A non-cardiac universal in-hospital postoperative acute kidney injury early warning detection system

Abstract

The application relates to a non-cardiac universal in-hospital postoperative acute kidney injury early warning detection system, a model and application. The system comprises a risk factor acquisition unit, an early warning analysis unit and an early warning processing unit. The risk factor acquisition unit is connected with a hospital clinical information system and a hospital laboratory information management system to acquire risk factors. The early warning analysis unit performs postoperative acute kidney injury early warning analysis according to the age, operation type, contrast agent, serum creatinine, urine protein, fasting blood glucose, anemia, blood neutrophil and lymphocyte count ratio (NLR) and D-dimer information of the patient in the risk factors acquired by the risk factor acquisition unit, and obtains early warning analysis results. The early warning processing unit sends the early warning analysis results to a hospital resource management system in a manner related to the probability and time of postoperative acute kidney injury of the patient in the early warning analysis results. The application constructs an early warning model, so that high-risk patients can be identified before operation.

Description

Technical Field

[0001] This invention relates to the field of early warning assessment technology for postoperative acute kidney injury, and to an early warning detection system, model and application for non-cardiac general inpatient postoperative acute kidney injury, and in particular to a system, method and device for establishing and applying an early warning model for postoperative acute kidney injury based on deep learning and medical statistics. Background Technology

[0002] The incidence of acute kidney injury (AKI) after major non-cardiac surgery ranges from 6.3% to 13.4%, making it a major cause of increased morbidity and mortality. Furthermore, AKI is associated with increased utilization of healthcare resources, including longer hospital and intensive care unit (ICU) stays and related costs.

[0003] Acute kidney injury (AKI) following non-cardiac surgery is a global problem with no effective treatment. Although clinicians are increasingly aware of the severity of AKI and its pathophysiology, there is still no definitive or effective treatment for the disease. Therefore, providing appropriate preventative measures and early intervention to prevent its occurrence and development are key aspects of AKI management. Numerous studies have attempted to construct prognostic models for AKI patients, primarily in the form of new scoring systems or electronic early warning systems. For example, Chen Yingying et al. enrolled 261 adult patients who developed AKI at Huashan Hospital affiliated with Fudan University in Shanghai from January 2009 to September 2011 (experimental group); and 102 adult patients who developed in-hospital AKI from October 2011 to March 2012 (validation group). Both groups were followed up for 90 days, and the 90-day mortality rate was recorded. Multivariate logistic regression analysis was used to identify independent risk factors for 90-day mortality in AKI patients, and a scoring system was developed. Multivariate logistic regression analysis identified five independent risk factors associated with 90-day prognosis of AKI: number of complications, use of vasoactive substances (dopamine), mechanical ventilation, blood urea nitrogen, and prealbumin. A scoring system was developed for the experimental group based on the sum of risk factor scores: total score ≤4 (low-risk group), 5–10 (intermediate-risk group), 11–16 (high-risk group), and 17–30 (very high-risk group). To explore the application value of an "electronic alarm system" for AKI in high-risk hospital wards, Wu et al. developed an e-alert system using a prospective randomized controlled trial. This system automatically diagnoses AKI based on serum creatinine levels and was implemented in intensive care units and departments focusing on cardiovascular diseases. Data analysis showed that the e-alert system is a reliable tool for diagnosing AKI.

[0004] Acute kidney injury (AKI) is a global health problem. The incidence of AKI worldwide ranges from 0.99% to 12%, and surgery is a definite precipitating factor. In fact, postoperative acute kidney injury (PO-AKI) is associated with a variety of poor short-term and long-term outcomes, including prolonged hospital stays, extended intensive care unit (ICU) stays, and high medical costs. Furthermore, PO-AKI is closely associated with increased mortality, progression of acute kidney disease or chronic kidney disease (CKD), and early graft loss. Patients with mild AKI who have undergone non-cardiac surgery, those recovering before discharge, or those with elevated serum creatinine levels below the diagnostic threshold still have an increased risk of poor prognosis. Currently, there is no effective treatment; therefore, we focus on prevention and early detection in high-risk patients. Serum creatinine, urine output, or biomarkers cannot identify high-risk patients for AKI early. Only a few studies have tested alarm models; however, most of them do not fully meet clinical needs.

[0005] Therefore, this study aims to establish a simple and easily accessible early warning model using big data to rapidly and accurately identify high-risk individuals prone to postoperative acute kidney injury (PO-AKI), reduce its incidence, and thus improve their organ and overall prognosis. Furthermore, this application also aims to evaluate the applicability, sensitivity, and specificity of this early warning model on the validation set.

[0006] CN114049952A discloses a machine learning-based intelligent prediction method for postoperative acute kidney injury (API), comprising: acquiring perioperative data of the patient to be tested; obtaining the PPI prediction result of the patient to be tested based on the perioperative data; wherein, the perioperative data is input into the prediction model to obtain the PPI prediction result of the patient to be tested; the prediction model includes several prediction sub-models, which are constructed using different machine learning networks; the PPI prediction result of the patient to be tested is calculated based on the model weights and prediction values ​​of the trained prediction sub-models; the training of the prediction model includes: determining the data weights of each sample data in the training set on each prediction sub-model, and training each prediction sub-model based on the data weights.

[0007] Machine learning and deep learning have been widely applied in the health sector. However, the lack of interpretability in deep learning remains a significant challenge. To address this issue and improve the interpretability of model variable selection, this application employs medical statistical methods to screen for relevant risk factors, and then utilizes the TabNet deep learning model to optimize the prediction accuracy of the early warning model.

[0008] Although deep learning lacks interpretability, its development momentum in the medical field is increasing. It is still necessary to combine it with medical statistics to help medical staff build a practical and reasonable early warning model to identify high-risk patients for postoperative acute kidney injury (PO-AKI).

[0009] Furthermore, on the one hand, there are differences in understanding among those skilled in the art; on the other hand, the applicant studied a large number of documents and patents when making this invention, but due to space limitations, not all details and contents were listed in detail. However, this does not mean that the present invention does not possess the features of these prior art. On the contrary, the present invention already possesses all the features of the prior art, and the applicant reserves the right to add relevant prior art to the background art. Summary of the Invention

[0010] To provide clinicians with more comprehensive and accurate quantitative values ​​of postoperative acute kidney injury (AKI) risk, existing technologies have developed solutions that use regression models to screen different effective variables. For example, patent document CN110827992A discloses a preoperative prediction method for AKI after hypertension surgery, including: acquiring preoperative clinical data related to AKI occurrence in patients; establishing a nomogram of the probability of AKI occurrence based on the preoperative clinical data and calculating a total risk score; calculating a predicted value of the probability of AKI occurrence in hypertensive patients based on the total risk score; and outputting the predicted value of the probability of AKI occurrence in hypertensive patients. This solution can obtain potentially relevant preoperative clinical data for AKI occurrence through univariate and multivariate logistic regression analysis using different screening modules, thereby assessing the prognostic value of the preoperative clinical data. However, in this solution, the variables related to AKI after screening still need to be analyzed again using a logistic regression model, which is significantly different from the initial warning model in the form of a rating scale in this invention.

[0011] To address the shortcomings of existing technologies, this invention provides a method and apparatus for constructing and applying a generalized early warning model for acute kidney injury after hospitalization in non-cardiac patients, aiming to solve at least one or more technical problems existing in the prior art.

[0012] To achieve the above objectives, this invention provides a general-purpose, non-cardiac postoperative acute kidney injury (APK) early warning detection system. The system includes a risk factor acquisition unit, an early warning analysis unit, and an early warning processing unit. The risk factor acquisition unit connects to the hospital's clinical information system to obtain information on the patient's age, surgical type, contrast agent, and anemia. It also connects to the hospital's laboratory information management system to obtain information on the patient's serum creatinine, urine protein, fasting blood glucose, neutrophil-to-lymphocyte ratio (NLR), and D-dimer. The early warning analysis unit performs an early warning analysis for APK based on the patient's age, surgical type, contrast agent, serum creatinine, urine protein, fasting blood glucose, anemia, and NLR and D-dimer information input from the risk factor acquisition unit, obtaining the early warning analysis results. The early warning processing unit obtains the early warning analysis results from the early warning analysis unit and sends them to the hospital resource management system in a manner correlated with the probability and time of APK in the patient as indicated in the early warning analysis results. This invention utilizes big data to establish a simple and easily accessible early warning model to identify high-risk individuals for postoperative acute kidney injury (PO-AKI), reducing its incidence and thus improving organ and overall prognosis. This invention employs at least nine risk factors to construct an early warning model applicable to general non-cardiac postoperative acute kidney injury patients, thereby identifying high-risk patients, especially preoperatively, and reducing short- and long-term mortality after non-cardiac surgeries.

[0013] According to a preferred embodiment, the risk factors associated with the early warning of non-cardiac postoperative acute kidney injury (PO-AKI) also include one or both of the following: surgical time and potassium. The risk factor acquisition unit further acquires the patient's surgical time through the hospital's clinical information system and / or the patient's serum potassium data through the hospital's laboratory information management system. Due to the ease of data acquisition, surgical time is incorporated into this invention to increase the predictive accuracy of the early warning model. Both hypokalemia and hyperkalemia are risk factors for PO-AKI; therefore, this invention also constructs an early warning model targeting the relationship between potassium and mortality and major cardiovascular events.

[0014] According to a preferred embodiment, the early warning analysis unit assigns variable values ​​to several risk factors acquired by the risk factor acquisition unit, and constructs the early warning model using a rating scale. This invention constructs the early warning model using a rating scale, which is simpler and clearer than formulaic forms. Furthermore, the factors in its rating scale are all routine items that are clinically available, facilitating rapid preoperative identification and even rapid identification in emergency situations.

[0015] According to a preferred embodiment, the early warning analysis unit uses a deep learning algorithm to train an early warning model composed of several risk factors, and the performance of the early warning model is judged by the area under the receiver operating characteristic (ROC) curve. This invention establishes, for the first time, an early warning model for acute kidney injury (AKI) after non-cardiac surgery. The performance of the early warning model is fully validated through internal evaluation of the modeling dataset and external evaluation of the validation dataset. Based on medical statistics, this invention combines deep learning to improve the prediction accuracy of the early warning model and ensure its interpretability. The results based on the ROC curve show that the model has good predictive performance. More importantly, the model involved in this invention contains 11 variables, which are simple and easy to use in clinical practice.

[0016] According to a preferred embodiment, the system further includes an early warning result display unit, used to display the early warning analysis results of the continuously predicted probability of acute kidney injury in non-cardiac general post-operative patients with acute kidney injury, in a manner correlated with the patient's post-operative time and corresponding post-operative time points. The early warning result display unit can display the results as a probability graph. The early warning result display unit of this invention generates a visual bar graph of the probability of acute kidney injury in non-cardiac general post-operative patients with acute kidney injury to show the impact of various risk factors on patients, providing insights into key factors affecting the prediction of medical outcomes.

[0017] This invention also relates to a pre-operative acute kidney injury warning model for non-cardiac patients. The warning model is based on the following factors of the patient: age, type of surgery, contrast agent, serum creatinine, urine protein, fasting blood glucose, anemia, blood neutrophil to lymphocyte ratio (NLR), and D-dimer.

[0018] According to a preferred embodiment, the steps for constructing the early warning model are as follows: assigning variable values ​​to identified factors related to postoperative acute kidney injury (ARDS) in non-cardiac patients; and constructing an early warning model based on a rating scale using several assigned factors with the aim of determining the prevalence of these factors. This invention uses deep learning to improve the predictive accuracy of the early warning model. Furthermore, to compensate for the limited interpretability of deep learning, this invention also uses medical statistics to identify risk factors.

[0019] This invention also relates to a computer-readable storage medium storing a program for assessing acute kidney injury (ARDS) after non-cardiac surgery in patients. The program for assessing ARDS after non-cardiac surgery is derived from an ARDS early warning model. This invention uses medical statistics to obtain a predictive model consisting of 11 variables, and then combines medical statistics with deep learning to input these variables into the established computer program for prediction.

[0020] This invention also relates to an early warning device for non-cardiac postoperative acute kidney injury, comprising a processor and a memory, the memory storing a program for assessing non-cardiac postoperative acute kidney injury in patients. The program for assessing non-cardiac postoperative acute kidney injury in patients is derived based on an early warning model for non-cardiac postoperative acute kidney injury.

[0021] This invention also relates to the application of patient age, type of surgery, contrast agent, serum creatinine, urine protein, fasting blood glucose, anemia, blood neutrophil-to-lymphocyte ratio (NLR), and D-dimer in the early warning of acute kidney injury after non-cardiac surgery. Attached Figure Description

[0022] Figure 1 This is a flowchart of the patient selection procedure provided by the present invention;

[0023] Figure 2 This is the ROC curve of the multi-factor logistic regression analysis model in the derivation set provided by this invention;

[0024] Figure 3 This is the ROC curve of the derivative-focused non-cardiac surgery AKI early warning score scale provided by the present invention;

[0025] Figure 4 This is the ROC curve of the multifactor logistic regression analysis model provided by this invention in the validation set;

[0026] Figure 5 This is the ROC curve of the AKI early warning score scale for non-cardiac surgery in the validation set provided by this invention;

[0027] Figure 6 This is the ROC curve of the early warning model trained by deep learning using all 11 risk factors provided by this invention;

[0028] Figure 7 The ROC curve of the early warning model provided by this invention, which utilizes all 11 risk factors and is trained by deep learning, has been validated externally.

[0029] Figure 8 This is the ROC curve of the early warning model with no surgical time trained by deep learning provided by this invention;

[0030] Figure 9 The ROC curve of the early warning model with no surgical time trained by deep learning provided by this invention has been externally validated.

[0031] Figure 10 This is the ROC curve of the low blood potassium warning model trained by deep learning provided by this invention;

[0032] Figure 11The ROC curve of the blood potassium deficiency early warning model trained by deep learning provided by this invention has been validated externally.

[0033] Figure 12 This is the ROC curve of the early warning model trained by deep learning, excluding operation time and blood potassium, provided by this invention;

[0034] Figure 13 The ROC curve of the early warning model trained by deep learning without operation time and blood potassium provided by this invention has been externally validated.

[0035] Figure 14 This is a schematic diagram of the hardware topology connection of an apparatus for constructing and applying an early warning model according to a preferred embodiment of the present invention.

[0036] List of reference numerals

[0037] 100: Early warning detection system; 110: Risk factor acquisition unit; 120: Early warning analysis unit; 130: Early warning processing unit; 200: Hospital clinical information system; 300: Hospital laboratory information management system; 400: Hospital resource management system. Detailed Implementation

[0038] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more definite definition of the scope of protection of the present invention. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments described herein.

[0039] This retrospective study utilized large datasets of non-cardiac surgery patients from the electronic medical record system of Xuanwu Hospital, Capital Medical University, to establish modeling and validation datasets. From the modeling datasets, this invention identified candidate risk factors and created an early warning model for postoperative acute kidney injury (PO-AKI); the model was then validated internally and externally. Subsequently, this invention employed the TabNet deep learning model to improve the predictive accuracy of the early warning model, and the model's performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC).

[0040] Of the 19,152 screened patients in the modeling set, 707 (3.69%) developed postoperative acute kidney injury (PO-AKI). This invention identified 11 independent risk factors for PO-AKI, with an AUC of 0.7695. A scoring scale was established by assigning different weights to these risk factors; the AUC of the AKI warning scoring scale was 0.7578. Of the 7,254 patients in the validation set, 195 (2.69%) developed PO-AKI. The AUC of the risk factors in the validation set was 0.7856, and the AUC of the AKI scoring scale was 0.7122. The 11 risk factors were substituted into the TabNet model for training; the AUC of the warning scoring scale reached 0.97 in the validation set and 0.93 in the test set.

[0041] This invention establishes, for the first time, an early warning model for acute kidney injury (AKI) following non-cardiac surgery and verifies its effectiveness. The variables in the model involved in this invention are simple and easily obtained. The combination of medical statistics and deep learning ensures both the interpretability of the model and improves prediction accuracy.

[0042] Example 1

[0043] This invention provides a method and apparatus for constructing and applying a generalized early warning model for acute kidney injury after hospitalization in non-cardiac patients, aiming to solve at least one or more technical problems existing in the prior art.

[0044] To achieve the above objectives, this invention provides a general-purpose, non-cardiac postoperative acute kidney injury early warning and detection system. For example... Figure 14As shown, the early warning detection system 100 includes a risk factor acquisition unit 110, an early warning analysis unit 120, and an early warning processing unit 130. The risk factor acquisition unit 110 is connected to the hospital clinical information system 200 to obtain information on the patient's age, surgical type, contrast agent, and anemia. The risk factor acquisition unit 110 is also connected to the hospital laboratory information management system 300 to obtain information on the patient's serum creatinine, urine protein, fasting blood glucose, neutrophil-to-lymphocyte ratio (NLR), and D-dimer. The early warning analysis unit 120 performs an early warning analysis for postoperative acute kidney injury based on the patient's age, surgical type, contrast agent, serum creatinine, urine protein, fasting blood glucose, anemia, and neutrophil-to-lymphocyte ratio (NLR) and D-dimer information input by the risk factor acquisition unit 110, and obtains the early warning analysis results. The early warning processing unit 130 obtains the early warning analysis results from the early warning analysis unit 120 and sends the results to the hospital resource management system 400 in a manner correlated with the probability and time of postoperative acute kidney injury in the patients identified in the early warning analysis results. This invention establishes an early warning model with simple and easily obtainable parameters using big data to identify high-risk individuals for postoperative acute kidney injury (PO-AKI), reduce its incidence, and thus improve organ and overall prognosis. This invention applies at least nine risk factors to construct an early warning model applicable to non-cardiac patients undergoing inpatient postoperative acute kidney injury, thereby identifying high-risk patients, especially preoperatively, and reducing short-term and long-term mortality after non-cardiac surgeries.

[0045] The risk factor acquisition unit 110 mainly consists of a high-performance server, possessing data processing capabilities and a stable operating environment. The risk factor acquisition unit 110 is equipped with a high-speed network interface card (NIC) for transmitting data in standardized formats (such as HL7, FHIR, etc.) via TCP / IP protocol to the hospital clinical information system 200 and / or the hospital laboratory information management system 300. It should be noted that the risk factor acquisition unit 110 of this invention employs encryption technology, access control, data backup, and other measures to ensure data security and privacy during data transmission and storage. The risk factor acquisition unit 110 acquires at least nine risk factors from the hospital clinical information system 200 and / or the hospital laboratory information management system 300, and sends several of these risk factors to the early warning analysis unit 120, which is used to construct an early warning model. The early warning analysis unit 120 and the risk factor acquisition unit 110 exchange data via an internal data bus. The early warning analysis unit 120 performs real-time or batch analysis on the input risk factors, generates early warning analysis results, and stores the analysis results in a specific database or cache for use by the early warning processing unit 130. The early warning processing unit 130 obtains the early warning analysis results from the early warning analysis unit 120 and sends the results to the hospital resource management system 400 in a manner that correlates the probability and time of postoperative acute kidney injury in the patients identified in the early warning analysis results. Through the above hardware components and data connection methods, the early warning detection system 100 of this invention can effectively integrate the data resources of existing systems in medical institutions, enabling real-time early warning of postoperative acute kidney injury in hospitalized patients, thereby improving medical quality and safety.

[0046] According to a preferred embodiment, the risk factors associated with the early warning of non-cardiac postoperative acute kidney injury (PO-AKI) also include one or both of the following: surgical time and potassium. The risk factor acquisition unit 110 further acquires the patient's surgical time through the hospital clinical information system 200 and / or the patient's serum potassium data through the hospital laboratory information management system 300. Due to the ease of data acquisition, surgical time is incorporated into this invention to increase the predictive accuracy of the early warning model. Both hypokalemia and hyperkalemia are risk factors for PO-AKI; therefore, this invention also constructs an early warning model targeting the relationship between potassium and mortality and major cardiovascular events.

[0047] According to a preferred embodiment, the early warning analysis unit 120 assigns variable values ​​to several risk factors acquired by the risk factor acquisition unit 110, and constructs an early warning model using a rating scale. This invention constructs an early warning model using a rating scale, which is simpler and clearer than formulas. Furthermore, the factors in its rating scale are all routine items that are clinically available, facilitating rapid preoperative identification and even rapid identification in emergency situations.

[0048] According to a preferred embodiment, the early warning analysis unit 120 uses a deep learning algorithm to train an early warning model composed of several risk factors, and judges the performance of the early warning model by the area under the receiver operating characteristic (ROC) curve. This invention establishes, for the first time, an early warning model for acute kidney injury (AKI) after non-cardiac surgery. The performance of the early warning model is fully validated through internal evaluation of the modeling dataset and external evaluation of the validation dataset. Based on medical statistics, this invention combines deep learning to improve the prediction accuracy of the early warning model and ensure its interpretability. The results from the ROC curve show that the model has good predictive performance. More importantly, the model involved in this invention contains 11 variables, which are simple and easy to use in clinical practice.

[0049] According to a preferred embodiment, the system further includes an early warning result display unit, used to display the early warning analysis results of the continuously predicted probability of acute kidney injury in non-cardiac general post-operative patients with acute kidney injury, in a manner correlated with the patient's post-operative time and corresponding post-operative time points. The early warning result display unit can display the results as a probability graph. The early warning result display unit of this invention generates a visual bar graph of the probability of acute kidney injury in non-cardiac general post-operative patients with acute kidney injury to show the impact of various risk factors on patients, providing insights into key factors affecting the prediction of medical outcomes.

[0050] This invention also relates to a pre-operative acute kidney injury warning model for non-cardiac patients. The warning model is based on the following factors of the patient: age, type of surgery, contrast agent, serum creatinine, urine protein, fasting blood glucose, anemia, blood neutrophil to lymphocyte ratio (NLR), and D-dimer.

[0051] According to a preferred embodiment, the steps for constructing the early warning model are as follows: assigning variable values ​​to identified factors related to postoperative acute kidney injury (ARDS) in non-cardiac patients; and constructing an early warning model based on a rating scale using several assigned factors with the aim of determining the prevalence of these factors. This invention uses deep learning to improve the predictive accuracy of the early warning model. Furthermore, to compensate for the limited interpretability of deep learning, this invention also uses medical statistics to identify risk factors.

[0052] This invention also relates to a computer-readable storage medium storing a program for assessing acute kidney injury (ARDS) after non-cardiac surgery in patients. The program for assessing ARDS after non-cardiac surgery is derived from an ARDS early warning model. This invention uses medical statistics to obtain a predictive model consisting of 11 variables, and then combines medical statistics with deep learning to input these variables into the established computer program for prediction.

[0053] This invention also relates to an early warning device for non-cardiac postoperative acute kidney injury, comprising a processor and a memory, the memory storing a program for assessing non-cardiac postoperative acute kidney injury in patients. The program for assessing non-cardiac postoperative acute kidney injury in patients is derived based on an early warning model for non-cardiac postoperative acute kidney injury.

[0054] This invention also relates to the application of patient age, type of surgery, contrast agent, serum creatinine, urine protein, fasting blood glucose, anemia, blood neutrophil-to-lymphocyte ratio (NLR), and D-dimer in the early warning of acute kidney injury after non-cardiac surgery.

[0055] This embodiment provides a detailed description of the medical statistical methods used in this invention:

[0056] 1. Research Methods

[0057] In this single-center retrospective cohort study, data from all non-cardiac surgery patients were obtained from the electronic medical record system of Xuanwu Hospital, Capital Medical University. The study was conducted as follows:

[0058] 1.1 A derived cohort was created using patient data from November 2015 to December 2020. This cohort allowed researchers to observe the incidence of acute kidney injury (AKI) within 7 days post-surgery. During this study, data from 38,146 patients were screened, and data from patients aged ≥18 years were collected. Patients undergoing urological or obstetric surgeries were excluded because urological surgeries can cause AKI through specific mechanisms, and the diagnostic criteria for AKI in pregnant women are not yet standardized. Secondly, data from patients undergoing thoracic surgery were excluded due to the small sample size. Based on these inclusion criteria, a total of 36,585 individuals were included. However, 2,418 patients were excluded due to preoperative end-stage renal disease, renal replacement therapy, or AKI within 30 days prior to surgery; additionally, 15,015 patients were excluded due to missing data. Finally, data analysis was performed on 19,152 eligible participants. The participant screening process is described in [link to study details]. Figure 1 .

[0059] 1.2 Researchers selected observed variables based on previous studies and performed univariate logistic regression analysis to identify potential risk factors associated with postoperative acute kidney injury (PO-AKI).

[0060] 1.3 Multivariate logistic regression analysis was performed based on potential risk factors to identify the risk factors most associated with postoperative acute kidney injury (PO-AKI) and to calculate their AUC.

[0061] 1.4 Based on the different weights of risk factors, establish an early warning model, present the early warning model in the form of a rating scale, and calculate the AUC.

[0062] 1.5 A validation set was built using data collected from January to December 2021 for external validation.

[0063] 1.6 The risk factors obtained from multivariate logistic regression are put into the TabNet model, and then internal and external validation is performed.

[0064] Specifically, baseline serum creatinine (Scr) level was defined as the value measured from 1 month before surgery to the day of surgery, and postoperative serum creatinine (Scr) level was defined as the value measured within 7 days after surgery. Postoperative serum creatinine (Scr) levels were compared with preoperative baseline values, and acute kidney injury (AKI) was diagnosed according to the following 2012 KDIGO criteria: an increase in serum creatinine (Scr) ≥0.3 mg / dL (≥26.5 μmol / L) or an increase in serum creatinine (Scr) ≥1.5 times (known or presumed) from baseline within 7 days after surgery.

[0065] 1.7 Statistical Methods

[0066] All statistical analyses were performed using SAS 9.4 software. To determine risk factors for postoperative acute kidney injury (PO-AKI), logistic regression was used for variable selection and model construction. For laboratory test indicators, the upper and lower limits of continuous values ​​were used as thresholds to transform continuous variables into categorical variables. Variables with internationally standardized limits were converted into categorical data according to their standards. The reference system was specified as the normal range of laboratory indicators and ordinal multiclass variables. For nominal multiclass variables, a chi-square test was performed on the incidence of acute kidney injury (AKI) between different groups. The group with the low incidence was subsequently designated as the reference system. Considering factors such as professional knowledge, work experience, univariate analysis results, and AUC, the selected factors were included in the multivariate logistic regression analysis. Finally, the optimized final model was established. The area under the receiver operating characteristic curve (AUC), cut-off value, sensitivity, specificity, and Youden index of the model were calculated. Understandably, metrics for evaluating the predictive performance of a model typically include positive predictive value (PPV), negative predictive value (NPV), and accuracy rate.

[0067] 1.8 The TabNet algorithm was used to train an early warning model based on 11 risk factors. The computer automatically split the original modeling dataset into a new training set and an internal validation set at an 8:2 ratio. The model was trained and internally validated, then externally validated using the original validation data. Receiver Operating Characteristic (ROC) curves were used to evaluate the model's performance, and the AUC was calculated as a measure of the model's predictive ability. Furthermore, feature importance analysis was performed using the TabNet model. The importance of each feature was calculated and ranked, and a visual bar chart was generated to show the relative contribution of each feature. This analysis provides insights into the key factors influencing the prediction of medical outcomes.

[0068] In summary, the method for constructing the postoperative AKI early warning model provided by this invention may include:

[0069] To obtain clinical data from several patients who did not undergo cardiac surgery;

[0070] Independent risk factors associated with postoperative AKI were identified using multivariate logistic regression analysis.

[0071] An initial early warning model is established based on the different weights of independent risk factors, and the initial early warning model is presented in the form of a rating scale.

[0072] Independent risk factors obtained from multivariate logistic regression analysis are input into a neural network model for deep learning to obtain a postoperative AKI early warning model.

[0073] In particular, before using multivariate logistic regression analysis to derive independent risk factors associated with postoperative AKI, univariate logistic regression analysis can be used to derive potential risk factors associated with postoperative AKI from clinical data.

[0074] 2. Research Results

[0075] 2.1 Basic Data Analysis

[0076] According to the KDIGO criteria, among the 19,152 cases in the modeling dataset, 707 cases (3.69%) developed acute kidney injury (AKI) postoperatively, and 18,445 cases (96.31%) did not develop AKI postoperatively. Their baseline data are shown in Table 1.

[0077] Table 1. Comparison of baseline data between postoperative AKI and non-AKI patients.

[0078]

[0079]

[0080]

[0081]

[0082] 2.2 Univariate logistic regression analysis of postoperative AKI risk factors

[0083] Univariate logistic regression analysis was performed to identify potential risk variables in the multivariate logistic regression model for postoperative acute kidney injury (PO-AKI). These potential risk variables may include basic patient information such as sex, age, body mass index (BMI), surgical department, hypertension classification, and chronic kidney disease (CKD) stage; perioperative examination results, i.e., various preoperative and postoperative physiological indicators, such as preoperative creatinine level, glomerular filtration rate (eGFR), uric acid level, blood urea nitrogen level, semi-quantitative urine protein classification, urine specific gravity, urine white blood cell count, urine red blood cell count, semi-quantitative urine glucose classification, urine pH, hemoglobin level, red blood cell count, hematocrit, white blood cell count, neutrophil count, lymphocyte count, neutrophil-to-lymphocyte ratio (NLR), monocyte count, and basophilia. The study included granulocyte count, platelet count, serum albumin, prealbumin, blood glucose, total cholesterol, high-density lipoprotein cholesterol (HDL-C), apolipoprotein A, creatine kinase, lactate dehydrogenase, aspartate aminotransferase, serum potassium, sodium, calcium, phosphorus, chloride, thrombin time, prothrombin time, activated partial thromboplastin time, international normalized ratio, prothrombin time activity, D-dimer, fibrinogen, lactate, partial pressure of carbon dioxide, partial pressure of oxygen, and oxygen saturation. Intraoperative circulatory data, including pulse, heart rate, and blood pressure, were also included. Intraoperative management data, such as operation time and contrast agent usage, were also included. A total of 54 variables were selected in this study.

[0084] 2.3 Multivariate logistic regression analysis of postoperative AKI risk factors

[0085] Variables with p-values ​​< 0.1 in univariate logistic regression analysis were included in multivariate logistic regression analysis to identify risk factors for postoperative acute kidney injury (PO-AKI) and their weights as predictors for the scoring scale. The multivariate logistic regression model included 11 independent risk factors for PO-AKI (p ≤ 0.05), as shown in Table 2. Specifically, the 11 independent risk factors for PO-AKI included the subject's age, surgical department, operation time, contrast agent use, serum creatinine level, semi-quantitative classification of urinary protein, serum potassium, fasting blood glucose, hemoglobin level, neutrophil-to-lymphocyte ratio, and D-dimer.

[0086] Table 2 Multivariate logistic regression analysis of postoperative AKI risk factors

[0087]

[0088]

[0089]

[0090] The receiver operating characteristic (ROC) curve was plotted based on the results of multivariate logistic regression analysis. The area under the ROC curve (AUC) was 0.7695. (See [link to relevant documentation]). Figure 2 To calculate the probabilities and classification tables of the logistic model, this study calculated the Youden index at different probability levels based on the sensitivity and specificity of the classification tables. At the maximum Youden index of 39.2, the model's sensitivity was 64.8% and its specificity was 74.4%. Preoperative serum potassium, NLR, and D-dimer levels, whether below or above normal, were risk factors for postoperative acute kidney injury (PO-AKI).

[0091] 2.4 Establish a pre-hospitalization scoring system and grading scale for postoperative acute kidney injury (PO-AKI).

[0092] For ease of use, a scoring scale is provided during the establishment of the early warning model in medical statistics. Multivariate logistic regression analysis was used to determine the weights of risk factors, and values ​​were assigned to the variables in the scoring scale. The scoring scale corresponds to scores of 2.5–3 points, designated as the gray zone. Therefore, a score <2.5 indicates a low risk of postoperative acute kidney injury (PO-AKI), with a prevalence of approximately 1%; a score of 2.5–3 represents the gray zone, with a prevalence of approximately 3%; and a score >3 indicates a high probability of PO-AKI, with a prevalence >11%. The lowest score on the scale is 0, and the highest score is 12.5, as detailed in Table 3.

[0093] Table 3. Early warning scoring scale for AKI after non-cardiac surgery based on reference values ​​and regression coefficients.

[0094]

[0095]

[0096]

[0097] See Figure 3 Receiver operating characteristic (ROC) curves for the rating scale were plotted, with an area under the ROC curve (AUC) of 0.7578 (P = 0.0001, OR = 2.758, 95% CI: 2.561–2.970). The maximum Youden index of the rating scale was 38, and the sensitivity and specificity of the rating scale were 69.9% and 68.1%, respectively.

[0098] 2.5 Validation of the early warning scoring system and scale for postoperative acute kidney injury (PO-AKI)

[0099] This study used electronic medical record data from Xuanwu Hospital, Capital Medical University in 2021 to validate the early warning scoring system and scale for postoperative acute kidney injury (PO-AKI). Baseline data from the modeling set and validation set were compared. The validation set included 7254 patients who underwent non-cardiac surgery, of whom 195 developed acute kidney injury (AKI) postoperatively (see Table 4).

[0100] Table 4 Comparison of baseline data between the derivation queue and the verification queue

[0101]

[0102]

[0103] When performing multivariate logistic regression analysis, the receiver operating characteristic (ROC) curve was plotted. The AUC of the validation set was 0.7856. See [link to relevant documentation]. Figure 4 The receiver operating characteristic (ROC) curve was plotted using the AKI rating scale. The AUC of the validation set was 0.7122. See [link to relevant documentation]. Figure 5 Similar to the AUC of the modeling set, external validation showed that the AKI rating scale had good predictive accuracy. The maximum Youden index of this rating scale was 30.5, with a sensitivity of 63.1% and a specificity of 67.4%. The corresponding scores for the rating scale ranged from 2.1 to 2.5. Therefore, this study adopted the optimal cutoff value of 2.5, which is similar to the results of the modeling set.

[0104] 2.6 The TabNet algorithm improves the accuracy of the early warning model. This study further provides models using different variables.

[0105] 2.6.1 The TabNet algorithm was used to train the early warning model composed of the above 11 independent risk factors (see...). Figure 6 ) and verification (see Figure 7The receiver operating characteristic (ROC) curves were plotted. Specifically, the AUC of the ROC curve for the warning model trained by deep learning was approximately 0.97; and the AUC of the ROC curve for the warning model trained by deep learning that passed external validation was approximately 0.93.

[0106] 2.6.2 The TabNet algorithm was used to train the early warning model that excluded operation time (the only intraoperative variable) (see...). Figure 8 ) and verification (see Figure 9 The receiver operating characteristic (ROC) curves were plotted. Specifically, the AUC of the ROC curve for the early warning model trained by deep learning that excludes the surgical time variable was approximately 0.98; and the AUC of the ROC curve for the early warning model trained by deep learning that excludes the surgical time variable and has been externally validated was approximately 0.88.

[0107] 2.6.3 Because potassium performs poorly in medical statistical validation, this study provides a method for training an early warning model that excludes potassium using the TabNet algorithm (see...). Figure 10 ) and verification (see Figure 11 Receiver operating characteristic (ROC) curves were plotted. Specifically, in the ROC curve of the early warning model trained by deep learning that excludes the blood potassium variable, the AUC was approximately 0.98; in the ROC curve of the early warning model trained by deep learning that excludes the blood potassium variable and passed external validation, the AUC was approximately 0.82.

[0108] 2.6.4 Training an early warning model that excludes surgical time and blood potassium levels using the TabNet algorithm (see...) Figure 12 ) and verification (see Figure 13 The receiver operating characteristic (ROC) curves were plotted. Specifically, the AUC was approximately 0.98 in the ROC curve of the early warning model excluding surgical time and serum potassium variables trained by deep learning; and approximately 0.81 in the ROC curve of the early warning model excluding surgical time and serum potassium variables after external validation.

[0109] 3. Research Discussion

[0110] Early and late acute kidney injury (AKI) following non-cardiac surgery is significantly associated with higher short-term and long-term mortality, and many of these patients could potentially avoid AKI. Therefore, identifying high-risk patients, especially preoperatively, is essential. However, to date, few studies have evaluated and validated early warning systems for postoperative acute kidney injury (PO-AKI). Therefore, this invention aims to establish a universally applicable inpatient AKI early warning system for non-cardiac surgeries that meets modern needs.

[0111] This study included experimental data from many other levels of surgery, such as emergency surgery. However, ICU patients were excluded. Secondly, among numerous statistical variables, urine output data were excluded due to their unreliability and the inability to guarantee the adaptation of the body to transient hypoperfusion or hypotension by medications, and intraoperative urine output had no impact on the inclusion of risk factors. On the other hand, considering that using serum creatinine (Scr) levels at admission might underestimate the incidence of acute kidney injury (AKI), it was recommended to use serum creatinine (Scr) within one month prior to admission. Furthermore, many factors, including severe illness, frailty, prolonged postoperative bed rest, and stress, are considered to cause a false normalization of renal function. Therefore, according to the KDIGO guidelines, if postoperative serum creatinine (Scr) concentration is lower than preoperative, this value will be used as the baseline for AKI identification in the following days. Other intraoperative variables were not used in the experimental data included in this study. However, operative time is an optional item due to its simplicity and practicality. This invention uses a TabNet model to compare the predictive accuracy of warning models with and without operative time.

[0112] The incidence of postoperative acute kidney injury (PO-AKI) varies across different subspecialties. The proportion of PO-AKI in the total PO-AKI also differs across subspecialties. Therefore, scores on the scoring scale for different subspecialties may need appropriate adjustment, or an intermediate formula with cross-site portability may be used to broaden the application of the early warning system of this invention. In the study involved in this invention, preoperative proteinuria was a risk factor for PO-AKI, similar to the findings of Tyler et al. or the NARA-AKI study. However, the pathophysiological explanation for this association is unclear, as it depends neither on baseline glomerular filtration rate (eGFR) nor on the presence of hypertension or diabetes, and infection, acute inflammation, or intraoperative hemodynamic changes cannot be demonstrated as the cause of this association. Researchers also found that failure to check for proteinuria preoperatively (urinalysis) is also a risk factor for PO-AKI, which warrants attention.

[0113] In the study presented in this invention, both hypokalemia and hyperkalemia are risk factors for postoperative acute kidney injury (PO-AKI), partially similar to the PO-AKI early warning model used at Xiangya Hospital. While many studies focus on the relationship between potassium and mortality and major cardiovascular events, the mechanism between potassium and PO-AKI has not been described. Since the same results were not obtained in the validation set, this invention uses the TabNet model to provide different early warning models with or without potassium as a parameter. Researchers found that the preoperative neutrophil-to-lymphocyte ratio (NLR) is a risk factor for PO-AKI, although this contradicts findings in other studies. Researchers were the first to discover that a decreased NLR is also a risk factor for PO-AKI. Kai Singbartl's review mentions a bidirectional interaction between acute kidney injury (AKI) and the immune system, with immune responses mediating kidney injury and AKI recovery; this correlation may partially explain the results presented in this invention.

[0114] Furthermore, researchers found that D-dimer levels below or above the normal range are also a risk factor for postoperative acute kidney injury (PO-AKI). D-dimer is not only associated with thromboembolic diseases, but has also been shown to increase in critically ill patients or those with acute kidney injury (AKI), such as contrast-induced AKI. Previous studies have also confirmed that D-dimer is associated with systemic inflammation or infection, thereby exacerbating inflammatory coagulation and fibrinolysis, ultimately promoting the occurrence of AKI in critically ill patients. However, this does not explain why D-dimer levels below the normal range are also a risk factor. Many previous studies have shown that age, contrast agents, preoperative anemia, diabetes, and chronic kidney disease (CKD) are major risk factors for AKI after non-cardiac surgery, all of which have been confirmed in the study presented in this invention. However, unlike other studies, this invention only used preoperative fasting blood glucose after admission; elevated fasting blood glucose is a risk factor. In the study described in this invention, statistical analysis revealed that preoperative serum creatinine (Scr), rather than glomerular filtration rate (eGFR) or blood urea nitrogen (BUN), was included as a risk factor in the early warning scoring scale. Notably, preoperative hemoglobin levels, if not examined, are also a risk factor for postoperative acute kidney injury (PO-AKI).

[0115] The early warning scale for non-cardiac surgery acute kidney injury (AKI) patients of this invention adopts a score scale format, which is simpler and clearer than formulaic formats. Furthermore, through multivariate logistic regression analysis, the scale only includes 11 variables, making it practical and quick. All variables are routine items readily available in clinical practice, facilitating rapid preoperative identification and even rapid identification in the emergency room. The AUC of this scoring scale on the derivation set and validation set are 0.7578 and 0.7122, respectively, indicating good predictive efficiency.

[0116] Researchers considered using deep learning to improve the predictive accuracy of early warning models. However, to compensate for the lack of interpretability in deep learning, they first used medical statistics to identify risk factors. Furthermore, deep learning faces another challenge when combined with medical statistics: how to directly identify the numerical range of a variable that constitutes a risk factor. In summary, the researchers found that both processes (internal and external validation) were accurate and preserved the interpretability of risk factors. They also found that deep learning can help quickly adjust predictive models based on different variable conditions. In this study, after adjustment and comparison, the researchers found that the predictive model consisting of 11 variables obtained through medical statistics performed best, reflecting that combining medical statistics with deep learning yields the best results. Although the predictive model after deep learning can no longer be presented as a simple scoresheet, it can still be used in established computer programs for prediction. Overall, this invention proposes a comprehensive deep learning workflow that combines data preprocessing, model training, evaluation, and feature importance analysis using the TabNet algorithm. It demonstrates the application of TabNet in medical outcome prediction, providing valuable insights for clinical decision-making.

[0117] It should be noted that this study has certain limitations: (1) This study is a single-center study, and subsequent multi-center prospective external validation of the rating scale is required; (2) This study is a retrospective study, and is therefore limited by existing data.

[0118] In another aspect, this invention provides a method for applying a postoperative AKI early warning model, or a method for predicting postoperative AKI based on a postoperative AKI early warning model, which can utilize the postoperative AKI early warning model constructed using the methods described in the embodiments of this invention. Specifically, the method for applying the postoperative AKI early warning model includes:

[0119] Obtain perioperative data from the subjects;

[0120] Several independent risk factors associated with postoperative acute kidney injury (AKI) in the perioperative data of the subjects were input into a deep learning-based postoperative AKI early warning model to obtain the predicted probability of the subjects developing AKI.

[0121] Specifically, the independent risk factors included 11 independent risk factors derived from multivariate logistic regression analysis: age, surgical department, operation time, contrast agent use, serum creatinine level, semi-quantitative classification of urinary protein, serum potassium, fasting blood glucose, hemoglobin level, neutrophil-to-lymphocyte ratio, and D-dimer.

[0122] Those skilled in the art will understand that, as long as the objectives of the present invention can be achieved, other steps or operations may be included before, after, or between the steps described above, for example, to further optimize and / or improve the method described in the present invention. Furthermore, although the method described in the present invention is shown and described as a series of actions performed sequentially, it should be understood that the method is not limited by the order. For example, some actions may occur in a different order than that described herein. Alternatively, one action may occur simultaneously with another action.

[0123] Example 2

[0124] In another aspect, the present invention provides an apparatus for constructing and applying a postoperative AKI early warning model, used to implement the method for constructing and applying a postoperative AKI early warning model as described in the above embodiments. The apparatus includes:

[0125] The data acquisition unit is used to acquire clinical data from a number of non-cardiac surgery patients.

[0126] The first screening unit is used to derive independent risk factors associated with postoperative AKI based on multivariate logistic regression analysis.

[0127] The model building unit is used to establish an initial early warning model based on the different weights of independent risk factors, and to present the initial early warning model in the form of a rating scale. Furthermore, the apparatus for constructing and applying a postoperative AKI early warning model provided by this invention also includes:

[0128] The deep learning unit is used to input the independent risk factors obtained from multivariate logistic regression analysis into the neural network model for deep learning, thereby obtaining a postoperative AKI early warning model.

[0129] Furthermore, the apparatus for constructing and applying a postoperative AKI early warning model provided by the present invention also includes:

[0130] The second screening unit is used to derive potential risk factors related to postoperative AKI from clinical data using univariate logistic regression analysis before deriving independent risk factors related to postoperative AKI based on multivariate logistic regression analysis. The apparatus for constructing and applying a postoperative AKI early warning model provided by this invention employs the method for constructing and applying a postoperative AKI early warning model as described in the above embodiments. The beneficial effects of the apparatus for constructing and applying a postoperative AKI early warning model provided by this invention are the same as the beneficial effects of the method for constructing and applying a postoperative AKI early warning model provided in the above embodiments, and other technical features in this apparatus are the same as those disclosed in the method of the above embodiments, and will not be repeated here.

[0131] It should be noted that the specific embodiments described above are exemplary. Those skilled in the art can devise various solutions inspired by the disclosure of this invention, and these solutions all fall within the scope of this invention and its protection. Those skilled in the art should understand that this specification and its accompanying drawings are illustrative and not intended to limit the scope of the claims. The scope of protection of this invention is defined by the claims and their equivalents. This specification contains multiple inventive concepts; terms such as "preferredly," "according to a preferred embodiment," or "optionally" indicate that the corresponding paragraph discloses an independent concept. The applicant reserves the right to file divisional applications based on each inventive concept.

Claims

1. A general-purpose, non-cardiac postoperative acute kidney injury early warning and detection system, characterized in that, The early warning detection system (100) includes: Risk Factor Acquisition Unit (110) The risk factor acquisition unit (110) is connected to the hospital clinical information system (200) to obtain information on the patient's age, type of surgery, contrast agent, and anemia from the hospital clinical information system (200). The risk factor acquisition unit (110) is connected to the hospital laboratory information management system (300) and acquires the patient's serum creatinine, urine protein, fasting blood glucose, blood neutrophil to lymphocyte ratio (NLR) and D-dimer information from the hospital laboratory information management system (300). The early warning analysis unit (120) performs early warning analysis on postoperative acute kidney injury based on the patient's age, surgical type, contrast agent, serum creatinine, urine protein, fasting blood glucose, anemia, blood neutrophil to lymphocyte ratio (NLR) and D-dimer information input by the risk factor acquisition unit (110), and obtains the early warning analysis results. The early warning processing unit (130) obtains the early warning analysis result from the early warning analysis unit (120) and sends the early warning analysis result to the hospital resource management system (400) in a manner that is associated with the probability and time of postoperative acute kidney injury of the patient in the early warning analysis result. Risk factors associated with early warning of acute kidney injury after non-cardiac surgery also include: operative time and one or both of potassium levels. The risk factor acquisition unit (110) also acquires the patient's operation time through the hospital clinical information system (200) and / or acquires the patient's serum potassium data through the hospital laboratory information management system (300); The early warning analysis unit (120) assigns variable values ​​to several risk factors obtained by the risk factor acquisition unit (110) and constructs an early warning model in the form of a rating scale. The early warning analysis unit (120) uses the TabNet algorithm to train the early warning model composed of several risk factors, and judges the performance of the early warning model by the area under the receiver operating characteristic curve.

2. The early warning detection system according to claim 1, characterized in that, The early warning detection system (100) further includes an early warning result display unit, used to display the early warning analysis results in a manner that correlates the continuously predicted probability of acute kidney injury in non-cardiac general postoperative patients with postoperative acute kidney injury with the patient's postoperative time and corresponding postoperative time point. The warning result display unit can display the results as a curve in the form of a probability graph.

Images