Machine learning model training method and apparatus, electronic device, and readable storage medium

Abstract

The present application relates to a kind of machine learning model training method, device, electronic equipment and readable storage medium, machine learning model training method is used to predict death risk, comprising the following steps: step S1, obtains training data set: obtain the multiple sample data of historical patient before entering ICU, extract the sample feature of the sample data, and the known survival outcome label corresponding to each historical patient;Step S2, feature repair: the sample feature in the pre-set day before patient enters ICU is calculated time weighted average, and the model input sample feature is obtained;Step S3, model initialization;Step S4, iterative training;Step S5, parameter update: according to the loss function value update the model parameter, and the death risk prediction model of training completion is obtained.The machine learning model training method for predicting death risk according to the embodiment of the present application can reduce the difference between the predicted death risk value and the known survival end label, and optimize the model training effect.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of medical data processing and clinical decision support technology, and more specifically, to a machine learning model training method, apparatus, electronic device and readable storage medium. Background Technology

[0002] Cardiac arrest refers to the sudden cessation of cardiac mechanical activity, leading to the interruption of blood circulation and oxygen supply to vital organs. Globally, out-of-hospital cardiac arrest accounts for approximately 15%–20% of all deaths, with an overall survival rate of only 10%–30%. Despite continuous advancements in cardiopulmonary resuscitation and intensive care techniques, the clinical course of cardiac arrest patients in the early stages of ICU care is highly heterogeneous and rapidly changing, making early and accurate prediction of mortality risk (or quantification of ultimate indicators) extremely difficult. Mechanical ventilation and other life-saving measures are crucial in the management of cardiac arrest and significantly influence patient outcomes. Therefore, there is an urgent clinical need for reliable risk stratification tools capable of identifying high-risk patients early and assisting in critical bedside decisions.

[0003] Currently, prognostic assessment for CA patients mainly relies on the following techniques: First, traditional prognostic indicators based on initial heart rhythm, return of spontaneous circulation (ROSC) time, whether it was witnessed arrest, underlying comorbidities, and biomarkers such as neuron-specific enolase (NSE), international ratio of prothrombin time (INR), and serum lactate; second, neurological function assessments represented by the Glasgow Coma Scale (GCS); third, conventional disease severity scoring scales such as SOFA, APACHE IV, and OASIS; and fourth, machine learning methods introduced in recent years, which can integrate high-dimensional clinical data and characterize complex relationships between variables, showing potential in predicting CA prognosis.

[0004] However, the aforementioned existing technologies still have the following drawbacks: (1) Most existing prognostic models and scoring scales rely on static measurements of patients at admission or within the first 24 hours, ignoring the dynamic evolution of physiological and neurological parameters over time. This makes it difficult to reflect the rapid changes in CA condition and limits their ability to predict patient outcomes. (2) Conventional disease severity scores (such as SOFA, APACHE IV, OASIS) have limited discriminative ability, and their consistency index (C-index) is usually only at a low level of 0.58 to 0.63, which is far from meeting the clinical demand for predictive accuracy; (3) Existing prognostic models based on machine learning generally face problems such as large data heterogeneity, easy overfitting and insufficient external validation. Moreover, most of them use data from the first day of admission for modeling, which cannot effectively capture the clinical trajectory of patients over time. (4) Existing machine learning models generally suffer from insufficient interpretability (i.e., the "black box" problem), and the model decision-making process is difficult for clinicians to understand and trust, which seriously hinders its promotion and application in real clinical scenarios; (5) Existing methods lack in-depth analysis of the nonlinear and time-varying relationship between clinical parameters and prognosis, and cannot reveal the dynamic evolution of risk with the duration of ICU hospitalization, making it difficult to guide the optimal timing and method of data collection and monitoring. (6) Most existing solutions can only provide a single, static risk estimate at a certain point in time, lacking the ability to dynamically monitor and continuously update longitudinal patient data.

[0005] In summary, there is an urgent need in this field to train a model for predicting the risk of death in ICU patients with cardiac arrest that balances predictive accuracy, clinical interpretability, and effective utilization of longitudinal dynamic features, thereby reducing the discrepancy between the predicted mortality risk value and the known end-of-survival label. Summary of the Invention

[0006] One objective of this invention is to provide a new technical solution for a machine learning model training method, apparatus, electronic device, and readable storage medium.

[0007] According to a first aspect of the present invention, a method for training a machine learning model for predicting mortality risk is provided, comprising the following steps: Step S1, obtaining a training dataset: obtaining multiple sample data of historical patients before admission to the ICU, extracting sample features of the sample data, and a known survival outcome label corresponding to each historical patient; Step S2, feature repair: calculating a time-weighted average of multiple sample features within a preset number of days before the patient's admission to the ICU to obtain model input sample features; Step S3, model initialization: constructing an initial machine learning model and initializing the model parameters of the initial machine learning model; Step S4, iterative training: inputting the sample features without time-weighting processing and the model input sample features obtained after time-weighting processing into the initial machine learning model to obtain a predicted mortality risk value; calculating a loss function value based on the difference between the predicted mortality risk value and the known survival outcome label; Step S5, parameter update: updating the model parameters based on the loss function value to meet the basic conditions and obtain a trained mortality risk prediction model.

[0008] Optionally, the sample characteristics include Glasgow Coma Scale score and duration of mechanical ventilation, wherein the Glasgow Coma Scale score is time-weighted.

[0009] Optionally, the sample characteristics further include: age, respiratory rate, peripheral capillary oxygen saturation, systolic blood pressure, heart rate, prothrombin time on the first day of admission, and hemoglobin on the first day of admission, wherein the respiratory rate, peripheral capillary oxygen saturation, systolic blood pressure, and heart rate are time-weighted.

[0010] Optionally, the sample data includes: (1) laboratory tests and disease severity scores within the first day of admission: including white blood cell count, platelet count, calcium, alanine aminotransferase, aspartate aminotransferase, anion gap, albumin, total bilirubin, creatinine, chloride, blood glucose, hematocrit, hemoglobin, lactate, potassium, activated partial thromboplastin time, international ratio of prothrombin time, prothrombin time, sodium, and SOFA, APACHE IV, OASIS scores; (2) demographic characteristics and vital signs collected within the preset number of days: age, sex, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, body temperature and respiratory rate; (3) arterial blood gas analysis indicators collected within the preset number of days before the patient enters the ICU: pH, inhaled oxygen concentration, arterial oxygen partial pressure, arterial carbon dioxide partial pressure, bicarbonate, peripheral capillary blood oxygen saturation; (4) ICU treatment information: ICU hospitalization duration, ICU survival status, and mechanical ventilation duration.

[0011] Optionally, the sample data is screened in two stages, LASSO regression and multivariate Cox regression, to obtain the sample characteristics.

[0012] Optionally, in step S1, clinical data is extracted from the three ICU databases of eICU-CRD, MIMIC-IV, and MIMIC-III as the sample data.

[0013] Optionally, the preset hour is 24 hours and the preset number of days is 3 days.

[0014] According to a second aspect of the present invention, a model training apparatus for predicting mortality risk is provided, comprising: a training dataset acquisition module, wherein the training dataset acquisition module acquires multiple sample data of historical patients, extracts sample features of the sample data, and a known survival outcome label corresponding to each historical patient; a feature repair module, wherein the feature repair module calculates a time-weighted average of multiple sample features within a preset number of days to obtain model input sample features; a model initialization module, wherein the model initialization module constructs an initial machine learning model and initializes the model parameters of the initial machine learning model; an iterative training module, wherein the iterative training module inputs the sample features that have not undergone time weighting and the model input sample features obtained after time weighting into the initial machine learning model to obtain a predicted mortality risk value; calculates a loss function value based on the difference between the predicted mortality risk value and the known survival outcome label; and a parameter update module, wherein the parameter update module updates the model parameters based on the loss function value to satisfy basic conditions, thereby obtaining a trained mortality risk prediction model.

[0015] According to a third aspect of the present invention, an electronic device is provided, comprising: a processor and a memory, wherein computer program instructions are stored in the memory, wherein when the computer program instructions are executed by the processor, the processor causes the processor to perform the method described in any of the preceding claims.

[0016] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the method described in any of the preceding claims.

[0017] One beneficial effect of this invention is that the machine learning model training method for predicting mortality risk according to this invention can be applied to the establishment of machine learning models for predicting mortality risk. These machine learning models can be applied to predicting mortality risk in the intensive care unit (ICU) for patients with cardiac arrest (CA). By calculating a time-weighted average of sample features within a preset number of days, this method solves the technical problem of highly irregular collection frequencies of vital signs and laboratory test data in the ICU clinical environment, resulting in a large number of sample values ​​and time point misalignments. This allows the machine learning model for predicting mortality risk to capture the time-cumulative effect of physiological indicators, effectively completing irregularly sampled data and fully preserving the interactive behavioral characteristics of the patient's physiological state dynamically changing over time. This overcomes the technical deficiency of traditional methods relying on static anesthesia data, leading to low accuracy, and reduces the difference between the predicted mortality risk value and the known end-of-survival label.

[0018] Other features and advantages of the present invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of these embodiments.

[0020] Figure 1 A flowchart illustrating a method for training a machine learning model for predicting mortality risk, according to some embodiments, is shown. Figure 2 A schematic diagram of the hardware structure of an electronic device according to some embodiments is shown.

[0021] Figure Labels Electronic device 100; processor 110; memory 120. Detailed Implementation

[0022] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0023] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0024] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0025] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0026] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0027] The following is a flowchart illustrating the training method for a machine learning model used to predict mortality risk according to an embodiment of this application, with reference to the accompanying drawings.

[0028] like Figure 1 As shown, the method for training a machine learning model for predicting mortality risk according to an embodiment of the present invention includes the following steps: Step S1: Obtain the training dataset: Acquire multiple sample data from historical patients, extract sample features from the sample data, and obtain the known survival outcome label for each historical patient. To ensure the measurability and data security of model training, the sample data obtained in this step can be anonymized. Optionally, the known survival outcome label can be set as a binary label; for example, the label for death occurring within the target time window (e.g., 28 days after ICU admission) can be set to 1, and the label for survival can be set to 0.

[0029] Step S2, Feature Repair: Calculate the time-weighted average of multiple sample features within a preset number of days to obtain the model input sample features. The time weighting is Time-Weighted Average (TWA). Due to the highly irregular collection frequency of vital signs and laboratory test data in the ICU clinical environment, there are numerous instances of misalignment between sampling values ​​and time points. Directly using raw descriptive data or simple computational methods would prevent the model from capturing the cumulative effect of physiological indicators over time. Step S2 performs feature repair by importing the Time-Weighted Average (TWA) system. The specific logic includes: using the duration of indicator measurement as weight, performing weighted integration and mean averaging on multiple sample features within a preset number of days (e.g., 72 hours before ICU admission). The TWA method not only effectively handles irregularly sampled data but also fully preserves the interactive behavioral characteristics of patients' dynamic physiological changes over time, overcoming the technical deficiency of traditional methods that rely on static data, leading to low model prediction accuracy.

[0030] Step S3, Model Initialization: Construct an initial machine learning model and initialize its parameters. Optionally, the initial machine learning model can employ a tree-based ensemble learning algorithm (such as XGBoost, LightGBM, Random Forest) or a deep neural network (such as a multilayer ensemble machine (MLP) or Transformer architecture). For example, using a gradient-based boosting tree model, initialize the configuration by setting hyperparameters such as the maximum tree depth, learning rate, and minimum sample size for leaf nodes. Proper initialization lays the foundation for subsequent synchronization of clinical parameters and depth.

[0031] Step S4, Iterative Training: Input the unweighted sample features and the time-weighted model input sample features into the initial machine learning model to obtain the predicted mortality risk value; calculate the loss function value based on the difference between the predicted mortality risk value and the known survival outcome label. For example, in a single forward propagation, the model can output a probability value between 0 and 1, i.e., the predicted mortality value. Moreover, a preset loss function can be used to determine the relationship between this predicted value and the known survival outcome label. Furthermore, using binary cross-entropy as the loss function is beneficial for solving binary classification mortality prediction problems.

[0032] Step S5, Parameter Update: Update the model parameters based on the loss function value to meet the basic conditions and obtain the trained mortality risk prediction model. For example, use gradient descent algorithms (such as Adam, SGD optimizers) to backpropagate the loss function value, calculate the model parameters, and perform iterative updates. Optionally, the preset training termination conditions include at least one of the following: reaching the maximum number of iterations (Epochs), the loss function value converging to a minimum and no longer significantly recovering, or the consistency index (C-index) on the validation set not improving for several consecutive iterations (i.e., the early stage of the early stopping mechanism). By introducing the early stopping mechanism and cross-validation mechanism, the over-testing problem faced by traditional medical machine learning models can be effectively overcome, ensuring that the final output mortality risk prediction model still has strong generalization ability and high prediction accuracy when faced with external, unseen clinical data.

[0033] Understandably, existing medical data collection suffers from issues of "non-sampling" and "missing values." For example, the data collection frequency in the ICU is highly inconsistent (e.g., respiratory rate is recorded every minute, while blood tests may only be performed once a day). Directly feeding raw data into the model would result in a massive amount of missing values. In this embodiment of the invention, TWA (Time-Based Analysis) performs integration and smoothing on the time axis rules, effectively filling in data gaps and reducing the statistical bias caused by traditional simple imputation (such as mean imputation and forward imputation). Furthermore, bedside instruments in the ICU often produce effects or transient anomalies (e.g., patients turning over causing SpO2 noise, resulting in a momentary drop in value to zero). Using TWA can effectively filter out these high-frequency transient noise points, preventing the model from disconnecting from these fault signals during training, thereby significantly reducing the risk of over-neighbor errors. Moreover, using TWA to extract features achieves effective dimensionality reduction of the data, improving model convergence and inference speed.

[0034] Therefore, the machine learning model training method for predicting mortality risk according to embodiments of the present invention can be applied to the establishment of machine learning models for predicting mortality risk, and these models can be applied to the prediction of mortality risk in intensive care units (ICUs) for patients with cardiac arrest (CA). The machine learning model training method for predicting mortality risk according to embodiments of the present invention, by calculating a time-weighted average of sample features within a preset number of days, can solve the technical problem of highly irregular collection frequencies of vital signs and laboratory test data in the ICU clinical environment, resulting in a large number of sample values ​​and time point misalignments. This allows the machine learning model for predicting mortality risk to capture the time-cumulative effect of physiological indicators, effectively completing irregularly sampled data and fully preserving the interactive behavioral characteristics of the patient's physiological state dynamically changing over time. This solves the technical deficiency of traditional methods relying on static anesthesia data, leading to low accuracy, and reduces the difference between the predicted mortality risk value and the known end-of-survival label.

[0035] According to one embodiment of the present invention, the sample features include Glasgow Coma Scale score and mechanical ventilation duration. After calculating the time-weighted average, the Time-Weighted Glasgow Coma Scale score (TWA-GCS) and mechanical ventilation duration can be obtained, respectively. The Glasgow Coma Scale score has undergone time-weighting processing. The TWA-GCS and mechanical ventilation duration play an important role in reducing the discrepancy between the predicted mortality risk value and the known survival end label in machine learning model training methods.

[0036] In some specific embodiments of the present invention, in addition to Glasgow Coma Scale score and duration of mechanical ventilation, sample characteristics also include: age, respiratory rate, peripheral capillary oxygen saturation, systolic blood pressure, heart rate, prothrombin time (PT) on the first day of admission, and hemoglobin on the first day of admission. Respiratory rate, peripheral capillary oxygen saturation, systolic blood pressure, and heart rate are time-weighted. By calculating the time-weighted average, time-weighted respiratory rate (TWA-RR), time-weighted peripheral capillary oxygen saturation (TWA-SpO2), time-weighted systolic blood pressure (TWA-SBP), and time-weighted heart rate (TWA-HR) can be obtained. Using the above sample characteristics can reduce the discrepancy between the predicted mortality risk value and the known survival end label in machine learning model training methods.

[0037] According to one embodiment of the present invention, the sample data includes: (1) Laboratory tests and disease severity scores within the first 24 hours of admission: including white blood cell count, platelet count, calcium, alanine aminotransferase (ALT), aspartate aminotransferase (AST), anion gap, albumin, total bilirubin, creatinine, chloride, blood glucose, hematocrit, hemoglobin, lactate, potassium, activated partial thromboplastin time (PTT), international ratio of prothrombin time (INR), prothrombin time (PT), sodium, as well as SOFA, APACHE IV, and OASIS scores. This embodiment comprehensively extracts individual laboratory test indicators reflecting multiple organ dysfunction syndrome (MODS) and macroscopic metabolic scales characterizing the global disease severity (such as SOFA, APACHE IV, etc.) as joint feature inputs. This not only fully covers the core pathologies of post-cardiac arrest syndrome (PCAS), such as liver and kidney injury, exhaustive acidosis, and activation of the trigger system, but also guides the model to focus on exploring the residual correlation between laboratory latent indicators and macroscopic scores by using expert experience scales as prior baseline features. This greatly improves the convergence of the model under complex and heterogeneous data and effectively solves the technical problem of low accuracy caused by traditional static velocity scoring tools, realizing personalized and dynamic quantification of mortality risk prediction.

[0038] (2) Demographic characteristics and vital signs collected within a preset number of days: age, sex, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, body temperature, and respiratory rate. By using demographic characteristics and vital signs, the model can establish a personalized risk assessment core for patients with different physiological backgrounds, avoiding absolute numerical judgments and significantly improving the model's generalization ability in different heterogeneous populations. Including systolic blood pressure, diastolic blood pressure, and mean arterial pressure in the feature space not only allows the model to learn the patient's perfusion pressure level, but also allows for in-depth analysis of the imbalance between cardiac pumping function and peripheral vascular pressure through implicit indirect combinations of the three (such as the size of the pulse pressure), accurately identifying early cardiogenic shock. Furthermore, by using body temperature characteristics within a preset number of days, the model can effectively track the implementation of clinical target temperature management (TTM) and the transient response to cough, further enhancing the early warning of the risk of irreversible damage to the nervous system. In addition, respiratory rate is usually the most sensitive and earliest indicator of stress (such as compensatory deep and rapid breathing in promoting acidosis). By combining time-weighted average (TWA) processing, transient rapid breathing caused by turning over or suctioning can be filtered out, leaving the true respiratory trend. In this embodiment, by extracting demographic characteristics such as age and gender, as well as dynamic signs such as heart rate, multidimensional blood pressure, and respiratory rate, a global feature matrix containing the patient's core physiological predictions and macroscopic hemodynamic status is constructed for the model. This matrix generates deep data and feature interaction with blood gas consumption characteristics, effectively overcoming the technical deficiency of traditional single-dimensional data analysis in capturing the hidden compensation in the early stage of shock, and significantly improving the consistency and accuracy of mortality risk prediction.

[0039] (3) Arterial blood gas analysis parameters collected within the preset number of days before the patient's admission to the ICU: pH, inhaled oxygen concentration (FiO2), arterial oxygen partial pressure (PaO2), arterial carbon dioxide partial pressure (PaCO2), and bicarbonate (HCO3-) - Peripheral capillary oxygen saturation (SpO2) was also considered. It is understood that the core injury mechanism of post-cardiac arrest syndrome (PCAS) is hypoxia-reperfusion injury following systemic hypoxia resuscitation. In this embodiment, the above sample data can comprehensively cover the factors assessing this injury. For example, FiO2, PaO2, and SpO2 can be jointly assessed to evaluate true oxygenation status; PaCO2 can accurately reflect cerebral blood flow and ventilation function; pH and HCO3... - Microcirculation perfusion and induced acidosis can be determined. This embodiment comprehensively considers pH, FiO2, PaO2, PaCO2, and HCO3. -In addition to core factors and oxygenation indicators such as SpO2, it not only comprehensively assesses the complex injury-reperfusion injury and acid-base imbalance in cardiac arrest patients, but also enables machine learning models to deeply explore the correlation between ventilation connection and multi-organ function conversion, thereby significantly improving the lateness and sensitivity of early mortality risk prediction.

[0040] (4) ICU treatment information: ICU hospital stay duration, ICU survival status, and mechanical ventilation duration. In this embodiment, when constructing the training dataset, ICU hospital stay duration, mechanical ventilation duration, and ICU hospital status can be extracted as multidimensional outcome labels, which can reflect the real treatment process. Monitoring ICU hospital stay duration reflects the overall repair cycle after multiple organ injuries in patients; mechanical ventilation duration directly reflects the recovery status of the patient's central nervous system and the degree of dependence on life support equipment. Moreover, during the model training stage, introducing the above data can increase the supervision signal for model training, enabling the finally trained model to output more granular risk assessment results.

[0041] In some specific embodiments of the present invention, sample data are screened through a two-stage process of LASSO regression and multivariate Cox regression to obtain sample features. In this embodiment, after two-stage screening using LASSO regression and multivariate Cox regression, nine independent predictive variables were finally determined, which constitute the core input features of the predictive model: Glasgow Coma Scale score, duration of mechanical ventilation, age, respiratory rate, peripheral capillary oxygen saturation, systolic blood pressure, heart rate, first-day prothrombin time (first-day PT), and first-day hemoglobin. Understandably, by employing a two-stage joint screening of sample features using LASSO regression and multivariate Cox regression, this mechanism, compared to conventional single feature selection methods, not only achieves sparsity reduction of high-dimensional data and eliminates collinearity noise from the bottom layer, thus preventing interference with model reconstruction, but also, after dimensionality reduction suggested by LASSO, introduces a multivariate Cox proportional hazards regression model in the second stage. Unlike conventional logistic regression, Cox regression can combine "survival status" with "ICU input duration / survival time" for analysis, eliminating the interference of confounding factors and rigorously screening out risk features that are independent and statistically significant for survival outcomes. In this embodiment, by using LASSO for large-scale initial feature screening and then using Cox regression for refined statistical calibration, the input dimension of the feature mechanism is effectively compressed, significantly reducing the computational cost of the machine learning model, while greatly improving the generalization performance of the final model on externally available data.

[0042] According to one embodiment of the present invention, in step S1, clinical data is extracted as sample data from three types of ICU databases: eICU-CRD, MIMIC-IV, and MIMIC-III. That is, the present invention extracts the above four types of clinical sample data as candidate features from large-scale ICU databases (eICU-CRD, MIMIC-IV, and MIMIC-III). This embodiment constructs a heterogeneous feature space with high geographical and institutional diversity by combining single-center (MIMIC) and multi-center (eICU-CRD) databases, overcoming the problems faced by traditional single-center databases from a fundamental mechanism perspective, and endowing the model with strong cross-institutional generalization ability and universality.

[0043] In some specific embodiments of the present invention, the preset hour is 24 hours and the preset number of days is 3 days. That is to say, the sample data includes: laboratory tests and disease severity scores within 24 hours of admission; demographic characteristics and vital signs collected within 3 days; and arterial blood gas analysis indicators collected within 3 days, which can accurately capture the longitudinal evolution of clinical parameters over time.

[0044] In other words, to capture the longitudinal evolution of clinical parameters over time, embodiments of the present invention can take the measurements of parameters such as heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, body temperature, respiratory rate, pH, FiO2, PaO2, PaCO2, SpO2, and GCS on days 1, 2, and 3 (a total of 3 days) after the patient is admitted to the ICU, and calculate their time-weighted average (TWA) using the following formula: Expanding, we have: TWA = (1×Day1 + 2×Day2 + 3×Day3) / (1 + 2 + 3) = (Day1 + 2·Day2 + 3·Day3) / 6 In the formula, Day i This represents the measured value of the clinical parameter on day i (i = 1, 2, 3, corresponding to the 1st, 2nd, and 3rd days after admission to the ICU); i represents the weight assigned to day i, which is equal to the number of days itself, i.e., the weight is 1 for day 1, 2 for day 2, and 3 for day 3, increasing linearly with the number of days, thus assigning higher weight to the measurement value that is closer to the assessment point (recent); the denominator Σi = 1 + 2 + 3 = 6 is the sum of the weights of each day, used to normalize the weighted result.

[0045] As shown in the formula above, the TWA in this embodiment is a linearly weighted 3-day weighted average, with the weight increasing with the number of days (1→2→3), maximizing the contribution of the most recent (day 3) measurement. This design reflects the clinical logic that "the more recent the physiological state, the more valuable it is for prognostic prediction." Compared to a single static measurement upon admission or an equally weighted (arithmetic) average, it can more accurately reflect the evolution of the patient's clinical state in the early ICU stage, thereby significantly improving the model's predictive ability.

[0046] In this embodiment, by extracting the time-weighted average (TWA) features of longitudinal clinical parameters of patients 3 days before admission to the ICU, and combining them with an interpretable machine learning model, it can be applied to the accurate prediction and interpretable output of the 28-day ICU mortality risk of cardiac arrest patients.

[0047] The present invention also discloses a model training device for predicting mortality risk, comprising: a training dataset acquisition module, a feature repair module, a model initialization module, an iterative training module, and a parameter update module.

[0048] Specifically, the training dataset acquisition module acquires multiple sample data from historical patients, extracts sample features from the sample data, and obtains the known survival outcome label for each historical patient. The feature repair module calculates the time-weighted average of multiple sample features within a preset number of days to obtain the model input sample features. The model initialization module constructs an initial machine learning model and initializes the model parameters of the initial machine learning model. The iterative training module inputs the sample features that have not undergone time weighting and the model input sample features obtained after time weighting into the initial machine learning model to obtain the predicted mortality risk value. Based on the difference between the predicted mortality risk value and the known survival outcome label, the loss function value is calculated. The parameter update module updates the model parameters based on the loss function value. Once the basic conditions are met, the trained mortality risk prediction model is obtained.

[0049] The model training device for predicting mortality risk according to embodiments of the present invention accurately reflects the evolution trend of a patient's clinical status in the early ICU stage through a time-weighted tool (TWA), thereby improving the training effect and efficiency of the model.

[0050] The present invention also provides a schematic diagram of the hardware structure of an electronic device.

[0051] like Figure 2 As shown, the electronic device 100 includes a processor 110 and a memory 120. The memory 120 is used to store an executable computer program, and the processor 110 is used to execute a machine learning model training method for predicting mortality risk, as described in any of the above method embodiments, under the control of the computer program.

[0052] Each module of the model training device for predicting mortality risk described above can be implemented by the processor 110 in this embodiment executing the computer program stored in the memory 120, or it can be implemented by other structures, which are not limited here.

[0053] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0054] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0055] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0056] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0057] Various aspects of the present invention are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0058] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processor of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner; thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0059] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0060] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. It will be known to those skilled in the art that implementation in hardware, implementation in software, and implementation using a combination of software and hardware are equivalent.

[0061] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.

Claims

1. A method for training a machine learning model for predicting mortality risk, characterized in that, Includes the following steps: Step S1: Obtain training dataset: Obtain multiple sample data of historical patients before admission to ICU, extract sample features of the sample data, and the known survival outcome label corresponding to each historical patient; Step S2, Feature Repair: Calculate the time-weighted average of multiple sample features within a preset number of days before the patient's admission to the ICU to obtain the model input sample features; Step S3, Model Initialization: Construct the initial machine learning model and initialize its parameters; Step S4, Iterative Training: Input the sample features that have not undergone time weighting and the model input sample features that have undergone time weighting into the initial machine learning model to obtain the predicted mortality risk value; Calculate the loss function value based on the difference between the predicted mortality risk value and the known survival end label; Step S5: Parameter Update: Update the model parameters according to the loss function value to meet the basic conditions and obtain the trained mortality risk prediction model.

2. The method for training a machine learning model for predicting mortality risk according to claim 1, characterized in that, The sample characteristics include Glasgow Coma Scale score and duration of mechanical ventilation, with the Glasgow Coma Scale score being time-weighted.

3. The method for training a machine learning model for predicting mortality risk according to claim 2, characterized in that, The sample characteristics also include: age, respiratory rate, peripheral capillary oxygen saturation, systolic blood pressure, heart rate, prothrombin time on the first day of admission, and hemoglobin on the first day of admission. The respiratory rate, peripheral capillary oxygen saturation, systolic blood pressure, and heart rate are time-weighted.

4. The method for training a machine learning model for predicting mortality risk according to claim 3, characterized in that, The sample data includes: (1) Laboratory tests and disease severity scores on the first day of admission: including white blood cell count, platelet count, calcium, alanine aminotransferase, aspartate aminotransferase, anion gap, albumin, total bilirubin, creatinine, chloride, blood glucose, hematocrit, hemoglobin, lactate, potassium, activated partial thromboplastin time, international ratio of prothrombin time, prothrombin time, sodium, as well as SOFA, APACHE IV, and OASIS scores; (2) Demographic characteristics and vital signs collected within the preset number of days: age, sex, heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, body temperature and respiratory rate; (3) Arterial blood gas analysis indicators collected within the preset number of days before the patient is admitted to the ICU: pH, inhaled oxygen concentration, arterial oxygen partial pressure, arterial carbon dioxide partial pressure, bicarbonate, and peripheral capillary blood oxygen saturation. (4) ICU treatment information: ICU hospitalization duration, ICU survival status, and mechanical ventilation duration.

5. The method for training a machine learning model for predicting mortality risk according to claim 4, characterized in that, The sample data were screened in two stages, using LASSO regression and multivariate Cox regression, to obtain the sample characteristics.

6. The method for training a machine learning model for predicting mortality risk according to claim 4, characterized in that, In step S1, clinical data is extracted from the three ICU databases, namely eICU-CRD, MIMIC-IV, and MIMIC-III, as the sample data.

7. The method for training a machine learning model for predicting mortality risk according to claim 4, characterized in that, The preset number of days is 3 days.

8. A model training device for predicting mortality risk, characterized in that, include: The training dataset acquisition module acquires multiple sample data of historical patients, extracts sample features from the sample data, and the known survival outcome label corresponding to each historical patient. The feature repair module calculates a time-weighted average of multiple sample features within a preset number of days to obtain the model input sample features; A model initialization module, which constructs an initial machine learning model and initializes the model parameters of the initial machine learning model; An iterative training module inputs the sample features that have not undergone time-weighted processing and the model input sample features that have undergone time-weighted processing into the initial machine learning model to obtain the predicted mortality risk value. Calculate the loss function value based on the difference between the predicted mortality risk value and the known survival end label; The parameter update module updates the model parameters according to the loss function value to meet the basic conditions and obtain the trained mortality risk prediction model.

9. An electronic device, characterized in that, include: A processor and a memory, wherein computer program instructions are stored in the memory, wherein when the computer program instructions are executed by the processor, the processor causes the processor to perform the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to perform the method of any one of claims 1-7.

Images