1
Machine Learning Approach for Medical Diagnosis Based on Prediction Model

Hemant Kasturiwale1*, Rajesh Karhe2 and Sujata N. Kale3

1Thakur College of Engineering and Technology, Kandivali (East), Mumbai, MS, India

2Shri Gulabrao Deokar College of Engineering, Jalgaon, MS, India

3Department of Applied Electronics, Sant Gadge Baba University, Amravati, MS, India

Abstract

The electrocardiography is the most crucial biosignals for critical analysis of the heart. The heart is the human body’s most vital and variety of control mechanisms that regulate the heart’s activities. The heart rate is an essential measure of cardiac function. The heart rate is represented as a time interval equal between two corresponding electrocardiogram (ECG) “R” peaks. The heart rate varies with the heart’s state. A machine learning technique is used to categorize the statistical parameters mentioned above to predict the individual’s physical state, including sleep, examination, and exercise, based on a physiologically important factor known as HRV. The chapter is focused on uses of manual classified data. Each hospital, clinic, and diagnostic center produces massive quantities of information such as patient records and test results to predict the presence of heart disease and provide care for the early stages. The results are validated and compared with predictions obtained from different algorithms. Classification and prediction are a mining technique that uses training data to construct a model, and then, that model is applied to test data to predict outcomes. Different algorithms are employed to disease datasets to diagnose chronic disease, and the findings have been positive. There is a need to establish an appropriate technique for the diagnosis of chronic diseases. This chapter discusses with insight various kinds of classification schemes for chronic disease prediction. Here, readers will come to choice know machine learning and classifiers made to get knowledge out of datasets.

Keywords: ECG, biosignals, machine learning, HRV, classification, prediction, cardiac diseases

1.1 Introduction

Biosignals are being used in various medical data, such as the electroencephalography (EEG), capturing electric fields created by brain cell activity, and magnetoencephalography (MEG) capturing magnet fields produced by electrical brain cell activity. The electrical stimulation comes from biological activity in various parts of the body. The most popular types of methods currently used to record biosignals in clinical research are described below, along with a brief overview of their functionality and related clinical application signals [1].

1.1.1 Heart System and Major Cardiac Diseases

The electrical activity generates the following types of signals:

• Magnetoencephalography (MEG) signals
• Electromyography (EMG) signals
• Electrooculography (EOG)signals
• Phonocardiography (PCG) signals
• Electrocorticography (ECoG) signals
• Electrocardiography (ECG or EKG) signals

Intervals between the waves are used as indicators of irregular cardiac operation, e.g., a prolonged PR interval from atrial activation to the start of ventricular activation may indicate cardiac failure [2, 3]. In addition, ECGs are used to study arrhythmias [4], coronary artery disease [5], and other heart failure disorders. In biosignals, the sampling frequency (or sampling rate) and the recording period are directly proportional to the data size and the data acquisition process speed. The ECG will be essential for the heart rhythm and disease research. The different heart conditions are as follows:

1. a) Arrhythmias
2. b) Coronary heart disease
3. c) Various types of heart blocks
4. d) Fibrillations
5. e) Congestive heart failure (CHF)
6. f) Myocardial infarction (MI)
7. g) Premature ventricular contraction (PVC)

1.1.2 ECG for Heart Rate Variability Analysis

Electrocardiogram (ECG) is a waveform pattern that describes the state of cardiac activity and cardiac safety. The ECG signal is non-stationary and non-linear. The ECG has a spectrum of frequencies between 0.05 and 100 Hz [6]. ECG analysis methods, including the heart rate variability (HRV), QRS identification, and ECG post-processing, have advanced considerably since device implementation. The word HRV reflects the interval difference between successive heartbeats.

1.1.3 HRV for Cardiac Analysis

The biomedical signal is an important health assessment parameter. For example, it has been used to detect and predict human stress [1], stroke, hypertension, sleep disorder, age, gender, and many more. The popular techniques to analyze the HRV fall into three categories as time domain, spectral or frequency domain based on fast Fourier transform (FFT) [7], and nonlinear methods consisting of Markov modeling, entropy-based metrics [8], and probabilistic modeling [9]. There are seven commonly used statistical time domain parameters [10] calculated from HRV segmentation during 5-min recording, comprising of RMSSD, SDNN, SDANN, SDANNi, SDSD, PNN50, and autocorrelation, which are considered for implementation. The HRV is also calculated by a device called PPA (peripheral pulse analyzer); it works based on pulses measured, which is different from HRV measurement using ECG. However, the focus would be on ECG-based HRV measurement, but the validation PPA-based method is considered [11]. Nonlinear measurement approaches aim to calculate the structure and complexity of the time series of RR intervals. HRV signals are non-stationary and nonlinear in nature. Analysis of HRV dynamics by methods based on chaos theory and nonlinear system theory is based on findings indicating that the processes involved in cardiovascular control are likely to interact with each other in a nonlinear manner. The more on indices (features/parameters) are discussed in Section 1.3.2.

1.2 Machine Learning Approach and Prediction

Learning is closely connected to (and sometimes overlaps with) quantitative statistics, which often concentrate on forecasting computers’ use. It has close connections with mathematical optimization, which provides the fields of methodology, theory, and implementation. The second sub-area focuses more on the study of exploratory data and is also known as non-monitored learning [2]. Unsupervised machine learning (ML) is also possible [11] and can be used to learn and construct baseline conduct profiles for different entities [12]. To gain knowledge of the past and to detect useful trends from massive, unstructured, and complex databases, machine learning algorithms use a range of statistical, probabilistic, and optimization methods [12]. These algorithms include automatic categorization of texts, network intrusion detection, junk e-mail filtering, credit-card fraud detection, consumer buying behavior, manufacturing optimization, and disease modeling. Most of these applications are performed using managed variants of the algorithms of ML rather than unattended [13].

The heart disease detail includes several features that predict heart disease. This large amount of medical data allowed data mining techniques to discover trends and diagnose patients. The historical medical data is very high, so it requires computational methods to process it. Data mining is a technique that removes the hidden pattern and uses as an analytical tool to analyze historical data. There are several different classification schemes for disease datasets. ML techniques are applied for classifying the statistical parameters above in a cardiological signal analysis to predict the RR interval estimate cannot be overemphasized. A precise method of calculation therefore needs to be developed. It is clear from the existing research theory that the conventional systems for chronic disease prediction are unable to establish reliable diagnostic systems as workers make it difficult to get correct responses and can minimize response time. Adaptive systems, by comparison, can increase the chances of success and can advise clinicians on care decisions. Current healthcare programmers can be enhanced by the efficient use of parallel classification systems, as they promote parallel implementation on multiple systems. Parallel classification systems also have a great potential to increase the predictive performance of diagnostic systems for chronic diseases [13, 14]. Here, classifiers are discussed out of the available are K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Ensemble AdaBoost (EAB), and Random Forest (RF).

1.3 Material and Experimentation

The proposed method comprises of two phases:

1. processing the enrolment database (PEP) and
2. Prediction (P).

Figure 1.1 shows that the research purpose types of database are created based on acquisition units. The standard database has varying sampling frequency which comprises of different age groups of male and female.

A total number of subjects and corresponding signal were acquired with different set conditions. This may comprises of female and male with varying age group with sampling frequency of 256 and 500 Hz [6, 15]. The model will be testing for cardiac HRV-based analysis with both the ECG and non-ECG (PPA). For the research purpose, the congestive heart failure, arrhythmia, sudden cardiac death, ventricular arrhythmia, CHF database data being considered along with externally obtained ECG and non-ECG.

1.3.1 Data and...