Sepsis patient blood stream infection risk auxiliary screening method and system based on neural network

Abstract

The application discloses a kind of based on neural network's sepsis patient blood stream infection risk auxiliary screening method and system, it is related to auxiliary medical technical field.The application uses patient multi-source time series data modeling, after sample equalization processing and multiple subset optimization construction training set, can alleviate sample imbalance problem, improve dataset representativeness;Fusion attention mechanism neural network can accurately mine time series characteristic key information, improve blood stream infection risk probability prediction accuracy;Through stratified threshold division risk grade, realize stratified screening, again in combination with the secondary check correction determination result of interpolation uncertainty and time series trend conflict coefficient, reduce the misjudgment of data missing, time series anomaly brings, finally output screening result with credibility mark, can provide quantitative, reliable sepsis blood stream infection auxiliary judgment basis for clinician, help early accurate identification high-risk patient, improve screening objectivity and clinical practicability.

Description

Technical Field

[0001] This invention relates to the field of assistive medical technology, and more specifically to a method and system for assistive screening of bloodstream infection risk in sepsis patients based on neural networks. Background Technology

[0002] Sepsis is a critical illness with high mortality and morbidity in the field of critical care medicine. Clinically, sepsis can be induced by severe local infection or complicated by bloodstream infection. The anti-infective treatment course, medication regimen, clinical intervention methods, and prognostic risks differ greatly between the two types of sepsis. Rapid and accurate identification of whether a sepsis patient has bloodstream infection is key to achieving precision diagnosis and treatment, reducing patient mortality, and is of great significance in avoiding missed or misdiagnosed cases in clinical practice.

[0003] Current clinical screening methods for bloodstream infections in sepsis patients have significant technical limitations. The gold standard is blood culture, but its testing cycle is long, ranging from 24 to 72 hours, which is unsuitable for the early treatment window of sepsis. Furthermore, factors such as preoperative antibiotic use and low pathogen load can easily lead to false negatives, making early and accurate diagnosis difficult. Traditional scoring systems such as SOFA and qSOFA can only assess the degree of organ damage in sepsis and cannot specifically identify bloodstream infections. Physicians rely solely on single static indicators and clinical experience, making it difficult to capture the hidden pathological features of dynamic changes in vital signs, resulting in high rates of missed and false positives.

[0004] Existing AI models mostly focus on the initial identification of sepsis, without being optimized for the bloodstream infection risk stratification scenario of patients diagnosed with sepsis. They also simply splice together single time-series vital sign data, failing to deeply mine the dynamic correlation features of multi-dimensional vital signs such as body temperature, heart rate, and blood pressure. The feature mining depth is insufficient, resulting in low early warning accuracy and high false positive and false negative rates, making it difficult to adapt to complex and real clinical scenarios.

[0005] In summary, current clinical practice lacks an intelligent method that can accurately and in real time extract multi-dimensional vital sign information from patients and rapidly screen for the risk of bloodstream infection in sepsis patients, thus failing to effectively address the pain points of missed and false diagnoses in manual screening. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for assisting in screening the risk of bloodstream infection in sepsis patients based on neural networks, in order to solve the problems existing in the background art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A neural network-based auxiliary screening method for bloodstream infection risk in sepsis patients includes the following steps: Acquire multi-source raw data of sepsis patients within a preset observation time window, label positive and negative samples of the multi-source raw data, and preprocess to construct a standardized time-series feature matrix; To balance the negative samples, all positive samples are combined with balanced negative samples to generate multiple training subsets, and the optimal training dataset is jointly selected using preset evaluation indicators. Construct a neural network model that integrates attention mechanisms, input standardized time-series data from the optimal training dataset, and output the probability of sepsis patients complicated with bloodstream infection; The risk probability output by the neural network is compared with the preset hierarchical threshold to classify multiple bloodstream infection risk levels; the risk judgment results are then verified and corrected by combining the interpolation uncertainty index and the temporal trend conflict coefficient, and bloodstream infection screening results with confidence labels are output.

[0008] Optionally, the multi-source raw data includes continuous time-series vital sign data, basic medical record features, inflammatory test indicators, and antimicrobial drug treatment records; a cross-system field mapping table is established to complete the unified mapping of synonymous fields, unit normalization conversion, and binding of timestamps to construct a unified time-series feature set; dual-constraint anomaly detection is performed on the unified time-series feature set: strong outliers are marked and missing values ​​are set based on clinical physiological thresholds, and outliers within the clinical range are judged as suspected missing values ​​based on the fluctuation amplitude of adjacent time series and treatment events; a feature-specific iterative imputation model is used to iteratively complete missing features, and the feature set to be imputed and the auxiliary feature set are divided according to the missing rate. The imputation model is trained and integrated feature set feature by feature and updated cyclically until the imputation difference is lower than the convergence threshold, and a complete standardized time-series feature matrix is ​​output.

[0009] Optionally, the missing value imputation process specifically includes: using a feature set-based iterative imputation model to iteratively complete the missing features; dividing the feature set to be imputed and the auxiliary feature set according to the missing rate; training the integrated imputation model feature by feature and updating iteratively until the imputation difference is lower than the convergence threshold; adding physiological consistency constraints during the imputation process; if the imputed value violates the physiological coupling relationship, performing projection correction and then backfilling the matrix.

[0010] Optionally, the negative sample balance processing includes: performing feature space clustering to divide all negative samples into multiple sample clusters, selecting representative samples from the center of each cluster to construct a basic undersampled negative subset; calculating the model influence score of each negative sample based on gradient loss, selecting negative samples with high contribution at the boundary to form a core negative sample set, and merging the basic undersampled negative subset and the core negative sample set to obtain a balanced negative sample set.

[0011] Optionally, the neural network model includes at least a temporal feature extraction layer and an attention-gated classification layer. The temporal feature extraction layer is used to extract temporal change features and pathological coupling features. The attention-gated classification layer uses pathological coupling features as a gating factor to perform weight modulation on temporal signs and characteristics, thereby suppressing noise fluctuation interference and amplifying continuous abnormal trend features.

[0012] Optionally, the calculation method for the pathological coupling features includes: Hemodynamic imbalance coupling index: based on logarithmically compressed weighted average of the rates of change of blood pressure, heart rate and blood oxygen over time, characterizing the degree of persistent circulatory disturbance; Systemic inflammatory burden index: constructed by fusion of temporal changes in procalcitonin, CRP, and leukocyte counts; Organ damage imbalance index: obtained by weighted fusion of time-series SOFA sub-items, creatinine, and bilirubin time-series characteristics logarithms and square roots; The three types of pathological coupling indices work together as attention gating factors to amplify the weight of continuous abnormal temporal signs and weaken the fluctuation interference caused by instantaneous monitoring noise.

[0013] Optionally, the correction of missed and false detections includes: pre-storing the uncertainty index of each feature imputation, identifying high-risk samples by the neural network, and statistically analyzing the proportion of high-uncertainty imputation features within the sample; if the proportion exceeds a preset threshold, automatically downgrading the risk level by one level and marking "data missing interference, blood culture recommended"; for low-risk samples identified by the neural network, if vital signs continuously deteriorate for more than a preset duration within the rolling time window, independently generating a missed detection prompt to alert physicians to hidden bloodstream infections, thus compensating for the time-series trend recognition defects of static threshold determination.

[0014] A neural network-based auxiliary screening system for bloodstream infection risk in sepsis patients, comprising: Sepsis Patient Data Processing Module: Used to acquire multi-source raw data of sepsis patients within a preset observation time window, label positive and negative samples of the multi-source raw data, and preprocess to construct a standardized time-series feature matrix; Sample Balance Processing Module: Used to balance negative samples, combining all positive samples with balanced negative samples to generate multiple training subsets, and jointly selecting the optimal training dataset using preset evaluation indicators. Neural network model building module: used to build a neural network model that incorporates an attention mechanism, taking standardized time-series data within the optimal training dataset as input and outputting the probability of sepsis patients having bloodstream infection. Risk Correction Hierarchical Output Module: This module compares the risk probability output by the neural network with a preset hierarchical threshold to classify multiple bloodstream infection risk levels. It combines the interpolation uncertainty index and the time-series trend conflict coefficient to perform secondary verification and correction on the risk judgment results, and outputs bloodstream infection screening results with confidence labels.

[0015] As can be seen from the above technical solution, compared with the prior art, the present invention provides a method and system for auxiliary screening of bloodstream infection risk in sepsis patients based on neural networks. It uses multi-source time-series data of patients for modeling, and after sample balancing and multi-subset optimization to construct a training set, it can alleviate the problem of sample imbalance and improve the representativeness of the dataset. The fusion of attention mechanism neural networks can accurately mine key information of time-series features, improving the accuracy of bloodstream infection risk probability prediction. Layered screening is achieved by dividing risk levels through hierarchical thresholds, and the judgment results are corrected by combining interpolation uncertainty and time-series trend conflict coefficients for secondary verification, reducing misjudgments and missed judgments caused by missing data and time-series anomalies. Finally, the system outputs screening results with confidence labels, providing clinicians with quantitative and reliable auxiliary judgment basis for sepsis bloodstream infection, helping to accurately identify high-risk patients in the early stages, and improving the objectivity and clinical applicability of screening. Attached Figure Description

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

[0017] Figure 1 This is a schematic diagram of the method flow provided by the present invention; Figure 2 This is a schematic diagram of the system structure provided by the present invention. Detailed Implementation

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

[0019] This invention discloses a neural network-based auxiliary screening method for bloodstream infection risk in sepsis patients, such as... Figure 1 As shown, it includes the following steps: S1. Standardized fusion preprocessing of multi-source time series data: Acquire multi-source raw data of sepsis patients within a preset observation time window. The multi-source raw data includes continuous time-series vital sign data, basic medical record features, inflammatory test indicators, and antimicrobial drug treatment records. Establish a cross-system field mapping table to complete the unified mapping of synonymous fields, unit normalization conversion, and binding of timestamps to construct a unified time-series feature set. Perform dual-constraint anomaly detection on the unified time-series feature set: mark strong outliers based on clinical physiological thresholds and co-sign missing values; determine whether outliers within the clinical range are marked as suspected missing values ​​by combining the fluctuation amplitude of adjacent time series with treatment events. Use a feature set-based iterative imputation model to iteratively complete missing features. Divide the feature set to be imputed and the auxiliary feature set according to the missing rate. Train the integrated imputation model feature by feature and update iteratively until the imputation difference is lower than the convergence threshold, and output a complete standardized time-series feature matrix. S2, Sample Class Imbalance Balance Optimization: Based on the standardized temporal feature matrix, positive samples (patients with sepsis complicated by bloodstream infection) and negative samples (patients with sepsis with simple local infection) are distinguished. Feature space clustering is performed on all negative samples to divide them into multiple sample clusters. Representative samples from the center of each cluster are selected to construct a basic undersampled negative subset. The model influence score of each negative sample is calculated based on gradient loss. Negative samples with high contribution at the boundary are selected to form a core negative sample set. The basic undersampled negative subset and the core negative sample set are merged to obtain a balanced negative sample set. All positive samples and balanced negative samples are combined according to multiple candidate positive and negative ratios to generate multiple training subsets. The optimal training dataset is jointly selected by the model's bloodstream infection identification recall rate and imputation consistency index. S3. Temporal deep feature mining neural network modeling: A two-layer neural network model integrating temporal attention gating is constructed, with standardized temporal vital sign sequences from the optimal training dataset as input. The first layer is a temporal feature extraction layer, which uses a sliding rolling time window to extract temporal features of the mean, fluctuation amplitude, and first-order difference rate of change of vital signs, and calculates three types of pathological coupling features: hemodynamic imbalance coupling index, systemic inflammatory load index, and organ damage imbalance index. The second layer is a multi-head attention gating classification layer, which uses pathological coupling features as gating factors to weight and modulate the temporal vital sign feature weights, suppressing short-term monitoring noise fluctuations and amplifying persistent abnormal temporal trend features. The network outputs the risk probability of sepsis patients complicated by bloodstream infection. S4. Bloodstream infection risk stratification output and correction of missed and false positives: The risk probability output by the neural network is compared with a preset stratification threshold to classify bloodstream infection risk levels as low, medium, high, and very high. High-risk warning samples are then re-verified using imputation uncertainty indicators and temporal feature conflict coefficients: if the high-risk determination of a sample relies on high-uncertainty imputation features, the risk level is automatically lowered and a suspicious indication is added; if a low-risk sample shows a continuous deterioration trend in vital signs, a false negative warning is generated; and bloodstream infection screening results with credibility annotations are output to assist physicians in reducing false negatives and false positives.

[0020] Furthermore, the iterative imputation process in S1 specifically includes: dividing the unified temporal feature set into features with high missing rates to be imputed and auxiliary features with low missing rates; for a single feature to be imputed, selecting samples without missing values ​​for that feature to construct a training subset, and training an ensemble regression model / classification model with the auxiliary features as input; backfilling the imputed prediction values ​​into the feature matrix, completing one round of traversal imputation according to the preset feature order, and calculating the global difference between the current round and the previous round of imputation results; terminating iterative imputation when the difference is less than the convergence threshold or the maximum number of iterations is reached; adding physiological consistency constraints during the imputation process, if the imputed values ​​violate the physiological coupling relationship of acid-base, perfusion, and electrolytes, performing projection correction before backfilling the matrix.

[0021] Furthermore, S1 also includes an information leakage shielding step: identifying diagnostic and treatment interventions such as antibiotic administration, fluid resuscitation, and activation of vasoactive drugs within the observation time window as leakage features, and setting an intervention impact time window based on the bloodstream infection label determination time; if the leakage feature recording time falls within the intervention impact time window, performing causal residual replacement processing on the leakage feature to eliminate the label information leakage bias caused by the physician's early intervention, and simultaneously writing the residual feature into the traceability field for retention.

[0022] Furthermore, in S2, the method for calculating the negative sample model influence score based on gradient loss is as follows: a shallow pre-trained neural network is used to make a preliminary prediction on the original standardized time-series feature matrix, and the gradient change magnitude of each negative sample on the model classification loss is calculated as the influence score. The higher the gradient magnitude, the more the sample is at the classification decision boundary, the stronger the differentiation effect on the bloodstream infection identification model boundary, and the more likely it is to be included in the core negative sample set.

[0023] Furthermore, the specific implementation method of positive and negative sample partitioning and balance processing in S2 for sample class imbalance balance optimization is as follows: (1) The sample source and label determination criteria adopt the hospital ICU / emergency department historical clinical database that is dedicated to offline modeling and has been desensitized. All samples in the database are pre-labeled with gold standard binary labels to form a supervised dataset. Among them, positive samples are patient samples diagnosed with sepsis and whose blood cultures have detected pathogens during the same course of the disease, and are uniformly labeled with label 1. Negative samples are patient samples diagnosed with sepsis but whose blood cultures are negative throughout the course of the disease and who only have pneumonia, abdominal abscess, or local infection foci of urinary tract infection, and are uniformly labeled with label 0. Each sample is bound to a corresponding standardized temporal feature matrix. The labels are manually pre-labeled by technicians in combination with blood culture test reports. New patients in the real-time inference stage are not labeled and the sample balancing process is not executed.

[0024] (2) The positive and negative samples are separated and divided by traversing all samples of the standardized time-series feature matrix. Based on the pre-stored labels, the positive set S+ and the negative set S- are divided into mutually exclusive positive sets. All positive sets S+ are completely retained without deletion. Only the negative set S-, which has a much larger number of positive samples and has class imbalance, is balanced and optimized.

[0025] (3) Negative sample stratification and equalization: The first step is to perform K-Means feature space clustering on the negative set S-time-series feature vector after standardization, divide it into multiple feature clusters, select the sample closest to the cluster center of each cluster and merge them to obtain the basic undersampled negative subset S-base, thus preserving the overall distribution diversity of negative samples; The second step is to pre-train a shallow basic classification network using the complete original dataset, calculate the gradient magnitude of the classification loss for each negative sample, and select samples whose gradient magnitude exceeds a preset threshold and are at the classification boundary to form the core negative sample set S-core. The third step is to merge S-base and S-core and remove duplicates to obtain a balanced negative sample set S-balance, which covers both routine sepsis samples and borderline samples that are easily confused.

[0026] (4) Optimal training dataset selection: Fix all positive sets S+, and generate multiple candidate training subsets by combining multiple positive and negative sample ratios of 1:1, 1:2, 1:3, and 1:4 with the balanced negative sample set S-balance. Each candidate subset is used to train the neural network independently, and the recall rate of bloodstream infection identification and the consistency of feature imputation are evaluated simultaneously. After comprehensive scoring, the dataset with the best comprehensive performance ratio is selected as the optimal training dataset.

[0027] (5) Stage Limitation Rules: This sample balance optimization process is only executed during the offline training stage of the model; during the clinical real-time inference stage, only the temporal features of the patients to be screened are input into the trained neural network, and sample clustering, gradient calculation, and sample matching screening operations are not carried out.

[0028] Furthermore, in S3, the pathological coupling index is calculated in the following ways: Hemodynamic imbalance coupling index: based on logarithmic compression weighting of the temporal blood pressure, heart rate, and blood oxygenation rate changes, representing the degree of persistent circulatory disturbance; Systemic inflammatory burden index: constructed by fusing temporal procalcitonin, CRP, and white blood cell temporal change slopes; Organ damage imbalance index: obtained by fusion of temporal SOFA sub-items, creatinine, and bilirubin temporal characteristics logarithmic and square root weighting; The three types of pathological coupling indices together serve as attention gating factors, amplifying the weight of persistent abnormal temporal signs and weakening the fluctuation interference caused by instantaneous monitoring noise.

[0029] Furthermore, the specific methods for correcting missed and false positives in S4 include: pre-storing the uncertainty index of each feature imputation, identifying high-risk samples by the neural network, and statistically analyzing the proportion of high-uncertainty imputation features within the sample; if the proportion exceeds a preset threshold, automatically downgrading the risk level by one level and marking it as "data missing interference, blood culture recommended"; for low-risk samples identified by the neural network, if vital signs continuously deteriorate for more than a preset duration within the rolling time window, an independent missed detection prompt is generated to alert physicians to hidden bloodstream infections, thus compensating for the time-series trend recognition defects of static threshold determination.

[0030] and Figure 1 Corresponding to the method shown, this invention also discloses a neural network-based auxiliary screening system for the risk of bloodstream infection in sepsis patients. Figure 1 The implementation of the method, specifically its structure, is as follows: Figure 2 As shown, it includes: Sepsis Patient Data Processing Module: Used to acquire multi-source raw data of sepsis patients within a preset observation time window, label positive and negative samples of the multi-source raw data, and preprocess to construct a standardized time-series feature matrix; Sample Balance Processing Module: Used to balance negative samples, combining all positive samples with balanced negative samples to generate multiple training subsets, and jointly selecting the optimal training dataset using preset evaluation indicators. Neural network model building module: used to build a neural network model that incorporates an attention mechanism, taking standardized time-series data within the optimal training dataset as input and outputting the probability of sepsis patients having bloodstream infection. Risk Correction Hierarchical Output Module: This module compares the risk probability output by the neural network with a preset hierarchical threshold to classify multiple bloodstream infection risk levels. It combines the interpolation uncertainty index and the time-series trend conflict coefficient to perform secondary verification and correction on the risk judgment results, and outputs bloodstream infection screening results with confidence labels.

[0031] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0032] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for assisting in screening the risk of bloodstream infection in sepsis patients based on neural networks, characterized in that, Includes the following steps: Acquire multi-source raw data of sepsis patients within a preset observation time window, label positive and negative samples of the multi-source raw data, and preprocess to construct a standardized time-series feature matrix; To balance the negative samples, all positive samples are combined with balanced negative samples to generate multiple training subsets, and the optimal training dataset is jointly selected using preset evaluation indicators. Construct a neural network model that integrates attention mechanisms, input standardized time-series data from the optimal training dataset, and output the probability of sepsis patients complicated with bloodstream infection; The risk probability output by the neural network is compared with the preset hierarchical threshold to classify multiple bloodstream infection risk levels; the risk judgment results are then verified and corrected by combining the interpolation uncertainty index and the temporal trend conflict coefficient, and bloodstream infection screening results with confidence labels are output.

2. The method for assisting in screening the risk of bloodstream infection in sepsis patients based on neural networks according to claim 1, characterized in that, The multi-source raw data includes continuous time-series vital signs data, basic medical record features, inflammatory test indicators, and antimicrobial drug treatment records. A cross-system field mapping table is established to complete the unified mapping of synonymous fields, unit normalization conversion, and binding of timestamps, constructing a unified time-series feature set. Dual-constraint anomaly detection is performed on the unified time-series feature set: strong outliers are marked and missing based on clinical physiological thresholds, and outliers within the clinical range are judged as suspected missing values ​​based on adjacent time-series fluctuation amplitude and treatment events. A feature-set iterative imputation model is used to iteratively complete missing features. The feature set to be imputed and the auxiliary feature set are divided according to the missing rate. The imputation model is trained and integrated feature by feature and updated cyclically until the imputation difference is lower than the convergence threshold, and a complete standardized time-series feature matrix is ​​output.

3. The method for assisting in screening the risk of bloodstream infection in sepsis patients based on neural networks according to claim 2, characterized in that, The missing value imputation process specifically includes: using a feature set-based iterative imputation model to iteratively complete missing features; dividing the feature set to be imputed and the auxiliary feature set according to the missing rate; training the integrated imputation model feature by feature and updating iteratively until the imputation difference is lower than the convergence threshold; adding physiological consistency constraints during the imputation process; if the imputed value violates the physiological coupling relationship, performing projection correction and then backfilling the matrix.

4. The method for assisting in screening the risk of bloodstream infection in sepsis patients based on neural networks according to claim 1, characterized in that, The negative sample balance processing specifically includes: performing feature space clustering to divide all negative samples into multiple sample clusters, selecting representative samples from the center of each cluster to construct a basic undersampled negative subset; calculating the model influence score of each negative sample based on gradient loss, selecting negative samples with high contribution at the boundary to form a core negative sample set, and merging the basic undersampled negative subset and the core negative sample set to obtain a balanced negative sample set.

5. The method for assisting in screening the risk of bloodstream infection in sepsis patients based on neural networks according to claim 1, characterized in that, The neural network model includes at least a temporal feature extraction layer and an attention-gated classification layer. The temporal feature extraction layer is used to extract temporal change features and pathological coupling features. The attention-gated classification layer uses pathological coupling features as a gating factor to perform weight modulation on temporal signs and characteristics, thereby suppressing noise fluctuations and amplifying continuous abnormal trend features.

6. The method for assisting in screening the risk of bloodstream infection in sepsis patients based on neural networks according to claim 5, characterized in that, The calculation methods for the pathological coupling features include: Hemodynamic imbalance coupling index: based on logarithmically compressed weighted average of the rates of change of blood pressure, heart rate and blood oxygen over time, characterizing the degree of persistent circulatory disturbance; Systemic inflammatory burden index: constructed by fusion of temporal changes in procalcitonin, CRP, and leukocyte counts; Organ damage imbalance index: obtained by weighted fusion of time-series SOFA sub-items, creatinine, and bilirubin time-series characteristics logarithms and square roots; The three types of pathological coupling indices work together as attention gating factors to amplify the weight of continuous abnormal temporal signs and weaken the fluctuation interference caused by instantaneous monitoring noise.

7. The method for assisting in screening the risk of bloodstream infection in sepsis patients based on neural networks according to claim 1, characterized in that, The specific correction of missed and false positives includes: pre-storing the uncertainty index of each feature imputation, identifying high-risk samples by the neural network, and statistically analyzing the proportion of high-uncertainty imputation features within the sample; if the proportion exceeds the preset threshold, automatically downgrading the risk level by one level and marking "data missing interference, blood culture recommended"; for low-risk samples identified by the neural network, if vital signs continuously deteriorate for more than the preset duration within the rolling time window, an independent missed detection prompt is generated to alert physicians to hidden bloodstream infections, thus compensating for the time-series trend recognition defects of static threshold determination.

8. A neural network-based auxiliary screening system for bloodstream infection risk in sepsis patients, characterized in that, include: Sepsis Patient Data Processing Module: Used to acquire multi-source raw data of sepsis patients within a preset observation time window, label positive and negative samples of the multi-source raw data, and preprocess to construct a standardized time-series feature matrix; Sample Balance Processing Module: Used to balance negative samples, combining all positive samples with balanced negative samples to generate multiple training subsets, and jointly selecting the optimal training dataset using preset evaluation indicators. Neural network model building module: used to build a neural network model that incorporates an attention mechanism, taking standardized time-series data within the optimal training dataset as input and outputting the probability of sepsis patients having bloodstream infection. Risk Correction Hierarchical Output Module: Used to compare the risk probability output by the neural network with a preset hierarchical threshold to classify multiple bloodstream infection risk levels; The risk assessment results are verified and corrected by combining the interpolation uncertainty index and the time series trend conflict coefficient, and the bloodstream infection screening results with confidence label are output.

Images