TWI836965B - Preoperative risk prediction system - Google Patents

Abstract

The present invention provides a preoperative risk prediction system, comprising: a database unit, which stores a plurality of historical data of a plurality of patients; and a processing unit, which is electrically connected to the database unit and includes a knowledge-based model, a time series overlay knowledge-base model and a machine learning model. Wherein, the plurality of historical data is input to the time series overlay knowledge-base model, which performs a time series overlay data processing on the plurality of historical data to generate a plurality of time series overlay knowledge-base data. Further, the plurality of time series overlay knowledge-base data is input to the knowledge-base model, which performs a data fusion processing on the plurality of time series overlay knowledge-base data to generate a plurality of fusion knowledge-base data. The plurality of fusion knowledge-base data is input to the machine learning model, which performs a knowledge base-based machine learning on the plurality of t fusion knowledge-base data to generate a plurality of output data, which includes the predicted length of hospital stay data, the predicted mortality data and the analysis of impact factors of each patient.

Description

Preoperative risk prediction system

The present invention relates to a risk prediction system, and in particular, to a pre-operative risk prediction system that uses a machine learning method based on knowledge base learning to accurately predict postoperative mortality and hospitalization days of patients before surgery.

Generally speaking, before the patient undergoes surgery, the anesthesiologist will conduct a preoperative anesthesia consultation. The doctor will review the patient's medical history, physical examination report and other relevant information. The patient will inform the doctor of his or her physical condition and medication during the preoperative anesthesia consultation, so that the anesthesiologist can assess the risks of anesthesia and surgery based on the above information and in combination with anesthesia risk grading tools such as ASAPS (American society of anesthesiologists physical status classification system) and POSPOM (preoperative score to predict postoperative mortality), thereby ensuring the safety of the patient during subsequent anesthesia and surgery.

However, most current anesthesia preoperative notes are usually made by doctors after obtaining relevant information about the patient, and then use the doctor's experience to conduct a risk assessment based on the obtained information. This method may affect the final result due to differences in doctors' personal experiences. Accuracy of judgment. In addition, using models to assist anesthesiologists before they complete risk assessment classification will require a predictive model that does not use anesthesia risk classification tools such as ASA classification as input data. In recent years, with the technological With the development of the medical industry, artificial intelligence (AI) related technologies such as machine learning (ML) and natural language processing (NLP) are gradually being used in the medical industry. Therefore, by applying artificial intelligence technology In the risk assessment of anesthesia preoperative notes, it will help assist physicians to perform preoperative risk assessment.

Therefore, how to design a pre-operative risk prediction system that can accurately predict the patient's postoperative mortality rate and length of hospital stay by using machine learning based on knowledge base learning to solve the technical problems of the previous technology is an important topic studied by the inventor of the present invention.

In view of this, the inventor has accumulated many years of research and practical experience in related fields and provides a pre-operative risk prediction system to improve the problems described in the previous technology.

In order to solve the problems existing in the above-mentioned prior art, the present invention provides a pre-operative risk prediction system, which includes: a database unit that stores a plurality of historical data of a plurality of patients; and a processing unit that is electrically connected to the database. unit, and includes time series overlay knowledge base model, knowledge base model (knowledge-based model), and machine learning model. Among them, a plurality of historical data are input into the time series overlay knowledge base model respectively, and the time series overlay knowledge base model performs time series overlay data processing based on the plurality of historical data to generate corresponding plurality of time series overlay knowledge base data, and a plurality of time series overlay knowledge base The data includes average length of stay and mortality data for each patient. Furthermore, a plurality of time series superimposed knowledge base data are input into the knowledge base model, and the knowledge base model performs data fusion processing based on a plurality of time series superimposed knowledge base data to generate a plurality of fused knowledge base data, and the plurality of fused knowledge base data includes each patient. average length of stay and mortality data. Moreover, a plurality of fused knowledge base data are input to the machine learning model, and the machine learning model is executed based on a plurality of fused knowledge base data. Machine learning based on the knowledge base generates a plurality of output data, and the plurality of output data includes the predicted length of stay, predicted mortality data and influencing factor analysis data for each patient.

Preferably, the plurality of historical data includes a plurality of data sets respectively from different medical institutions.

Preferably, the time series superposition knowledge base model is configured to extract the surgical codes, death markers, and years of multiple historical data from multiple data sets, and perform mortality statistics for each surgical code in each year to output multiple time series superposition knowledge base data to establish a knowledge base. In the knowledge base, multiple time series superposition knowledge base data marked as the mth year output by the time series superposition knowledge base model will be superimposed on multiple time series superposition knowledge base data marked as the m-1th year, where m is a positive integer.

Preferably, the knowledge base model is connected to the knowledge base, and the knowledge base model is configured as follows: identifying the source medical institutions of the plurality of historical data, and performing pre-fusion based on the past data of the source medical institutions in the knowledge base; based on The past data of medical institutions other than the source medical institution in the knowledge base are pre-fused with surgery codes; or data is filled based on the average mortality rate last year of the source medical institution or medical institutions other than the source medical institution. .

Preferably, the machine learning model is configured to extract multiple feature parameters from multiple fused knowledge base data, and the multiple feature parameters include at least one selected from International Statistical Classification of Diseases and Related Health Problems (ICD) codes, surgery codes, historical average mortality rates of each surgery, Diagnosis Related Groups (DRG) codes, blood draw items, gender, age, number of hospital stays before surgery, number of surgeries in a single hospital stay, and hospitalization department.

Preferably, the pre-surgery risk prediction system of the present invention further includes a data pre-processing model, which is configured to perform data pre-processing on a plurality of historical data in each data set to fill in a plurality of historical data. Missing values in historical data, standardize the plurality of historical data based on the training data set, encode the plurality of historical data using the training data set, and output the plurality of historical data processed by the data to the knowledge base model .

Preferably, the pre-surgery risk prediction system of the present invention further includes a user interface for inputting historical data of patients before surgery, thereby obtaining data including the patient's predicted length of stay, predicted mortality data and influencing factor analysis data. Output data.

Preferably, the time series stacking knowledge base model, the knowledge base model, and the machine learning model are selected from at least one of the group consisting of a gradient boosting decision tree model, a histogram gradient boosting decision tree model, a random forest tree model, and a long short-term memory model.

To sum up, the pre-surgery risk prediction system of the present invention includes a data pre-processing model, a knowledge base model, a time-series overlay knowledge base model, and a machine learning model, thereby using different algorithms to target various data from different medical institutions. Centralized historical data for data processing. Specifically, historical data are input into the time series overlay knowledge base model, and the time series overlay knowledge base model performs time series overlay data processing based on the historical data to generate corresponding time series overlay knowledge base data. The time series overlay knowledge base data includes the average length of stay of each patient. and mortality data. Moreover, the time series superimposed knowledge base data is continuously input to the knowledge base model, and the knowledge base model performs data fusion processing based on the time series superimposed knowledge base data to generate fused knowledge base data. The fused knowledge base data includes the average length of stay and death of each patient. rate data. Finally, the fused knowledge base data is input into the machine learning model, and the machine learning model performs knowledge base-based machine learning based on the fused knowledge base data to generate output data. The output data includes the predicted length of stay, predicted mortality data, and data for each patient. Impact factor analysis data can be used as a reference for physicians in pre-operative anesthesia notes. Thereby, the pre-surgery risk prediction system of the present invention can accurately predict the postoperative hospitalization days and death of patients through machine learning methods based on knowledge bases. rate, and according to the experimental results, the pre-operative risk prediction system of the present invention has an accuracy of more than 90% in predicting the length of stay and mortality.

Through the above configuration, the preoperative risk prediction system of the present invention can cooperate with the preoperative anesthesia consultation system to assist doctors in performing preoperative risk assessment, thereby reducing errors in human judgment and improving the accuracy of preoperative risk assessment, which is conducive to ensuring the safety of subsequent anesthesia and surgery for patients.

1: Preoperative risk prediction system

10: Database unit

111: First data set

112: Second data set

113:Third data set

20: Processing unit

21: Time series superposition knowledge base model

22: Knowledge Base

23: Knowledge base model

24: Machine learning model

241: Output data

25: Data pre-processing model

30: User Interface

Figure 1 is a schematic architectural diagram of a pre-surgery risk prediction system according to an embodiment of the present invention; Figure 2 is a schematic flow chart of time-series superposition data processing of a time-series superposition knowledge base model of the pre-surgery risk prediction system according to an embodiment of the present invention.

Figure 3 is a schematic flow chart of the data fusion processing of the knowledge base model of the pre-operative risk prediction system according to an embodiment of the present invention; Figure 4 is a schematic block diagram of the pre-operative risk prediction system according to an embodiment of the present invention; Figure 5 is a receiver operating characteristic curve (ROC) diagram for mortality prediction of the pre-operative risk prediction system according to an embodiment of the present invention; Figure 6 is a pre-operative risk according to an embodiment of the present invention Receiver operating characteristic curve (ROC) diagram of the prediction system regarding the prediction of hospitalization days; Figures 7A to 7C are the pre-surgery risk prediction system according to an embodiment of the present invention regarding whether to use the knowledge base in histogram gradient boosting decision-making Schematic diagram of the comparison results of the bootstrap method on the Histogram Gradient Boosting Decision Tree (HGBT) model; and Figures 8A to 8C show the pre-surgery risk prediction system according to an embodiment of the present invention on whether to use the knowledge base on the multiple sampling method (bootstrap) on the gradient boosting decision tree (Gradient Boosting Decision Tree 300, GBT300) model. Schematic diagram of the comparison results of the multiple sampling method (bootstrap).

In order to help the examiner understand the technical features, content and advantages of the present invention and the effects it can achieve, the present invention is described in detail as follows with the accompanying drawings and in the form of an embodiment. The drawings used therein are only for illustration and auxiliary explanation purposes, and may not be the true proportions and precise configurations after the implementation of the present invention. Therefore, the proportions and configurations of the attached drawings should not be interpreted to limit the scope of rights of the present invention in actual implementation.

It should be understood that although the terms "first", "second", etc. may be used in the present invention to describe various elements, components, regions, layers and/or parts, these elements, components, regions, layers and/or parts should not be limited by these terms. These terms are only used to distinguish one element, component, region, layer and/or part from another element, component, region, layer and/or part. Therefore, the "first element", "first component", "first region", "first layer" and/or "first part" discussed below may be referred to as "second element", "second component", "second region", "second layer" and/or "second part" without departing from the spirit and teachings of the present invention.

Additionally, the terms "comprises" and/or "includes" refer to the presence of stated features, regions, integers, steps, operations, elements and/or parts, but do not exclude the presence of one or more other features, regions, integers, steps, operations , elements, parts and/or combinations thereof.

Unless otherwise defined, all terms (including technical and scientific terms) used in the present invention have the same meanings as commonly understood by persons of ordinary skill in the art to which the present invention belongs. It is understood that those terms as defined in commonly used dictionaries should be interpreted as having definitions consistent with their meanings in the context of the relevant art and the present invention, and will not be interpreted as idealized or overly formal meanings unless expressly defined as such herein.

The present invention will be described in further detail below with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and are therefore not intended to limit the present invention.

Please refer to Figures 1 to 6, Figures 7A to 7C, and Figures 8A to 8C, Figure 1 is a schematic diagram of the architecture of a preoperative risk prediction system according to an embodiment of the present invention; Figure 2 is a schematic diagram of the flow of time-series superposition data processing of a time-series superposition knowledge base model of a preoperative risk prediction system according to an embodiment of the present invention; Figure 3 is a schematic diagram of the flow of data fusion processing of a knowledge base model of a preoperative risk prediction system according to an embodiment of the present invention; Figure 4 is a schematic block diagram of a preoperative risk prediction system according to an embodiment of the present invention; Figure 5 is a receiver operating characteristic curve (receiver operating characteristic) for mortality prediction of a preoperative risk prediction system according to an embodiment of the present invention. curve, ROC) graph; Figure 6 is a receiver operating characteristic curve (ROC) graph for hospital stay prediction of a preoperative risk prediction system according to an embodiment of the present invention; Figures 7A to 7C are schematic diagrams of comparison results of whether to use the knowledge base in a multiple sampling method (bootstrap) on a histogram gradient boosting decision tree (HGBT) model of a preoperative risk prediction system according to an embodiment of the present invention; and Figures 8A to 8C are schematic diagrams of comparison results of whether to use the knowledge base in a multiple sampling method (bootstrap) on a gradient boosting decision tree (Gradient Boosting Decision Tree 300, GBT300) model of a preoperative risk prediction system according to an embodiment of the present invention.

As shown in Figure 1, according to an embodiment of the present invention, a pre-surgery risk prediction system 1 is provided, which includes: a database unit 10, which stores a plurality of historical data of a plurality of patients; and a processing unit 20, which electrically It is connected to the database unit 10 and includes a time series superposition knowledge base model 21, a knowledge base model (knowledge-based model) 23, and a machine learning model 24. Among them, a plurality of historical data are respectively input to the time series overlay knowledge base model 21, and the time series overlay knowledge base model 21 performs time series overlay data processing based on the plurality of historical data, and generates a corresponding plurality of time series overlay knowledge base data, and a plurality of time series overlay knowledge base data. The knowledge base data includes average length of stay and mortality data for each patient. Furthermore, a plurality of time series superposed knowledge base data are input to the knowledge base model 23, and the knowledge base model 23 performs data fusion processing based on a plurality of time series superimposed knowledge base data to generate a plurality of fused knowledge base data, and the plurality of fused knowledge base data includes Data on the average length of stay and mortality of each patient. Finally, the plurality of fused knowledge base data are input to the machine learning model 24, and the machine learning model 24 uses the plurality of fused knowledge base data to perform knowledge base-based machine learning to generate a plurality of output data 241, and the plurality of output data 241 includes each The patient's predicted hospitalization days data, predicted mortality data and influencing factor analysis data.

Specifically, the predicted hospitalization days data can be the number of days the patient is expected to be hospitalized after surgery, such as 3 days, 5 days, etc., and the predicted mortality data can be death 24 hours after surgery, death 48 hours after surgery, and death 72 hours after surgery. Mortality rates for conditions such as death and in-hospital death. In addition, the influencing factor analysis data can be influencing factors learned by each model that are associated with post-operative mortality prediction, which can include heart-related test data (for example, myocardial enzymes, B-type natriuretic diuretics) Peptides (B-type natriuretic peptide, BNP, etc.), kidney-related test data (for example, creatinine (creatine), blood urea nitrogen (BUN), etc.), coagulation-related test data (international normalized ratio (international normalized ratio, INR), and international diseases and related health issues International Statistical Classification of Diseases and Related Health Problem (ICD) codes, etc.

Specifically, the plurality of historical data include the first data set 111, the second data set 112 and the third data set 113 respectively from different medical institutions. The first data set 111 is a data set of major hospitals that use the preoperative risk prediction system 1 of the present invention to assist doctors in performing preoperative risk assessment, that is, an internal data set, and the second data set 112 and the third data set 113 are from Data sets from other hospitals, i.e. external data sets. In the following, the first data set 111, the second data set 112 and the third data set 113, which are divided into internal data sets and external data sets, will be used to verify the accuracy of the pre-surgery risk prediction system 1, so it will not be discussed here. Repeat. Among them, the first data set 111, the second data set 112 and the third data set 113 may include, for example, the age of the patient, pre-operative blood test data, the number of surgeries in this hospitalization, the date of past operations, and the number of operations performed before the operation. Numerical data of hospitalization days, as well as preoperative diagnosis codes including comorbidities (for example, 5 ICD codes in sequence), scheduled surgery codes (PCS codes), and hospitalization diagnosis related group (Diagnosis Related Groups, DRG) codes , surgical department code, and gender and other category information.

In this embodiment, the first data set 111 is a data set from Kaohsiung Medical University Chung-Ho Memorial Hospital (KMUH), and the second data set 112 and the third data set 113 are respectively It is a data set from Kaohsiung Municipal Hsiao-Kang Hospital (KMHK) and Kaohsiung Municipal Ta-Tung Hospital (KMTTH), but the present invention is not limited thereto. In other embodiments, data from other medical institutions may be further used.

In addition, in this embodiment, the multiple historical data are the data of patients after de-identification, which excludes patients who are under the age of 20 at the time of surgery, excludes patients who receive local anesthesia, and excludes data of outpatient surgeries that are not performed in the operating room, thereby increasing the accuracy of subsequent predictions of postoperative mortality and length of hospital stay. In addition, the electronic medical records of patients record in-hospital death within 30 days after surgery and cross-check the death date of the death statistics file of the Taiwan National Health Insurance Database within 30 days after surgery, and then encode it into a binary coded file (binary data), where 0 means the patient survives and 1 means the patient dies, so the predicted endpoint will cover patients who die in the hospital and die outside the hospital after surgery.

For example, in this embodiment, the first dataset 111 of the Kaohsiung Medical University Chung Ho Memorial Hospital (KMUH) as an internal dataset includes 87,907 patients with a total of 118,442 surgeries, and excludes 7,629 surgeries in which the patients were minors, thereby obtaining 82,573 patients with a total of 110,813 surgeries. The above multiple historical data can be further divided into the Kaohsiung Medical University Affiliated Chung Ho Memorial Hospital (KMUH) training dataset (including 75724 surgeries), the Kaohsiung Medical University Affiliated Chung Ho Memorial Hospital (KMUH) validation dataset (including 16750 surgeries), and the Kaohsiung Medical University Affiliated Chung Ho Memorial Hospital (KMUH) test dataset (including 18339 surgeries). Furthermore, the second dataset 112 of Kaohsiung Municipal Hospital (KMHK) as an external dataset includes 22,512 patients, multiple historical data of 27,731 surgeries, and 1,491 surgeries in which the patients were minors were excluded, thereby obtaining 21,195 patients, multiple historical data of 26,240 surgeries, and the above multiple historical data can be used as the training dataset of Kaohsiung Municipal Hospital (KMHK) (including 26,240 surgeries). Similarly, the third dataset 113 of Kaohsiung Municipal Tatung Hospital (KMTTH) as an external dataset contains multiple historical data of 227,822 patients who underwent 27,859 surgeries, and 826 cases where the patients were minors during the surgery were excluded, resulting in multiple historical data of 22,066 patients who underwent 27,033 surgeries. The above multiple historical data can be used as the training dataset of Kaohsiung Municipal Tatung Hospital (KMTTH) (including 27,033 surgeries).

Furthermore, the preoperative risk prediction system 1 of the present invention further includes a data preprocessing model 25, which performs data preprocessing on a plurality of historical data, wherein the data preprocessing includes the following steps: missing data imputation, data splitting, data normalization, and encoding categorical features, which will be described in detail below. The plurality of historical data after data preprocessing are continuously input into the knowledge base model 23.

Specifically, the first data set 111 (internal data set), the second data set 112 and the third data set 113 (external data set) from different medical institutions in the historical data are processed through the data pre-processing model 25. Pre-processing, and transmit the plurality of historical data pre-processed to the knowledge base model 23, which will calculate the total annual number of operations, the total number of death cases, and the cumulative average death of each surgery from different medical institutions statistical data such as rate, annual average mortality rate, average mortality rate for all surgical procedures, and standard deviation of mortality rates for each surgical procedure.

Further, in the data pre-processing operation of the data pre-processing model 25, the filling of missing data values includes filling the missing values of the digital data in the historical data with -1, and the values exceeding the upper limit or lower limit are filled with the maximum possible limit. Values are filled in, and the category data in multiple historical data are coded according to the ICD and DRG dictionary files. If the items are not stored in the dictionary files or are missing, they are coded into another category. In the data preprocessing operation of the data preprocessing model 25, data segmentation includes segmenting historical data according to time series, so as to prevent the data calculated for individual cases in the historical data from containing repeated information in the past. In the data pre-processing operation of the data pre-processing model 25, data standardization includes standardizing the digital data in the historical data based on the selected training data set into a mean of 0, and the data is expressed as a positive and negative standard deviation. method, and continue to adjust the data to the range of 0 to 1, and the data sets from different medical institutions are standardized based on the same training data set. Moreover, in the data preprocessing operation of data preprocessing model 25, category item coding includes all coding An encoding dictionary is established based on the selected training data set, and is subsequently encoded using nominal encoding.

The preoperative risk prediction system 1 of the present invention divides the first data set 111 into an internal data set, and divides the second data set 112 and the third data set 113 into external data sets, so as to input the multiple historical data of the first data set 111, the second data set 112 and the third data set 113 into the time series superposition knowledge base model 21 for time series superposition data processing, so as to generate time series superposition knowledge base data with internal and external data distinction and establish a knowledge base 22 accordingly. Similarly, the preoperative risk prediction system 1 of the present invention divides the first data set 111 into an internal data set, and divides the second data set 112 and the third data set 113 into external data sets, so as to input the multiple historical data of the first data set 111, the second data set 112 and the third data set 113 into the data preprocessing model 25 for data preprocessing, which is conducive to the subsequent verification of the accuracy of the preoperative risk prediction system 1 of the present invention.

In addition, the data generated by the data pre-processing model 25 will be continuously input into the knowledge base model 23 to perform data fusion processing, and the multiple fused knowledge base data generated by the data fusion processing will be continuously input into the machine learning model 24 to generate output data 241, thereby effectively reducing the impact of noise in the input historical data and improving the accuracy of the final output result. Thus, when doctors conduct preoperative anesthesia assessment, they can further refer to the predicted hospitalization days data, predicted mortality data and influencing factor analysis data in the output data 241 generated by the preoperative risk prediction system 1 of the present invention, thereby reducing the human judgment errors that may be caused by preoperative risk assessment based solely on the doctor's experience, and can improve the accuracy of preoperative risk assessment.

Further, the time series overlay knowledge base model 21 is configured to: retrieve the operation codes, death marks, and year categories of the plurality of historical data in the plurality of data sets, and perform mortality statistics of each operation code in each year to output The knowledge base 22 is established by superimposing knowledge base data on multiple time series. Among them, in the knowledge base 22, the plurality of time series superposition knowledge base materials marked as the mth year output by the time series superposition knowledge base model 21 will be superimposed on the plurality of time series superposition knowledge base materials marked as the m-1th year, m is a positive integer.

Specifically, as shown in FIG. 2, the time series superposition knowledge base model 21 receives multiple historical data from the database model, and the time series superposition knowledge base model 21 performs the time series superposition data processing on the received historical data as follows. First, the time series superposition knowledge base model 21 determines the source medical institution to which each historical data belongs, and performs mortality statistics on multiple historical data by surgery code and year, and sorts them by year.

Moreover, the time-series superposition knowledge base model 21 uses surgery codes as categories to perform loop calculations on the plurality of sorted historical data. First, it is determined whether there are annual vacancies in the plurality of historical data of this surgery code. If there are annual vacancies, then Deduct the data of n vacant years, calculate the average mortality rate corresponding to this surgery code (where n is a positive integer) using (10-n) years as the moving pane, and output the time series overlay knowledge base data. Or, if there are no annual vacancies, calculate the average mortality rate corresponding to this surgery code using a moving pane of 10 years, and output the time series overlay knowledge base data. Finally, the time series superposition knowledge base data marked as year m is superimposed to the time series superposition knowledge base data marked year m-1, and the knowledge base 22 is established based on the time series superposition knowledge base data.

Further, the knowledge base model 23 is connected to the knowledge base 22, and the knowledge base model 23 is configured to identify the source medical institutions of the plurality of historical data, and perform pre-processing with surgical codes based on the past data of the source medical institutions in the knowledge base 22. Fusion; pre-fusion with procedure codes based on historical data from medical institutions other than the source medical institution in knowledge base 22; or based on last year's average mortality rate at the source medical institution or at medical institutions other than the source medical institution. to fill in the data.

Specifically, as shown in FIG. 3 , the knowledge base model 23 receives a plurality of pre-processed historical data from the data pre-processing model 25, and receives time-series superposition knowledge base data processed by time-series superposition from the knowledge base 22, and the data fusion processing performed by the knowledge base model 23 on the received data can be divided into three blocks, namely internal knowledge fusion, external knowledge fusion, and knowledge filling. Specifically, internal knowledge fusion is to first obtain the past knowledge base of the source medical institution to which each data belongs (i.e., pre-processed historical data and time-series superimposed knowledge base data), and perform early fusion with the knowledge base of this hospital using the surgery code. If there are missing values in the data after fusion, external knowledge fusion will be performed. External knowledge fusion is to obtain the knowledge base of other medical institutions, and perform early fusion with the knowledge base of other hospitals using the surgery code. If there is no past data for the surgery corresponding to the historical data, knowledge filling will be performed. Knowledge filling is to obtain and use the average mortality rate of the same hospital in the previous year for data filling. If there is no average mortality rate data in the previous year, fill in -1 to represent unknown information.

The fused knowledge base data generated through the data fusion processing of the knowledge base model 23 is continuously input into the machine learning model 24, and the machine learning model 24 performs machine learning based on the knowledge base based on the plurality of fused knowledge base data to generate a plurality of output data 241, The plurality of output data 241 include the predicted hospitalization days data, predicted mortality data and influencing factor analysis data of each patient. Using the fused knowledge base data processed by the knowledge base model 23 to perform knowledge base-based machine learning can effectively improve the accuracy of the pre-surgery risk prediction system 1 of the present invention, which will be described in detail below.

Furthermore, the fused knowledge base data input into the composite deep machine learning model 24 extracts a plurality of feature parameters from the machine learning model 24, and machine learning is performed based on the extracted plurality of feature parameters, thereby improving the accuracy of the pre-operative risk prediction system 1 of the present invention. Among them, the plurality of feature parameters may include at least one selected from the International Statistical Classification of Diseases and Related Health Problems (ICD) code, surgery code, the average annual mortality rate of each surgery, the total number of surgeries for each surgery, the inpatient diagnosis group (DRG) code, blood draw items (e.g., coagulation index (INR), inflammation index (CRP), or white blood cell index (WBC)), gender, age, body mass index, number of hospitalization days before surgery, number of surgeries in a single hospitalization and hospitalization department, whether smoking, whether having a history of diabetes, hypertension, cardiovascular disease, cirrhosis, malignant tumors, etc., and anesthesia risk level, but the present invention is not limited thereto. In other embodiments, the type of feature parameters captured by the composite deep learning model can be set according to the actual situation.

After the machine learning model 24 calculates, the output data 241 including the predicted values of postoperative mortality and hospital stay can be output. The evaluation metrics adopt the area under the curve (AUC) and calculate the AUPRC (Area Under the Precision-Recall Curve) to analyze the positive prediction rate. The calculation of the area under the curve (AUC) is shown in the following formula (1):

Where B is the positive sample, W is the negative sample, x is the predicted value in the predicted positive sample, y is the predicted value in the negative sample, and all possible orderings of x and y are the Cartesian product B × W.

In addition, the above-mentioned time series superposition knowledge base model 21, knowledge base model 23, and machine learning model 24 can be selected from a decision tree (Decision Trees, DT) model, a gradient boosting decision tree (Gradient Boosting Tree, GBT) model, histogram gradient Any of the machine learning models such as the histogram-based gradient boosting tree model, the random forest tree (Random Forest, RF) model, and the long short-term memory (Long Short-Term Memory, LSTM) model.

As shown in FIG. 4 , the preoperative risk prediction system 1 of the present invention further includes a user interface 30 for inputting historical data of patients before surgery, thereby obtaining output data 241 including patient predicted hospital stay data, predicted mortality data, and influencing factor analysis data through data processing of each model in the preoperative risk prediction system 1. For example, after identity authentication, the doctor can input the patient's name through the user interface 30 to check the patient's past medical history, physical examination report and other related data from the database unit 10, and can drive the processing unit 20 to execute the above data processing process, thereby generating output data 241 including the patient's predicted hospitalization days data, predicted mortality data and influencing factor analysis data, which can be used as reference data for doctors to perform preoperative anesthesia consultation and help improve the accuracy of preoperative risk assessment.

The specific method of the clinical trial of the preoperative risk prediction system 1 of the present invention is as follows: first, after the patient is scheduled for surgery and hospitalized, the conventional preoperative anesthesia consultation procedure is carried out before the surgery, and the patient is informed of the clinical trial process of the preoperative risk prediction system 1 of the present invention, and after obtaining the patient's consent, the hospital's medical record system is marked that the patient participates in the clinical trial of the preoperative risk prediction system 1 of the present invention. In addition, the relevant data stored in the medical record system is de-identified using an encryption method for identifiable data items, and after de-identification is completed, it is transmitted to the database unit 10 for storage.

In addition, if the total number of surgeries for a single procedure is less than 10, it means that the average number of surgeries for this procedure is less than 2 per year over the past 7 years. This is a rare or new procedure in clinical practice, which will affect the learning of the model. Therefore, data with less than 10 surgeries will be excluded.

The processing unit 20 of the preoperative risk prediction system of the present invention can automatically perform data processing based on the data in the database unit 10, and encrypt the generated predicted hospital stay data and predicted mortality data. For example, the time series superposition knowledge base model 21 and the data pre-processing model 25 can have the function of data encryption, and transmit the encrypted predicted hospital stay data and predicted mortality data to an independent server in the hospital's medical record system for sealing. In this way, the data can be kept sealed during the clinical trial process until the patient is discharged from the hospital, so that all clinical trial participants (doctors, patients, nurses and analysts) cannot know the predicted hospitalization days data and predicted mortality data generated by the pre-operative risk prediction system 1 of the present invention, thereby ensuring that doctors, patients and other related personnel will not be affected by the above data and can perform anesthesia and surgery according to medical routine.

In addition, after the patient is discharged from the hospital, the data such as the length of hospitalization will also be encrypted and transferred to an independent server in the hospital's medical record system for storage. Finally, the data is unlocked two months after the patient is discharged from the hospital. At this time, the analyst logs into the independent server in the medical record system after identity authentication to obtain the above data for comparison and evaluation of the pre-surgery risk prediction system 1 of the present invention. Accuracy.

As shown in Figure 5, when the patient's mortality rate is evaluated by the preoperative risk prediction system 1 of the present invention, the area under the curve of the receiver operating characteristic curve (Area Under the Receiver Operating Characteristic curve, AUROC) can be as high as 0.929, and the accuracy of predicting mortality is about 92%. Furthermore, as shown in Figure 6, when the patient's hospital stay is evaluated by the preoperative risk prediction system 1 of the present invention, the area under the curve of the receiver operating characteristic curve (AUROC) can be as high as 0.913, and the accuracy of predicting mortality is about 91%. It can be seen that the preoperative risk prediction system 1 of the present invention has excellent accuracy in evaluating postoperative hospital stay and mortality.

Further, in this embodiment, the histogram gradient boosting decision tree (HGBT) model and the gradient boosting decision tree (Gradient Boosting Decision Tree 300, GBT300) are used as the main models, and Kaohsiung Medical University The data set of the University-affiliated Chung Ho Memorial Hospital (KMUH) was used for internal validation, and the data sets of Kaohsiung Municipal Tatong Hospital (KMTTH) and Kaohsiung Municipal Xiaogang Hospital (KMHK) were used for external validation.

Specifically, as shown in Figures 7A to 7C , Figure 7A is a compound prediction of a pre-surgery risk prediction system on whether to use a knowledge base on a Histogram Gradient Boosting Decision Tree (HGBT) model according to an embodiment of the present invention. Schematic diagram of comparison results of sampling method (bootstrap). Among them, Figure 7A shows the results of internal verification using Kaohsiung Medical University Chung Ho Memorial Hospital (KMUH) as the test data set. As a result, Figure 7B shows the results of external validation using Kaohsiung Municipal Xiaogang Hospital (KMHK) as the test data set, and Figure 7C shows the results of external validation using Kaohsiung Municipal Datong Hospital (KMTTH) as the test data set. Part A of Figures 7A to 7C is the area under curve (AUC) performance of the histogram gradient boosting decision tree (HGBT) model using the knowledge base, and part B is the histogram using the knowledge base. The area under the Precision-Recall Curve (PRAUC) performance of the gradient boosting decision tree model (HGBT).

Specifically, as shown in Part A of Figure 7A, in the case of using Kaohsiung Medical University Chung Ho Memorial Hospital (KMUH) as the test data set for internal validation, there is a histogram gradient boosting decision tree model using the knowledge base ( The AUC surface of HGBT) is about 0.950, and the AUC surface of the Histogram Gradient Boosted Decision Tree (HGBT) model without using the knowledge base is about 0.945. As shown in Part A of Figure 7B, in the case of external validation using Kaohsiung Municipal Xiaogang Hospital (KMHK) as the test data set, the AUC performance of the Histogram Gradient Boosting Decision Tree (HGBT) model using the knowledge base is approximately is 0.944, and the AUC surface of the Histogram Gradient Boosting Decision Tree (HGBT) model without using the knowledge base is approximately 0.939. As shown in Part A of Figure 7C, in the case of external validation using Kaohsiung Municipal Datong Hospital (KMTTH) as the test data set, the AUC performance of the Histogram Gradient Boosting Decision Tree (HGBT) model using the knowledge base is approximately is 0.947, and the AUC surface of the Histogram Gradient Boosting Decision Tree (HGBT) model without using the knowledge base is approximately 0.945. It can be seen that when using the histogram gradient to boost the decision tree model, combining the knowledge base and historical data with the machine learning model can achieve better accuracy.

Specifically, as shown in Figures 8A to 8C, Figure 8A shows a pre-surgery risk prediction system according to an embodiment of the present invention, regarding whether to use the knowledge base on the gradient boosting decision tree (GBT300) compound sampling method on the model. (bootstrap) comparison results diagram. Among them, Figure 8A shows the results of internal verification using Kaohsiung Medical University Chung Ho Memorial Hospital (KMUH) as the test data set, and Figure 8B shows The results of external validation using Kaohsiung Municipal Xiaogang Hospital (KMHK) as the test data set, and Figure 8C shows the results of external validation using Kaohsiung Municipal Datong Hospital (KMTTH) as the test data set. Part A of Figures 8A to 8C is the area under curve (AUC) performance of the gradient boosting decision tree (GBT300) model using the knowledge base, and part B is the gradient boosting decision making using the knowledge base. Area Under the Precision-Recall Curve (PRAUC) performance of the tree (GBT300) model's precision recall curve.

Specifically, as shown in Part A of FIG. 8A, when the Chung Ho Memorial Hospital of Kaohsiung Medical University (KMUH) was used as the test data set for internal validation, the AUC performance of the gradient boosting decision tree (GBT300) model with the knowledge base was approximately 0.949, and the AUC surface of the histogram gradient boosting decision tree model without the knowledge base was approximately 0.945. As shown in Part A of FIG. 8B, when the Kaohsiung Municipal Xiaogang Hospital (KMHK) was used as the test data set for external validation, the AUC performance of the gradient boosting decision tree (GBT300) model with the knowledge base was approximately 0.955, and the AUC surface of the histogram gradient boosting decision tree model without the knowledge base was approximately 0.934. As shown in Part A of Figure 8C, when using Kaohsiung Municipal Tatung Hospital (KMTTH) as the test data set for external validation, the AUC performance of the gradient boosting decision tree (GBT300) model with the knowledge base is about 0.937, and the AUC surface of the histogram gradient boosting decision tree model without the knowledge base is about 0.941. It can be seen that when using the gradient boosting decision tree (GBT300) model, combining the knowledge base and historical data with the machine learning model can have a better accuracy.

In summary, the preoperative risk prediction system of the present invention includes a data pre-processing model, a knowledge base model, a time series superposition knowledge base model, and a machine learning model, thereby using different algorithms to process historical data in various data sets from different medical institutions. Specifically, the historical data is input into the time series superposition knowledge base model, and the time series superposition knowledge base model performs time series superposition data processing based on the historical data to generate corresponding time series superposition knowledge base data, which includes the average hospitalization days and mortality data of each patient. Furthermore, the time-series superimposed knowledge base data is continuously input into the knowledge base model, and the knowledge base model performs data fusion processing based on the time-series superimposed knowledge base data to generate fused knowledge base data, which includes the average hospital stay data and mortality data of each patient. Finally, the fused knowledge base data is input into the machine learning model, and the machine learning model performs knowledge base-based machine learning based on the fused knowledge base data to generate output data, which includes the predicted hospital stay data, predicted mortality data, and influencing factor analysis data of each patient, for doctors to use as reference data for preoperative anesthesia consultation. Thus, the preoperative risk prediction system of the present invention can accurately predict the length of hospital stay and mortality rate of patients after surgery through machine learning based on knowledge base. According to experimental results, the preoperative risk prediction system of the present invention has an accuracy rate of more than 90% in predicting the length of hospital stay and mortality rate.

Through the above configuration, the preoperative risk prediction system of the present invention can cooperate with the preoperative anesthesia consultation system to assist doctors in performing preoperative risk assessment, thereby reducing errors in human judgment and improving the accuracy of preoperative risk assessment, which is conducive to ensuring the safety of subsequent anesthesia and surgery for patients.

The above is only illustrative and not restrictive. Any equivalent modifications or changes that do not depart from the spirit and scope of the present invention shall be included in the appended patent scope.

1: Preoperative risk prediction system

10: Database unit

111:First data set

112: First data set

113: The third data set

20: Processing unit

21: Time series superposition knowledge base model

22: Data Statistical Model Knowledge Base

23: Composite deep learning knowledge base model

24:Machine learning model

241:Output data

25: Data preprocessing model

Claims (8)

A pre-operative risk prediction system, which includes: a database unit that stores a plurality of historical data of a plurality of patients; and a processing unit that is electrically connected to the database unit and includes a time-series superimposed knowledge base model, A knowledge base model (knowledge-based model), and a machine learning model; wherein, the plurality of historical data are respectively input to the time series overlay knowledge base model, and the time series overlay knowledge base model performs time series superposition based on the plurality of historical data Data processing to generate corresponding plurality of time series overlay knowledge base data, the plurality of time series overlay knowledge base data include the average length of stay data and mortality data of each patient; and the plurality of time series overlay knowledge base data are input into the knowledge base A database model, and the knowledge base model performs data fusion processing based on the plurality of time series superimposed knowledge base data to generate a plurality of fused knowledge base data. The plurality of fused knowledge base data includes the average length of stay and mortality rate of each patient. data; and the plurality of fused knowledge base data are input to the machine learning model, and the machine learning model performs knowledge base-based machine learning based on the plurality of fused knowledge base data to generate a plurality of output data, and the plurality of output data Contains the predicted hospitalization days data, predicted mortality data and influencing factor analysis data for each patient; and the plurality of historical data are sorted in a time unit, and in the time series overlay data processing, if the plurality of historical data If there is no vacancy in any of the time units, the corresponding average mortality data is calculated using a first moving pane, and then the plurality of time series overlay knowledge base data is output. If the plurality of historical data are If there is any gap in the historical data within the time unit, then the data of the n gaps in the time unit will be deducted, and the corresponding average mortality data will be calculated by subtracting a second moving pane of n from the first moving pane. Then, the plurality of time series superposed knowledge base data are output, where n is a positive integer. The pre-surgery risk prediction system as described in claim 1, wherein the plurality of historical data includes a plurality of data sets respectively from different medical institutions. A pre-operative risk prediction system as described in claim 2, wherein the time series superposition knowledge base model is configured to: extract the surgery codes, death markers, and years of the plurality of historical data in the plurality of data sets, and perform mortality statistics for each surgery code in each year, so as to output the plurality of time series superposition knowledge base data to establish a knowledge base; wherein, in the knowledge base, the plurality of time series superposition knowledge base data marked as the mth year output by the time series superposition knowledge base model are superimposed on the plurality of time series superposition knowledge base data marked as the m-1th year, and m is a positive integer. The pre-operative risk prediction system as described in claim 3, wherein the knowledge base model is connected to the knowledge base, and the knowledge base model is configured to: identify a source medical institution of the plurality of historical data, and perform pre-fusion with the surgery code based on the past data of the source medical institution in the knowledge base; perform pre-fusion with the surgery code based on the past data of medical institutions other than the source medical institution in the knowledge base; or perform data filling based on the average mortality rate of the source medical institution or medical institutions other than the source medical institution last year. The pre-surgery risk prediction system as described in claim 1, wherein the machine learning model is configured to: retrieve a plurality of feature parameters from the plurality of fusion knowledge base data. The number is used to perform machine learning based on the knowledge base, and the plurality of feature parameters include selected from the International Statistical Classification of Diseases and Related Health Problems (ICD) code, surgery code, and historical data of each surgery. Average mortality rate, hospitalization diagnosis related group (DRG) code, blood draw items, gender, age, length of stay before surgery, number of surgeries in a single hospitalization, and at least one of the inpatient departments. The preoperative risk prediction system as described in claim 2 further comprises a data preprocessing model, which is configured to perform data preprocessing on the plurality of historical data in each of the data sets to fill in missing values in the plurality of historical data, and to perform standardization on the plurality of historical data using a training data set as a standard, and to encode the plurality of historical data using the training data set, and output the plurality of historical data that have undergone data preprocessing to the knowledge base model. The pre-operative risk prediction system as described in claim 1, further comprising a user interface for inputting the historical data of the patient before surgery, thereby obtaining data including the predicted length of stay and predicted mortality of the patient. and the output data of the impact factor analysis data. The pre-operative risk prediction system as described in claim 1, wherein the time series superposition knowledge base model, the knowledge base model, and the machine learning model are selected from at least one of the group consisting of a gradient boosting decision tree model, a histogram gradient boosting decision tree model, a random forest tree model, and a long short-term memory model.