Artificial Intelligence and Machine Learning in Healthcare Applications
Fares Alshammari
|
Department of Health Informatics, College of Public Health and Health Informatics, University of Ha'il, Ha'il, Kingdom of Saudi Arabia. |
Abstract
Artificial intelligence (AI) is a field of computer science concerned with the simulation of human intelligence by smart machines and computational rationality. AI is about to transform medical practice. With the help of this technology, a doctor can examine and treat the patient without visiting any hospital or clinic. Thus, this technology is now accessible to provide online services to patients. Any patient complaints may be rapidly addressed with various health issues. AI has been studied in several fields of medical practice and health science, including population health, precision medicine, and natural language treatment. Our objective is to review the applications of AI in the field of health science research. In this research, how AI helps solving the difficult medical problems through extensive research and development is demonstrated.
Keywords: Artificial, intelligence, machines, computer
Introduction
In recent years, AI has utility and effectiveness in solving many real-world computation-intensive problems, such as intelligent transportation systems, agriculture, education, Chatbot, Sentiment Analysis, Prediction, and especially healthcare systems [1-3], etc. AI has been introduced into the field of medicine to keep a digital medical record and perform examinations by utilizing intelligent technologies. It has solutions, specifically in targeted therapy, single-composition medicament, and personalized cures. AI simulates human intelligence characteristics through the use of machine-like computer systems. It’s also able to rapidly analyze, conclude, predict, learn, and even correct itself. Also, it can plan images, recognize speech, and learn a specific trait. AI systems train certain datasets to better outcomes and accurately solve complicated issues [4, 5]. AI is an innovative technology that aims to help the surgeon through medication [6], treatment, and surgery [7]. Its principal application for complicated cases is to improve decision-making. Besides, this technology detects, investigates, tracks, and controls infections in hospitals [8]. It can also develop and optimize the online appointment platform for patients. Moreover, to serve humanity, AI will be useful in all medical areas.
AI behaves as any human being science. It quickly captures text, images, medical data, and bioinformatics. AI-based machines can understand human languages to decide without making any errors which helps AI-assisted surgery robots in better quality and results [7, 9, 10]. In rural areas there is a lack of health care providers and this technology may be successfully utilized for filling this gap. It ameliorates the quality of medical studies to meet an urgent demand in rural areas [11, 12]. Indeed, AI increases the efficiency of health professionals. It also ameliorates the quality of health services with a lower cost. Accordingly, physicians are advised towards an accurate diagnosis [13, 14]. Furthermore, it was an important function in scanning technologies like 3D scanners, magnetic resonance imaging, computed tomography, and X-rays [15]. AI creates data through various virtual media and regularly communicates with the patient, therefore, a better decision will be made [1]. Patients profit from timely and accurate decisions [16-18]. AI appeared as a very good technology to be used for higher life expectancy.
This technology is very useful. It is flexible, adaptable and it recognizes patterns and computes fast [19]. Many superb computer scientists have carried out AI for 30 years. In the last decades, medical doctors have utilized AI to overcome medical issues. More recently, a significant AI branch has been developed which is called Machine Learning (ML) [1].
The main purpose of this review paper is the exploration of the AI and ML in the evolution of Health science. In this work, we provided the latest developed applications in various intelligent medical field based on AI and ML. We focused on the role and impacts of previous technologies on the most important aspects of health science, such as health recommender systems, electroencephalography analysis, disease identification and diagnosis, intelligent robot surgery, and image recognition technology. To achieve this aim, the structure of the review is as follows. In section 2, we define ML and the most used categories. In section 3, a detailed literature review of AI and ML techniques applied to health care are conducted. In section 4 a discussion is provided. Finally, the 5th section is a conclusion.
Machine learning
Figure 1: Classification of ML techniques
ML enables machines to be smarter without the humans’ intervention. The machines can learn from previous experiences and observations, when focusing on the design, analysis, development, and implementation of methods for accessing and using data. Thus, observations like instructions, direct experience, or examples are necessary to make a good decision [20].
ML automatically evaluates medical results and accurately presents them [21, 22]. The algorithm decisions can be made, including supervised learning, unsupervised learning, semi-supervised and reinforced learning by using ML. ML is used to identify disease probability [23-25]. It actually saves patients’ record for better treatment [26-28].
ML is subcategorized into three subtypes, as shown in figure 1. First, the system in supervised learning infers a function from labeled training data. Second, the system in unsupervised learning tries to infer the unlabeled data structure. Third, the system in reinforcement learning interacts with a dynamic environment.
Supervised learning is an automatic learning approach derived from the data of labeled training. These latter are made up of a set of training samples. Each example of the supervised training data set includes one input vector, a pair of input objectives, as well as one preferred output value. An algorithm, within supervised ML, will analyze the training data. Then an inferred function named a classifier will be made. Indeed, such an inferred function must predict the output value in an accurate way for any suitable input object. The learning algorithm will reasonably generalize the training data to situations unobserved previously. Supervised learning is divided into two types of learning tasks: classification and regression. The classification models seek to predict distinct classes, like the example of blood groups, whereas the regression models try and seek to predict numerical values. Some famous examples of supervised learning are decision trees, linear regression, random forest, logistic regression (LR), and Support Vector Machines (SVM).
In contrast, unsupervised ML concerns a situation that tries to discover a structure that is hidden in unmarked data. Because the examples provided to the learner are not labeled. In general, no error or reward signal exists for evaluating a potential solution. A reward signal can be an important factor in distinguishing supervised and unsupervised ML. In the field of statistics, unsupervised learning is linked to the estimation of density. We look forward to learn how to inherit the structure of our data without using explicitly supplied labels. In addition, unsupervised learning is utilized for dimensionality reduction, association analysis, and clustering. The latter is to separate data in meaningful subclasses. These latter are named clusters, according to the similarity between data objects, the data objects likeliness is an important criterion for clustering. Few examples of such algorithms include auto-encoders, principal component analysis, and k-means clustering.
It is a new type of ML that recently got much attention. In reinforcement learning, machines are not provided with examples of correct input-output pairs. On the other hand, a method is provided for the machine with the target of quantifying its performance as a reward signal. Reinforcement learning methods are like animals’ and humans’ learning. Machines try multiple and various things and are rewarded when doing something good.
This kind of learning is useful when solution spaces are large and infinite, and when machines can be thought of as agents that interact within their environment. Video games, like AlphaGo or others, fit this task because scores work well as rewards. The machine will go on learning by simulating which patterns maximizes its reward.
The distinct outputs of both principal types of ML methods are illustrated and presented in figure 2. Labeled samples with output and input data in supervised learning are used for the development of a model approximating the relationship between these values. The developed model is used for making future specific output predictions which provided a set of input values. The output value in regression analysis is continuous. On the other hand, for conventional classification tasks, the test case will be classified by the output into a specific class. In unsupervised learning, the algorithm will just use the input data to propose the natural structure present globally in the data points, without the use of one specific output for prediction. Similar to the outputs of supervised learning, the outputs of unsupervised learning are grouped into more discrete or continuous clusters.
Figure 2: Differences between machine learning methods.
Machine Learning in the Medical Field
ML has shown its utility and effectiveness in solving many real-world computation-intensive problems, such as chatbot, sentiment analysis, finance prediction, medical field, and search engines [29]. In medicine field, ML facilitates different highly complicated and time-consuming tasks. We will briefly discuss the most recent research trends as well as the activities related to health care based on ML.
A Health Recommender System (HRS) contains several phases according to the basic architectures of health information systems as depicted in figure 3. Cloud computing, an increase in data rates, high-performance internet of things (IOT) devices, and advanced sensors are four significant concern keys for a successful HRS in health informatics.
HRS led an extensive use of AI and ML techniques in advanced health care systems. These are known as health intelligence [29-32]. These techniques have played an essential role in disease diagnosing [33], medical imaging, social media analytics, and cure prediction, for disease [34, 35]. For chronic disease diagnosis and monitoring, HRSs have an essential function of a continuous monitoring and supporter. The authors in [36] proposed an intelligent smartphone-based HRS, to monitor patients depression and mental disorders.This device provides treatment if necessary. They divided 1047 volunteers’ data into a testing set and a training one. Then, they constructed a depression prediction model through the use of a decision tree and support vector machine (SVM) algorithms.
Figure 3: Health recommender system architecture.
In [37], an HRS,especially for patients with diabetes, was designed to ameliorate life quality by helping a patient to predict disease risks and reliable health recommendations. Also, a precise prediction model was built for diagnosing the risks linked to diabetics via the application of several classifications utilizing decision tree algorithms for prescribing more precise medical advice through the application of unified collaborative filtering about some external features, medical history, etc. They applied the decision tree to build one model that predicts, and diagnoses diseases and their risks. One overall random forest model was developed utilizing some decision trees. As a result, the HRS presented in [37] proved to be more efficient in terms of recall and accuracy utilizing the random forest algorithm in comparison with the other algorithms like the REP-tree and the decision stump.
The decision tree algorithm has been utilized for regression and classification of problems and can be utilized for HRSs on its own or combined with other supervised classifiers. The authors in [38-40] studied smartphones based on wrist-worn motion sensors to identify several complex daily actions such as eating, smoking, and drinking coffee. In [38], three different classifiers, which were Naive Bayes, decision tree, and K-Nearest Neighbor (KNN) were utilized to function with various window sizes to recognize a complex or simple activity. In [39], the authors achieved a good accuracy to classify stationary, running, and walking activities. The writers of [40] deployed both SVM and decision trees in their framework.
Deep learning, as an ML sub-field has been founded upon algorithms for learning multiple representation levels with the objective of modeling complex relationships among data [41], as well as the neural network theory while using multiple processing layers. Generally, these latter will learn the abstract representations of input data. Such algorithms will extract low and high-level features, in an automatic manner, needed for classification. Convolutional Neural Networks (CNNs) which are deep learning algorithms and their inputs are images, extract the features of the objects in such images and differentiate them from one another. Also, the need for pre-processing in CNNs is less than the other classification algorithms. CNN's represent very well-known techniques utilized in computer vision systems and image recognition. In [42], they proposed a novel approach to detect an automatic way of diagnosing the Myocardial Infarction (MI) using electrocardiogram (ECG) signals. In their work, the authors implemented a CNN algorithm to automatically detect normal and MI ECG beats. The writers also achieved an average accuracy of 93.53% through the use of ECG beats with noise and 95.22% without noise. This suggested algorithm could perfectly detect the unknown ECG signals even with noise. Thus, such a good system can be introduced in hospitals to support the clinicians' diagnosis of MI. The proposed system has been good and advantageous for the early diagnosis of cardiovascular diseases.
Existing studies based on SVMs [43,44], logistic regression [45], CNN's [46], Neural Networks (NNs) [47], and others [48], have proposed and developed HRSs for people suffering from dementia. This has caused memory loss to patients whose number has been rising worryingly worldwide. The authors in [49] provided an ML state-of-the-art review. The ML approaches were applied to health informatics for what related to dementia care. Indeed, the writers collected and reviewed the existing scientific methodologies. They identified the relevant issues and challenges in case they had faced big health data. ML demonstrated promising applications for the analysis of neuroimaging data concerning dementia care. In general, little effort has been made to take advantage of heterogeneous integrated information using sophisticated ML approaches.
Electroencephalography (EEG) has been a vital method to identify certain health conditions in patients starting from the date of its discovery. Researchers have proposed several ML methods developed with bioengineering applications for EEG analysis. Table 1 discusses and presents different ML methods developed for EEG analysis.
In paper [50], the authors reviewed literature within the period 1988-2018 to capture earlier as well as present EEG classification methods in multiple applications. Accordingly, they determined the overall effectiveness of every ML method. The writers also noticed that all the primary methods utilized for ML were applied in some EEG classification forms. These methods include decision tree, naïve Bayes (NB), SVMs, CNNs, logistic regression, KNNs, recurrent neural network (RNN), and random forest. Each method has its accuracy within respective applications. From another perspective, higher overall classification accuracy will be achieved when some methods are combined and implemented correctly. The authors in [50] presented a comprehensive summary of the ML applications utilized for EEG analysis. They gave an overview of every method and its applications.
Table 1: ML methods for EEG analysis
Machine Learning Method |
Application |
Authors |
Accuracy Results |
Logistic Regression |
Epilepsy Classification |
[51] |
95.88% |
Linear Regression |
Robust ECG Artifact Removal |
[52] |
98.11% |
Logistic Regression |
Motion Discrimination |
[53] |
77.0% |
RNN, CNN, neural networks, Logistic Regression, |
Identification of Automatic Abnormal EEG |
[54] |
3.47% Better accuracy was obtained with RNN |
SVM |
Emotion Classification |
[55] |
96.83% |
SVM |
Multimodal Facial Recognition |
[56] |
82.75% |
SVM, KNN |
Multiple Sclerosis Detection |
[57] |
93.08% |
SVM |
Alcohol Use Disorder Detection |
[58] |
98.0% |
Naive Bayes |
EEG Classification |
[59] |
81.07-91.60% |
Naive Bayes |
Major Depressive Disorder |
[60] |
93.6% |
Naive Bayes |
Brain Activity Classification |
[61] |
87% |
epileptic seizure detection |
[62] |
100% |
|
Random Forest |
Characterize and Quantify Tonic Thermal Pain |
[63] |
89.45% |
ML is effective in predicting the infection risk within patients, while creating ahead of time alarms, hence helping medical teams respond quickly. Early diagnosis provides patients with appropriate care, enables patients to minimize the damage produced by disease or isolate and avoids the spread risk.
In [8], the authors surveyed the state of the art in what concerns AI-based infection prediction through the use of a systematic literature review. They reviewed 101 relevant documents published within the period 2003-2019. The objective was to study the papers where AI and ML were utilized to predict infections in patients by the means of physiological data as features. The writers described all the whole review process carefully, and eight databases were taken into account indexing most of the literature which were in different scholarly formats. Indeed, these authors concluded that the most usually focused infection was by a considerable margin of the sepsis, followed by the infections of both Clostridium difficile and surgical site. Most studies used AI and ML techniques. The logistic regression, SVMs, naive Bayes, and random forest were the most common ones. The authors showed that the automatic diagnosis of any infectious disease utilizing ML was well documented within the medical literature. Table 2 discusses and presents different ML methods developed for predicting infections utilizing AI.
Table 2: ML methods for Predicting Infections
Paper |
Infection |
ML Methods |
Features |
Accuracy |
[64] |
Sepsis |
LR, SVM, ANN |
Temperature, respiratory rate heart rate, blood pressure |
93% |
[65] |
Sepsis |
LR |
oxygen saturation, temperature, respiratory rate heart rate, |
73.98% |
[66] |
Sepsis |
KNN |
Microbiology data |
94.55% |
[67] |
Sepsis |
CNN, SVM, LR |
oxygen saturation, temperature, respiratory rate, heart rate, Lab Test |
87.5% |
[68] |
Surgical site infection |
SVM |
oxygen saturation, temperature, respiratory rate heart rate, Lab Test |
not mentioned |
[69] |
General Infection |
SVM, KNN, LR |
Temperature, respiratory rate heart rate |
90.2% |
[70] |
Surgical site infection |
SVM, LR |
Laboratory tests |
86% |
[71] |
Sepsis |
LR |
Temperature, respiratory rate heart rate, blood pressure |
78.9% |
The identification and diagnosis of diseases are the main motivations that the medical field can benefit from largely. ML can be used to help doctors save time through the detection of diseases in their early stages. Cancer detection is an area that has been greatly studied in ML research. Microsoft launched in 2010 "nnerEye" [72], which was an ML project that could detectbrain tumors and identify its stage in minutes, instead of a lot of hours by humans. Other research has to concentrate on the detection of breast cancer [73] to facilitate the earlier detection of the disease and the analysis of breast cancer diagnoses.
Parkinson's disease can be described as a neurological movement disorder. The accelerometer signals captured by portable sensors tied into every patient can be beneficial for monitoring this illness. The writers in [74], developed such a system. They compared standard ML pipelines to CNN-based deep learning. The experimental results showed that deep learning was better than the other state-of-the-art ML algorithms, in terms of classification rate, by at least 4.6 %. The authors also discussed deep learning disadvantages and advantages of movement assessment based on the sensor. They concluded that deep learning was a promising method.
Dyslexia is a learning disability that affects nearly 10% of the world population. Identifying dyslexic children at an early phase is extremely important. Researchers forwarded various respective techniques to identify dyslexic children. Research found in the literature has basically utilized SVMs [75,76], naive Bayes [77,78], logistic regression [79], CNN's [80], KNNs [81] and linear discriminant analysis as ML algorithms for classifying participants. SVMs have generally been the most commonly used algorithm. The problem has been in essence a binary classification problem (identifying dyslexic and non-dyslexic users). The authors in [82] reviewed existing dyslexia detection techniques which had utilized ML approaches.
Decision tree algorithms were applied in a successful way to diagnose MR imaging Alzheimer’s disease and schizophrenia and to also classify lung nodules [83,84]. Specifically, these types of algorithms were utilized in cardiac imaging to predict cardiac risks and mortality [85,86].
Robots have been greatly utilized in surgery. Previously, discrete robots with reduced mobility assisted surgeons in clinical practice. With the constant progress of AI and medical technologies and the appearance of intelligent and sophisticated robots, old robots have been gradually replaced by flexible characteristics and good environmental adaptability. These intelligent robots are expected to become a very significant force of future surgery [7, 87]. They are manufactured to gradually adapt to the direction of future surgical development [88].
The authors of the papers [88,89] showed that using various ML approaches would enable the improvement of the accuracy of lengthy robot-assisted surgeryprediction, hence it increases the utilization of this resource.
Image recognition technology is defined as a method for recognizing images to analyze and process them by computer. This is an important AI technology. Indeed, it has been based on deep learning [90]. The authors in [91] showed that deep learning had an essential part in applying the image recognition technology for detecting and identifying lesions. In 2017, a study [90] used CNNs to identify cancerous breast lesions. The accuracy of CNN recognition was better. In another deep learning-based study [92], intelligent cervical image recognition could help doctors to diagnose cervical cancer early with a precision rate of around 90%.
Many researchers have claimed that several difficulties were encountered when applying image recognition technology. For example, the learning model within multilayer neural convolution needed much data, and the efficiency of computers required further improvement. Moreover, high-performance supercomputers were not popular. Therefore, in the future more investigations will be needed to resolve issues related to hardware, technology integration as well as optimization algorithms [90, 91].
There are a lot of other sectors where ML is effectively used, such as recording and storing medical data [93,94], managing all medication systems [95,96], analyzing different tests [94], and correct diagnosis and treatment [97,98].
Discussion
The application of AI and ML in health care science is a very common area of research.Thousands of papers have been published in this field each year since 2012. Considerable effort is needed for researchers to keep up with the latest developments in this area, and dedicated literature reviews are necessary to save researchers time. The number of articles published in this area is essential for an in-depth review, so we proposed two criteria for selecting the articles to be examined in this study: i) journal articles are preferred to conference articles; ii) recent articles are preferred to older articles; about multiple similar articles, those with higher citation were mentioned, and new contribution and better data sets were selected for review.
Through this review paper, we found a difficulty to compare studies. This is because several ML methods are trained and validated on different datasets in a variety of studies, making results difficult to be compared with other studies. As a result, almost all studies reported better performance compared to other studies. A study presented in [99] proved that AdaBoost is the most excellent ML method. But, in other papers [100] it is demonstrated that the ANN method is more precise than AdaBoost, SVM, and logistic regression. Other studies [55] and [58] demonstrated that the SVM was more precise than both ANN and naive Bayes. Nevertheless, the authors in [51, 52] asserted that logistic regression and linear regression classifiers had better predictive results compared to the SVM classifier. Additional conflicts appeared in other researchers' works, their writers confirmed that SVM combined with naive Bayes is more accurate than ANNs and decision trees [62]. A suitable workspace with a large-scale open-source dataset is necessary to improve comparability between studies by examining different methods on a standardized dataset.
Conclusion
In this research paper, a literature review of AI and mainly ML techniques for the health science and medical field is conducted. ML can help to supervise and decide a suitable treatment for patients. It evaluates images without any medical doctors, clinicians, or surgeons. ML-based technologies provide an assessment that facilitates predicting medical emergencies. It is useful to offer any medical consultation with a digital application. Practically, implementing such technology will improve the accuracy and performance of the diagnosis and treatment. We can apply it to reduce the medical cost and prevent diseases. It is also used to decrease unnecessary hospital appointments and answer patient questions. ML can identify problems during the unavailability of the doctor. It is helpful in the identification of the origin of the disease, and it can provide better medication for the patient. It accelerates the clinical examinations to produce a decisive result. ML creates the analytical algorithms of various features from the patientsdata, which are useful for providing knowledge about the patient and the disease degree.
Funding: This study was self-funded.
Conflict of Interest: The author declares that there is no conflict of interest.
References