, supplementary data, or differences in standards. This application performs neighborhood repair on diagnostic and treatment data that is not marked as the target and has anomalies in the continuity of basic credibility to obtain the final credibility. Based on this, it constructs a risk feature vector by hierarchical gating and weighted aggregation of the diagnostic and treatment data sequence. This weakens or removes low-credibility data and prioritizes the adoption of high-credibility data. As a result, even when there are lags, misalignments or asynchronous disturbances in the diagnostic and treatment data, it can still generate more stable and reliable needle removal risk prediction results, improve the adoptability of early warning prompts and support the consistency of hierarchical observation and treatment decisions after needle removal. Attached Figure Description

[0010] The disclosure of this invention is illustrated with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of this invention. Wherein: Figure 1 This is a flowchart of the needle removal risk prediction method based on data credibility constraints proposed in this invention; Figure 2 This is a structural block diagram of the needle removal risk prediction system based on data credibility constraints proposed in this invention. Detailed Implementation

[0011] It is readily understood that, based on the technical solution of this invention, those skilled in the art can propose various interchangeable structural methods and implementations without altering the essential spirit of the invention. Therefore, the following detailed embodiments and accompanying drawings are merely illustrative examples of the technical solution of this invention and should not be considered as the entirety of the invention or as limitations or restrictions on the technical solution of this invention.

[0012] Reference Figure 1 As shown, this application proposes a method for predicting needle removal risk based on data credibility constraints, including: Obtain the baseline time of needle removal and nursing scenario for the target patient, and extract the preset retrospective time window, diagnosis and treatment data set and intervention event type set and their effect rules based on the nursing scenario; It should be noted that the baseline time is the reference time corresponding to the risk assessment of needle removal for the target patient. It can be the time when the arterial access removal event (cannula / needle removal) occurs, or the current assessment time or the planned assessment trigger time in the expected task. It can be directly obtained from nursing records, event reports, or device alarm timestamps. The nursing scenario is a contextualized identifier of the patient's nursing process and monitoring conditions, used to limit the data update cycle, data source, and intervention rules, such as arterial pressure monitoring catheter care under intensive care, observation / postoperative recovery of arterial sheath removal after interventional treatment, emergency observation, and can also be extended to other invasive vascular access or intravenous therapy related nursing scenarios. For example, the backtracking time window ensures that it covers the diagnosis and treatment data chain from the most recent intervention event to the baseline time, while avoiding the introduction of historical noise that is irrelevant to the current risk by an excessively long window; typical values ​​are: 12h in ICU / intensive care, 2h~6h for arterial sheath removal observation / postoperative recovery after intervention, and 2h for emergency observation; to ensure that there are at least enough observation points for each data type within the window, the example ensures that the backtracking time window exceeds 6 times the preset update cycle; The data set of diagnostic and therapeutic data may include: vital signs (heart rate, blood pressure, preferably including invasive arterial pressure and waveform quality related indicators, etc.), sedation-agitation assessment (such as RASS or similar scales), tubing / monitoring device status (such as arterial pressure channel alarm, transducer zeroing / calibration and leveling status, flushing / three-way valve opening and closing status, pressure bag status, waveform quality / blockage indication, etc.), puncture site and fixation / dressing status (bleeding degree grading, hematoma range / progression, dressing change, pressure bandage or hemostatic device status / review, etc.), nursing rounds and behavioral event records, etc. The set of intervention event types may include: sedation / analgesia related interventions (initiation of drug administration, dosage increase, dosage decrease, discontinuation), restraint and safety protection interventions (restraint activation, release, adjustment, etc.), arterial tubing and monitoring device maintenance (flushing, transducer zeroing / calibration, leveling, three-way valve opening / closing / channel switching, pressure bag handling, alarm handling, etc.), hemostasis and puncture site handling (pressure start / end, pressure bandage adjustment, hemostasis device activation / release verification, dressing change, etc.), and nursing operation record types (turning over and patting the back, suctioning, etc.). For each type of intervention event, its action rules are preset in the nursing scenario, which are used to provide: the influence window, the set of applicable diagnostic and therapeutic data types, and the penalty intensity; wherein the influence window ranges from [5min, 6h], the penalty intensity is a normalized intensity, and the range is [0.1, 0.95]; the set of applicable diagnostic and therapeutic data types is used to limit the intervention event to only apply penalties or invalidation to the data types that are clinically related to it; For example, when the intervention event type is arterial line and monitoring device maintenance, the action rule can be: the impact window is 20 minutes, the penalty intensity is 0.3, and the set of applicable diagnostic and therapeutic data types includes line / monitoring device status, and may include vital signs (such as invasive arterial pressure) that are disturbed synchronously with waveform quality; when the intervention event type is hemostasis and puncture site treatment, the action rule can be: the impact window is 2 hours, the penalty intensity is 0.8, and the set of applicable diagnostic and therapeutic data types includes puncture site and fixation / dressing status, vital signs, and nursing rounds and behavioral event records. Furthermore, the action rules are given by the rule table of the nursing scenario index, and their formulation follows the following: determining the impact window based on the duration of the intervention event's impact; determining the set of applicable diagnostic and treatment data types based on the data types covered or synchronously disturbed by the intervention event's impact path; determining the penalty intensity based on the level of damage to the representativeness of past data caused by the intervention event (such as state mutation, record misalignment probability, gap probability), and allowing configuration updates according to differences in nursing scenarios; Extract the diagnosis and treatment data sequence and intervention event sequence within the retrospective time window, and align them to a unified timeline with the baseline time as the reference point; For each piece of diagnostic and treatment data in the diagnostic and treatment data sequence, the basic reliability is obtained by integrating the quality of the diagnostic and treatment data with the time-sensitivity decay relative to the baseline time. Based on the action rules, determine the influence window, applicable diagnostic and therapeutic data types, and associated penalty intensity of each intervention event in the intervention event sequence, and mark the diagnostic and therapeutic data that are earlier than the intervention event and match the applicable diagnostic and therapeutic data types within the influence window as the action objects; Based on the completeness of the treatment data type of the affected object relative to the treatment rule, and the trend consistency of all treatment data within the affected window, feedback correction is performed on the penalty intensity; Based on the revised penalty intensity and the relative time distance between the corresponding target and the intervention event, penalty weights or hard failure markers are applied to each target to update the basic reliability of the corresponding diagnosis and treatment data; For diagnostic and treatment data that are not marked as targets and have abnormal continuity of basic credibility, neighborhood repair is performed based on adjacent updated basic credibility. The updated or neighborhood-repaired basic credibility is recorded as the final credibility. Based on this, the diagnostic and treatment data sequence is gated or weighted to construct a risk feature vector, which is then input into the risk prediction model to obtain the expected risk of needle removal. It should be noted that the risk prediction model is a deep learning model, which uses a multilayer perceptron neural network: the risk feature vector is used as input, and each dimension of the feature is input into the network input layer. After nonlinear mapping through at least two fully connected hidden layers, the output layer gives the expected value of bleeding / hematoma risk after arterial access removal; the expected risk value takes a continuous value from 0 to 1, and the larger the value, the higher the risk level of bleeding / hematoma after removal in the corresponding nursing scenario. During the training phase, the system constructs training samples from historical nursing data: for each training sample, a baseline time and nursing scenario are selected, and the diagnosis and treatment data sequence and intervention event sequence are extracted within the retrospective time window. The basic credibility is calculated, and the penalty correction and neighborhood repair are completed under the constraint of the intervention event action rule to obtain the final credibility. Then, hierarchical gating and weighted aggregation are performed on the diagnosis and treatment data sequence to form a risk feature vector. Furthermore, training labels Generated from bleeding-related event annotations after arterial access removal: Within a preset future time period after the baseline time, if a predefined bleeding / bleeding or hematoma progression event occurs in nursing records, event reports, or equipment alarms (including but not limited to puncture site bleeding reaching a preset grading threshold, escalation of continuous pressure / compression bandaging, activation or adjustment of hemostatic devices, hematoma expansion records, escalation of treatments triggered by bleeding risk, or physician assessment records, etc.), then... Otherwise ; to pair the samples Input risk prediction model, where, For risk feature vectors, To train labels, network parameters are iteratively optimized and updated (to improve model output). With tags The risk prediction model is obtained by gradually reducing the differences. The preset future time period is a prediction window starting from the baseline time, used to define the time range corresponding to the expected value of bleeding / hematoma risk after arterial access removal. The preset future time period can be set in the range of 10 minutes to 24 hours according to the nursing scenario, so as to cover the main observation and treatment time scale of bleeding / hematoma risk after arterial access removal in the nursing process. Its value is set based on the density of intervention events in the scenario, the update rhythm of diagnosis and treatment data, and the minimum observation requirements of the nursing process, and keeps the training phase and the online inference phase consistent. For example, the expected value of post-extraction bleeding / hematoma risk ranges from [0,1], with a higher value indicating a higher risk of post-extraction bleeding / bleeding or hematoma progression. The nursing strategy level is output based on the range into which the expected post-extraction bleeding / hematoma risk falls, and a corresponding preventative treatment intensity is generated. No strategy reduces the prescribed procedures or minimum observation requirements. Low risk can be defined as: expected post-extraction bleeding / hematoma risk less than 0.05; medium risk can be defined as: expected post-extraction bleeding / hematoma risk greater than or equal to 0.05 and less than 0.15; high risk can be defined as: expected post-extraction bleeding / hematoma risk greater than 0.15. Furthermore, when providing risk-based care strategies, the system should prompt "Follow standard hospital procedures": Low-risk nursing strategies may include: recommending that the frequency of rounds / checks be 1.0 times that of routine rounds, and performing routine checks on tubing fixation and puncture sites; The nursing strategy for medium-risk patients may include: recommending that the frequency of rounds / checks be 2.0 times that of the routine, recommending that one additional check of key points be added (including check of fixed status, pipeline routing and equipment alarm handling), and suggesting that the patient be assessed for changes in agitation / cooperation. High-risk nursing strategies may include: recommending that the frequency of rounds / reviews be 3.0 times the routine, recommending two additional key point reviews, and outputting a prompt such as "recommend strengthening local compression, review of hemostatic devices, and medical-nursing collaborative assessment"; when increased bleeding at the puncture site, expansion of hematoma, abnormal fluctuations in vital signs, or coagulation-related abnormalities are monitored or recorded, generating and outputting escalation reminder information (including but not limited to prompting the doctor to notify or initiating the in-hospital escalation treatment process) to reduce the risk of delayed detection of bleeding / bleeding or hematoma progression after removal; Through the above technical solution, this embodiment constructs a retrospective time window, a set of diagnostic and treatment data types, and a set of intervention event types and their rules of action under the constraints of the nursing scenario. On a unified time axis, it applies a feedback-correctable penalty and a recoverable repair mechanism to the past data within the intervention impact window, thereby achieving dynamic constraints and structured screening of the credibility of diagnostic and treatment data. This reduces the misleading effect of intervention and recording issues on the expected risk of bleeding / hematoma after removal, and improves the reliability and interpretability of risk feature construction.

[0013] In an optional embodiment, the basic reliability is obtained by fusing the quality of diagnostic and treatment data with the time-sensitivity decay relative to the baseline time, specifically including: Based on the nursing scenario, a preset update cycle corresponding to the data type of diagnosis and treatment is extracted. The preset update cycle is set based on the clinical routine update frequency of this type of diagnosis and treatment data in the nursing scenario and its effective freshness span for the risk of needle removal. Its value range is set according to different scenarios. For example, for the monitoring / infusion status scenario, the typical value can be 10 minutes, and for the nursing assessment / round record scenario, it can be 2 hours. Within the retrospective time window, the sampling interval sequence of this diagnostic data type is statistically analyzed and the sampling interval dispersion is calculated, and a sampling reliability factor is generated based on it; For example, the sampling reliability factor is ,in, The sampling interval dispersion, It is a factor constant with a value range of [1,4], and a typical value is 2; The time interval between each diagnosis and treatment data point and the reference time point are calculated. The data points are then normalized according to the preset update cycle parameters to obtain the initial timeliness. The initial timeliness is then adjusted using the sampling reliability factor to obtain the actual timeliness. The formula for calculating the initial aging time is: ; In the formula, For actual timeliness, For the initial validity period, As a matter of timeliness common sense and with a value range of [1,3], the typical value is 2; by amplifying the timeliness structure inversely by sampling reliability, the irregularity of sampling is transformed into a stronger timeliness penalty, suppressing the false trends caused by gaps / delays in nursing records from the input side; Extract missing completeness indicators, data source consistency indicators, and outlier interpretability indicators from each diagnosis and treatment data, and generate data quality scores; It should be noted that the three types of indicators respectively characterize whether the diagnosis and treatment data is missing, whether it is from the same source, and whether it is interpretable, and together they determine the quality. For example, after normalizing the missing data integrity index, the data source consistency index, and the outlier interpretability index, they are weighted and added together in a 4:3:3 ratio to obtain the data quality score. The timeliness decay coefficient is determined based on the actual timeliness, and then integrated with the data quality score to obtain the basic credibility. For example, the aging attenuation coefficient is The basic credibility is obtained by weighting the time-sensitivity decay coefficient and the data quality score in a 6:4 ratio. Through the above technical solution, this embodiment achieves interpretable quantification of the basic credibility of diagnosis and treatment data by jointly modeling data quality and timeliness decay under the constraints of nursing scenarios and introducing the inverse amplification of sampling reliability on timeliness penalty, thereby providing stable input for subsequent intervention penalties, misalignment identification and feature gating.

[0014] In an optional embodiment, before performing feedback correction on the penalty intensity, an offset determination is made for the time points of each diagnosis and treatment data in the diagnosis and treatment data sequence, specifically including: Starting from the endpoint furthest from the reference time in the retrospective time window, the extended time period is obtained by extending it along a unified time axis in a direction furthest from the reference time. The endpoint of the extension is the time point of the first intervention event in the extension direction or the time point of the diagnosis and treatment data furthest from the starting point within the extended time period. In order to avoid the template mapping being unstable due to the lack of intervention reference points near the endpoint furthest from the reference time in the retrospective time window, the extension is made to the first intervention event or the furthest diagnosis and treatment point in the direction furthest from the reference time to form a stable initial alignment segment. Preset treatment templates are generated based on preset update cycles corresponding to nursing scenarios and treatment data types. These templates include preset time point ranges and allowable offset thresholds for each treatment data type. The allowable offset threshold is a preset percentage of the preset update cycle, with the preset percentage ranging from 10% to 50%, and a typical value being 25%. It should be noted that the preset diagnosis and treatment template serves as a temporal reference framework for a given nursing scenario. It is structured according to the data types of diagnosis and treatment, defining the preset time point range and allowable offset threshold for each data type within the retrospective time window. This is used for mapping and comparing actual diagnosis and treatment data with the theoretical time frame, supporting time offset determination and completeness statistics. For periodically updated data types, the preset time point range represents the regular update cycle. For event-driven or non-periodic data, a differentiated configuration strategy is adopted (which can be set to no preset time point range or zero expected frequency), activating subsequent processing only when an event is triggered, achieving a flexible adaptation mechanism without mandatory temporal constraints. Within the retrospective time window, the time periods between events are divided according to adjacent intervention events. Based on the action rules corresponding to each time period, the preset time point range of the preset diagnosis and treatment template is updated sequentially to obtain the intervention impact diagnosis and treatment template. Through the two-level template architecture that evolves from the preset diagnosis and treatment template to the intervention impact diagnosis and treatment template, the misalignment judgment is upgraded from a single time regularity dimension to an explicit judgment that incorporates the correlation between the intervention event and the disturbance of the recording and collection rhythm, effectively reducing the probability of misalignment caused by changes in the actual nursing process. It should be noted that the intervention impact diagnosis and treatment template is a dynamic temporal framework constructed based on the preset diagnosis and treatment template and combined with the set of intervention event types and their role rules. It is used to characterize the changes in the recording temporal pattern and the effective reference range of data caused by nursing treatment before and after the intervention event. Specifically, through a time-based processing mechanism divided by adjacent intervention events, the baseline temporal constraints of the preset template are dynamically adjusted according to the intervention response rules. This ensures that the preset time point range or expected frequency constraints of relevant diagnosis and treatment data types within the intervention sensitive period are adapted to the actual nursing rhythm, avoiding misjudgments caused by reasonable temporal deviations. In the extended time period and the time period between each event, the diagnosis and treatment data are mapped to the intervention impact diagnosis and treatment template according to the data type, and the time offset relative to the range of the most recent preset time point is calculated. If the time offset exceeds the allowable offset threshold, the corresponding diagnostic data will be marked as a suspected error site. Through the above technical solution, this embodiment constructs a diagnosis and treatment template that includes an allowable offset threshold, and forms an intervention impact diagnosis and treatment template under the constraints of the intervention event action rule, thereby enabling interpretable determination of misalignment of diagnosis and treatment data time points, and providing a reliable basis for subsequent adjustment of penalty intensity and credibility correction.

[0015] In an optional embodiment, after marking the corresponding diagnostic data as a suspected misdiagnosis site, the method further includes adjusting the penalty intensity, specifically including: Within the time periods between each event, diagnostic and treatment data that were not marked as suspected misalignments and whose mappings satisfy the intervention impact diagnostic and treatment template are selected to form a timely set; The diagnostic and treatment data corresponding to each diagnostic and treatment data type in the time set are sorted by time to form a type time series subsequence. Based on this type time series subsequence, a trend sign sequence and an inflection point sequence are constructed. The trend sign sequence can be represented by the difference sign between two adjacent points (rising / falling / stable). The inflection point sequence can be represented by local maxima / minimum points or the position where the trend sign changes. For each suspected error site, extract the trend symbol sequence and inflection point sequence of the diagnostic and treatment data type to which it belongs in the timely set, and calculate the local trend consistency index between the suspected error site and the trend symbol sequence and inflection point sequence. Based on the local trend consistency index, suspected misalignment points are classified as true record points or true misalignment points; Remove the mark of the real record point as the target, and keep its basic credibility unchanged and regard it as the updated basic credibility; The penalty intensity corresponding to the actual misalignment point is adjusted by a preset multiple. If the adjusted penalty intensity value exceeds the preset intensity threshold, a hard failure mark is applied to the actual misalignment point. For example, the preset multiple ranges from 1.2 to 3, with a typical value of 2; the preset intensity threshold ranges from 0.75 to 0.95, with a typical value of 0.85. For diagnostic and treatment data other than true recording points and true misalignment points, the target markers and penalty intensity should remain unchanged; except for true recording points and true misalignment points, the target markers and penalty intensity should remain unchanged to avoid over-correction and expanding the scope of influence. It should be noted that the introduction of an adaptive reference mechanism within a time period using the punctual set as a reference adapts to the differences in the recording rhythm and fluctuation patterns of different events in the same nursing scenario, giving robustness to the misalignment judgment and intensity adjustment time period, and effectively avoiding false penalties caused by fixed thresholds. Through the above technical solution, this embodiment performs secondary screening of suspected misaligned sites by using the trend structure consistency based on the punctual set, and strengthens the penalty for real misaligned sites by a preset multiple and performs hard failure when the preset intensity threshold is exceeded, thereby reducing the risk of systematic misjudgment caused by misaligned data entering feature construction.

[0016] In an optional embodiment, feedback correction is performed on the penalty intensity based on the completeness of the treatment data type of the affected object relative to the application rule, and the trend consistency of all treatment data within the affected window. Specifically, this includes: Based on the preset update cycle and intervention impact diagnosis and treatment template, an expected arrival list is generated in each impact window. The expected arrival list is used to record the expected arrival times of each type of applicable diagnosis and treatment data within each preset time point range. It should be noted that the expected number of arrivals is determined by the preset update cycle and the preset time range of the intervention impact on the diagnosis and treatment template. That is, within the impact window, the expected number of arrivals corresponding to each preset time range is usually 0 or 1 (it can be 2 if necessary to cover dual-source acquisition); the length of the impact window is configured by the action rule, typically 5 min-6 h (for example, intervention events such as sedation / restraint have a longer-lasting impact on the validity of diagnosis and treatment data). Within the influence window, the actual number of times the target object falls within each preset time point range is counted. The actual misalignment points are not included in the actual number of arrivals. The completeness index is calculated based on the actual number of arrivals and the expected arrival list, and a gap location set is generated. All diagnostic and treatment data within the affected window are weighted and aggregated according to basic reliability, and the weighted change direction of each diagnostic and treatment data type between adjacent preset time points is calculated. It should be noted that the weighted aggregation criterion is as follows: for the same type of medical data within the range of two adjacent preset time points, the representative value is obtained by weighting according to the basic confidence level (such as weighted mean / weighted median), and then the adjacent representative values ​​are compared to determine the direction of weighted change (rising / falling / stable). The trend consistency index is calculated based on the weighted change direction consistency ratio of each diagnostic and treatment data type and the distribution density of the gap location set. This index is then integrated with the completeness index to form a feedback correction factor. This feedback correction factor is used to correct the penalty intensity associated with the intervention event in the action rule. It should be noted that the directional consistency ratio is used to characterize whether different diagnostic and treatment data types show a consistent trend response within the influence window. The directional consistency ratio is calculated as follows: the weighted change direction of each diagnostic and treatment data type is used as the voting item, the percentage of types that are consistent with the dominant direction within the window is counted, and the weighted effective sample size of the type in the adjacent range is used as the voting weight to reduce the impact of low-confidence or sparse data on the consistency judgment. The process of obtaining the distribution density of the gap location set is as follows: the influence window is discretized according to the preset time point range, the number of times the gap location set appears in each preset time point range is counted, and the degree of gap aggregation is characterized by the concentration of the number of gap occurrences (such as the maximum gap ratio or the peak value of the number of gaps within a unit range). The more aggregated the gaps are, the higher the distribution density is, which indicates that the continuous support available for trend judgment is worse, thus applying a stronger reduction to the direction consistency ratio, thereby obtaining the trend consistency index, whose value is limited to [0,1]. The more consistent the trend direction and the more dispersed the gaps are, the larger the trend consistency index is. Furthermore, the trend consistency index and the completeness index are fused to generate a feedback correction factor, with a fusion ratio of 6:4. The completeness index is used to characterize the coverage of the expected number of arrivals and the actual number of arrivals within the influence window, while the trend consistency index is used to characterize the consistency of the trend response based on the coverage. Through the joint feedback correction mechanism of completeness and trend consistency, the single dependence of the penalty intensity on the prior configuration of the action rule is removed, and it is instead constrained by the true support within the dynamic data interval. This effectively suppresses feature degradation caused by penalty overload in scenarios with high missing rates in nursing records. Through the above technical solution, this embodiment generates a list of expected arrivals based on a template and combines it with actual arrivals to form a completeness constraint. At the same time, it uses the trend consistency within the influence window to provide feedback correction on the penalty intensity, thereby improving the adaptability and stability of the intervention penalty and reducing erroneous corrections caused by data gaps or trend confusion.

[0017] In an optional embodiment, for diagnostic and treatment data that is not marked as an object of action and has abnormal continuity in basic confidence, neighborhood repair is performed based on adjacent updated basic confidence levels, specifically including: It should be noted that abnormal continuity of baseline confidence refers to an abnormal situation in which the baseline confidence of diagnostic and treatment data does not conform to the local continuous change pattern relative to the updated baseline confidence of preceding and subsequent data. This includes, but is not limited to, sudden jumps in baseline confidence, local breaks, or abnormal deviations from the interpolation results relative to the neighborhood. For example, the deviation of the basic credibility of the diagnosis and treatment data from the anchor interpolation credibility can be used as a judgment index: when the deviation is greater than a preset deviation threshold, it can be determined that the diagnosis and treatment data has an abnormal continuity of basic credibility; wherein, the preset deviation threshold can be in the range of 0.1-0.3, and a typical value can be 0.2; The nearest real record points of the diagnosis and treatment data on a unified time axis are extracted as anchor points. When a real record point on either side is missing, the diagnosis and treatment data corresponding to the nearest updated basic confidence level on that side is used as a replacement anchor point. The time span and basic confidence level gradient between anchor points are calculated. The basic confidence level gradient is the ratio of the difference in basic confidence level between two anchor points to the time span between anchor points. Based on the basic confidence of the anchor points and the time span between the anchor points, the interpolation confidence of the diagnosis and treatment data is obtained by interpolating at its time points. For example, the interpolation confidence level of this diagnostic data at its point in time is: ; In the formula, To assess the interpolation reliability of the diagnostic data to be repaired at its point in time, , To ensure the credibility of the anchor point foundation, , For anchor time, The time required for repair; In the timely set corresponding to the time period between events to which the diagnosis and treatment data belongs, extract the trend symbol sequence that is consistent with the data type of the diagnosis and treatment data, and extract local trend segments according to the time span; It should be noted that, within the time period of this event, the same type of trend symbol sequence is taken and local segments are extracted according to the anchor point time span to obtain overall monotonic / rise-then-fall / fall-then-rise pattern information, providing pattern priors for the construction of the constraint field; A trend constraint field is constructed based on the basic confidence gradient and local trend segments. The direction of change of the basic confidence of the diagnosis and treatment data is uniformly projected onto the trend constraint field to obtain the projection consistency. It should be noted that the process of constructing the trend constraint field is as follows: the sign of the basic confidence gradient is used as the magnitude and direction constraint of the confidence change, and the monotonicity characteristics and extreme value distribution pattern determined by the local trend segment are used as the morphological constraint of the confidence change, so as to obtain the set of basic confidence change directions allowed within the time span. The direction of change of the baseline confidence level relative to the anchor point interpolation confidence level of the diagnostic data is taken as the projection direction. This projection direction is then mapped to the set of allowed directions. When the projection direction belongs to the set of allowed directions, it is considered a consistent projection and recorded as high consistency. When the projection direction does not belong to the set of allowed directions, it is replaced with a direction in the set of allowed directions that is consistent with the baseline confidence level gradient direction. This is considered a consistent projection and recorded as low consistency. This yields the projection consistency degree used to generate the trend consistency correction coefficient. Specifically, based on the fundamental credibility of this diagnostic and treatment data With interpolation confidence The sign of the difference determines the direction of change in the baseline reliability of the diagnostic data as the direction to be projected. ;like ,but For rising; if ,but For decrease; if ,but For stability; Within the punctual set corresponding to the time period between events to which the diagnostic data belongs, a trend symbol sequence consistent with the data type of the diagnostic data is extracted, and local trend segments are truncated according to the time span between anchor points. Based on these local trend segments, the monotonicity characteristics and extreme value distribution patterns are determined, and combined with the direction of the basic confidence gradient, a set of permissible directions corresponding to the trend constraint field is generated. : When a local trend segment is a monotonically rising pattern, let ={Rising, Stable}; When a local trend segment is a monotonically declining pattern, let ={Declining, Stable}; When the local trend segment is a pattern of first rising and then falling or first falling and then rising, let ={rising, falling, stable}, with the direction of the basic confidence gradient as the preferred direction; like Then it is considered a uniform projection, let ;like Then the direction to be projected will be replaced with The direction that is consistent with the direction of the basic confidence gradient is considered an inconsistent projection, let ;when When multiple non-stationary directions are present and the basic confidence gradient direction is stationary, the direction that appears more frequently in the local trend segment is preferred as the replacement direction. A trend consistency correction coefficient is generated based on the deviation between the projection consistency and the basic confidence of the diagnosis and treatment data relative to the anchor interpolation confidence. A weighted smoothing is then performed on the basic confidence of the diagnosis and treatment data to obtain the repaired confidence. The repaired confidence is recorded as the final confidence. For example, the formula for calculating the trend consistency correction coefficient is as follows: ; In the formula, This is a trend consistency correction coefficient. For projection consistency, The deviation is the degree of deviation; where, , It is a non-zero minimum constant; The expression for calculating the repair credibility is: ,in, To restore credibility; Through the above technical solution, this embodiment constructs anchor points with real record points and establishes a trend constraint field by combining the trend pattern of the timely set. Then, it uses consistency projection and deviation to generate correction coefficients for weighted smoothing repair, thereby achieving robust repair of diagnosis and treatment data that have not been marked as the target and whose basic credibility has continuity gaps or abnormal deviations, reducing the impact of credibility holes on subsequent feature construction.

[0018] In an optional embodiment, the updated or neighborhood-repaired baseline confidence level is recorded as the final confidence level, and the diagnostic data sequence is gated or weighted based on this final confidence level to construct a risk feature vector, specifically including: Based on the final credibility and the preset credibility interval, the diagnosis and treatment data are divided into an effective participation set, a weak participation set and a masked set. Attention-weighted aggregation is performed on the effective participation set, decay aggregation is performed on the weak participation set, and the data in the masked set is removed when constructing the risk feature vector. For example, the preset confidence interval is , The value range is 0.3-0.5, with a typical value of 0.4; The value range is 0.7-0.9, with a typical value of 0.8; Furthermore, when When, it is divided into a valid participant set; when When, it is divided into a weakly participating set; when When, it is divided into a shielding set; The weights for decay aggregation can be: ,in, Number the diagnostic and treatment data with weak participation; Within each time period between events, based on the temporal distribution characteristics of the diagnostic and treatment data types and their corresponding final credibility, time-specific risk sub-features are constructed respectively, and the risk sub-features are fused to form a risk feature vector; Through the above technical solution, this embodiment achieves hierarchical gating based on the final credibility and the preset credibility threshold. Attention aggregation, decay aggregation or elimination are adopted for different credibility levels. Time-specific risk sub-features are constructed and fused within the time period between events to reduce the pollution of risk features by low-credibility data and improve the stability and interpretation consistency of risk feature vectors.

[0019] See Figure 2 As shown, this solution proposes a needle removal risk prediction system based on data credibility constraints to implement the aforementioned needle removal risk prediction method based on data credibility constraints, including: The rule preset module is used to obtain the baseline time of needle removal and nursing scenario of the target patient, and extract the preset retrospective time window, diagnosis and treatment data type set and intervention event type set and their effect rules according to the nursing scenario; The timeline alignment module is used to extract the diagnosis and treatment data sequence and intervention event sequence within the backtracking time window and align them to a unified timeline with the baseline time as the reference point. The credibility calculation module is used to calculate the basic credibility of each piece of medical data in the medical data sequence by combining the quality of the medical data with the time-sensitivity decay relative to the baseline time. The object marking module is used to determine the influence window, applicable diagnostic and treatment data types and their associated penalty intensity of each intervention event in the intervention event sequence according to the action rules, and to mark the diagnostic and treatment data that are earlier than the intervention event and match the applicable diagnostic and treatment data types within the influence window as the action objects; The feedback correction module is used to perform feedback correction on the penalty intensity based on the completeness of the data type of the affected object relative to the action rule, and the trend consistency of all data types affecting the window. The update execution module is used to apply penalty weights or hard failure flags to each target based on the corrected penalty intensity and the relative time distance between the target and the intervention event, so as to update the basic reliability of the corresponding diagnosis and treatment data. The aggregation prediction module is used to perform neighborhood repair on diagnostic and treatment data that are not marked as targets and have abnormal continuity of basic credibility. The updated basic credibility is recorded as the final credibility, and the diagnostic and treatment data sequence is gated or weighted and aggregated according to it to construct a risk feature vector. The vector is then input into the risk prediction model to obtain the expected risk of needle removal.

[0020] In another embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above embodiments.

[0021] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps described above.

[0022] In one embodiment, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the steps described above.

[0023] The technical scope of this invention is not limited to the content described above. Those skilled in the art can make various modifications and variations to the above embodiments without departing from the technical concept of this invention, and all such modifications and variations should fall within the protection scope of this invention.

Claims

1. A method for predicting needle removal risk based on data credibility constraints, characterized in that, The method includes: Obtain the baseline time of needle removal and nursing scenario for the target patient, and extract the preset retrospective time window, diagnosis and treatment data set and intervention event type set and their effect rules based on the nursing scenario; Extract the diagnosis and treatment data sequence and intervention event sequence within the retrospective time window, and align them to a unified timeline with the baseline time as the reference point; For each piece of diagnostic and treatment data in the diagnostic and treatment data sequence, the basic reliability is obtained by integrating the quality of the diagnostic and treatment data with the time-sensitivity decay relative to the baseline time. Based on the action rules, determine the influence window, applicable diagnostic and therapeutic data types, and associated penalty intensity of each intervention event in the intervention event sequence, and mark the diagnostic and therapeutic data that are earlier than the intervention event and match the applicable diagnostic and therapeutic data types within the influence window as the action objects; Based on the completeness of the treatment data type of the affected object relative to the treatment rule, and the trend consistency of all treatment data within the affected window, feedback correction is performed on the penalty intensity; Based on the revised penalty intensity and the relative time distance between the corresponding target and the intervention event, penalty weights or hard failure markers are applied to each target to update the basic reliability of the corresponding diagnosis and treatment data; For diagnostic and treatment data that are not marked as targets and have abnormal continuity of basic credibility, neighborhood repair is performed based on adjacent updated basic credibility. The updated or neighborhood-repaired basic credibility is recorded as the final credibility. Based on this, the diagnostic and treatment data sequence is gated or weighted to construct a risk feature vector, which is then input into the risk prediction model to obtain the expected risk of needle removal.

2. The method according to claim 1, characterized in that, The basic reliability is obtained by integrating the quality of diagnostic and treatment data with the time-sensitivity decay relative to the baseline time, specifically including: Based on the nursing scenario, extract the preset update cycle corresponding to the data types of diagnosis and treatment; Within the retrospective time window, the sampling interval sequence of this diagnostic data type is statistically analyzed and the sampling interval dispersion is calculated, and a sampling reliability factor is generated based on it; The time interval between each diagnosis and treatment data point and the reference time point are calculated. The data points are then normalized according to the preset update cycle parameters to obtain the initial timeliness. The initial timeliness is then adjusted using the sampling reliability factor to obtain the actual timeliness. Extract missing completeness indicators, data source consistency indicators, and outlier interpretability indicators from each diagnosis and treatment data, and generate data quality scores; The timeliness decay coefficient is determined based on the actual timeliness, and then integrated with the data quality score to obtain the basic credibility.

3. The method according to claim 2, characterized in that, Before applying feedback correction to the penalty intensity, the time points of each diagnosis and treatment data point in the diagnosis and treatment data sequence are offset, specifically including: Starting from the endpoint of the retrospective time window that is far from the reference time, the extended time period is obtained by extending it along a unified time axis in a direction far from the reference time. The endpoint of the extension is the time point of the first intervention event in the extension direction or the time point of the diagnosis and treatment data that is farthest from the starting point within the extended time period. Preset diagnosis and treatment templates are generated based on the preset update cycle corresponding to the nursing scenario and diagnosis and treatment data types. These templates include the preset time point range and allowed offset threshold for each diagnosis and treatment data type. Within the retrospective time window, the time periods between events are divided according to adjacent intervention events. Based on the action rules corresponding to each time period, the preset time point range of the preset diagnosis and treatment template is updated sequentially to obtain the intervention impact diagnosis and treatment template. In the extended time period and the time period between each event, the diagnosis and treatment data are mapped to the intervention impact diagnosis and treatment template according to the data type, and the time offset relative to the range of the most recent preset time point is calculated. If the time offset exceeds the allowed offset threshold, the corresponding diagnostic data will be marked as a suspected error site.

4. The method according to claim 3, characterized in that, After marking the corresponding diagnostic data as suspected errors, the penalty intensity is also adjusted, specifically including: Within the time periods between each event, diagnostic and treatment data that were not marked as suspected misalignments and whose mappings satisfy the intervention impact diagnostic and treatment template are selected to form a timely set; The diagnostic and treatment data corresponding to each type of diagnostic and treatment data in the time set are sorted by time to form a type time series subsequence, and a trend symbol sequence and an inflection point sequence are constructed based on the type time series subsequence; For each suspected error site, extract the trend symbol sequence and inflection point sequence of the diagnostic and treatment data type to which it belongs in the timely set, and calculate the local trend consistency index between the suspected error site and the trend symbol sequence and inflection point sequence. Based on the local trend consistency index, suspected misalignment points are classified as true record points or true misalignment points; Remove the mark of the real record point as the target, and keep its basic credibility unchanged and regard it as the updated basic credibility; The penalty intensity corresponding to the actual misalignment point is adjusted by a preset multiple. If the adjusted penalty intensity value exceeds the preset intensity threshold, a hard failure mark is applied to the actual misalignment point. For diagnostic and treatment data other than actual record points and actual misalignment points, the target labeling and penalty intensity remain unchanged.

5. The method according to claim 4, characterized in that, Based on the completeness of the treatment data type of the affected object relative to the application rule, and the trend consistency of all treatment data within the affected window, feedback adjustments are made to the penalty intensity, specifically including: Based on the preset update cycle and intervention impact diagnosis and treatment template, an expected arrival list is generated in each impact window. The expected arrival list is used to record the expected arrival times of each type of applicable diagnosis and treatment data within each preset time point range. Within the influence window, the actual number of times the target object falls within each preset time point range is counted. The actual misalignment points are not included in the actual number of arrivals. The completeness index is calculated based on the actual number of arrivals and the expected arrival list, and a gap location set is generated. All diagnostic and treatment data within the affected window are weighted and aggregated according to basic reliability, and the weighted change direction of each diagnostic and treatment data type between adjacent preset time points is calculated. The trend consistency index is calculated based on the weighted change direction consistency ratio of each diagnostic and treatment data type and the distribution density of the gap location set. This index is then integrated with the completeness index to form a feedback correction factor. This feedback correction factor is used to correct the penalty intensity associated with the intervention event in the action rule.

6. The method according to claim 4, characterized in that, For diagnostic and treatment data that are not marked as targets and exhibit abnormal continuity in basic confidence, neighborhood repair is performed based on adjacent updated basic confidence levels. This includes: Extract the nearest real record points from the preceding and following timelines of the diagnosis and treatment data on a unified timeline as anchor points. If a real record point is missing on either side, use the diagnosis and treatment data corresponding to the nearest updated basic credibility on that side as a replacement anchor point. Calculate the time span and basic credibility gradient between anchor points. Based on the basic confidence of the anchor points and the time span between the anchor points, the interpolation confidence of the diagnosis and treatment data is obtained by interpolating at its time points. In the timely set corresponding to the time period between events to which the diagnosis and treatment data belongs, extract the trend symbol sequence that is consistent with the data type of the diagnosis and treatment data, and extract local trend segments according to the time span; A trend constraint field is constructed based on the basic confidence gradient and local trend segments. The direction of change of the basic confidence of the diagnosis and treatment data is uniformly projected onto the trend constraint field to obtain the projection consistency. A trend consistency correction coefficient is generated based on the deviation between the projection consistency and the basic confidence of the diagnosis and treatment data relative to the anchor interpolation confidence. A weighted smoothing is then performed on the basic confidence of the diagnosis and treatment data to obtain the repaired confidence. The repaired confidence is recorded as the final confidence.

7. The method according to claim 6, characterized in that, The updated or neighborhood-repaired baseline confidence level is recorded as the final confidence level. Based on this, the diagnostic data sequences are gated or weighted to construct a risk feature vector, specifically including: Based on the final credibility and the preset credibility interval, the diagnosis and treatment data are divided into an effective participation set, a weak participation set and a masked set. Attention-weighted aggregation is performed on the effective participation set, decay aggregation is performed on the weak participation set, and the data in the masked set is removed when constructing the risk feature vector. Within each event time period, based on the temporal distribution characteristics of the diagnostic and treatment data types and their corresponding final credibility, time-specific risk sub-features are constructed, and the risk sub-features are fused to form a risk feature vector.

8. A needle removal risk prediction system based on data credibility constraints, characterized in that, A method for predicting needle removal risk based on data credibility constraints as described in any one of claims 1-7, comprising: The rule preset module is used to obtain the baseline time of needle removal and nursing scenario of the target patient, and extract the preset retrospective time window, diagnosis and treatment data type set and intervention event type set and their effect rules according to the nursing scenario; The timeline alignment module is used to extract the diagnosis and treatment data sequence and intervention event sequence within the backtracking time window and align them to a unified timeline with the reference time as the reference point; The credibility calculation module is used to calculate the basic credibility of each diagnosis and treatment data in the diagnosis and treatment data sequence by combining the quality of the diagnosis and treatment data with the time-related decay relative to the baseline time. The object marking module is used to determine the influence window, applicable diagnostic and treatment data types and their associated penalty intensity of each intervention event in the intervention event sequence according to the action rules, and to mark the diagnostic and treatment data that are earlier than the intervention event and match the applicable diagnostic and treatment data types within the influence window as the action objects; The feedback correction module is used to perform feedback correction on the penalty intensity based on the completeness of the data type of the affected object relative to the action rule, and the trend consistency of all data types affecting the window. The update execution module is used to apply penalty weights or hard failure flags to each target based on the corrected penalty intensity and the relative time distance between the target and the intervention event, so as to update the basic reliability of the corresponding diagnosis and treatment data. The aggregation prediction module is used to perform neighborhood repair on diagnostic and treatment data that are not marked as targets and have abnormal continuity of basic credibility. The updated basic credibility is recorded as the final credibility, and the diagnostic and treatment data sequence is gated or weighted and aggregated according to it to construct a risk feature vector. The vector is then input into the risk prediction model to obtain the expected risk of needle removal.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

Images