Automatic labeling method and system for predicting the risk of septic shock

Abstract

This application provides an automatic annotation method and system for predicting the risk of bloodstream infection-related septic shock, belonging to the field of smart healthcare. It addresses the problem that existing models often rely on labeled sepsis or septic shock events as input, and the annotation process depends on manual work by clinical experts, resulting in high subjectivity and low efficiency. The method includes: identifying suspected infection events and determining suspected infection time anchors based on the target patient's medication order data and pathogen data; identifying suspected infection events by detecting the matching results of antimicrobial drug administration records and culture sampling records under preset time constraints; constructing sepsis and septic shock assessment observation windows based on the suspected infection time anchors to determine the onset time of bloodstream infection-related septic shock; and automatically generating sample labels for training or evaluating the risk prediction model based on the onset time of bloodstream infection-related septic shock and preset risk prediction task configuration.

Description

Technical Field

[0001] This application relates to the field of smart healthcare technology, and in particular to an automatic labeling method and system for predicting the risk of septic shock. Background Technology

[0002] Predictive research on bloodstream infection (BSI)-related septic shock is a key scientific issue for the accurate identification of sepsis. In clinical predictive research on sepsis, existing models mostly rely on labeled sepsis or septic shock events as input. The labeling process often depends on manual review by clinical experts, which suffers from high subjectivity, low efficiency, and difficulty in scalable reproduction. Therefore, automatically, accurately, and consistently identifying key time points of bloodstream infection-related septic shock from massive, heterogeneous electronic medical record time-series data, and constructing observation and prediction windows suitable for machine learning models, remains a challenge and a gap in current research. Summary of the Invention

[0003] This application provides an automatic annotation method and system for predicting the risk of septic shock. It can solve the problem that in sepsis-related clinical prediction studies, existing models mostly rely on labeled sepsis or septic shock events as input. The annotation process often depends on manual review by clinical experts, which has the problems of strong subjectivity, low efficiency and difficulty in large-scale reproduction.

[0004] The first aspect of this application provides an automatic annotation method for predicting the risk of septic shock, including: Based on the medication order data and pathogen data of the target patients, suspected infection events are identified and suspected infection time anchors are determined. The identification of suspected infection events includes detecting the matching results of antimicrobial drug administration records and culture sampling records under a preset time constraint relationship. Based on the suspected infection time anchor point, a sepsis determination observation window and a septic shock determination observation window are constructed to determine the onset time of bloodstream infection-related septic shock. Based on the onset time of bloodstream infection-related septic shock and the preset risk prediction task configuration, sample labels and corresponding observation time windows and prediction time windows are automatically generated for training or evaluating the risk prediction model. The sample labels are used to characterize whether bloodstream infection-related septic shock occurs within the prediction time window.

[0005] A second aspect of this application provides an automatic labeling system for predicting the risk of septic shock, comprising: The identification unit is used to identify suspected infection events and determine the time anchor of suspected infection based on the medication order data and pathogen data of the target patient. The identification of suspected infection events includes detecting the matching results of antimicrobial drug administration records and culture sampling records under a preset time constraint relationship. The determination unit is used to construct a sepsis determination observation window and a septic shock determination observation window based on the suspected infection time anchor point, so as to determine the onset time of bloodstream infection-related septic shock. An automatic labeling unit is used to automatically generate sample labels and corresponding observation time windows and prediction time windows for training or evaluating risk prediction models based on the onset time of bloodstream infection-related septic shock and the preset risk prediction task configuration. The sample labels are used to characterize whether bloodstream infection-related septic shock occurs within the prediction time window.

[0006] A third aspect of this application provides an electronic device, including a memory and a processor, wherein the processor is used to execute a computer program stored in the memory to implement the steps of the above-described automatic labeling method for predicting the risk of septic shock.

[0007] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described automatic labeling method for predicting the risk of septic shock.

[0008] In summary, the automatic labeling method for predicting the risk of septic shock provided in this application uses time anchors as the main thread to organize evidence scattered across multiple systems into a computable event chain. First, a suspected infection event is constructed using the time constraint relationship between antimicrobial drugs and culture sampling, and a suspected infection time anchor is determined. Then, using this anchor as a benchmark, a sepsis assessment observation window is constructed under fixed rules, and the sepsis onset time is determined using changes in organ function scores or alternative criteria. Next, a septic shock assessment observation window is constructed based on the sepsis onset time, and within this window, the onset time of bloodstream infection-related septic shock is determined by combining fluid resuscitation adequacy, hemodynamic support medication, and abnormal tissue perfusion indicators. Finally, an observation time window and a prediction time window are automatically generated using the shock onset time as the target event time point, and sample labels for training or evaluation are generated. Transforming the complex clinical event timing that originally relied on manual inference into an executable automatic annotation process significantly reduces annotation costs, improves annotation consistency and traceability, and enables data from different medical institutions, different wards, and different levels of informatization to generate structurally consistent labels and window configurations, directly supporting the large-scale training, cross-center validation, and online deployment of risk prediction models.

[0009] Correspondingly, the systems, electronic devices, and computer-readable storage media provided in the embodiments of the present invention also have the above-mentioned technical effects. Attached Figure Description

[0010] Figure 1 A flowchart illustrating a possible automatic labeling method for predicting the risk of septic shock, provided as an embodiment of this application; Figure 2 A schematic structural block diagram of a possible automatic labeling system for predicting the risk of septic shock, provided for an embodiment of this application; Figure 3 A schematic diagram of the hardware structure of a possible automatic labeling system for predicting the risk of septic shock, provided for an embodiment of this application; Figure 4 A schematic structural block diagram of a possible electronic device provided in an embodiment of this application; Figure 5 This is a schematic structural block diagram of a possible computer-readable storage medium provided for embodiments of this application. Detailed Implementation

[0011] This application provides an automatic annotation method and related equipment for predicting the risk of septic shock. It can solve the problem that in sepsis-related clinical prediction studies, existing models mostly rely on labeled sepsis or septic shock events as input. The annotation process often depends on manual review by clinical experts, which has the problems of strong subjectivity, low efficiency and difficulty in large-scale reproduction.

[0012] Please see Figure 1 The flowchart below illustrates an automatic annotation method for predicting the risk of septic shock, as provided in this application embodiment. Specifically, it may include: S110-S130.

[0013] S110, based on the medication order data and pathogen data of the target patient, identify suspected infection events and determine the suspected infection time anchor point. The identification of suspected infection events includes detecting the matching results of antimicrobial drug administration records and culture sampling records under a preset time constraint relationship.

[0014] S120, using the suspected infection time anchor point as a benchmark, construct a sepsis determination observation window and a septic shock determination observation window to determine the onset time of bloodstream infection-related septic shock.

[0015] S130, based on the onset time of bloodstream infection-related septic shock and the preset risk prediction task configuration, automatically generate sample labels and corresponding observation time windows and prediction time windows for training or evaluating the risk prediction model. The sample labels are used to characterize whether bloodstream infection-related septic shock occurs within the prediction time window.

[0016] In some examples, the construction of sepsis assessment observation windows and septic shock assessment observation windows based on the suspected infection time anchor point to determine the onset time of bloodstream infection-related septic shock includes: Based on the suspected infection time anchor point, a sepsis determination observation window is constructed, and the sepsis onset time is determined within the sepsis determination observation window based on changes in organ function scores or alternative criteria. Based on determining the time of sepsis onset, a septic shock assessment observation window is constructed, and within the septic shock assessment observation window, the onset time of bloodstream infection-related septic shock is determined based on the combined conditions of adequate fluid resuscitation, hemodynamic support medication, and abnormal tissue perfusion indicators.

[0017] For example, in order to enable the automatic labeling of bloodstream infection-related septic shock to be based on traceable and aligned multi-source evidence, the system obtains the electronic medical record data of the target patient from the target medical institution's information system. The occurrence time of the shock event is not a single field that can be directly read, but is scattered in the time series of medication, testing, monitoring and treatment records. Therefore, it is necessary to aggregate the scattered evidence into a unified data structure to support subsequent time constraint matching and joint condition determination. In practice, the system retrieves at least the following data for each target patient: medication orders, microbial culture or pathogen detection data, vital sign monitoring data, laboratory test data, and treatment records. For each record, a standardized event table is created by extracting the patient identifier, event type, event occurrence time, event value, and data source identifier. Medication orders prioritize retaining the actual execution time and dosage information to avoid bias caused by using only the prescription time. Microbial culture or pathogen detection data retains the sampling time, sample type, and result reporting time. Vital sign monitoring data retains the collection frequency and missing segment information. Laboratory test data retains both sampling time and reporting time, prioritizing the sampling time as the clinical status timestamp. Treatment records retain the time points for key events such as fluid resuscitation volume, mechanical ventilation, central venous catheterization, and resuscitation records. During the import phase, the system performs unified timestamp processing, including time zone unification, cross-system clock offset correction, and merging of duplicate records, while retaining the original timestamps and converted records for traceability. This provides a complete set of evidence covering infection assessment, organ function changes, and hemodynamic support for subsequent automatic labeling. For example, information such as the first dose of antibiotics administered to the same patient, blood culture sampling, lactate retesting, decrease in mean arterial pressure, and norepinephrine initiation are all organized into a calculable timeline, thereby avoiding missed or incorrect judgments due to relying solely on data from a single system.

[0018] For example, in order to stably locate the infection assessment initiation time in the absence of an explicit infection start time field, the system identifies suspected infection events and determines the suspected infection time anchor based on medication order data and microbial culture or pathogen detection data. Clinically, the treatment of suspected bloodstream infections is usually accompanied by two highly consistent actions: antimicrobial drug administration and culture sampling. There is a modelable constraint relationship between the two in time. Using this relationship, a consistent definition of suspected infection events can be constructed from real-world data. In practice, the system retrieves the target patient's antimicrobial drug administration records and culture sampling records. The antimicrobial drug administration records prioritize the actual time of the first dose administration, while the culture sampling records prioritize the actual sampling time of blood culture or pathogen detection. The system sets preset time constraints and performs matching judgments. For example, when a culture sampling time is detected within a first preset time period before an antimicrobial drug administration time or within a second preset time period after the administration time, the system determines that the drug administration record and the sampling record constitute a valid match, thereby identifying a suspected infection event. When there are multiple administrations and multiple samplings, the system can select the first pair of records that meet the matching relationship as evidence of suspected infection and write the matching candidate pair and its time difference into the traceability field for auditing. The suspected infection time anchor can be determined according to preset rules as the earlier of the sampling time and the first dose administration time or the time closer to the start of emergency assessment. At the same time, the system filters the antimicrobial drug categories to exclude false triggers caused by perioperative prophylactic medications or non-infection-related medications. Transforming subjective and difficult-to-unify suspected infection initiation points into reproducible time anchors allows subsequent sepsis and shock windows to be constructed using the same reference time. For example, when a patient has blood cultured first and then receives the first dose of broad-spectrum antibiotics within one hour, the anchor point can fall on the sampling time to cover changes in the patient's condition before and after sampling. When a patient first receives emergency antibiotics and then has blood cultured again, the anchor point can still be identified through reverse time constraints, thereby reducing missed labeling and improving cross-center consistency.

[0019] For example, in order to objectively locate the time point of organ dysfunction in the context of infection, the system constructs a sepsis judgment observation window based on the suspected infection time anchor point, and determines the sepsis onset time based on changes in organ function scores or alternative criteria within this window. The key feature of sepsis is the impairment of infection-related organ function, which often manifests as significant changes in multiple indicators relative to the baseline in a short period of time. Using score changes can reduce misjudgments caused by occasional abnormalities in single indicators, and introducing alternative criteria can improve coverage when scores cannot be fully calculated. In practice, the system establishes an observation window around the suspected infection time anchor point. For example, it can extend forward to a preset baseline period and backward to a preset assessment period. Within the baseline period, the system calculates the baseline value of organ function scores. The baseline value can be the lowest score or a steady-state statistical value to reduce the impact of occasional fluctuations. Within the assessment period, the system aggregates the scoring elements required for vital signs, laboratory tests, and treatment records according to timestamps, calculates the organ function scores that change over time, and obtains the score increment relative to the baseline. When the score increment first reaches a preset threshold, the corresponding time point is determined as the sepsis onset time, and the changes in key indicators that triggered the threshold are written into the evidence field, such as increased creatinine, decreased platelets, increased bilirubin, decreased consciousness, or persistently low blood pressure. When scoring elements are missing, making it impossible to reliably calculate the score, the system calls alternative criteria for judgment. The alternative criteria can be composed of a combination of the persistence of available abnormal vital signs, abnormal key laboratory indicators, and treatment events, and the time point that first meets the judgment conditions is also taken as the onset time. Therefore, the system can stably output the sepsis initiation point even when the data integrity of multiple centers is inconsistent. This reduces the risk of mistaking previous chronic organ dysfunction for new sepsis, while preserving the true characteristics of rapid deterioration in the emergency course. For example, if a patient experiences a simultaneous deterioration in creatinine and platelet count four hours after the anchor point, causing the score increment to reach the threshold, the system can mark this moment as the sepsis onset time, thus providing a starting point that is closer to the pathological process for subsequent shock assessment.

[0020] For example, to combine disparate signals such as decreased blood pressure, use of vasopressors, or insufficient perfusion into a more specific determination of septic shock, the system constructs a septic shock determination observation window based on determining the time of sepsis onset. Within this window, the onset time of bloodstream infection-related septic shock is determined based on the combined conditions of adequate fluid resuscitation, hemodynamic support medication, and abnormal tissue perfusion indicators. Shock cannot be defined by hypotension alone, but should be reflected in the presence of circulatory support requirements and perfusion abnormalities even with adequate resuscitation. By combining conditions, both noise-induced false triggering and interventional medication-induced false triggering can be reduced simultaneously. In practice, the system uses the sepsis onset time as the starting point and extends forward by a preset duration to form a shock assessment observation window. Within this window, the system analyzes the volume and time distribution of fluid resuscitation to assess the adequacy of resuscitation. The adequacy of resuscitation can be determined based on rules such as the cumulative crystalloid or total fluid intake reaching a threshold within the preset assessment period, or the presence of records of continuous fluid resuscitation. The system can also be normalized according to body weight parameters. Simultaneously, the system analyzes hemodynamic support medication records to identify the initiation time, duration, and dosage information of vasoactive drugs, forming candidate shock trigger time points. Candidate points can be the first initiation time of vasoactive drugs or the time when the maintenance dose is reached. The system further searches for abnormal tissue perfusion indicators near candidate points and imposes a persistence requirement. Abnormal tissue perfusion indicators include at least lactate exceeding a threshold, mean arterial pressure consistently below a threshold, significantly reduced urine output, or skin perfusion-related records. Single abnormal measurements can be excluded based on time series continuity. When the system determines that the conditions for adequate resuscitation, hemodynamic support medication, and abnormal tissue perfusion are simultaneously met within the same time neighborhood, the earliest time point meeting the combined conditions is identified as the onset time of bloodstream infection-related septic shock, and an evidence chain consisting of fluid resuscitation evidence, medication evidence, and perfusion evidence is output. This allows for earlier and more consistent identification of the shock initiation point and reduces the risk of mislabeling. For example, if a patient requires norepinephrine to maintain mean arterial pressure even after cumulative fluid resuscitation meets the adequacy standard, and lactate levels rise upon retesting, the system can label the earliest time meeting the combined conditions as the shock onset time. However, in cases of transient sedation-related hypotension leading to short-term use of vasopressors but without elevated lactate and insufficient resuscitation evidence, the combined conditions will not be triggered, thus improving labeling specificity.

[0021] For example, to ensure consistent input and output definitions for training samples of the risk prediction model across different patients and centers, the system automatically generates sample labels and corresponding observation and prediction time windows based on the onset time of bloodstream infection-related septic shock and the preset risk prediction task configuration. Training samples must clearly define the scope of observation data extraction and the scope of predicted event judgment, and the observation window must not contain information after the event occurs to avoid label leakage, thereby ensuring the authenticity and reliability of the model performance evaluation. In specific implementation, the system reads the preset risk prediction task configuration, which includes at least the length of the observation time window, the length of the prediction time window, the window positioning rules relative to the event time, and sample selection rules. The system uses the shock onset time as the target event time point, extracts a preset length forward to form the observation time window for extracting model input features, and defines a preset length backward to form the prediction time window for determining label semantics. Sample labels are assigned based on whether the shock onset time falls within the prediction time window; if it does, it is marked as an event that has occurred; otherwise, it is marked as an event that has not occurred. The system simultaneously checks the observation time... The system checks whether the combined criteria for shock determination have been met within the observation window, or whether there are strong warning events such as continued use of vasopressors and significantly elevated lactate levels. If so, the sample is removed or the window start and end are adjusted according to the configured rules to avoid introducing post-event information into the observation window. The system ultimately outputs a structured annotation result table, which includes at least the patient identifier, suspected infection time anchor, sepsis onset time, shock onset time, observation window start and end, prediction window start and end, label value, and evidence traceability fields. It also allows generating multiple sets of samples under different task configurations for the same patient to adapt to different applications such as emergency early warning, ward monitoring, or ICU early intervention. Thus, label generation is standardized into a configurable process, ensuring the comparability of training data and supporting adjustments to the lead time according to clinical needs. For example, when it is necessary to predict whether shock will occur in the next six hours six hours in advance, the system automatically generates an observation window that does not include information on vasopressors after shock occurs, thereby avoiding inflated training performance and improving the reliability of the model after deployment.

[0022] In some examples, identifying suspected infection events and determining suspected infection time anchors includes: If the target patient is detected to have a first time of antibiotic administration, determine whether there is a culture sampling time within a first preset time period before the first time of antibiotic administration to determine a first time constraint relationship; If a culture sampling time is detected for the target patient, it is determined whether there is a first administration time of antimicrobial drug within a second preset time period after the culture sampling time, so as to determine the second time constraint relationship. The time of first administration of the antimicrobial drug or the time of culture sampling that satisfies the time constraint relationship is determined as the time anchor point of the suspected infection, and the target patient is marked as a suspected infected patient to enter the subsequent automatic labeling process.

[0023] For example, in order to reliably identify the initiation time of infection assessment when electronic medical records lack an explicit suspected infection start time field, the system prioritizes the first administration of antimicrobial drugs as the trigger entry and searches backward for culture and sampling records. In clinical practice, the path for suspected bloodstream infection usually involves first performing etiological sampling and then initiating empirical antimicrobial treatment as soon as possible. The two have a relatively close sequential relationship in time. This relationship can be used to distinguish infection-related actions from other non-infection-related medications or tests. In practice, the system identifies the first administration time of antimicrobial drugs in medication orders and execution records, preferably using the actual execution time rather than the prescription time, and limits the scope of antimicrobial drugs, including only therapeutic antimicrobial drugs and excluding records that may cause accidental triggering, such as perioperative prophylaxis, skin tests, or topical medications. Subsequently, the system uses the first administration time as a reference point and searches for sampling times for microbial culture or pathogen detection within a first preset time period before it. The sampling time can preferentially select blood culture sampling time, and can also allow sampling times for other sterile body fluid cultures or rapid pathogen detection as candidate evidence. The first preset time period can be configured to range from several hours to several tens of hours to adapt to the emergency department's sampling before administration process. At the same time, for cases where the same patient has multiple samplings, the system selects the sampling record closest to the first administration and located within the time window according to the time distance priority principle, and forms a pair of matching evidence with the first administration record, thereby determining that the first time constraint relationship is established. Therefore, it is possible to identify suspected infection events with high consistency in a large number of real cases where sampling is performed before medication is administered, reducing the risk of misdiagnosing non-infectious prophylactic medication as infection based solely on antibiotic administration; for example, if a patient has a blood culture taken after being admitted to the emergency department, and then the first dose of broad-spectrum antibiotic is administered within one to two hours, this matching relationship can be stably triggered and provide an accurate time reference for the subsequent construction of the sepsis window.

[0024] For example, to avoid only covering emergency scenarios where sampling is performed before administration and missing real-world scenarios where administration is performed before sampling or administration is delayed after sampling, the system, upon detecting a culture sampling time for the target patient, further searches backwards using the sampling time as the entry point for the first administration record of the antimicrobial drug. Clinically, there may be changes in the time sequence due to priority of rescue, patient transfer, or delays in the execution of medical orders, but as long as the sampling and subsequent antimicrobial treatment are associated within a reasonable time span, they can still be regarded as the same suspected infection event. In practice, the system identifies the culture sampling time from microbial culture or pathogen detection data, prioritizing the actual sampling time over the result reporting time. Priorities can be set for sampling types, such as blood culture over sputum culture, and then over urine culture, to improve targeting of bloodstream infection scenarios. Subsequently, using this sampling time as a reference point, the system searches for the first-dose time of antimicrobial drugs within a second preset time period. This first-dose time also prioritizes the actual dosing time recorded in the execution record and performs therapeutic filtering on antimicrobial drugs. The second preset time period can be configured to be longer than the first preset time period to cover situations where antimicrobial treatment is delayed after sampling. The system can also add supplementary verification conditions, such as requiring the indication or diagnostic field of the medication order to be related to infection, or requiring changes in infection-related test indicators after sampling, to reduce erroneous matching between incidental sampling and irrelevant dosing. When multiple dosing records exist, the system selects the earliest first-dose record of therapeutic antimicrobial drugs within the second preset time period after sampling to form matching evidence with that sampling record, thereby confirming the validity of the second time constraint relationship. This can significantly improve the recall rate of suspected infection identification, especially for high-risk emergency patients. For example, if a patient is admitted to the emergency room and is immediately treated with antibiotics based on clinical experience, and blood cultures are collected after the patient is stabilized, or if a patient is sampled first but medication is delayed for several hours due to bed turnover, the patient can still be included in the suspected infection population through the second time constraint relationship, thereby reducing missed labeling and increasing the coverage of subsequent sepsis and shock labeled samples.

[0025] For example, when the first or second time constraint relationship is established, the system determines the time of first administration of antimicrobial drugs or the time of culture sampling that satisfies the time constraint relationship as the suspected infection time anchor point, and marks the target patient as a suspected infection patient to enter the subsequent automatic labeling process. The suspected infection time anchor point is used as a unified time reference point for the construction of the observation window for subsequent sepsis determination. The selection of the anchor point should be as close as possible to the actual starting point of infection assessment and treatment initiation to reduce window drift between different cases and provide a basis for the comparability of subsequent labels. In practical implementation, if the first time constraint is valid, the system can use the culture sampling time or the first antibiotic administration time as the anchor point. The earlier occurrence time can be prioritized to cover early changes between sampling and administration, or a time point more representative of infection management initiation can be selected based on institutional configuration. If the second time constraint is valid, the system can similarly select an anchor point between the sampling time and the first administration time according to rules, and write this anchor point and triggering evidence into a traceability field. The traceability field includes at least the matching relationship type, the first or second preset duration parameter, the matched administration record identifier, the matched sampling record identifier, and the time difference between the two. When both types of time constraints are valid simultaneously or multiple candidate matches exist, the system can use rules such as minimum time difference priority, blood culture priority, or therapeutic broad-spectrum antibiotic priority to determine the final anchor point, and retain unselected candidate pairs as alternative evidence. After determining the anchor point, the system marks the target patient as a suspected infection patient and enters the subsequent automatic labeling process for sepsis and shock determination. Patients who do not meet any time constraints are either excluded from this process or directed to another path to avoid noise labeling. Therefore, the suspected infection identification results are transformed into structured anchor points and patient stratification markers, making the input population more focused and the time reference more consistent in the subsequent automatic labeling process. This reduces mislabeling caused by including non-infected patients or patients with insufficient evidence of infection. For example, for a certain patient, the system not only gives a suspected infection conclusion, but also provides the anchor point time and its source evidence, which is convenient for clinical or research personnel to verify. It also ensures that the subsequent sepsis onset time and shock onset time are derived under the same anchor point system, improving the consistency and interpretability of the labels.

[0026] In some examples, the step of constructing a sepsis assessment observation window based on the suspected infection time anchor point, and determining the sepsis onset time within the sepsis assessment observation window based on changes in organ function scores or alternative criteria, includes: The sepsis determination observation window is formed by extending a preset time forward and backward from the suspected infection time anchor point; Within the sepsis assessment observation window, the sepsis onset time is determined based on whether the increase in the target patient's organ function score relative to the baseline score reaches a preset threshold. The baseline score is the lowest or median score within a preset baseline time period prior to the sepsis assessment observation window. In cases where organ function scores are missing or cannot be calculated, alternative criteria are used to determine the timing of sepsis onset. These alternative criteria include rules based on combinations of abnormal vital signs or simplified scoring.

[0027] For example, in order to limit the determination of sepsis to the range of suspected infection events and reduce the probability of mistaking previous chronic organ dysfunction for new sepsis, the system expands the sepsis determination observation window forward and backward by a preset time period centered on the suspected infection time anchor point. The suspected infection anchor point represents the time reference point for infection assessment and treatment initiation. The observation window with a fixed structure built around this reference point can interpret subsequent changes in organ function as new changes in the context of infection, while making the annotations of different patients and different centers have a comparable time alignment. In practice, the system reads the window parameter configuration. The forward extension duration is used to cover the baseline and early changes in the patient's condition before the anchor point, and the backward extension duration is used to cover the time period during which organ dysfunction may occur after the infection assessment is initiated. The forward extension duration can be set to a range of several hours to more than ten hours to cover monitoring and testing data before and after admission to the emergency room or in the early stages of admission to the resuscitation room. The backward extension duration can be set to a range of more than ten hours to tens of hours to cover the process of organ dysfunction in most sepsis cases during the emergency room and early hospitalization stages. When constructing the window, the system uses the timestamp of the event table as the standard, and maps vital sign monitoring, laboratory test sampling, treatment events and scoring elements to the time axis of the window in a unified manner. It also corrects for cross-system clock offsets and records the relationship between the window start and end times and the anchor point time for traceability. This windowing process eliminates the reliance on subjective starting points for sepsis diagnosis, instead establishing a stable correlation with suspected infection events. For example, if a patient has mild renal dysfunction before the anchor point but experiences a rapid rise in creatinine and a sustained drop in blood pressure after the anchor point, this window can more clearly distinguish between the baseline state and the new deterioration, thereby improving the accuracy and consistency of sepsis initiation location.

[0028] For example, in order to transform the occurrence of organ dysfunction from complex evidence collected by multiple indicators and asynchronously into a calculable time point, the system determines the sepsis onset time based on whether the increment of the target patient's organ function score relative to the baseline score reaches a preset threshold within the sepsis assessment observation window. The organ function score can aggregate changes in dimensions such as renal function, liver function, coagulation, circulation, respiration and neurology into a unified scale, thereby reducing misjudgments caused by occasional abnormalities of a single indicator, and emphasizing newly developed or aggravated organ dysfunction after infection rather than previous chronic abnormalities through the increment relative to the baseline. In practice, the system sets a preset baseline time period before the suspected infection time anchor point. This baseline time period can be configured to be several hours to several tens of hours before the anchor point, used to capture the patient's relatively stable state before the infection assessment is initiated. Within this baseline time period, the system calculates organ function scores and determines the baseline score. The baseline score can be the lowest score within this time period to reflect the optimal baseline state, or the median score to reduce the impact of occasional abnormalities. The specific method can be configured according to the scenario and written into the traceability field. Subsequently, the system calculates the score time series in the sepsis determination observation window according to time progression. This time progression can adopt an event-driven approach, that is, when key elements are updated, the score is recalculated and the score timestamp is bound to the element sampling time. At the same time, for irregular test sampling, the most recent valid value is used within a reasonable validity period to avoid score series breakage. The system calculates the score increment for each time point and compares it with a preset threshold. When the score increment reaches the threshold for the first time, the time point is determined as the sepsis onset time, and the key element changes that triggered the increment are recorded, such as creatinine rising from the baseline level to a new level, platelet count falling to a new range, and mean arterial pressure continuously decreasing and requiring circulatory support. Therefore, it can provide a stable starting point for sepsis even in the case of noise and uneven sampling in multi-center data. For example, if a patient experiences an increase in creatinine and a decrease in platelets within four hours after a suspected infection anchor point, and the score increment reaches the threshold for the first time, the system marks this moment as the sepsis onset time. This is earlier and more consistent than directly reading the diagnosis time, and is also more suitable for time alignment of subsequent shock determination window construction and risk prediction samples.

[0029] For example, considering the significant differences in the level of informatization, availability of test items, and data loss patterns among different medical institutions, organ function scores may be missing or uncalculated due to the unavailability of key elements. Relying solely on scores would prevent a large number of patients from completing automatic annotation. Therefore, when scores are missing or uncalcifiable, the system switches to alternative criteria to determine the timing of sepsis attacks. More readily available vital signs and some key test indicators can be used to construct executable judgment rules to ensure the coverage of the automatic annotation process. Simultaneously, the system maintains the specificity of the judgment as much as possible through combined conditions and persistent constraints. In specific implementation, the system first checks the computability of the scores. For example, switching is triggered when the proportion of missing score elements within the observation window exceeds a threshold, key tests are missing for a long period, or the data format does not meet the calculation requirements. After switching, the system uses alternative criteria to determine the timing of sepsis attacks. Alternative criteria can include rules for abnormal combinations of vital signs, such as persistent hypotension or mean arterial pressure consistently below a threshold within a preset continuous time period, accompanied by a persistently increasing respiratory rate or a persistently decreasing blood oxygen saturation, combined with evidence reflecting insufficient organ perfusion, such as decreased consciousness or significantly reduced urine output. Alternative criteria can also include… The system includes simplified scoring rules, calculating a simplified organ function score using only a limited set of available data, and locating the onset time based on the incremental increase relative to the baseline reaching a threshold. To reduce misjudgment, the system incorporates event consistency checks. For example, if an abnormality in vital signs is accompanied by records of rapid fluid resuscitation, oxygen escalation, or transfer to intensive care, the confidence level of the judgment is increased. If the abnormality occurs only once and recovers rapidly without supporting evidence, the judgment is delayed or the event is not triggered. Ultimately, the system uses the time when the alternative criteria are first met as the sepsis onset time and outputs the triggered abnormal combination, duration, and supporting evidence. This switching mechanism significantly improves cross-center adaptability and sample coverage. For instance, if a center lacks arterial blood gas analysis or some tests, resulting in the inability to calculate the score, but its monitoring data is complete, the system can still locate the sepsis onset time through a combination of signals such as persistent hypotension and respiratory distress. This avoids insufficient training sample size due to missing labels while maintaining sensitivity and interpretability for real organ dysfunction events.

[0030] In some examples, the process of constructing a septic shock assessment observation window based on determining the time of sepsis onset, and determining the time of bloodstream infection-related septic shock onset within the septic shock assessment observation window based on a combination of factors including adequacy of fluid resuscitation, hemodynamic support medication, and abnormal tissue perfusion indicators, includes: Within the observation window for determining septic shock, the hemodynamic support medication record is detected to obtain the vasoactive drug initiation time, and the vasoactive drug initiation time is used as a candidate shock time point. For the candidate shock time points, determine whether the total amount of fluid resuscitation within the preset resuscitation assessment time period reaches a preset threshold to determine the adequacy of fluid resuscitation; Determine whether the hemodynamic parameters near the candidate shock time point meet the low perfusion threshold condition, and determine whether the lactate level near the candidate shock time point exceeds the preset lactate threshold. When the conditions for the use of vasoactive drugs, adequate fluid resuscitation, and hypoperfusion and lactate abnormalities are met simultaneously, the candidate shock time point is determined as the onset time of bloodstream infection-related septic shock.

[0031] For example, in order to extract the time of occurrence of septic shock from diagnostic texts or subjective records and transform it into an automatically locatable event time point, the system constructs a septic shock judgment observation window after determining the time of sepsis onset, and detects hemodynamic support medication records within this window to obtain the vasoactive drug initiation time as a candidate shock time point. The initiation of vasoactive drugs usually means that the clinical team has determined that the patient has a circulatory support requirement that is difficult to maintain through general treatment, which is one of the most operational structured time markers in the process of shock evolution. In practice, the system uses the sepsis onset time as the starting point and extends it forward by a preset duration to form a shock assessment observation window. Within this window, it analyzes medication orders and execution records, identifies the type of vasoactive drug, and extracts its first start time. Vasoactive drugs include at least norepinephrine, dopamine, epinephrine, and vasopressin. The system prioritizes the infusion pump start time or nursing execution time as the start time to reduce the premature bias caused by using only the order inception time. At the same time, it sets a duration filter for records of short-term trials or rapid discontinuation, such as requiring continuous infusion for more than a preset minimum duration or dose increase within a preset time period, to avoid misjudging short-term corrective medication as shock. When multiple vasoactive drugs exist or there are multiple start-stop cycles, the system can determine the candidate shock time point based on the start time that first meets the continuity condition, and record the corresponding drug name, initial dose, route, and duration as traceability fields. This provides a stable candidate time point, allowing subsequent assessments of resuscitation adequacy and perfusion abnormalities to be conducted around the same moment, thereby reducing the starting point drift caused by inconsistent durations of hypotension in different cases. For example, if a patient initiates norepinephrine several hours after a sepsis attack due to a sustained decrease in mean arterial pressure, the candidate shock time point can be accurately placed at the pump initiation moment, facilitating further verification of whether conditions such as adequate resuscitation and elevated lactate are met.

[0032] For example, to distinguish between transient hypotension caused by insufficient resuscitation and shock that still requires vasoactive support after resuscitation, the system determines whether the total amount of fluid resuscitation within a preset resuscitation assessment period reaches a preset threshold for candidate shock time points to determine the adequacy of fluid resuscitation. The determination of septic shock emphasizes that there is still insufficient circulatory perfusion after volume resuscitation. In electronic medical records, fluid resuscitation is usually scattered in medical orders, execution and intake / output records, and needs to be aggregated and calculated near candidate time points to form comparable resuscitation evidence. In practice, the system uses the candidate shock time point as the center or a preset time period before the candidate shock time point as the resuscitation assessment time period. The resuscitation assessment time period can be configured to a range of several hours and can be set according to the emergency resuscitation pathway. The system extracts the implemented volume of crystalloid fluids, colloid fluids, and other volume expansion fluids from fluid resuscitation orders and execution records, infusion pump records, and nursing intake and output records. It also removes duplicate records within the same time period and merges them. For orders without a clear implemented volume, it infers the volume based on the execution mark and routine fluid preparation rules and writes it into the traceability field. The system converts the implemented volume of various resuscitation fluids within the assessment time period into standard resuscitation volume according to preset rules and sums them up. The conversion rules can include the equivalent volume expansion coefficient of different fluid types in the calculation, or normalize the total volume according to weight parameters to adapt to patients of different body types. When the total resuscitation volume reaches a preset threshold or meets the configurable rules such as continuous fluid resuscitation at a certain rate, the adequacy of fluid resuscitation is determined, and the calculated total resuscitation volume, the start and end of the assessment time period, the fluid composition, and the data source are recorded as evidence. This reduces the probability of mislabeling patients in the early resuscitation stage as being in shock and makes the labeling more consistent with clinical definitions. For example, if a patient receives only a small amount of maintenance fluid before the candidate time point but starts taking a small dose of vasopressor due to transient hypotension, the resuscitation adequacy will not be established and shock labeling will not be triggered, thus reducing mislabeling. Conversely, if a patient has completed multiple rapid fluid resuscitations before the candidate time point and the total amount has reached the threshold, but still requires continuous vasopressor support, then the resuscitation adequacy is established, providing key support for shock determination.

[0033] For example, in order to simultaneously include the need for circulatory support and the presence of tissue insufficiency in the judgment and exclude transient blood pressure drops caused only by sedation, anesthesia or other non-perfusion reasons, the system judges whether hemodynamic parameters meet the low perfusion threshold conditions near the candidate shock time point and whether lactate level exceeds the preset lactate threshold. Tissue insufficiency is often reflected by persistent hypotension or difficulty in maintaining mean arterial pressure, and further reflects perfusion and metabolic imbalance at the cellular level through elevated lactate. The combination of the two can improve the specificity and interpretability of shock judgment. In practice, the system defines a hemodynamic verification time neighborhood before and after the candidate shock time point. This neighborhood can be configured to range from tens of minutes to several hours. It extracts mean arterial pressure, systolic blood pressure, diastolic blood pressure, heart rate, etc. from vital sign monitoring data to form a continuous sequence. The system performs quality control on the sequence, eliminating obvious measurement noise and requiring that it meet the persistence condition. For example, the mean arterial pressure must be below the low perfusion threshold for more than a preset minimum duration or repeat at multiple sampling points to avoid single-point abnormalities. At the same time, the system extracts the lactate sampling result that is closest in time to the candidate time point from the laboratory test data, and prioritizes the sampling time rather than the reporting time for alignment. If there are multiple lactate results, the earliest result after the candidate time point and within the allowable time difference or the result closest before and after the candidate time point is selected for judgment. If necessary, the validity period of the lactate result is set to avoid using premature historical results. When the hemodynamic indicators meet the low perfusion threshold condition and the lactate level exceeds the preset lactate threshold, the low perfusion and lactate abnormality conditions are determined to be established, and the specific index value, duration, lactate value, and sampling time of the trigger threshold are recorded as evidence. This significantly reduces misjudgments caused by transient blood pressure fluctuations or non-infection-related medications, and makes the labeling more closely reflect the pathological process of tissue hypoperfusion. For example, if a patient has a persistently low mean arterial pressure and elevated lactate near a candidate time point, it indicates that hypoperfusion is real and shock labeling is more reliable. If a patient's blood pressure drops due to transient sedation but lactate does not rise and the drop in blood pressure is not sustained, then the conditions for hypoperfusion and abnormal lactate do not hold, thus avoiding mislabeling.

[0034] For example, to integrate evidence from multiple sources into an automated labeling result that better conforms to the clinical definition of bloodstream infection-related septic shock, the system adopts a combined condition rule in determining candidate shock time points. When the conditions of vasoactive drug use, adequate fluid resuscitation, and hypoperfusion and lactate abnormalities are met simultaneously, the candidate shock time point is determined as the onset time of bloodstream infection-related septic shock. A single condition is easily misjudged due to missing data, noise, or atypical treatment, while the combined condition requires that the treatment behavior be consistent with physiological evidence, which can significantly improve the credibility of shock initiation labeling and make the labeling result have an interpretable chain of evidence. In practice, the system performs conditional assessments for each candidate shock time point: the condition for vasoactive drug use is determined by the duration of vasoactive drug initiation; the condition for adequate fluid resuscitation is determined by the total resuscitation volume reaching a threshold within the resuscitation assessment period; and the conditions for hypoperfusion and lactate abnormalities are determined by hemodynamic parameters consistently below the threshold and lactate exceeding the threshold. When all three conditions are met simultaneously, the system outputs the candidate time point as the shock onset time, along with the corresponding evidence fields. These evidence fields include at least the initiation and duration of vasopressors, the total resuscitation volume and its calculation interval, the duration of hypoperfusion, and the lactate sampling time and result. If multiple candidate time points meet the conditions, the system can use the earliest time point that meets the combined conditions as the shock onset time to reflect the earliest moment when shock enters an identifiable state, while retaining subsequent time points that meet the conditions for disease progression analysis. If lactate evidence is missing due to untimely lactate detection, the system can mark the candidate point as pending confirmation rather than directly labeling it, to avoid excessive mislabeling in the event of missing data. Therefore, this joint judgment strategy can provide a more stable and clinically consistent shock onset time in a noisy real electronic medical record environment, reducing both mislabeling and omissions. For example, if a patient starts norepinephrine after adequate fluid resuscitation and experiences elevated lactate while the mean arterial pressure remains low, the shock onset time can be stably labeled as the time when the vasopressor is started. The generated label is more suitable for risk prediction model training and cross-center validation, and the basis for the label can be clearly explained during review.

[0035] In some examples, the automatic generation of sample labels and corresponding observation and prediction time windows for training or evaluating risk prediction models includes: Using the onset time of bloodstream infection-related septic shock as the target event time point, a time period of preset length is extracted before it as the observation time window, and a time period of preset length is extracted after it as the prediction time window. When the target event time point is detected to have occurred or there are labeling conditions that conflict with the target event time point within the observation time window, the sample is removed or the observation time window and prediction time window are redefined for the sample. The sample, along with its corresponding suspected infection time anchor, sepsis onset time, bloodstream infection-related septic shock onset time, the observation time window, and the prediction time window, are written into a structured annotation result table for use in training, validating, or deploying a bloodstream infection-related septic shock risk prediction model.

[0036] For example, to ensure that the range of input data and the semantics of output labels for the risk prediction task remain consistent across different patients and to avoid information leakage during the model training phase, the system uses the onset time of bloodstream infection-related septic shock as the target event time point, extracts a preset length of time period before the event as the observation time window, and extracts a preset length of time period after the event as the prediction time window. The samples of the risk prediction model should clearly define the historical information that the model can only see and the future time range that the model needs to predict. Furthermore, the observation time window must be located entirely before the occurrence of the target event, so that the patterns learned by the model come from the precursors of the event rather than information after the event. In practice, the system reads the preset risk prediction task configuration, which includes at least the observation time window length, the prediction time window length, the positioning method of the time window relative to the target event time point, and the time accuracy requirements. The system determines the time window boundary with the target event time point as the center, extracts the start and end times of the observation time window forward and the start and end times of the prediction time window backward, and maps the time window onto the event table time axis to extract the corresponding vital sign sequence, test sequence, and treatment sequence features. During the extraction process, the system prioritizes using the sampling time or monitoring collection time as the feature timestamp and aligns the test report delays to avoid mistaking the report time as the clinical state occurrence time. For cases where the data sampling frequency within the window is inconsistent, the system can divide the window into several fixed-length time segments and calculate statistics and trend quantities within each segment to obtain a stable input representation independent of the window length. This achieves standardization and comparability in sample construction. For example, when the task is configured to use continuous observation data after admission to predict whether shock will occur in the next six hours, all samples will construct the observation and prediction range with their respective shock onset time as a reference, so that each sample in the model training set expresses the same type of lead prediction problem, thereby improving the consistency of model evaluation and reducing performance fluctuations caused by inconsistent window definitions.

[0037] For example, to avoid mistakenly including samples that have already entered a state of shock or contain evidence of shock after it has occurred in the pre-prediction task, the system checks whether the target event time point has occurred or whether there are labeling conditions that conflict with the target event time point within the observation time window after generating the observation time window and prediction time window. If a conflict occurs, the sample is removed or the observation time window and prediction time window are redefined. If the observation time window contains strong prompting information such as evidence of joint judgment of shock, continued use of vasopressors, and significant increase in lactate, the model will generate artificial performance by using post-event information during training, which cannot be reproduced during deployment, thus leading to distortion of the warning threshold and clinical unreliability. In practice, the system performs conflict detection on the observation time window. Conflict conditions include at least: the occurrence of a time point within the observation time window that meets the joint criteria for shock determination; the initiation of vasoactive drugs within the observation time window that lasts for a preset duration and is accompanied by evidence of hypoperfusion; or the emergence of another set of evidence chains earlier than the predetermined shock onset time, causing the target event time point to no longer be the earliest time point to meet the criteria. The system can also detect whether there are strong indicative events at the end of the observation time window that highly overlap with the target event time point. For example, if lactate sampling time is at the end of the observation window and the value is extremely high, accompanied by persistent hypotension, it indicates that shock may have occurred but the recording is delayed. When the conflict condition is met, the system adopts a processing strategy according to the configuration: if the conflict means that the sample does not meet the task definition for predicting future occurrences, the sample is removed and the reason for removal is recorded; if the conflict only indicates that the window boundary is unreasonable but can still meet the task definition through adjustment, the observation time window and the prediction time window are redefined. For example, the observation time window is moved forward as a whole or the length of the observation time window is shortened to ensure that the observation window does not contain evidence of the event occurring, while keeping the length of the prediction time window unchanged and recalculating the labels. This effectively prevents training data leakage and improves the temporal consistency of sample labels. For example, if a patient has already started vasopressors and has elevated lactate levels within the observation window before the shock onset time labeled by the system, failure to remove or adjust this information will cause the model to learn obvious posterior information such as vasopressor initiation. Checking and processing this information ensures that the model training is closer to real clinical early warning scenarios, thereby improving the transferability and reliability of performance after deployment.

[0038] For example, to enable the automatic annotation results to be directly consumed by the modeling process and to support cross-center review and traceability, the system writes the sample and its corresponding suspected infection time anchor, sepsis onset time, bloodstream infection-related septic shock onset time, observation time window, and prediction time window into a structured annotation result table. This solidifies the scattered event chain derivation results into a unified structured output, enabling consistent call to sample extraction and feature generation during the training phase, reproducible evaluation by window playback during the validation phase, and real-time sliding window prediction and label definition consistent control monitoring during the deployment phase. In practice, the system generates a unique identifier for each sample and writes it into a labeling results table. This table includes at least the patient identifier, sample generation time, suspected infection time anchor and its matching evidence identifier, sepsis onset time and its triggering evidence summary, shock onset time and its combined conditional evidence summary, observation time window start and end, prediction time window start and end, label value, and quality control fields. Quality control fields may include window completeness, key indicator missing rate, whether window relocation occurred, whether it is a sample to be confirmed, and the code for the reason for removal, to support subsequent sample screening and sensitivity analysis. The system can also establish a link between the labeling results table and the record identifiers in the original event table, enabling rapid location of specific medication records, sampling records, lactate results, and blood pressure sequence fragments that trigger the judgment during model review. This structured results table significantly improves data governance and modeling efficiency. Training and evaluation personnel do not need to repeatedly parse complex original medical record data; they can directly extract samples and construct features based on the results table. Simultaneously, clinical reviewers can quickly check why a sample was labeled as a high-risk event based on the evidence fields, thereby reducing communication costs in cross-center collaboration and improving the consistency of label definitions during model development and deployment.

[0039] In some cases, considering that in real-world data, the culture sampling time is not always equivalent to the infection onset time, the following reasons apply. This is because different sampling sites reflect different pathogen loads. For catheter-related bloodstream infections, catheter-collected blood often has a higher pathogen load, making it easier to report a positive culture earlier; peripheral blood may report a positive culture later or not at all. Different culture bottle types have different growth characteristics: aerobic, anaerobic, and fungal bottles have different detection thresholds and growth curves. If multiple bottle types from the same sampling show consistency in time, it is more likely to reflect true bacteremia; if only one bottle type is late to positive and lacks consistency, it is more likely to be a contaminated or low-load state. The positive time interval can serve as indirect evidence between the bacterial load in vivo and the sampling time. Under the same laboratory system, a shorter time interval from sampling to reporting a positive result usually corresponds to a higher initial bacterial load or is closer to the peak of bacteremia. This makes the positive time interval useful for constraining the anchor point from being erroneously shifted. Therefore, in some examples, the identification of suspected infection events and determination of the suspected infection time anchor point also includes: Acquire information about the culture bottle type, culture-positive indication time, and culture sampling site related to the target patient's blood culture; determine the positive time interval from the culture sampling time to the culture-positive indication time based on the culture-positive indication time information, and calculate a culture growth consistency score based on the culture bottle type information and the positive time interval; determine whether the culture sampling site includes peripheral blood sampling sites and blood sampling sites related to vascular catheters, and calculate a site difference score based on the difference in the positive time interval corresponding to different sampling sites; If the culture growth consistency score meets the preset conditions and the site difference score meets the preset conditions, the culture sampling time of the earlier sampling site corresponding to the site difference score is determined as the corrected suspected infection time anchor point, or the suspected infection time anchor point is backtracked to a preset backtracking time earlier than the culture sampling time to obtain the corrected suspected infection time anchor point, and the subsequent automatic labeling process is entered based on the corrected suspected infection time anchor point.

[0040] For example, the difference between the time required for a positive culture report and the sampling site can be transformed into calculable evidence to determine whether a suspected infection anchor should remain in situ and be traced back, or whether a specific sampling site should be prioritized as a more credible anchor. The sampling time, culture bottle type, positive culture system reporting time, and sampling site identifiers (peripheral vein, central venous catheter, arterial catheter sampling, catheter tip correlation, etc.) for each blood culture record are extracted from the testing system. Multiple bottle records under the same medical order are aggregated to form a single sampling set, avoiding the mixing of cultures from different time points into the same evidence. For each bottle in the single sampling set, the time interval from sampling time to positive report time is calculated. Consistency is assessed within this set; for example, whether the positive time intervals of multiple bottles are concentrated within a reasonable range, whether aerobic and anaerobic bottles show a similar positive report sequence, and whether the presence of fungal bottles or special bottle types is consistent with the clinical scenario of fungemia in immunosuppressed patients (fungal bottles reporting positive first or together with other bottles are more credible). The output of this step is not a simple positive or negative result, but a quantitative or grading result of the strength of consistency that can be used for subsequent judgment. If the same sampling includes both peripheral and catheter sites, the difference in the positive time interval between the two types of sites is compared. Generally, catheter-related bloodstream infections are more likely to show a significantly shorter positive time interval at catheter sites than at peripheral sites, and this difference is stable. If peripheral sites report positive first or are more consistent, while catheter sites show isolated, delayed positive results, it suggests possible contamination or non-dominant evidence at the catheter site. Anchor point correction is performed based on evidence of growth consistency and site difference. If consistency is strong and catheter sites report positive significantly earlier, and there is a clinical background of catheter placement, the catheter sampling time is determined as a more credible suspected infection anchor point, or existing anchor points are aligned with the catheter sampling time. If consistency is weak and there is only a single bottle with delayed positive results, and site differences are unstable, the anchor point is not established earlier, and the weight of culture evidence in anchor point determination can be reduced to avoid mistaking contamination as the starting point of infection. If multiple sampling batches exist, the batch that best explains the changes in organ function scores within the subsequent sepsis assessment window is selected as the anchor point source. Using the corrected anchor points as a benchmark, observation windows for sepsis and septic shock are constructed. Then, the onset time and sample slicing are completed according to the claimed system. This elevates the positivity of blood cultures to a higher level of positive evidence quality grading. This reduces the likelihood of treating isolated, delayed positive cases as strong evidence and lowers the probability of mistakenly including contaminated cases in the infection anchor point. Introducing site differences into anchor point selection allows for earlier and more accurate time anchor points for catheter-related infections. For predicting bloodstream infection-related septic shock, setting anchor points too late can cause the observation and prediction windows to shift backward, resulting in truncated or misaligned early warning signals learned by the model.

[0041] In some cases, considering that lactate is one of the key pieces of evidence in shock assessment, the accuracy of lactate values ​​in electronic medical records is significantly affected by pre-analysis processes. The longer the delay between specimen collection and machine testing, the more likely cellular metabolism is to alter lactate readings. Factors such as specimen transport temperature control, anticoagulation treatment, and collection tube type can cause systematic biases. Delays are more likely to occur during peak clinical periods or during cross-floor transfers, and these delays do not reflect patient deterioration but rather workflow overload. Therefore, in some examples, the determination of the onset time of bloodstream infection-related septic shock based on abnormal tissue perfusion indicators within the septic shock assessment observation window also includes: Obtain the specimen collection time, specimen receipt time, and laboratory testing time from the blood gas or lactate test records of the target patient; determine a specimen processing delay index based on the time interval between the specimen collection time and the laboratory testing time, and assign a confidence weight to the lactate level or perform a delay correction on the lactate level to obtain a corrected lactate level based on the specimen processing delay index; when determining the onset time of bloodstream infection-related septic shock, use the combined satisfaction of the corrected lactate level and the low perfusion threshold condition near the candidate shock time point as the basis for judging the tissue perfusion abnormality index, and increase the lactate abnormality judgment threshold or require additional tissue perfusion abnormality evidence to confirm the onset time of bloodstream infection-related septic shock if the confidence weight is lower than a preset threshold.

[0042] For example, the time delay from collection to in-vitro testing can be used as a quality indicator for lactate evidence. Lactate levels can be weighted or numerically corrected for confidence when determining shock, making the consistency between lactate evidence and low perfusion evidence a triggering condition, thus reducing false positives. For each lactate or blood gas record, the specimen collection time, specimen receipt time, and in-vitro testing time should be extracted. If the receipt time is missing, the laboratory receiving time or barcode scanning time can be used instead. The delay from collection to in-vitro testing should be calculated and categorized into high, medium, and low confidence levels according to preset rules. This grading does not require reliance on a single threshold; a piecewise function can be used, so that the longer the delay, the lower the evidence weight. Lactate evidence can be weighted without changing the lactate value, but it should be treated as low-weight evidence when determining shock, requiring stronger evidence of other perfusion abnormalities to trigger. Correction can also be performed, adjusting lactate based on the delay time to obtain corrected lactate, which can then be used in threshold determination. The correction function can be obtained by fitting historical data, making it self-adaptive within the hospital. Near the candidate shock time point, the occurrence of abnormal lactate levels should be temporally consistent with evidence of perfusion abnormalities such as hypotension, decreased urine output, abnormal skin temperature gradient, and prolonged capillary refill. When lactate levels are considered low-confidence, the trigger threshold is raised, for example, requiring the abnormal lactate level to persist across two consecutive tests, or requiring the simultaneous presence of multiple perfusion abnormalities. Thus, by explicitly modeling laboratory process factors, delays in the testing process are avoided from being mistaken for changes in patient condition. This directly reduces false positives based on lactate as triggering evidence, especially during peak emergency room hours or nighttime testing queues.

[0043] In some cases, the onset time of vasoactive drugs is often considered as a candidate time point for shock in automatic labeling. However, in real clinical practice, there are numerous technical medication events, such as tubing obstruction, leakage, or rerouting leading to pressure alarms, requiring nurses to briefly administer vasoactive drugs to maintain blood pressure or test pathways. Alternatively, vasoactive drugs may be briefly used and then quickly discontinued without establishing sustained titration or accompanied by worsening perfusion evidence. These events appear as onset times in medical orders and administration records, but their duration is extremely short, and failure to eliminate them can result in numerous false-positive shock initiations. Therefore, in some examples, the step of detecting hemodynamic support medication records within the septic shock assessment observation window to obtain the onset time of vasoactive drugs and using it as a candidate shock time point also includes: Obtain the infusion pump operation log related to the infusion of the vasoactive drug, which includes pressure curves, alarm event records, and infusion rate change records. If a pressure alarm event or abnormal fluctuation in the pressure curve is detected near the start time of the vasoactive drug, determine whether the vasoactive drug meets the conditions for low dose, short duration, or rapid discontinuation within a preset short period to identify a suspected non-pathological shock drug administration event. When a suspected non-pathological shock drug administration event is identified, remove the vasoactive drug start time corresponding to the suspected non-pathological shock drug administration event from the candidate shock time points, or update the candidate shock time points to the vasoactive drug start time that meets the conditions for continuous titration upregulation. Based on the updated candidate shock time points, continue to perform a joint condition determination of fluid resuscitation adequacy and abnormal tissue perfusion indicators to determine the onset time of the bloodstream infection-related septic shock.

[0044] For example, the infusion pump log can be incorporated into the decision-making process, enabling the system to identify noisy drug administration and shift candidate time points from short-term trials to truly sustained titration increases at decompensated moments. The vasoactive drug initiation time is aligned with the pressure curve, alarm event records, and infusion rate change records in the infusion pump log onto the same timeline. Near the vasoactive drug initiation time, it is assessed for characteristics such as pressure alarms, sudden pressure curve changes, frequent alarms and de-escalations. If these are present, the time period is marked as a tubing stability risk period, serving as a basis for subsequent elimination or downweighting. The infusion duration, maximum infusion rate, presence of rapid discontinuation, and lack of sustained titration increases are checked. If the characteristics of short duration, low dose, and rapid discontinuation are met, and there is no sustained deterioration in arterial pressure, no continuous increase in lactate, and no significant decrease in urine output during the same period, it is determined to be a technically short-term drug administration event, and the candidate shock time point is eliminated. In subsequent time periods, the timing of the onset or accelerated upregulation of vasoactive drugs that meet the criteria for continuous titration upregulation is identified and used as new candidate shock time points. The onset time is then confirmed based on a combination of fluid resuscitation adequacy and perfusion abnormalities. This directly reduces the error of misjudging short-term trials as the onset of shock, a common mistake in high-intensity nursing scenarios that severely damages training data by mislabeling numerous non-shock samples as pre-shock samples, leading to model learning bias.

[0045] In some cases, considering the common problem of future information leakage in automatic labeling, certain key evidence may only become available clinically after the event occurs, such as culture results, imaging reports, and certain test verification results. Retrospective datasets contain timestamps for this evidence, but if the algorithm directly uses information prior to the result or report time, it may inadvertently introduce future information into the current judgment. Such labels may appear accurate in retrospective assessments, but are unreproducible in prospective predictions because the model was trained using information that is not available in advance in reality. Therefore, in some examples, the construction of an observation window based on the suspected infection time anchor and the determination of the onset time of bloodstream infection-related septic shock also includes: The system associates visibility timestamps with various electronic medical record time-series data records of the target patient. These visibility timestamps characterize the earliest time in the clinical process when the corresponding data record can be used for decision-making. For any candidate shock time point or candidate sepsis onset time point, the system limits the execution of corresponding judgment rules to data records whose visibility timestamps are earlier than or equal to the candidate time point, thereby obtaining candidate judgment results under visibility constraints. Counterfactual removal or delay verification is performed on key evidence that triggers the candidate judgment results. Counterfactual removal or delay verification includes removing key evidence while keeping the other evidence unchanged or delaying the visibility timestamp of the key evidence to a preset duration before re-executing the judgment rules. When counterfactual removal or delay verification causes the candidate judgment result to no longer meet the preset conditions, the candidate time point is rejected or the candidate time point is moved to a time point that can stably trigger the judgment rules under visibility constraints, thereby outputting the bloodstream infection-related septic shock onset time and corresponding sample label that meet the visibility constraints.

[0046] For example, the earliest visible time can be defined for different data sources: medical orders and medication records are usually visible in near real-time; vital signs monitoring can be visible in near real-time based on the collection time; test results are visible at the report generation time, not the sampling time; positive cultures and pathogen identification are visible at the time the testing system first indicates a positive result or the report is released; and imaging reports are visible at the time the report is approved. When the system obtains a candidate sepsis onset time or candidate shock time point, it only allows the use of data with a visibility timestamp no later than that candidate time point to execute the judgment rules, obtaining a judgment result under visibility constraints. Identify key evidence that triggers candidate judgments, such as abnormal lactate levels, continuously titrated vasoactive drugs, and organ function score jumps. Counterfactual removal can be performed, temporarily removing the evidence and recalculating whether it can still trigger. Alternatively, counterfactual delay can be performed, artificially postponing the visibility timestamp of the evidence for a certain period and recalculating whether it can still trigger. If it cannot trigger after removal or delay, it indicates that the candidate time point is highly dependent on the evidence, and it is necessary to check whether the evidence was truly visible at that time. When a candidate time point cannot be stably triggered under visibility constraints, that time point is rejected or shifted until a time point that can be triggered solely by the visible evidence at that time is found, which is then output as the final onset time. This directly eliminates the labeling bias inherent in retrospective, advance-knowledge-of-results models, making the automatic labeling results closer to the available labels for prospective scenarios, thereby improving the consistency of the risk prediction model in real-world deployment environments.

[0047] In some cases, the key to the onset of shock lies not in the first dose, but in the decompensation transition that requires sustained hemodynamic support and escalating doses. In real-world scenarios, multiple vasoactive drugs may be used concurrently, sequentially or cross-titrated. The onset time of a single drug is often influenced by the procedure and cannot represent the pathophysiological transition. Mapping multiple drugs to an equivalent dose sequence and identifying the inflection point from the stable phase to the sustained upregulation phase on this equivalent dose sequence is closer to the moment of decompensation. This inflection point must be verified by evidence of resuscitation adequacy and perfusion abnormalities to avoid misdiagnosing prophylactic medication as shock. Therefore, in some examples, the determination of the onset time of bloodstream infection-related septic shock based on hemodynamic support medication within the septic shock assessment observation window also includes: The system acquires infusion records of multiple vasoactive drugs for the target patient within the observation window for septic shock assessment, and converts the infusion rates of each vasoactive drug into a unified equivalent dose-time series based on a preset equivalent dose mapping relationship. It then performs change point detection or piecewise fitting on the equivalent dose-time series to obtain the dose inflection point time characterizing the transition from a stable phase to a continuously increasing phase of the equivalent dose, and determines the dose inflection point time as a candidate shock time point or uses it to update the candidate shock time point. Near the dose inflection point time, it determines whether the adequacy of fluid resuscitation and abnormal tissue perfusion indicators meet preset conditions. When the hemodynamic support medication conditions, the adequacy of fluid resuscitation conditions, and the abnormal tissue perfusion indicator conditions are simultaneously met, the dose inflection point time is determined as the onset time of bloodstream infection-related septic shock.

[0048] For example, the start time, rate adjustment records, and stop time of medications such as norepinephrine, epinephrine, vasopressin, dopamine, or dobutamine can be extracted and unified into a sequence representation with a fixed time granularity. Based on a preset mapping relationship, the rate of each drug is converted into a unified equivalent dose unit, forming an equivalent dose sequence that changes over time. For non-linear equivalence relationships such as vasopressin, piecewise mapping or in-hospital experience conversion tables can be used to ensure clinically acceptable interpretation. Change point detection or piecewise fitting is performed on the equivalent dose sequence to identify the time point where the dose changes from relatively stable or intermittent adjustment to continuous upward adjustment with an increasing rate of increase. The criterion for the inflection point should reflect continuity; for example, the equivalent dose maintains an upward trend for a period of time after the inflection point, and it is not a one-off short-term surge.

[0049] Near the inflection point, assess whether fluid resuscitation has met the preset adequacy criteria and examine hypoperfusion parameters and lactate abnormalities. Only when hemodynamic support enters a sustained upward phase, and evidence of resuscitation adequacy and perfusion abnormalities is met, is the inflection point defined as the time of shock onset. Therefore, the initiation of medication is replaced by the turning point where a structural change in medication demand occurs. This significantly improves the temporal accuracy of the shock onset, allowing the observation window to more accurately cover the prodromal signals before the turning point.

[0050] The above describes the automatic labeling method for predicting the risk of septic shock in the embodiments of this application. The following describes the automatic labeling system for predicting the risk of septic shock in the embodiments of this application.

[0051] Please see Figure 2 One embodiment of the automatic labeling system for predicting the risk of septic shock described in this application may include: The identification unit 201 is used to identify suspected infection events and determine suspected infection time anchors based on the medication order data and pathogen data of the target patient. The identification of suspected infection events includes detecting the matching results of antimicrobial drug administration records and culture sampling records under a preset time constraint relationship. The determination unit 202 is used to construct a sepsis determination observation window and a septic shock determination observation window based on the suspected infection time anchor point, so as to determine the onset time of bloodstream infection-related septic shock. The automatic labeling unit 203 is used to automatically generate sample labels and corresponding observation time windows and prediction time windows for training or evaluating risk prediction models based on the onset time of bloodstream infection-related septic shock and the preset risk prediction task configuration. The sample labels are used to characterize whether bloodstream infection-related septic shock occurs within the prediction time window.

[0052] above Figure 2 The automatic labeling system for predicting the risk of septic shock in this application embodiment has been described from the perspective of modular functional entities. The automatic labeling system for predicting the risk of septic shock in this application embodiment will now be described in detail from the perspective of hardware processing. Please refer to [link to relevant documentation]. Figure 3 One embodiment of the automatic labeling system 300 for predicting the risk of septic shock in this application includes: The system includes an input device 301, an output device 302, a processor 303, and a memory 304, wherein the number of processors 303 can be one or more. Figure 3 Taking a processor 303 as an example. In some embodiments of this application, the input device 301, output device 302, processor 303, and memory 304 can be connected via a bus or other means, wherein... Figure 3 Taking the example of a connection between China and Israel via a bus.

[0053] Specifically, the processor 303 executes the above steps by calling the operation instructions stored in the memory 304.

[0054] By calling the operation instructions stored in memory 304, processor 303 is also used to execute... Figure 1 Any of the methods in the corresponding embodiments.

[0055] Please see Figure 4 , Figure 4 A schematic diagram illustrating an embodiment of the electronic device provided in this application.

[0056] like Figure 4As shown, this application provides an electronic device, including a memory 304, a processor 303, and a computer program 411 stored in the memory 304 and executable on the processor 303. When the processor 303 executes the computer program 411, it performs the above steps.

[0057] In practical implementation, when processor 303 executes computer program 411, it can achieve... Figure 1 Any of the corresponding implementation methods in the embodiments.

[0058] Since the electronic device described in this embodiment is the device used to implement an automatic labeling system for predicting the risk of septic shock in the embodiments of this application, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this embodiment based on the method described in the embodiments of this application. Therefore, how the electronic device implements the method in the embodiments of this application will not be described in detail here. Any device used by those skilled in the art to implement the method in the embodiments of this application is within the scope of protection of this application.

[0059] Please see Figure 5 , Figure 5 This is a schematic diagram illustrating an embodiment of a computer-readable storage medium provided in this application.

[0060] like Figure 5 As shown, this embodiment provides a computer-readable storage medium 500 on which a computer program 511 is stored, which performs the above steps when executed by a processor.

[0061] By calling the operation instructions stored in memory 304, processor 303 is also used to execute... Figure 1 Any of the methods in the corresponding embodiments.

[0062] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An automatic annotation method for predicting the risk of septic shock, characterized in that, include: Based on the medication order data and pathogen data of the target patients, suspected infection events are identified and suspected infection time anchors are determined. The identification of suspected infection events includes detecting the matching results of antimicrobial drug administration records and culture sampling records under a preset time constraint relationship. Based on the suspected infection time anchor point, a sepsis determination observation window and a septic shock determination observation window are constructed to determine the onset time of bloodstream infection-related septic shock. Based on the onset time of bloodstream infection-related septic shock and the preset risk prediction task configuration, sample labels and corresponding observation time windows and prediction time windows are automatically generated for training or evaluating the risk prediction model. The sample labels are used to characterize whether bloodstream infection-related septic shock occurs within the prediction time window.

2. The method according to claim 1, characterized in that, The process of constructing sepsis and septic shock assessment observation windows based on the suspected infection time anchor point is used to determine the onset time of bloodstream infection-related septic shock, including: Based on the suspected infection time anchor point, a sepsis determination observation window is constructed, and the sepsis onset time is determined within the sepsis determination observation window based on changes in organ function scores or alternative criteria. Based on determining the time of sepsis onset, a septic shock assessment observation window is constructed, and within the septic shock assessment observation window, the onset time of bloodstream infection-related septic shock is determined based on the combined conditions of adequate fluid resuscitation, hemodynamic support medication, and abnormal tissue perfusion indicators.

3. The method according to claim 1, characterized in that, The identification of suspected infection events and determination of suspected infection time anchor points include: If the target patient is detected to have a first time of antibiotic administration, determine whether there is a culture sampling time within a first preset time period before the first time of antibiotic administration to determine a first time constraint relationship; If a culture sampling time is detected for the target patient, it is determined whether there is a first administration time of antimicrobial drug within a second preset time period after the culture sampling time, so as to determine the second time constraint relationship. The time of first administration of the antimicrobial drug or the time of culture sampling that satisfies the time constraint relationship is determined as the time anchor point of the suspected infection, and the target patient is marked as a suspected infected patient to enter the subsequent automatic labeling process.

4. The method according to claim 2, characterized in that, The process of constructing a sepsis assessment observation window based on the suspected infection time anchor point, and determining the sepsis onset time within the sepsis assessment observation window based on changes in organ function scores or alternative criteria, includes: The sepsis determination observation window is formed by extending a preset time forward and backward from the suspected infection time anchor point; Within the sepsis assessment observation window, the sepsis onset time is determined based on whether the increase in the target patient's organ function score relative to the baseline score reaches a preset threshold. The baseline score is the lowest or median score within a preset baseline time period prior to the sepsis assessment observation window. In cases where organ function scores are missing or cannot be calculated, alternative criteria are used to determine the timing of sepsis onset. These alternative criteria include rules based on combinations of abnormal vital signs or simplified scoring.

5. The method according to claim 2, characterized in that, Based on determining the time of sepsis onset, a septic shock assessment observation window is constructed. Within this window, the onset time of bloodstream infection-related septic shock is determined based on a combination of factors including adequacy of fluid resuscitation, hemodynamic support medication, and abnormal tissue perfusion indicators. This includes: Within the observation window for determining septic shock, the hemodynamic support medication record is detected to obtain the vasoactive drug initiation time, and the vasoactive drug initiation time is used as a candidate shock time point. For the candidate shock time points, determine whether the total amount of fluid resuscitation within the preset resuscitation assessment time period reaches a preset threshold to determine the adequacy of fluid resuscitation; Determine whether the hemodynamic parameters near the candidate shock time point meet the low perfusion threshold condition, and determine whether the lactate level near the candidate shock time point exceeds the preset lactate threshold. When the conditions for the use of vasoactive drugs, adequate fluid resuscitation, and hypoperfusion and lactate abnormalities are met simultaneously, the candidate shock time point is determined as the onset time of bloodstream infection-related septic shock.

6. The method according to claim 1, characterized in that, The automatic generation of sample labels and corresponding observation and prediction time windows for training or evaluating risk prediction models includes: Using the onset time of bloodstream infection-related septic shock as the target event time point, a time period of preset length is extracted before it as the observation time window, and a time period of preset length is extracted after it as the prediction time window. When the target event time point is detected to have occurred or there are labeling conditions that conflict with the target event time point within the observation time window, the sample is removed or the observation time window and prediction time window are redefined for the sample. The sample, along with its corresponding suspected infection time anchor, sepsis onset time, bloodstream infection-related septic shock onset time, the observation time window, and the prediction time window, are written into a structured annotation result table for use in training, validating, or deploying a bloodstream infection-related septic shock risk prediction model.

7. The method according to claim 1, characterized in that, The identification of suspected infection events and determination of suspected infection time anchor points include: Obtain information on the target patient's blood culture-related information, including culture bottle type, time of positive culture indication, and culture sampling site. Based on the culture positive indication time information, the positive time interval from the culture sampling time to the culture positive indication time is determined, and the culture growth consistency score is calculated based on the culture bottle type information and the positive time interval. Determine whether the culture sampling sites include peripheral blood collection sites and blood collection sites related to vascular catheters, and calculate the site difference score based on the difference in the positive time intervals corresponding to different sampling sites; If the culture growth consistency score meets the preset conditions and the site difference score meets the preset conditions, the culture sampling time of the earlier sampling site corresponding to the site difference score is determined as the corrected suspected infection time anchor point, or the suspected infection time anchor point is backtracked to a preset backtracking time earlier than the culture sampling time to obtain the corrected suspected infection time anchor point, and the subsequent automatic labeling process is entered based on the corrected suspected infection time anchor point.

8. An automatic labeling system for predicting the risk of septic shock, characterized in that, include: The identification unit is used to identify suspected infection events and determine the time anchor of suspected infection based on the medication order data and pathogen data of the target patient. The identification of suspected infection events includes detecting the matching results of antimicrobial drug administration records and culture sampling records under a preset time constraint relationship. The determination unit is used to construct a sepsis determination observation window and a septic shock determination observation window based on the suspected infection time anchor point, so as to determine the onset time of bloodstream infection-related septic shock. An automatic labeling unit is used to automatically generate sample labels and corresponding observation time windows and prediction time windows for training or evaluating risk prediction models based on the onset time of bloodstream infection-related septic shock and the preset risk prediction task configuration. The sample labels are used to characterize whether bloodstream infection-related septic shock occurs within the prediction time window.

9. An electronic device, characterized in that, The electronic device includes at least one processor and at least one memory connected to the processor, wherein the processor is configured to call program instructions in the memory to execute the automatic labeling method for predicting the risk of septic shock as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the storage medium to perform an automatic labeling method for predicting the risk of septic shock as described in any one of claims 1 to 7.

Images