An ARDS complicated with AKI patient subtype typing device based on a Transformer model and application thereof

Abstract

The application discloses application of a group of clinical indexes in preparation of a product for subtyping of ARDS patients complicated with AKI, and constructs a Transformer model for subtyping of ARDS patients complicated with AKI and an ARDS patient complicated with AKI subtyping device based on the Transformer model in combination with the clinical indexes. The ARDS patient complicated with AKI subtyping device can realize accurate and automatic subtyping of ARDS patients complicated with AKI, effectively divides the ARDS patients complicated with AKI into three subtypes with different disease progressions, and then realizes individualized treatment of ARDS patients complicated with AKI with different disease progressions, and the subtyping result has excellent generalization and clinical interpretability.

Description

Technical Field

[0001] This application relates to the fields of medical artificial intelligence and clinical big data analysis technology, and in particular to a subtyping device for ARDS patients with AKI based on the Transformer model and its application. Background Technology

[0002] Acute respiratory distress syndrome (ARDS) and acute kidney injury (AKI) are common and serious complications in intensive care units (ICUs), and the combined occurrence of both (ARDS-AKI) significantly increases patient mortality and treatment difficulty. Both ARDS and AKI exhibit high heterogeneity, meaning that different patients show significant differences in their etiology, pathophysiological mechanisms, clinical manifestations, disease trajectories, and treatment responses. This heterogeneity often limits the effectiveness of a "one-size-fits-all" treatment strategy. Therefore, precise subtyping of patients to tailor treatment strategies for different subtypes has become a key challenge in achieving precision medicine in critical care.

[0003] Currently, clinical ARDS classification mainly relies on the Berlin definition, which is based on static indicators at a single time point, such as the oxygenation index (P / F ratio), chest imaging, and positive end-expiratory pressure (PEEP) level. This definition fails to effectively capture the dynamic evolution of the disease and is insufficient to reflect deeper pathophysiological differences. While AKI has staging standards such as KDIGO, it faces similar problems. Traditional clustering methods, such as Latent Profile Analysis (LPA) or Latent Category Analysis (LCA), and their extensions for handling longitudinal data, such as Latent Category Mixture Model (LCMM), have been attempted for subtyping ARDS or AKI. These methods can discover latent categorical structures in the data based on cross-sectional or longitudinal data. However, these methods have significant limitations: First, they typically handle longitudinal trajectories or survival outcomes separately, failing to jointly model both within a unified model framework, thus limiting their ability to discover subtypes with significant survival differences; second, traditional clustering methods may yield good typing results on the training set, but their generalization ability is often insufficient, and they perform inconsistently on the test set or external validation set, resulting in poor consistency of typing results; third, traditional methods have limited processing capabilities for complex data that is prevalent in ICU environments, characterized by high dimensionality, multimodality, large amounts of missing data, and a combination of horizontal static features (such as demographic information and underlying diseases) and vertical dynamic features (such as daily vital signs and laboratory indicators).

[0004] In recent years, deep learning technology, especially the Transformer architecture, has achieved great success in natural language processing and time series analysis due to its powerful sequence modeling and feature extraction capabilities. However, applying it to ICU medical data with small sample sizes and high missing rates presents significant challenges, including how to ensure the model's classification performance on such data, how to ensure its output is consistent with important clinical endpoints (such as survival ranking), and how to improve the model's interpretability to meet clinical needs.

[0005] Furthermore, existing research largely focuses on the classification of single-organ syndromes (such as ARDS or AKI alone). For ARDS and AKI, two common organ failures that may involve pathophysiological interactions, the consistency and correlation across organ classifications have not been fully explored. Revealing the correspondence between latent categories of lung-kidney syndromes is of significant scientific and clinical value for understanding the common mechanisms of multi-organ failure and developing synergistic treatment strategies.

[0006] Therefore, there is an urgent need for an ARDS-AKI patient classification method and system that can comprehensively utilize longitudinal big data from multi-center ICUs, effectively handle data gaps and heterogeneity, jointly model disease trajectories and survival risks, and use advanced deep learning technology to improve classification generalization ability and accuracy, ultimately revealing cross-organ associations and providing guidance for precision treatment. Summary of the Invention

[0007] The purpose of this invention is to overcome the above-mentioned shortcomings of the prior art and provide a subtyping device for ARDS patients with AKI based on the Transformer model and its application.

[0008] The primary objective of this invention is to provide a set of clinical indicators for the application in the preparation of products for the classification of patients with ARDS complicated by AKI.

[0009] The second objective of this invention is to provide a method for constructing a Transformer model for subtyping ARDS patients complicated with AKI.

[0010] A third objective of this invention is to provide a Transformer model constructed using the above-described construction method.

[0011] The fourth objective of this invention is to provide a subtyping device for ARDS patients with AKI based on the Transformer model.

[0012] The fifth objective of this invention is to provide an electronic device.

[0013] To achieve the above objectives, the present invention is implemented through the following solution: This invention claims protection for the use of a set of clinical indicators in the preparation of products for classifying patients with ARDS complicated by AKI. These clinical indicators include age, height, weight, body mass index, predicted weight, sex, presence of sepsis, presence of trauma, history of multiple transfusions, presence of aspiration, presence of pneumonia, history of elective surgery, presence of HIV, presence of leukemia or lymphoma, presence of solid tumors, immunosuppression, presence of liver failure or cirrhosis, presence of diabetes, history of chronic kidney disease, body temperature, heart rate, systolic blood pressure, diastolic blood pressure, respiratory rate, ventilator-set respiratory rate, blood oxygen saturation, positive end-expiratory pressure, fluid output, fluid intake, platelet count, serum creatinine, pH value, partial pressure of oxygen, partial pressure of carbon dioxide, oxygen concentration, and oxygenation index.

[0014] This invention also claims protection for a method of constructing a Transformer model for subtyping ARDS patients with AKI, comprising the following steps: S1. Obtain the clinical indicators of each sample in the training set; each sample in the training set is an ARDS patient complicated with AKI; the clinical indicators are the clinical indicators shown above. For numerical clinical indicators, the corresponding values ​​are obtained directly, and for binary clinical indicators, 0 / 1 corresponds to yes / no or male / female. S2. Preprocess the clinical indicators of each sample in the training set obtained in step S1 to obtain the standardized clinical indicators of each sample in the training set. During preprocessing, the mean and standard deviation of clinical indicators for each sample in the training set were recorded; S3. Using the standardized clinical indicators of each sample in the training set obtained in step S2 as the input of the Transformer model, and using the subtype classification of ARDS patients with AKI as the prediction target, the Transformer model is trained in combination with the subtypes of each sample in the training set to obtain a Transformer model that can perform subtype classification of ARDS patients with AKI. The subtypes of each sample in the training set are obtained by the Joint Latent Class Model (JointLCMM), and the subtypes include recovering, non-progressive, and progressive.

[0015] Preferably, the clinical indicators of each sample in the training set in step S1 are the clinical indicators of each sample in the training set for 8 consecutive days.

[0016] More preferably, the consecutive 8 days refer to the period from day 0 to day 7 after the onset of illness or enrollment of each sample in the training set.

[0017] The preprocessing in step S2 involves preprocessing the clinical indicators of each sample in the training set obtained in step S1 for 8 consecutive days, on a daily basis.

[0018] Preferably, the preprocessing in step S2 is as follows: the clinical indicators of each sample in the training set obtained in step S1 are sequentially cleaned, missing values ​​are imputed, and Z-score is standardized to obtain the standardized clinical indicators of each sample in the training set.

[0019] More preferably, the data cleaning includes: removing clinical indicators in the training set that exceed the 0.02 quantile and 0.98 quantile, and eliminating samples with a clinical indicator missing rate > 20%.

[0020] More preferably, the missing value imputation is performed using the predicted mean matching algorithm in the multiple imputation method.

[0021] When constructing the Transformer model for subtyping ARDS patients with AKI, the clinical indicators of each sample in the training set were processed through systematic data cleaning rules and advanced missing value imputation methods. The influence of different clinical indicator dimensions was eliminated through standardization, which provided high-quality and standardized input for the construction of the Transformer model and laid the foundation for the accuracy of the model.

[0022] Preferably, the method for obtaining the subtypes of each sample in the training set in step S3 is as follows: The standardized clinical indicators of each sample in the training set are used as input to the Joint Latent Class Model (JLCMM). The JLCMM classifies each sample based on the standardized clinical indicators of each sample to obtain the subtype of each sample in the training set.

[0023] Among them, the JointLatent Class Model (JointLCMM) uses an unsupervised approach to jointly model standardized clinical indicators of each sample within a latent class framework. This results in each subtype in the typing results having not only unique longitudinal trajectory features (clinical indicator features) but also corresponding to different survival risks, thereby enhancing the clinical interpretability and practicality of the typing.

[0024] In step S3, when training by combining the subtypes of each sample in the training set, the classification performance of the model during training is evaluated by at least one of the kappa coefficient and F1 value.

[0025] Preferably, in step S3, the recovery type refers to ARDS patients with AKI whose disease condition has improved, the non-progressive type refers to ARDS patients with AKI whose disease activity is relatively quiescent, and the progressive type refers to ARDS patients with AKI whose disease activity is persistent. Among them, the recovery survival outcome is better than the non-progressive survival outcome, which is better than the progressive survival outcome.

[0026] This invention also claims protection for the Transformer model constructed by any of the above-described construction methods.

[0027] The present invention also claims protection for a subtyping device for ARDS patients with AKI based on the Transformer model, including a data acquisition module, a data processing module, a subtyping module and a result output module; The data acquisition module is used to acquire clinical indicators of patients with ARDS complicated with AKI; the clinical indicators are the clinical indicators shown above. For numerical clinical indicators, the corresponding values ​​are directly acquired, and for binary clinical indicators, 0 / 1 corresponds to yes / no or male / female. The data processing module standardizes the clinical indicators of ARDS patients with AKI obtained by the data acquisition module based on the mean and standard deviation of the clinical indicators of each sample in the training set in the construction method described above, and obtains standardized clinical indicators of ARDS patients with AKI. The subtyping module is the Transformer model mentioned above. It takes the standardized clinical indicators of ARDS patients with AKI obtained by the data acquisition module as input, and the Transformer model outputs the subtyping results of ARDS patients with AKI. The subtypes include remission type, non-progressive type, and progressive type; The result output module is used to output the results obtained by the fractal module.

[0028] Preferably, the data acquisition module is used to acquire clinical indicators of ARDS patients complicated with AKI for 7 consecutive days.

[0029] Preferably, the data processing module standardizes the clinical indicators of ARDS patients with AKI for 7 consecutive days.

[0030] The present invention also claims protection for an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to cause the electronic device to perform the aforementioned means.

[0031] The present invention also claims protection for a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described apparatus.

[0032] Compared with the prior art, the present invention has the following beneficial effects: This invention provides the application of a set of clinical indicators in the preparation of products for the subtyping of ARDS patients with AKI. Furthermore, it constructs a Transformer model for subtyping ARDS patients with AKI and a subtyping device based on the Transformer model, using these clinical indicators. The subtyping device enables accurate and automated subtyping of ARDS patients with AKI, effectively classifying them into three subtypes with different disease progressions. This allows for personalized treatment for patients with ARDS patients at different disease progressions, and the subtyping results exhibit excellent generalization and clinical interpretability. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the construction and use of the Transformer model for classifying ARDS patients with AKI, as shown in Example 1. Figure 2 This is a 28-day Kaplan-Meier survival analysis graph of three different subtypes of ARDS complicated with AKI in the data cohort of Example 1; Figure 3 This is a 60-day Kaplan-Meier survival analysis graph of three different subtypes of ARDS complicated with AKI in the data cohort of Example 1; Figure 4 This is a 28-day Kaplan-Meier survival analysis diagram of ARDS patients with three different subtypes in the experimental group and control group 2 in Example 2; Figure 5 This is a 60-day Kaplan-Meier survival analysis of ARDS patients with three different subtypes in the experimental group and control group 2 in Example 2. Detailed Implementation

[0034] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Unless otherwise specified, the experimental methods used in the following embodiments are conventional methods; the materials and reagents used, unless otherwise specified, are commercially available.

[0035] Example 1: A Transformer model for classifying ARDS patients with AKI The flowchart illustrating the construction and use of the Transformer model for classifying ARDS patients with AKI in this embodiment is as follows: Figure 1 As shown, the details are as follows: I. Sample Acquisition Data cohort acquisition: Data on acute respiratory distress syndrome (ARDS) patients from experiments conducted between 1996 and 2005 (including ARMA, LASRS, ALVEOLI, and FACTT experiments) in the ARDSNet database (N=2632) and data on ARDS patients admitted to the ICU of the First Affiliated Hospital of Guangzhou Medical University between 2018 and 2024 (N=671) were extracted. Patients were excluded according to exclusion criterion 1, resulting in a data cohort containing data on 857 ARDS patients with AKI. All patient data were collected from Day 0 to Day 7 after enrollment (8 days of data). Exclusion criterion 1 is: (1) patient age < 18 years; (2) missing patient survival outcome; (3) missing patient hospital stay; (4) indicators with a missing rate > 40% in patient data; (5) non-acute kidney injury (AKI) patients; II. Training the Transformer Model 1. Experimental Methods (1) Acquisition of sample data Clinical parameters of ARDS patients with AKI were collected from Day 0 to Day 7 (8 consecutive days after onset or enrollment) in the data cohort, including age (years), height (cm), weight (kg), and body mass index (kg / m²). 2 Predicted weight (kg), sex, sepsis, trauma, multiple transfusions, aspiration, pneumonia, elective surgery, HIV, leukemia or lymphoma, solid tumor, immunosuppression, liver failure or cirrhosis, diabetes, chronic kidney disease, body temperature (°C), heart rate (beats / minute), systolic blood pressure (mmHg), diastolic blood pressure (mmHg), respiratory rate, ventilator-set respiratory rate, oxygen saturation (%), positive end-expiratory pressure (cmH2O), fluid output (mL / 24h), fluid intake (mL / 24h), platelet count (102). 9 The parameters included: serum creatinine (μmol / L), pH value, partial pressure of oxygen (mmHg), partial pressure of carbon dioxide (mmHg), oxygen concentration (%), and oxygenation index (PaO2 / FiO2 ratio, mmHg).

[0036] Simultaneously, the survival outcomes (28-day mortality rate, 60-day mortality rate, and ICU stay) of each ARDS patient with AKI in the data queue were obtained.

[0037] (2) Data preprocessing For the clinical indicators of each ARDS patient with AKI in the data cohort from Day 0 to Day 7, the clinical indicators of each day were cleaned, missing values ​​were filled and standardized to obtain the standardized clinical indicators of each ARDS patient with AKI in the data cohort. The cleaning process specifically involves: for numerical indicators in the clinical indicators of each ARDS patient with AKI in the data queue, values ​​exceeding the 0.02 quantile and 0.98 quantile are considered outliers and are processed for missing values. Then, samples with a clinical indicator missing rate of more than 20% are removed to obtain the cleaned clinical indicators of each ARDS patient with AKI in the data queue. Missing value imputation specifically involves using the Predicted Mean Matching (PMM) algorithm in the Multiple Imputation Method (MICE) to perform multiple imputation on the missing values ​​in the cleaned clinical indicators of each ARDS patient with AKI in the data queue, thereby obtaining the missing value imputed clinical indicators of each ARDS patient with AKI in the data queue. Standardization specifically involves: after imputing missing values ​​for continuous clinical indicators in each ARDS-AKI patient in the data cohort, Z-score standardization is performed with a mean of 0 and a standard deviation of 1 to obtain standardized clinical indicators for each ARDS-AKI patient in the data cohort; during the Z-score standardization process, the mean and standard deviation of each clinical indicator in each day's data for each ARDS-AKI patient in the data cohort are recorded.

[0038] (3) Joint latent class model (JointLCMM) classification The standardized clinical indicators of each ARDS patient with AKI in the data cohort obtained in step (2) were used as the input features of the Joint Latent Class Model (JointLCMM). The minimum class data of BIC (Bayesian Information Criterion) = 3 and the proportion of minimum latent class samples > 5% were used as the benchmarks. The subtype classification (3 types, recovery type, non-progressive type and progressive type) of each ARDS patient with AKI in the data cohort was used as the classification target of the JointLCMM model. Combined with the survival outcome of each ARDS patient in the data cohort, the ARDS patients with AKI in the data cohort were divided into recovery type (Class 1, 537 patients), non-progressive type (Class 2, 118 patients) and progressive type (Class 3, 202 patients), and the subtype labels (Class 1 to Class 3) of each ARDS patient in the data cohort were obtained. The recovery type refers to ARDS patients with AKI whose disease condition has improved; the non-progressive type refers to ARDS patients with AKI whose disease activity is relatively quiescent; and the progressive type refers to ARDS patients with AKI whose disease activity is persistent. Among them, the recovery survival outcome is better than the non-progressive survival outcome, which is better than the progressive survival outcome.

[0039] Next, for patients with ARDS complicated with AKI of three different subtypes in the data cohort, Kaplan-Meier survival analyses were performed at 28 days and 60 days, respectively, based on the survival outcomes of patients with ARDS complicated with AKI.

[0040] Next, the external validation cohort was processed using the method shown above to obtain standardized clinical indicators and classification labels for each ARDS patient in the external validation cohort.

[0041] (4) Training of the Transformer model The data cohort (containing 857 ARDS patients with AKI) was then divided into a training cohort and a validation cohort at an 8:2 ratio, and standardized clinical indicators and subtype labels for each ARDS patient in the training and validation cohorts were obtained.

[0042] Based on the standardized clinical indicators and classification labels of ARDS patients with AKI in the training cohort, the Transformer model was trained and validated using a validation cohort, as detailed below: Standardized clinical indicators (mostly time-series feature data) of each ARDS patient with AKI in the training cohort were obtained and feature vector sequences of each ARDS patient with AKI in the training cohort were constructed (sequence length = 8 days, feature dimension = d, where d is the number of multidimensional time-series features).

[0043] The feature vector sequences of each ARDS patient with AKI in the training queue were used as the input features of the Transformer model. The classification results of ARDS patients with AKI were used as the prediction target of the model. The model was trained by combining the classification labels of each ARDS patient with AKI in the training queue to obtain the trained Transformer model (a Transformer model for subtyping ARDS patients with AKI). The model was then validated using a validation queue. The Kappa, Recall, Precision, Accuracy, F1 score, and ROC-AUC scores for the recovery (Class 1), non-progressive (Class 2), and progressive (Class 3) subtypes were recorded during the validation process.

[0044] The hyperparameters during the Transformer model training process are set as follows: d_model=16, nhead=4, num_layers=4, dropout=0.3.

[0045] 2. Experimental Results After classifying ARDS patients with AKI in the data cohort based on a joint latent class model, the 28-day Kaplan-Meier survival analysis of the three different subtypes of ARDS with AKI is shown in the figure below. Figure 2 As shown in the figure, the 60-day Kaplan-Meier survival analysis of patients with ARDS complicated by AKI in three different subtypes is as follows. Figure 3 As shown.

[0046] The results showed that after classifying the ARDS patients with AKI in the data cohort using the Joint Latent Class Model (JointLCMM), the survival curves of the three subtypes of ARDS patients with AKI were significantly separated in both the 28-day and 60-day KM analyses (Log-rank p < 0.0001). The recovery type (Class 1) had the lowest mortality rate (1 death in the 28-day survival analysis, mortality rate 0.2%; 21 deaths in the 60-day survival analysis, mortality rate 3.9%), the progression-free type (Class 2) had a moderate mortality rate (85 deaths in the 28-day survival analysis, mortality rate 72%; 116 deaths in the 60-day survival analysis, mortality rate 98.3%), and the progressive type (Class 3) had the highest mortality rate (200 deaths in the 28-day survival analysis, mortality rate 99.0%; 200 deaths in the 60-day survival analysis, mortality rate 99.0%). There were significant differences among the three subtypes of ARDS patients with AKI.

[0047] This indicates that the Joint Latent Class Model (JointLCMM) is accurate in classifying ARDS patients with AKI, effectively dividing them into three subtypes and providing accurate classification labels for subsequent Transformer model training.

[0048] When using a validation queue for validation, the Transformer model had the following Kappa values: 0.650 (0.547-0.749), Recall: 0.802 (0.744-0.860), Precision: 0.833 (0.784-0.882), Accuracy: 0.802 (0.738-0.860), F1 score: 0.810 (0.753-0.863), ROC-AUC for Class 1 (recovery type): 0.871 (0.820-0.917), ROC-AUC for Class 2 (no progression type): 0.794 (0.672-0.908), and ROC-AUC for Class 3 (progression type): 0.843 (0.779-0.899).

[0049] The results showed that the trained Transformer model had an accurate ability to distinguish between recovery type (Class 1), non-progressive type (Class 2), and progressive type (Class 3) when subtyping ARDS patients with AKI. It was able to classify ARDS patients with AKI into different subtypes, and thus achieve personalized treatment for ARDS patients with AKI with different disease progression.

[0050] Example 2: Testing a Transformer model for classifying ARDS patients with AKI I. Experimental Methods Acquisition of the test cohort: Patients in the MIMIC IV database (N=364627) and the eICU-CRD database (N=200859) were excluded according to exclusion criterion 2 to obtain ARDS patients in the MIMIC IV database (N=3054) and the eICU-CRD database (N=3900). Exclusion criteria 2: (1) PEEP < 5 cmH2O; (2) PaO2 / FiO2 ration > 300 mmHg; (3) PEEP and PaO2 / FiO2 ration data missing; (4) not admitted to the ICU; (5) not admitted to the ICU for the first time or the same patient.

[0051] Next, ARDS patient data from experiments conducted between 2007 and 2013 in the ARDSNet database (including the EDEN, OMEGA, SAILS, and ALTA experiments) (N=1909), ARDS patients in the MIMIC IV database (N=3054), and ARDS patients in the eICU-CRD database (N=3900) were excluded according to exclusion criterion 1 shown in Example 1, resulting in a test cohort containing 1124 ARDS patients with AKI.

[0052] Experimental Group: Clinical indicators (Day 0–Day 7) of each ARDS-AKI patient in the test cohort were obtained (the clinical indicators shown in the sample data acquisition in Example 1, Day 0–Day 7 after enrollment or onset). These indicators were then standardized using Z-scores, combining the mean and standard deviation of each clinical indicator from the daily data of each ARDS-AKI patient in the training cohort obtained in Example 1. This yielded standardized clinical indicators for each ARDS-AKI patient in the test cohort. These standardized clinical indicators were then input into the Transformer model trained in Example 1 to obtain subtype classification results for each ARDS-AKI patient in the test cohort. A 28-day Kaplan-Meier survival analysis was performed on each patient based on their subtype classification and survival outcomes. Simultaneously, Kappa, Recall, Precision, Accuracy, F1 scores, and ROC-AUC values ​​for the recovery (Class 1), non-progressive (Class 2), and progressive (Class 3) subtypes were recorded when testing the trained Transformer model using the test cohort.

[0053] Control group 1: For each patient in the test cohort, Kdigo standard staging was performed according to the existing technology (KDOQI US commentary on the 2012 KDIGO clinical practice guideline for glomerulonephritis, PMID: 23871408), and 28-day KM survival curves were generated based on the Kdigo standard staging results and the survival outcomes of each patient.

[0054] II. Experimental Results When the trained Transformer model was tested using a test queue, the trained Transformer model (experimental group) divided the ARDS patients in the test into 688 recovering cases (Class 1), 147 non-progressive cases (Class 2), and 289 progressive cases (Class 3).

[0055] The 28-day Kaplan-Meier survival analysis of three different subtypes of ARDS in the experimental group and control group 1 is shown in the figure below. Figure 4 As shown in the figure, the 60-day Kaplan-Meier survival analysis of ARDS patients with three different subtypes in the experimental group and control group 1 is as follows. Figure 5 As shown.

[0056] The results show that the trained Transformer model shown in Example 1 of this invention can distinguish ARDS patients with AKI into three different subtypes. These three subtypes show significant differences in clinical characteristics and are significantly better than the current KDIGO staging standard in distinguishing between 28-day and 60-day mortality.

[0057] This demonstrates that the trained model shown in Embodiment 1 of the present invention can distinguish ARDS patients with AKI into different subtypes, thereby enabling personalized treatment for ARDS patients with AKI at different disease progressions.

[0058] Example 3: A subtyping device for ARDS patients with AKI based on the Transformer model A subtyping device for ARDS patients with AKI based on the Transformer model is characterized by including a data acquisition module, a data processing module, a subtyping module, and a result output module; The data acquisition module is used to acquire 8 consecutive days of clinical indicators for patients with ARDS complicated with AKI; the clinical indicators are the clinical indicators shown in claim 1. For numerical clinical indicators, the corresponding values ​​are directly acquired, and for binary clinical indicators, 0 / 1 corresponds to yes / no or male / female. The data processing module uses the average value and standard deviation of each clinical indicator in the daily data of each ARDS-AKI patient in the training queue obtained in Example 1 to perform Z-score standardization on the clinical indicators of ARDS-AKI patients obtained by the data acquisition module for 8 consecutive days, so as to obtain the standardized clinical indicators of ARDS-AKI patients. The subtyping module is the trained Transformer model shown in Example 1. The standardized clinical indicators of ARDS patients with AKI obtained by the data acquisition module are used as the input of the trained Transformer model. The trained Transformer model outputs the subtyping results of ARDS patients with AKI. The subtypes include remission type, non-progressive type, and progressive type; The result output module is used to output the results obtained by the fractal module.

[0059] Example 4 An electronic device An electronic device includes a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to cause the electronic device to perform the ARDS-AKI patient subtyping device based on the Transformer model shown in Embodiment 3.

[0060] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, other variations or modifications can be made based on the above description and ideas, and it is neither necessary nor possible to exhaustively describe all implementation methods here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. The application of a set of clinical indicators in the preparation of products for classifying patients with ARDS complicated by AKI, characterized in that, The clinical indicators include age, height, weight, body mass index, predicted weight, sex, presence of sepsis, presence of trauma, history of multiple blood transfusions, presence of aspiration, presence of pneumonia, history of elective surgery, presence of HIV, presence of leukemia or lymphoma, presence of solid tumors, presence of immunosuppression, presence of liver failure or cirrhosis, presence of diabetes, history of chronic kidney disease, body temperature, heart rate, systolic blood pressure, diastolic blood pressure, respiratory rate, ventilator-set respiratory rate, blood oxygen saturation, positive end-expiratory pressure, fluid output, fluid intake, platelet count, serum creatinine, pH value, partial pressure of oxygen, partial pressure of carbon dioxide, oxygen concentration, and oxygenation index.

2. A method for constructing a Transformer model for subtyping ARDS patients complicated with AKI, characterized in that, Includes the following steps: S1. Obtain clinical indicators for each sample in the training set; each sample in the training set is an ARDS patient complicated with AKI; the clinical indicators are the clinical indicators shown in claim 1. For numerical clinical indicators, the corresponding values ​​are obtained directly, and for binary clinical indicators, 0 / 1 corresponds to yes / no or male / female. S2. Preprocess the clinical indicators of each sample in the training set obtained in step S1 to obtain the standardized clinical indicators of each sample in the training set. During preprocessing, the mean and standard deviation of clinical indicators for each sample in the training set were recorded; S3. Using the standardized clinical indicators of each sample in the training set obtained in step S2 as the input of the Transformer model, and using the subtype classification of ARDS patients with AKI as the prediction target, the Transformer model is trained in combination with the subtypes of each sample in the training set to obtain a Transformer model that can perform subtype classification of ARDS patients with AKI. The subtypes of each sample in the training set are obtained by the Joint Latent Class Model (JointLCMM), and the subtypes include recovering, non-progressive, and progressive.

3. The construction method according to claim 2, characterized in that, The clinical indicators for each sample in the training set mentioned in step S1 are the clinical indicators for each sample in the training set for 8 consecutive days.

4. The construction method according to claim 2, characterized in that, The preprocessing described in step S2 is as follows: the clinical indicators of each sample in the training set obtained in step S1 are sequentially cleaned, missing values ​​are imputed, and Z-score is standardized to obtain the standardized clinical indicators of each sample in the training set.

5. The construction method according to claim 4, characterized in that, The data cleaning includes: Clinical indicators in the training set that exceed the 0.02 quantile and 0.98 quantile are removed, and samples with a clinical indicator missing rate > 20% are excluded.

6. The construction method according to claim 4, characterized in that, The missing value imputation is performed using the predicted mean matching algorithm in the multiple imputation method.

7. The construction method according to claim 2, characterized in that, The method for obtaining the subtypes of each sample in the training set in step S3 is as follows: The standardized clinical indicators of each sample in the training set are used as input to the Joint Latent Class Model (JLCMM). The JLCMM classifies each sample based on the standardized clinical indicators of each sample to obtain the subtype of each sample in the training set.

8. The Transformer model constructed by any one of the construction methods described in claims 2 to 7.

9. A subtyping device for ARDS patients with AKI based on the Transformer model, characterized in that, It includes a data acquisition module, a data processing module, a classification module, and a result output module; The data acquisition module is used to acquire clinical indicators of patients with ARDS complicated with AKI; the clinical indicators are the clinical indicators shown in claim 1. For numerical clinical indicators, the corresponding values ​​are directly acquired, and for binary clinical indicators, 0 / 1 corresponds to yes / no or male / female. The data processing module standardizes the clinical indicators of ARDS patients with AKI obtained by the data acquisition module based on the mean and standard deviation of the clinical indicators of each sample in the training set in the construction method shown in claim 3, so as to obtain standardized clinical indicators of ARDS patients with AKI. The subtyping module is the Transformer model as described in claim 8. It takes the standardized clinical indicators of ARDS patients with AKI obtained by the data acquisition module as input, and the Transformer model outputs the subtyping results of ARDS patients with AKI. The subtypes include remission type, non-progressive type, and progressive type; The result output module is used to output the results obtained by the fractal module.

10. An electronic device, characterized in that, It includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the apparatus of claim 9.

Images