KR102421172B1 - Smart Healthcare Monitoring System and Method for Heart Disease Prediction Based On Ensemble Deep Learning and Feature Fusion - Google Patents

Abstract

A smart healthcare monitoring method and system for heart disease prediction based on ensemble deep learning and shape fusion are presented. The smart healthcare monitoring method for heart disease prediction based on ensemble deep learning and shape fusion proposed in the present invention includes the steps of collecting data on heart disease patients through wearable sensor measurement and electronic medical testing, and FRF (Framingham Risk Functions) with EMR (Framingham Risk Functions) Electronic Medical Record), combining data collected through FRF and wearable sensor measurements using a shape fusion approach to generate medical data about heart disease, selecting shapes based on information acquisition techniques, and conditionally Calculating shape weights based on probability and training to predict heart disease of a patient through an ensemble deep learning classifier using shape weights based on a shape selected based on an information acquisition technique and a conditional probability.

Description

Smart Healthcare Monitoring System and Method for Heart Disease Prediction Based On Ensemble Deep Learning and Feature Fusion

The present invention relates to a smart healthcare monitoring method and system for heart disease prediction based on ensemble deep learning and shape fusion.

Accurate prediction of heart disease is essential to effectively treating patients with heart disease before a heart attack occurs. This goal can be achieved using an optimal machine learning model with rich medical data on heart disease. Recently, various systems based on machine learning have been proposed to predict and diagnose heart disease. However, these systems cannot handle high-dimensional data sets because there is no smart framework that can use other data sources for heart disease prediction. Existing systems also utilize existing techniques to select a shape from a data set and calculate a general weight for that shape based on its importance. These methods have also failed to improve the performance of heart disease diagnosis.

Automatic prediction of heart disease is one of the most essential and difficult health problems in the real world. Heart disease affects the functioning of blood vessels and causes coronary artery infections, weakening patients, especially adults and the elderly. The World Health Organization (WHO) has found that cardiovascular disease causes more than 18 million deaths each year. In addition, the United States spends $1 billion per day treating heart disease. The leading cause of death in the United States is heart disease, such as stroke, heart attack, and high blood pressure. Therefore, early prediction of heart disease is very important for effectively treating heart disease patients before a heart attack or stroke occurs.

Cardiovascular disease can be identified by performing medical tests and using wearable sensors. However, extracting valuable risk factors for heart disease from electronic medical tests is difficult as doctors seek to diagnose patients quickly and accurately. These electronic medical records (EMRs) are unstructured and continue to grow in size due to daily medical testing. Currently, wearable sensors are also used to continuously monitor a patient's body, both internally and externally, to detect heart disease. However, wearable sensor data for predicting heart disease is damaged by signal artifacts such as missing values and noise, resulting in poor system performance and inaccurate results. First of all, the use of wearable sensors and EMR together is an important and difficult task when monitoring patients with heart disease. Second, extracting relevant and meaningful features from data is a challenge for predicting heart disease. Therefore, there is a need for an intelligent system that can automatically fuse information extracted from both sensor data and EMR, analyze the extracted data to identify hidden symptoms of heart disease, and predict heart disease before a heart attack occurs.

Currently, several systems for predicting and diagnosing cardiovascular disease using data mining techniques and hybrid models have been proposed. Data mining techniques extract risk factors from unstructured text data. Also, a hybrid model is the integration of two different methods that work better together than any one separate method. These hybrid models basically involve two main steps. In the first step, a shape selection or shape weighting approach is applied to select a subset of shapes or to identify shape weights. In the second step, a subset of shapes or weights are used as input to a classifier that predicts heart disease. However, the data sets collected for heart disease contain relevant features along with many redundant and irrelevant features. Both duplicate and unrelated features cause confusion and noise in the definition of the target class. Handling of these shapes is not only time consuming, but also affects the accuracy of classification. In addition, the existing heart disease diagnosis system is based on a general shape weighting method. This method assigns equal weights to each shape for all classes. However, they use ambiguous combinatorial operations, which can establish the significance of the distinguishing features, and the lack of theoretical support increases the mean square error (MSE) and reduces the accuracy of the predictive model. Therefore, it is necessary to remove unnecessary shapes and give specific weights to the shapes before applying the classification model.

Wearable sensors and EMR play an important role in healthcare monitoring systems for patients with heart disease. However, extracting shapes from sensor data and EMR and combining them to transform them into structured data is a difficult task. Moreover, selecting shapes from structured data and then assigning them valuable weights is another challenge for machine learning (ML)-based systems. Therefore, we first look at a wearable sensor-based heart disease diagnosis system, and then focus on extracting information from text data and then shape fusion. We also present a brief review of the identification of geometric significance of healthcare data in the area of predicting heart disease.

In recent years, various systems using wearable sensors have been proposed to improve the process of predicting heart disease. Al-Makhadmeh and Tolba presented a wearable medical device-based system that collects details about patients with heart disease before and after a heart attack. They sent the collected data to a medical system, then utilized shape extraction techniques and deep learning models for important shape extraction and accurate classification. However, this system has limitations in efficient shape extraction and shape weighting approach. In addition, 23 attributes are used for training, increasing system complexity and dimensionality. A system called HealthFog was provided to automatically treat people with heart disease using Internet of Things (IoT) devices and deep learning models. The main purpose of this system is to competently manage heart disease data from IoT devices. Moreover, another novel framework for decision support system has been proposed for disease detection in patients. The system integrates medical data and data collected from wearable medical devices. In addition, it utilizes a multi-hierarchical approach in conjunction with deep learning for disease diagnosis. A hybrid recommendation system based on IoT devices was presented to diagnose heart disease. The system recommends a diet plan and physical activity according to the condition of a person with heart disease. However, this recommendation is based on a simple rule that requires semantic information to extract more sensitive information about the patient for an accurate recommendation. An IoT-based disease prediction system using a fuzzy neural classifier is presented. This system provides a new framework for mHealthcare monitoring systems for the diagnosis of serious diseases. In addition, a three-layer framework with ML models is presented for storing and processing data from cardiac patients using wearable devices. The first layer collects physiological data from sensors, the second layer stores medical data in the cloud, and the third layer predicts heart disease using logistic regression analysis. However, this system conducts an experiment using seven shapes, which is insufficient to accurately determine the health status of a person with heart disease. Also, the data pre-processing step is missing, making the data preparation for the ML classifier difficult to understand.

The extraction of valuable features from medical text data and the fusion of sensor data is another key issue for heart disease diagnosis systems. The shape extraction technique retrieves meaningful information from medical big data. The process of data fusion merges disparate data sources to create related and valuable data. Recently, various models have been proposed that extract features from medical text data and combine sensor data with other data. A real-time system based on an ML classifier was provided to predict heart disease. The system adopts two approaches as shape extraction (unification and mitigation) to extract healthcare data sets of important shapes. A new and powerful structure was proposed in mine EMR for the diagnosis of heart disease. The system uses word embedding models for shape identification and long-term short-term memory (LSTM) for heart failure prediction. In addition, another system used the results of ultrasound scans to predict mortality rates for heart disease patients in hospitals. The system uses text mining techniques to extract shapes and then applies a deep learning model for mortality prediction. A new text mining approach is presented to extract information from EMR. The authors of this system use a rules-based engine that automates information extraction and decision-making tasks. However, many medical records contain unstructured data, making it difficult for rule-based engines to process such data. Heart disease risk scores are also identified using the patient's unstructured EMR. In this system, the authors use text mining techniques to extract heart-critical factors from unstructured EMR and calculate heart disease risk scores for diabetic patients.

The wearable sensor is used to detect the user's feelings about music recommendations. In this system, the authors utilized shape-level fusion, which independently extracts shapes from each sensor and then combines the extracted shapes for emotion detection. Moreover, in another system, the authors used both environmental sensors and wearable body sensors for emotion detection. They applied all three levels of fusion (data, shape, and decision) to investigate the impact of the environment on human health. An iterative convolutional neural network (RCNN)-based system has been presented to predict disease risk. This system extracts features from structured and unstructured data of the patient and fuses them using a deeply reliable network to improve the accuracy of RCNN.

The technical task to be achieved by the present invention is to provide a new smart healthcare monitoring method and system for predicting heart disease using deep learning and functional fusion. The shape fusion method combines sensor data and functions extracted from electronic medical records to generate medical data, and the information acquisition technique eliminates irrelevant and redundant shapes and selects important shapes, which reduces the computational burden and improves system performance. can be improved In addition, the conditional probabilistic approach further improves system performance by calculating specific shape weights for each class.

In one aspect, the smart healthcare monitoring method for heart disease prediction based on ensemble deep learning and shape fusion proposed by the present invention includes collecting data on patients with heart disease through wearable sensor measurement and electronic medical test, Framingham Risk (FRF). Functions) from Electronic Medical Record (EMR), combining data collected through FRF and wearable sensor measurements using a shape fusion approach to generate medical data on heart disease, and shape based on information acquisition techniques train to predict the patient's heart disease through the ensemble deep learning classifier using the shape weights based on the shape and conditional probability selected based on the shape and conditional probability selected based on the conditional probability and calculating the shape weight based on the conditional probability. include

The step of extracting FRF from EMR uses a text mining technique and a rule-based engine, applies morphology and lemmatization algorithms to all unstructured data, identifies the lemma of each word, and extracts the unstructured text. We perform tokenization that separates chunks, and extracts FRFs using an N-gram approach.

The step of selecting a shape based on the information acquisition technique and calculating the shape weight based on the conditional probability selects a shape whose importance is measured according to the classification task using the information acquisition technique, and uses entropy to measure the system uncertainty. to find the difference between the leading and trailing entropy of two given individual variables.

Selecting the shape based on the information acquisition technique and calculating the shape weight based on the conditional probability determines the shape significance for each class, uses the maximize and sum functions to identify the overall shape significance for all classes, and , compute a specific shape weight for each class using a probabilistic approach.

In another aspect, the smart healthcare monitoring system for heart disease prediction based on ensemble deep learning and shape fusion proposed in the present invention collects data on patients with heart disease through a wearable sensor measurement, a gateway device, and a gateway device. A database that stores data on patients with heart disease and FRF (Framingham Risk Functions) extracted from EMR (Electronic Medical Record), using a shape fusion approach, combines data collected through FRF and wearable sensor measurements, When medical data is generated, a shape is selected based on the information acquisition technique, and shape weights are calculated based on the conditional probability, the ensemble deep learning classifier using the shape weights based on the shape selected based on the information acquisition technique and the conditional probability It includes a health condition prediction and disease diagnosis engine that predicts a patient's heart disease through

According to embodiments of the present invention, medical data is generated by combining sensor data and functions extracted from electronic medical records through a shape fusion method, and an important shape is selected by eliminating irrelevant and redundant shapes through an information acquisition technique, This can reduce the computational burden and improve system performance. In addition, the conditional probabilistic approach further improves system performance by calculating specific shape weights for each class. The proposed system can obtain an accuracy of 98.5%, which is higher than that of the existing system.

1 is a framework of a Smart Healthcare Monitoring System (SHMS) according to an embodiment of the present invention.
2 is a flowchart illustrating a method of operating SHMS according to an embodiment of the present invention.
3 is a diagram illustrating a workflow of SHMS according to an embodiment of the present invention.
4 is a view for explaining an FRF extraction module according to an embodiment of the present invention.
5 is a diagram for explaining the fusion of FRF extracted from text data and sensor data according to an embodiment of the present invention.
6 is a diagram for explaining a conceptualized ensemble deep learning model according to an embodiment of the present invention.
7 is a diagram illustrating an ontology used to make a recommendation for a person with a heart disease according to an embodiment of the present invention.

The present invention proposes a smart healthcare monitoring method and system for predicting heart disease using deep learning and shape fusion approaches. First, shape fusion methods combine sensor data and features extracted from electronic medical records to generate valuable medical data. Second, the information acquisition technique eliminates irrelevant and redundant shapes and selects important shapes, which reduces the computational burden and improves system performance. In addition, the conditional probabilistic approach further improves system performance by calculating specific shape weights for each class. Finally, an ensemble deep learning model is trained to predict heart disease. The proposed system is evaluated with heart disease data and compared with traditional classifiers based on shape fusion, shape selection and weighting techniques. The proposed system achieves an accuracy of 98.5%, which is higher than that of the existing system. These results show that the proposed system is more effective in predicting heart disease compared to other state-of-the-art methods of the prior art. Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a framework of a Smart Healthcare Monitoring System (SHMS) according to an embodiment of the present invention.

The proposed ensemble deep learning and shape fusion-based smart healthcare monitoring system (SHMS) for heart disease prediction is a gateway device 130, database 140, health condition prediction and disease diagnosis engine 150. include

The gateway device 130 collects data on a person with a heart disease through wearable sensor measurement.

The database 140 stores data on the heart disease patient collected through the gateway device and Framingham Risk Functions (FRF) extracted from the Electronic Medical Record (EMR).

The health state prediction and disease diagnosis engine 150 uses a shape fusion approach to combine data collected through FRF and wearable sensor measurement and generate medical data about heart disease, select a shape based on the information acquisition technique, If the shape weight is calculated based on the conditional probability, the patient's heart disease is predicted through the ensemble deep learning classifier using the shape weight based on the shape selected based on the information acquisition technique and the conditional probability.

The health state prediction and disease diagnosis engine 150 combines data collected through FRF and wearable sensor measurement using a shape fusion approach, and a data generator 151 that generates medical data related to heart disease, based on an information acquisition technique Through the ensemble deep learning classifier using the shape selection and weight calculation unit 152 that selects a shape and calculates the shape weight based on the conditional probability It includes a learning unit 153 that trains to predict the patient's heart disease.

The proposed smart healthcare monitoring system (SHMS) is divided into different layers to thoroughly describe the information base of each step. Then, we present the structure of the ensemble deep learning model and ontology, which is adopted by SHMS to predict patients' heart disease and recommend diet plans and activities (154).

SHMS has two main data sources. The first source is a wireless body sensor network (WBSN) 110 . Another data source is EMR 120 . The system uses WBSN based medical sensors to monitor electrocardiogram (ECG), electrodes (EEG), heart rate (EMG), blood pressure (BP), location, activity, respiratory rate, blood sugar, oxygen saturation, and daily health monitoring of the patient. Collect internal and external physiological data such as cholesterol levels. EMR provides patient observation reports, medical records, smoking records, diabetes records and detailed clinical examinations. After detecting the data from the heart disease patient, the proposed system transmits the data to the relevant gateway device. A variety of devices can be used to collect and forward the sensed data for further processing. In this system, physiological data is transmitted via Bluetooth and WiFi devices. Both the detected data and the EMR are stored securely in a database that is considered medical big data. The purpose of SHMS is to predict a patient's disease risk based on the data collected. Therefore, using the health status prediction and disease diagnosis engine, heart disease is predicted based on the structured and unstructured data collected. These engines include data fusion through the data generation unit 151, shape selection and pre-processing through the weight calculation unit 152, deep learning-based heart disease prediction through the learning unit 153, ontology-based recommendations 154, etc. 4 consists of steps. In the first step, the shapes extracted from structured data and unstructured data are fused using the proposed fusion method. The next step is where data mining techniques are used to pre-process the data. This step includes data filtering, normalization, selection of valuable shapes using data mining techniques, and shape weighting using conditional probabilities. In the third step, the pre-processed data is passed to a deep learning classifier trained on the heart disease data set for the final prediction of heart disease. In the fourth step, the ontology is used to recommend a diet plan or activity according to the patient's health status.

2 is a flowchart illustrating a method of operating SHMS according to an embodiment of the present invention.

The proposed smart healthcare monitoring method for heart disease prediction based on ensemble deep learning and shape fusion includes collecting data on heart disease patients through wearable sensor measurement and electronic medical testing (210), and FRF (Framingham Risk Functions) to EMR (Framingham Risk Functions). Extracting from Electronic Medical Record (220), combining data collected through FRF and wearable sensor measurement using a shape fusion approach and generating medical data about heart disease (230), based on the information acquisition technique To predict the patient's heart disease through the ensemble deep learning classifier using the shape weights based on the shape selected based on the shape and the conditional probability selected based on the shape selected based on the shape and the conditional probability based on the step 240 of selecting a shape and calculating the shape weight based on the conditional probability ( 240 ) and training 250 .

In step 210 , data on a person with a heart disease is collected through wearable sensor measurement and electronic medical testing.

In step 220, Framingham Risk Functions (FRF) are extracted from Electronic Medical Record (EMR). At this time, a token that uses a text mining technique and a rule-based engine, applies a morphology and a lemmatization algorithm to all unstructured data, identifies the lemma of each word, and divides the unstructured text into small chunks. , and extract the FRF using an N-gram approach.

In step 230, a shape fusion approach is used to combine data collected through FRF and wearable sensor measurements and generate medical data regarding heart disease.

In step 240, a shape is selected based on the information acquisition technique, and shape weights are calculated based on the conditional probability. The information acquisition technique is used to select a shape whose significance is measured according to the classification task, and to use entropy to measure the system uncertainty, find the difference between the leading entropy and the trailing entropy of two given individual variables. Determine shape significance for each class, use maximize and sum functions to identify overall shape significance for all classes, and calculate specific shape weights for each class using a probabilistic approach.

In step 250, the ensemble deep learning classifier is trained to predict the patient's heart disease using the shape weights based on the shape selected based on the information acquisition technique and the conditional probability.

3 is a diagram illustrating a workflow of SHMS according to an embodiment of the present invention.

The work flow of the proposed SHMS as shown in FIG. 3 is data collection layer, data fusion and feature extraction layer, data preprocessing layer, disease prediction and recommendation. Recommendation Layer) and 4 sequential layers are used.

There are two main pieces of data in SHMS. The first data is sensor data 310 collected from a wireless body sensor network (WBSN). Another data is EMR data 320 collected from EMR. The system uses WBSN based medical sensors to monitor electrocardiogram (ECG), electrodes (EEG), heart rate (EMG), blood pressure (BP), location, activity, respiratory rate, blood sugar, oxygen saturation, and daily health monitoring of the patient. Collect internal and external physiological data such as cholesterol levels. EMR provides patient observation reports, medical records, smoking records, diabetes records and detailed clinical examinations. After detecting the data from the heart disease patient, the proposed system transmits the data to the relevant gateway device. A variety of devices can be used to collect and forward the sensed data for further processing. In this system, physiological data is transmitted via Bluetooth and WiFi devices.

The proposed SHMS considers two types of heart disease prediction data. That is, patient physiological data and EMR as shown in FIG. 3 . Patient physiological data is collected with the help of wearable sensors. Two types of sensors are used to collect physiological data: medical sensors and activity sensors. Medical sensors include respiratory rate sensors, oxygen saturation sensors, blood pressure sensors, cholesterol level sensors, glucose level sensors, temperature sensors, EEG sensors and electrocardiogram sensors. They connect to the patient's body and collect physiological data without interruption. Wearable watches are also used to record physical activity and heart rate. The patient's physical activity provides valuable information in predicting heart disease. The model uses a networked device to transmit physiological data to a personal database to analyze heart conditions. In the data collection layer, ID is assigned to all sensor data (eg, S1 is EEG sensor data) as shown in FIG. 3 . The system treats these identities as in-process features. Sensor data is also displayed in columns with ID and numeric values for further processing.

Moreover, unstructured EMR of patients is collected to identify risk factors for predicting heart disease. EMR includes laboratory reports, medical records, questions and observations, allergies and medications, and personal statistics. By analyzing EMR, it is possible to extract FRF, which can provide useful information for disease prediction. FRF includes age, sex, presence of diabetes, smoking history, cholesterol level, BP, heart rate, excretion (EF), body mass index (BMI), obesity (if any), and diet. However, the volume of EMR data is usually large, and each record consists of data with highly dimensional variance variables. Therefore, text mining method is used to effectively extract FRF from unstructured text data.

4 is a view for explaining an FRF extraction module according to an embodiment of the present invention.

As shown in FIG. 4, information is extracted from the unstructured EMR through a separate module called the FRF extraction module. The unstructured EMR is assigned to the FRF extraction module to identify and extract risk factors associated with heart disease. This module is divided into two sub-modules: structured FRF extraction and categorical FRF extraction. In structured FRF extraction, the system extracts factors whose values are already available in a structured form. For example, the system extracts age, height, heart rate, EF, gender, BP and BMI along with the values of the structured fields. In categorical FRF extraction, the system extracts risk factors whose values appear in different classes (eg, history of diabetes, history of smoking, family history of coronary artery disease [CAD]). These factors are extracted using text mining techniques and a rule-based engine. The text mining technique consists of three main steps. In the first step, morphological and lemmatization algorithms are applied to all unstructured data to identify the lemma of each word. The next step is tokenization, which breaks down complex unstructured text into smaller chunks. In the third step, risk factors are extracted using an N-gram approach. In general, the value of a risk factor is represented by two or three adjacent words. Therefore, bigram and trigram are used to extract 2 and 3 adjacent factor words, respectively. Abbreviations indicating gender, age, cholesterol levels and emissions, and rules are created to identify specific characteristics. For example, the EMR may include the phrase 48y. A value of 48 is extracted as the patient's age based on a series of abbreviations. Categorical FRF extraction obtains information about a patient's diabetes history, family CAD history, and smoking history based on a set of rules. In order to understand the history mentioned above, separate word bags were created based on the terms diabetes, CAD, and smoking. This module filters records with the help of a rules-based engine to remove records that do not contain information about heart disease, diabetes and/or smoking. It also eliminates risk factors without numerical value. The result of this module is an FRF with values or categories in a structured format.

5 is a diagram for explaining the fusion of FRF extracted from text data and sensor data according to an embodiment of the present invention.

As shown in FIG. 5 , the fusion of FRF extracted from text data and sensor data will be discussed. Fusion is the process of merging different data sources to create more valuable and relevant data for classification. There are three levels of fusion: data level, shape level, and decision level. Data-level fusion combines disparate data from disparate sources that match each other. Data fusion can be classified into a configuration level and a decision-making level. At the feature level, features are individually retrieved from different data sets and then merged to create a set of features best suited for prediction. In the decision stage, the decisions of various processes are considered in order to increase the accuracy of the system. In general, data-level fusion is undesirable because it contains a lot of redundant data. In contrast, the morphology level contains enough information to identify disease risk. The proposed system performs both data level and shape level fusion. The workflow for data fusion is shown in FIG. 5 . First, the patient's physiological data are collected using a sensor as previously described, and the FRF is extracted from the EMR. The sensor data is then fused with other extracted FRFs, such as age, gender, diabetes history, and smoking history. Finally, the sensor data and the extracted shape are converted into a comma-separated value (CSV) file for easy parsing. In this way, the system finds the best combination of features related to heart disease. The main goal of the proposed system is to use sensor data with appropriate and low-dimensional features extracted from EMR for heart disease prediction. However, the sensor data may have missing values and the extracted shape constitutes relevant information, which reduces the prediction accuracy and increases the shape dimension. In addition, they increase the complexity and the need for memory for classification. Therefore, data pre-processing is applied before actual processing, which improves the quality of the data and saves time and memory.

Data preprocessing is the most essential step before applying ML algorithms. Direct application of real-world data for prediction tasks is impossible because it tends to be noisy, incomplete, and inconsistent. Therefore, a pre-processing step is applied to effectively represent the data for heart disease prediction. Data preprocessing includes missing data filtering, normalization, shape selection and shape weighting.

Data collected using wearable sensors and data extracted from EMR are useless and contain misleading information. Wearable sensor data for predicting heart disease can be corrupted by signal artifacts such as missing values and noise, resulting in poor prediction accuracy or inaccurate results. In addition, data extracted from EMR are assumed to be missing when they do not contain at least one value. Information may be missing or the FRF value may not be recorded due to the failure of the text mining technique to recognize the FRF value in the extracted data. The present invention filters data using a well-known filtering approach called Kalman filtering. This filter cleans the data by removing noise, overwrites, and inconsistencies. It also utilizes two unsupervised filters in the data filtering phase: RemoveUseless and ReplaceMissing values. The first filter removes unnecessary attributes with a variance of up to 90%. The second filter replaces all missing values in the structured data set with the mean and median values of the existing data using the following equation:

(1)

(One)

,

,

,

및

는 각각 X= {"나이", "콜레스테롤", "성별", "심박수", "CAD history", "흡연 이력"}과 같은 형상,

={0, 2}의 특징의 범주 수준, 패턴 넘버, 범주

내에서 형상

의

번째 패턴, 범주

내에서 형상

의 평균을 각각 나타낸다. 이 작업에서 는 범주

내에서 형상

의 결측값을 대체한다. 또한, 웨어러블 센서 기반 생성 데이터의 한계를 극복하기 위해 추출된 FRF 값을 활용한다. 예를 들어, 누락된 값을 데이터 세트의 현재 FRF 속성 값으로 대체한다.here

,

,

,

and

X = {"age", "cholesterol", "gender", "heart rate", "CAD history", "smoking history"}, respectively.

Category level, pattern number, category of the feature in ={0, 2}

shape within

of

second pattern, category

shape within

represents the average of each. In this task, the category

shape within

replace missing values of In addition, the extracted FRF value is utilized to overcome the limitations of wearable sensor-based generated data. For example, replace missing values with the values of the current FRF attribute in the data set.

는 많은 형상을 포함하고 있으며, 모든 형상은 다른 수치값을 포함하고 있어 계산 과정 중 어려움이 증가한다. 따라서 정규화 기법은 0과 1 사이의 범위에서 데이터 집합

를 정규화하고 심장병예측의 연산 과정 중 수치적 복잡성을 줄이는 데 사용된다. 데이터 정규화에 다양한 방법을 활용할 수 있다. 제안된 시스템에서는 잘 알려진 최소-최대 정규화 방법을 사용한다. 이 방법은 다음과 같은 식을 사용하여 원래 데이터 세트

의 숫자 값 DV를 [0, 1] 간격 내에

으로 표시한다. heart disease data set

contains many shapes, and all shapes contain different numerical values, increasing the difficulty during the calculation process. Therefore, the regularization technique is a set of data in the range between 0 and 1.

It is used to normalize and reduce the numerical complexity during the computational process of heart disease prediction. Various methods are available for data normalization. The proposed system uses a well-known min-max normalization method. This method uses the expression

Numeric values of dv within the interval [0, 1]

indicated as

(2)

(2)

,

,

,

는 전체 데이터 집합에서 각각 정규화된 데이터 값, 원본 데이터 값, 최소 데이터 값 및 최대 데이터 값이며 new_max와 new_min은 변환된 데이터 집합의 범위를 나타낸다. new_max=1과 new_min=0을 사용한다. 이 방법을 사용하여 모든 형상의 값은 간격 [0, 1] 내에 있다.here

,

,

,

is the normalized data value, the original data value, the minimum data value, and the maximum data value in the entire data set, respectively, and new_max and new_min represent the range of the transformed data set. Use new_max=1 and new_min=0. Using this method, the values of all shapes are within the interval [0, 1].

에는 27개의 속성이 있다. 그들 중 몇몇만이 질병을 주어진 범주 중 하나로 분류하는데 유용하다. 시스템은 데이터 세트의 기능의 중요성에 기초하여 특정 문제에 대해 배울 수 있다. 제안된 시스템에서는 정보 이득(Information Gane; IG)을 이용하여 분류 과제에 따라 중요성을 측정하는 형상을 선택한다. 제안된 시스템은 시스템 불확실성을 측정하기 위해 엔트로피를 사용한다. 그것은 다음 식에서와 같이 주어진 두 개의 개별 변수인 A와 B의 선행 엔트로피와 후행 엔트로피 사이의 차이를 발견한다.Patient records usually consist of many extraneous features that reduce the accuracy of predictions. However, extracting meaningful information from medical records, reducing noise by removing extraneous features, and accurately predicting heart disease with a limited number of features are all challenging tasks. Before applying the predictive model, it is essential to remove the noisy data, select useful features that will give accurate results, and reduce the complexity and dimensionality of the data set. Therefore, shape selection is an important step in improving the clarity of the data and reducing the training time of ensemble deep learning models. Various shape selection methods are used for medical data sets, such as sequential future selection, weighted least squares, coarse aggregation, and univariate shape selection. The information acquisition method that affects the prediction result by removing the noise shape is utilized. Structured data sets for predicting heart disease

has 27 attributes. Only a few of them are useful for classifying a disease into one of the given categories. A system can learn about a particular problem based on the importance of the function of the data set. The proposed system selects a shape that measures importance according to the classification task using the information gain (IG). The proposed system uses entropy to measure system uncertainty. It finds the difference between the leading and trailing entropy of two separate variables, A and B, given as in the following equation.

(3)

(3)

Here, A and B are separate random variables, and the preceding entropy of shape A can be calculated using Equation 4.

(4)

(4)

는

의 선행 확률을 나타낸다. 후행 엔트로피 B를 받은 후 A의 조건부 엔트로피는 식 5를 사용하여 계산할 수 있다.here

Is

represents the leading probability of . After receiving the trailing entropy B, the conditional entropy of A can be calculated using Equation 5.

(5)

(5)

IG can be measured by putting Equations 4 and 5 into Equation 3:

(6)

(6)

The proposed system estimates the importance of each shape for the heart disease prediction task using Equation 6. The system measures the IG for each feature and then discards the least significant features. Reduce the features by removing one feature at a time until the degradation is over.

Shape weighting is a method of assigning a weight to each shape based on its importance. This is different from feature selection, which completely removes redundant and irrelevant features from the training data set. Various methods have been used for shape weighting, such as analysis stratification process and rule-based weighting. This method assigns equal weights to each shape for all classes, called the normal weight of the shape. In general shape weighting, shape importance is first determined for each class. We then use the maximize and sum functions to identify overall shape significance for all classes. However, existing approaches have some limitations. One of the main issues is that they are based on uncertain combinatorial operation, which can establish the importance of distinguishing features. In addition, these existing approaches lack theoretical support, increasing MSE and decreasing the accuracy of predictive models. The present invention uses a probabilistic approach to achieve specific shape weights for each class. It's a more sophisticated approach than normal weighting. For predictive tasks, shape weights should be class-specific. This improves the performance of the predictive model by learning the different meanings of shapes for each class. The weight matrices of the general shape and the specific shape are shown in Table 1 and Table 2, respectively.

<Table 1>

<Table 2>

,

,

, ....,

을 n개의 형상 변수라 하면,

는 형상 벡터 〈

,

,

, ....,

〉에 의해 나타나는 인스턴스로서, 여기서

는

의 값을 나타낸다. 예를 들어

의 특정 형상 가중치는 다음 식을 사용하여 계산할 수 있다.In Table 1, each shape shares the same weight for all classes, whereas in Table 2 each shape shares a different weight for all classes.

,

,

, ....,

Let be n shape variables,

is the shape vector q

,

,

, ....,

An instance represented by >, where

Is

represents the value of for example

The specific shape weight of can be calculated using the following equation.

(7)

(7)

와

는 각각 클래스 변수 값과 클래스 c에 대한 형상값

의 형상 가중치를 나타낸다.

의 값은 형상값

과 관련이 있으며, 여기서 각 형상값에 서로 다른 가중치가 할당된다.

의 범위는 0부터 1까지이며, 이는 심장병 예측 과제에 대한 형상값

의 중요성을 나타낸다. 이 접근법은 심장병 형상의 값을 보다 관리하기 쉬운 형태로 변환하는데 유용하다. 형상값에 대한 이러한 가중치를 획득하는 주된 목적은 깊이 있는 학습 모델을 가진 초기 가중치로 사용하여 더 나은 예측 결과를 얻기 위한 것이다. here

Wow

is the class variable value and the shape value for class c, respectively.

represents the shape weight of .

The value of is the shape value

, where different weights are assigned to each shape value.

The range of is from 0 to 1, which is the shape value for the heart disease prediction task.

indicates the importance of This approach is useful for transforming the values of heart disease shape into a more manageable form. The main purpose of obtaining these weights for shape values is to use them as initial weights with a deep learning model to obtain better prediction results.

6 is a diagram for explaining a conceptualized ensemble deep learning model according to an embodiment of the present invention.

The fourth layer of the proposed SHMS includes the ensemble deep learning model conceptualized as the framework shown in FIG. This model is a feed-forward network using an error back propagation and gradient algorithm for binary classification of heart disease. First, we use a different number of shapes from the Cleveland data set to train the model, and then use them to predict the results of the real-time input data presented in Fig. 6 . We used a boosting algorithm called LogitBoost as a meta-learning classifier, which boosts the deep learning model to achieve high accuracy. Based on the embodiments of the present invention and literature review, the boosting algorithm is more suitable for processing noisy data than the AdaBoost algorithm. It is designed to reduce bias and variance to improve classification performance.

The ensemble deep learning model consists of five layers: an input layer, three hidden layers, and an output layer. The input layer consists of 16 nodes according to the number of shapes in the data set. It is essential to evaluate the hidden layers of the neural model using different numbers of nodes. In the proposed work, we consider a fully connected hidden layer of 20 nodes, which produces consistent performance. Attributes are assigned to the deep neural network model using the input layer. These attributes are then propagated to the three hidden layers by multiplying their weights by values. Calculate the weighted sum and add a bias to process the input data at the hidden layer node as shown in Equation 8.

(8)

(8)

,

및

는 각각 입력 데이터, 노드 간 가중치 및 바이어스를 나타낸다. 이후

는 정류된 선형유닛(ReLU) 활성화 함수를 사용하여 변환되고 변환된 데이터를 출력 노드로 가져와 심장병을 예측한다. 딥러닝 모델의 다른 파라미터는 학습율(0.03)을 포함하며, 아담 최적기를 사용한다. 출력 계층의 크기는 2진법 분류(심장병의 존재 또는 심장병의 부재)의 결과를 나타내는 2개의 노드이다. here

,

and

denotes input data, inter-node weights and biases, respectively. after

is transformed using a rectified linear unit (ReLU) activation function and brings the transformed data to the output node to predict heart disease. Another parameter of the deep learning model includes the learning rate (0.03) and uses the Adam optimizer. The size of the output layer is two nodes representing the result of binary classification (presence or absence of heart disease).

7 is a diagram illustrating an ontology used to make a recommendation for a person with a heart disease according to an embodiment of the present invention.

온톨로지 언어(OWL)의 여러 단계를 사용하여 개발되었다. 첫째, 클래스가 구성된다. 둘째, 데이터 속성과 오브젝트 속성이 구현되어 클래스 간의 관계를 결정하는 데 활용된다. 마지막으로, 추론 규칙은 Semantic Web Rule Language 규칙을 사용하여 권고를 위해 작성된다. 제안된 온톨로지를 세 가지 주요 이유로 사용한다. 첫째로, 대부분의 심장병 예측 시스템은 의사들로 하여금 환자를 위한 어떤 종류의 조언을 위해 의료 리포트를 재확인하게 한다. 따라서 병원이나 원격지에서 자동적으로 환자에게 식이요법 계획이나 활동을 제안할 수 있는 새로운 권고 시스템이 필요하다. 둘째, 권고 모듈은 사전 진단 및 진단 심장 질환자에게 조언하기 위한 전문적인 지식으로 사용될 수 있는 도메인 지식을 요구한다. 셋째, 심장병의 증상은 다른 질병과 비슷하여 의사를 혼란스럽게 할 수 있으므로 권고 중에 혼란을 없애기 위한 규칙 기반 시스템이 필요하다. 도 7은 심장 질환자를 위한 권고안을 만들기 위하여 사용되는 제안된 온톨로지를 제시한다. An ontology is a collection of classes (shapes) and relationships between them. Ontology uses vocabulary to share knowledge from a specific domain among other modules. The definition and construction procedure of an ontology are extensively discussed in the prior art. The proposed ontology is

It was developed using several levels of the ontology language (OWL). First, the class is constructed. Second, data properties and object properties are implemented and utilized to determine the relationship between classes. Finally, inference rules are written for recommendations using Semantic Web Rule Language rules. We use the proposed ontology for three main reasons. First, most heart disease prediction systems allow doctors to double-check medical reports for some kind of advice for their patients. Therefore, there is a need for a new recommendation system that can automatically suggest a diet plan or activity to a patient in a hospital or remote location. Second, the recommendation module requires domain knowledge that can be used as professional knowledge to advise patients with pre-diagnosis and diagnosis heart disease. Third, the symptoms of heart disease can be similar to other diseases, which can confuse doctors, so a rule-based system is needed to eliminate confusion during recommendations. 7 presents the proposed ontology used to make recommendations for heart disease patients.

The ontology-based recommendation model is the most essential part of the proposed SHMS. The developed ontology consists of facts and rules. This fact includes the shape extracted from sensor data, EMR, treatment, symptoms, and laboratory test results, as shown in FIG. 7 . Rules include SWRL rules, which are expert knowledge that provide recommendations to patients based on predicted health status outcomes. These rules are gathered from online papers and cardiologists on all types of heart disease. In the present invention, 56 rules were designed to accurately recommend activities and diet plans for patients with heart disease. Some of the SWRL rules are thoroughly discussed in this previously published paper [48]. However, in order to understand the use of all the rules, the following rules are explained.

rule:

Description: If X is a patient with heart disease and BMI, HR, and exercise results are high, normal, and no, respectively, the recommendation is physical activity. This rule recommends that the patient increase physical activity so that the calories burned are burned.

A deep learning model first uses patient data to predict heart disease. After prediction, the system identifies the patient's gender. The background is that the recommendations for men with heart disease are different from those for women. In addition, the system finds the age of the patient and then identifies the group (young, adult, old) to which that age belongs. The model then recommends a dietary plan or activity (eg, quit smoking, increased physical activity, weight control, reduced meat consumption) based on the patient's gender, age, and predicted outcomes. The module also calls rescue and emergency services if the value of the extracted shape is too high and the predicted result is negative.

According to an embodiment of the present invention, a new information framework for processing both sensor data and medical records using ensemble deep learning and shape fusion techniques for heart disease prediction is proposed.

An FRF extraction module was developed to detect and extract low-dimensional risk factors for heart disease from unstructured EMR. In addition, shape-level fusion is performed to combine both the sensor data and the extracted FRF for heart disease prediction.

The IG approach is proposed for shape selection that reduces noise by reducing the complexity and dimensionality of the data set by removing extraneous features. It also utilizes a conditional probabilistic approach that computes specific shape weights for each class (class) to increase prediction accuracy.

A boosting algorithm called LogitBoost is used as a meta-learning classifier that reduces bias and variance and achieves high accuracy by activating deep learning models. In addition, an ontology based on Semantic Web Rule Language (SWRL) rules has been developed to automatically recommend diet plans or activities for people with heart disease.

The proposed system improves the performance and achieves a high accuracy of 98.5% in predicting heart disease compared to other state-of-the-art methods.

The device described above may be implemented as a hardware component, a software component, and/or a combination of the hardware component and the software component. For example, devices and components described in the embodiments may include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), It may be implemented using one or more general purpose or special purpose computers, such as a programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to execution of the software. For convenience of understanding, although one processing device is sometimes described as being used, one of ordinary skill in the art will recognize that the processing device includes a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that may include For example, the processing device may include a plurality of processors or one processor and one controller. Other processing configurations are also possible, such as parallel processors.

The software may comprise a computer program, code, instructions, or a combination of one or more thereof, which configures a processing device to operate as desired or is independently or collectively processed You can command the device. The software and/or data may be any kind of machine, component, physical device, virtual equipment, computer storage medium or device, to be interpreted by or to provide instructions or data to the processing device. may be embodied in The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic such as floppy disks. - includes magneto-optical media, and hardware devices specially configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like.

As described above, although the embodiments have been described with reference to the limited embodiments and drawings, various modifications and variations are possible by those skilled in the art from the above description. For example, the described techniques are performed in a different order than the described method, and/or the described components of the system, structure, apparatus, circuit, etc. are combined or combined in a different form than the described method, or other components Or substituted or substituted by equivalents may achieve an appropriate result.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

Claims (8)

collecting, by the gateway device, data on the heart disease patient through wearable sensor measurement and electronic medical testing;
extracting, by the health state prediction and disease diagnosis engine, Framingham Risk Functions (FRF) from Electronic Medical Record (EMR);
combining, by the health state prediction and disease diagnosis engine, data collected through FRF and wearable sensor measurements using a shape fusion approach, and generating medical data on heart disease;
selecting, by the health state prediction and disease diagnosis engine, a shape based on an information acquisition technique, and calculating shape weights based on conditional probability; and
Training the health state prediction and disease diagnosis engine to predict a patient's heart disease through an ensemble deep learning classifier using a shape weight based on a shape selected based on an information acquisition technique and a conditional probability;
A smart healthcare monitoring method for heart disease prediction based on ensemble deep learning and shape fusion, including According to claim 1,
The step of the health state prediction and disease diagnosis engine extracting the FRF from the EMR comprises:
It uses text mining techniques and rule-based engines, and applies morphology and lemmatization algorithms to all unstructured data to identify the canonical form (lemma) of each word, and tokenization that separates unstructured text into small chunks. and extracting the FRF using the N-gram approach.
Smart healthcare monitoring method for heart disease prediction based on ensemble deep learning and shape fusion. According to claim 1,
The health state prediction and disease diagnosis engine selects a shape based on an information acquisition technique, and calculates shape weights based on conditional probability,
Using the information acquisition technique, we select a shape to measure importance according to the classification task, and use entropy to measure system uncertainty to find the difference between the leading and trailing entropy of two given individual variables.
Smart healthcare monitoring method for heart disease prediction based on ensemble deep learning and shape fusion. According to claim 1,
The health state prediction and disease diagnosis engine selects a shape based on an information acquisition technique, and calculates shape weights based on conditional probability,
Determine shape significance for each class, use maximize and sum functions to identify overall shape significance for all classes, and calculate specific shape weights for each class using a probabilistic approach.
Smart healthcare monitoring method for heart disease prediction based on ensemble deep learning and shape fusion. a gateway device that collects data on a person with heart disease through wearable sensor measurement;
a database for storing data on patients with heart disease collected through the gateway device and Framingham Risk Functions (FRF) extracted from EMR (Electronic Medical Record);
Using a shape fusion approach, combining data collected through FRF and wearable sensor measurements and generating medical data on heart disease, selecting shapes based on information acquisition techniques, and calculating shape weights based on conditional probability, information A health state prediction and disease diagnosis engine that predicts a patient's heart disease through an ensemble deep learning classifier using shape weights based on shape selected based on acquisition technique and conditional probability
Smart healthcare monitoring system for heart disease prediction based on ensemble deep learning and shape fusion. 6. The method of claim 5,
Health condition prediction and disease diagnosis engine,
a data generating unit that combines data collected through FRF and wearable sensor measurements using a shape fusion approach and generates medical data on heart disease;
a shape selection and weight calculation unit that selects a shape based on an information acquisition technique and calculates a shape weight based on a conditional probability; and
A learning unit that trains to predict a patient's heart disease through an ensemble deep learning classifier using a shape weight based on a shape selected based on an information acquisition technique and a conditional probability
Smart healthcare monitoring system for heart disease prediction based on ensemble deep learning and shape fusion. 7. The method of claim 6,
The shape selection and weight calculation unit,
Using the information acquisition technique, we select a shape to measure importance according to the classification task, and use entropy to measure system uncertainty to find the difference between the leading and trailing entropy of two given individual variables.
Smart healthcare monitoring system for heart disease prediction based on ensemble deep learning and shape fusion. 7. The method of claim 6,
The shape selection and weight calculation unit,
Determine shape significance for each class, use maximize and sum functions to identify overall shape significance for all classes, and calculate specific shape weights for each class using a probabilistic approach.
Smart healthcare monitoring system for heart disease prediction based on ensemble deep learning and shape fusion.

Images