TWI891506B - Method and system of cardiogenic shock death risk prediction - Google Patents

Abstract

A system of cardiogenic shock death risk prediction includes a processor that is configured to perform the following operations: obtaining multiple training data, each training data including a survival status and multiple medical records, each medical record including multiple characteristic parameters; for each training data, using a machine learning algorithm to fill in missing values in the medical records of the training data; using a machine learning algorithm to establish a cardiogenic shock death risk prediction model based on the training data after filling in missing values; using model interpretability technology to obtain a contribution of each of the characteristic parameters to the cardiogenic shock death risk prediction model, and selecting multiple major risk factors from the characteristic parameters based on the contributions.

Description

Method and system for predicting the risk of death from cardiogenic shock

The present invention relates to a risk prediction method and system, and in particular to a method and system for predicting the risk of death from cardiogenic shock.

Cardiogenic shock is shock caused by abnormal heart function. Patients often suffer from an inability to maintain adequate blood flow, leading to organ hypoxia and damage, ultimately resulting in death. Cardiogenic shock is a common cause of death among patients in intensive care units (ICUs). Therefore, using machine learning methods to extract information related to shock and death from electronic healthcare records (EHRs) has become a major research focus in critical care. However, the accuracy of predicting the risk of death from cardiogenic shock remains to be improved.

Therefore, improving the accuracy of predicting the risk of death from cardiogenic shock has become one of the issues that relevant technical fields seek to address.

Therefore, the purpose of the present invention is to provide a method and system for predicting the risk of death from cardiogenic shock that can overcome at least one shortcoming of the prior art.

Therefore, the present invention provides a method for predicting the risk of death due to cardiogenic shock, which is implemented by a computer system and includes the following steps: (A) obtaining a plurality of training data, each of which corresponds to a patient and includes a survival status of the patient and a plurality of medical records of the patient within a specified period of time, each of which includes a plurality of characteristic parameters related to the patient's physiological information and the treatment measures taken by the patient; (B) for each training data, using (C) using a machine learning algorithm to impute missing values in the medical records of the training data; (D) using a machine learning algorithm to establish a cardiogenic shock mortality risk prediction model based on the training data after missing value imputation; (E) using model interpretability techniques to obtain a contribution of each of the feature parameters to the cardiogenic shock mortality risk prediction model; and (E) selecting a plurality of key risk factors from the feature parameters based on the contributions.

Therefore, the present invention provides a cardiogenic shock death risk prediction system comprising a storage device and a processor.

The storage device stores a plurality of training data. Each training data corresponds to a patient and includes a vital status of the patient and a plurality of medical records of the patient within a specified period of time. Each medical record includes a plurality of characteristic parameters related to the patient's physiological information and the treatment measures administered to the patient.

The processor is connected to the storage device and includes a data pre-processing module, a model training module, and a model interpretation module.

The data pre-processing module is used to fill in missing values in the medical records of each training data using a machine learning algorithm.

The model training module is used to establish a cardiogenic shock mortality risk prediction model based on the training data after missing value filling using a machine learning algorithm.

The model interpretation module utilizes model interpretability techniques to obtain the contribution of each of the characteristic parameters to the cardiogenic shock mortality risk prediction model and, based on the contributions, selects a plurality of key risk factors from the characteristic parameters.

The effectiveness of the present invention lies in: first, by including the medical records of each patient within the specified time period in the training data, the cardiogenic shock mortality risk prediction model trained based on the training data can more fully capture the patient's physiological information and treatment measures that change dynamically over time, compared to previous models trained based on data at a single time point, thereby improving the prediction accuracy of the cardiogenic shock mortality risk prediction model. Secondly, by using a machine learning algorithm to fill missing values in the medical records before training, compared to traditional missing value filling methods that do not use machine learning, missing values can be filled more accurately based on the correlations between the values in the training data. This helps provide higher quality and more complete training data for training the cardiogenic shock mortality risk prediction model, and also improves the prediction accuracy of the cardiogenic shock mortality risk prediction model. Furthermore, by utilizing model interpretability techniques to obtain these key risk factors, not only can the cardiogenic shock mortality risk prediction model be adjusted based on these key risk factors to improve its accuracy, but it can also allow medical staff to understand the correlation between various physiological information and a patient's risk of death from cardiogenic shock, thereby adjusting patient care based on these key risk factors to improve prognosis.

1: Cardiogenic shock death risk prediction system

11: Storage device

12: Processor

121: Data pre-processing module

122: Model training module

123: Model Interpretation Module

124: Prediction Module

S21~S24: Steps

Other features and benefits of the present invention will be more clearly demonstrated in the accompanying drawings, wherein: FIG1 is a block diagram illustrating an embodiment of the cardiogenic shock mortality risk prediction system of the present invention; and FIG2 is a flow chart illustrating how a processor of the embodiment executes a cardiogenic shock mortality risk prediction program.

Before the present invention is described in detail, it should be noted that similar components are represented by the same reference numerals in the following description.

Referring to FIG1 , an embodiment of the cardiogenic shock death risk prediction system 1 of the present invention includes a storage device 11 and a processor 12 connected to the storage device 11.

The storage device 11 can be implemented as, for example, but not limited to, any type of fixed or removable random access memory (RAM), hard disk drive (HDD), solid state drive (SSD), or similar device, or a combination thereof, and is used to store multiple training data sets. Each training data set corresponds to a patient and includes the patient's vital status and multiple medical records for the patient over a specified period of time. Each medical record includes multiple characteristic parameters related to the patient's physiological information and the treatments administered to the patient.

In this embodiment, each training data includes the medical records of the corresponding patient recorded every two hours within twenty-four hours after a time point (i.e., the designated time period is twenty-four hours), and the survival status indicating whether the corresponding patient died within one hundred and sixty-eight hours (i.e., seven days) after the time point. The characteristic parameters of each medical record include age, sex, height, weight, 24-hour urine output, Acute Physiology and Chronic Health Evaluation II (APACHE II) value, body mass index, heart rate, body temperature, mean arterial blood pressure (ABP mean), arterial blood gas (ABG) pH, ABG oxygen partial pressure, ABG carbon dioxide partial pressure, ABG bicarbonate concentration, ABG lactate value, ABG glucose value, hematocrit, platelet concentration, B-type Natriuretic Peptide (BNP), and ABG urea-stimulating peptide (BNP). Peptide concentration, C-reactive protein concentration, creatinine concentration, positive end expiratory pressure (PEEP), inspired fraction of oxygen Oxygen (abbreviated as FiO2), the total amount of fluid infused every two hours, the total amount of fluid output every two hours, a ventilator usage parameter indicating whether the corresponding patient uses a ventilator, a dopamine usage parameter indicating whether the corresponding patient uses dopamine, a dobutamine usage parameter indicating whether the corresponding patient uses dobutamine, a norepinephrine usage parameter indicating whether the corresponding patient uses norepinephrine, an epinephrine usage parameter indicating whether the corresponding patient uses epinephrine, and a vasopressin usage parameter indicating whether the corresponding patient uses vasopressin.

In this embodiment, the processor 12 includes a data pre-processing module 121, a model training module 122, a model interpretation module 123, and a prediction module 124. The operation of each module in the processor 12 will be described in detail below.

Below, referring to FIG. 2 , a detailed description will be given of how the processor 12 executes a cardiogenic shock death risk prediction process. The process includes steps S21 to S24.

In step S21, for each training data set, the data pre-processing module 121 uses a machine learning algorithm to perform missing value imputation on the medical records in the training data.

More specifically, for each training data set, the data pre-processing module 121 first uses the Last Observation Carried Forward (LOCF) method to fill in missing values for the corresponding feature parameters in the medical records recorded later in the training data using the values of the feature parameters in the medical records recorded earlier in the training data. Next, for each training data set, the data pre-processing module 121 uses the Miss Forest algorithm based on random forests to fill in the remaining missing values in the medical records in the training data. This is because even after using the LOCF method to impute missing values, some individual characteristic parameters in the patient's medical records still had no values recorded within the specified time period, making them impossible to impute using the LOCF method. Considering the possible correlation between the values of these characteristic parameters, the Miss Forest algorithm was used to fill in the remaining missing values to improve the completeness of the training data.

In step S22, the model training module 122 uses a machine learning algorithm based on the training data after missing value filling to establish a cardiogenic shock mortality risk prediction model.

In this embodiment, the model training module 122 uses a 5-fold cross-validation method, an Adam optimizer, and a binary focal loss function. It utilizes a long short-term memory (LSTM) neural network to build a cardiogenic shock mortality risk prediction model based on the missing value-filled training data. The cardiogenic shock mortality risk prediction model includes an LSTM layer and a fully connected layer. The LSTM layer includes four LSTM units, and the fully connected layer includes four neurons.

In step S23, the model interpretation module 123 utilizes model interpretability technology to obtain the contribution of each of the characteristic parameters to the cardiogenic shock mortality risk prediction model and, based on the contributions, selects a plurality of key risk factors from the characteristic parameters.

In this embodiment, the model interpretation module 123 utilizes the SHAP (SHapley Additive exPlanations) interpreter to obtain the Shapley value of each of the characteristic parameters as the contribution, and selects the characteristic parameters with the top ten highest Shapley values as the primary risk factors.

It is worth noting that the model interpretation module 123 also outputs these key risk factors, enabling medical staff to understand which factors have the greatest impact on patients. This ensures that the information learned by the cardiogenic shock mortality risk prediction model is consistent with clinical expertise, thereby achieving effective risk prediction.

In step S24, the prediction module 124 obtains multiple pending medical records of a patient within the specified time period and, based on the pending medical records, utilizes the cardiogenic shock mortality risk prediction model to obtain a prediction result indicating the patient's mortality risk.

In this embodiment, the prediction module 124 utilizes the cardiogenic shock death risk prediction model based on the patient's 24-hour waiting medical records to obtain a prediction result indicating the patient's risk of death from cardiogenic shock within seven days.

In summary, first, by including each patient's medical records within the specified time period in the training data, the cardiogenic shock mortality risk prediction model trained based on the training data can more fully capture the patient's dynamically changing physiological information and treatment measures over time than previous models trained based on data from a single time point, thereby improving the prediction accuracy of the cardiogenic shock mortality risk prediction model. Secondly, by using a machine learning algorithm to fill missing values in the medical records before training, compared to traditional missing value filling methods that do not use machine learning, missing values can be filled more accurately based on the correlations between the values in the training data. This helps provide higher quality and more complete training data for training the cardiogenic shock mortality risk prediction model, and also improves the prediction accuracy of the cardiogenic shock mortality risk prediction model. Furthermore, by utilizing model interpretability techniques to obtain these key risk factors, not only can these key risk factors be used to subsequently adjust the cardiogenic shock mortality risk prediction model to improve its accuracy, but they can also help medical staff understand the correlation between various physiological information and a patient's risk of death from cardiogenic shock, thereby adjusting patient care based on these key risk factors to improve prognosis. Therefore, the purpose of this invention can be truly achieved.

However, the above descriptions are merely examples of the present invention and should not be construed to limit the scope of the present invention. Any simple equivalent variations and modifications made within the scope of the patent application and the contents of the patent specification are still covered by the present patent.

1: Cardiogenic shock death risk prediction system

11: Storage device

12: Processor

121: Data pre-processing module

122: Model training module

123: Model Interpretation Module

124: Prediction Module

Claims (6)

A method for predicting the risk of death from cardiogenic shock is implemented using a computer system and comprises the following steps: (A) obtaining a plurality of training data, each training data corresponding to a patient and including the patient's survival status and a plurality of medical records of the patient within a specified time period, each medical record including a plurality of characteristic parameters indicating the patient's physiological information and the treatment measures administered to the patient; (B) for each training data, using the last observed value estimation method and the Miss Forest algorithm based on random forests to fill in missing values in the medical records of the training data; (C) Using the training data after missing value imputation, a machine learning algorithm is used to establish a prediction model for the risk of death from cardiogenic shock. The prediction model comprises a long-term and short-term memory neural network layer and a fully connected layer. (D) Using model interpretability techniques, a contribution of each of the feature parameters to the prediction model for the risk of death from cardiogenic shock is obtained. (E) Based on the contributions, a plurality of key risk factors are selected from the feature parameters. The method for predicting the risk of death due to cardiogenic shock as described in claim 1 further comprises, after step (C), the following steps: (F) obtaining multiple pending medical records of a patient to be tested within the specified time period; and (G) obtaining a prediction result indicating the magnitude of the patient's risk of death using the cardiogenic shock risk of death prediction model based on the pending medical records. A method for predicting the risk of death due to cardiogenic shock as described in claim 1, wherein, in step (A), the characteristic parameters include a gender value, an intake oxygen partial pressure, a platelet concentration, a lactate value from arterial blood gas analysis, a ventilator usage parameter indicating whether the corresponding patient is using a ventilator, a positive end-expiratory pressure, a mean arterial pressure, a creatinine concentration, and an arterial blood gas analysis bicarbonate concentration. A cardiogenic shock mortality risk prediction system comprises: A storage device storing a plurality of training data, each training data corresponding to a patient and including the patient's survival status and a plurality of medical records of the patient within a specified period of time, each medical record including a plurality of characteristic parameters indicating the patient's physiological information and the treatment measures administered to the patient; and A processor connected to the storage device and comprising: A data pre-processing module for performing missing value imputation on the medical records of each training data using a last observed value estimation method and a random forest-based Miss Forest algorithm; A model training module is configured to establish a cardiogenic shock mortality risk prediction model using a machine learning algorithm based on the training data after missing value imputation. The cardiogenic shock mortality risk prediction model comprises a long-term and short-term memory neural network layer and a fully connected layer. A model interpretation module is configured to utilize model interpretability techniques to obtain a contribution of each of the feature parameters to the cardiogenic shock mortality risk prediction model and, based on the contributions, select a plurality of key risk factors from the feature parameters. A cardiogenic shock death risk prediction system as described in claim 4, wherein the processor further includes a prediction module for obtaining multiple pending medical records of a patient to be tested within a specified period of time, and using the cardiogenic shock death risk prediction model to obtain a prediction result indicating the patient's death risk based on the pending medical records. A cardiogenic shock death risk prediction system as described in claim 4, wherein the characteristic parameters include a gender value, an intake oxygen partial pressure, a platelet concentration, a lactate value from arterial blood gas analysis, a ventilator usage parameter indicating whether the corresponding patient is using a ventilator, a positive end-expiratory pressure, a mean arterial pressure, a creatinine concentration, and an arterial blood gas analysis bicarbonate concentration.