Introduction
Now-a-days, researchers are increasingly looking into new and innovative techniques with the help of information technology to overcome the rapid surge in healthcare costs facing the community. Research undertaken in the past has shown that artificial intelligence (AI) tools and techniques can aid in the diagnosis of disease states and assessment of treatment outcomes. This has been demonstrated in a number of areas, including: help with medical decision support system, classification of heart disease from electrocardiogram (ECG) waveforms, identification of epileptic seizure from electroencephalogram (EEG) signals, ophthalmology to detect glaucoma disease, abnormality in movement pattern (gait) recognition for rehabilitation and potential falls risk minimization, assisting functional electrical stimulation (FES) control in rehabilitation setting of spinal cord injured patients, and clustering of medical images (Begg et al., 2003; Garrett et al., 2003; Masulli et al., 1998; Papadourokis et al., 1998; Silva & Silva, 1998).
Recent developments in information technology and AI tools, particularly in neural networks, fuzzy logic and support vector machines, have provided the necessary support to develop highly efficient automated diagnostic systems. Despite plenty of future challenges, these new advances in AI tools hold much promise for future developments in AI-based approaches in solving medical and health-related problems. This article is organized as follows: Following an overview of major AI techniques, a brief review of some of the applications of AI in healthcare is provided. Future challenges and directions in automated diagnostics are discussed in the summary and conclusion sections.
Artificial Intelligence Techniques
There have been a number of artificial intelligence (AI) tools developed over the past decade or so (cf., Haykin, 1999; Keckman, 2002). Many of these have found their applications in medical and health-related areas. Commonly applied AI techniques can be listed as:
• Neural networks
• Fuzzy logic
• Support vector machines
• Genetic algorithms
• Hybrid systems
In the following, we give a brief overview of neural networks, fuzzy logic and the relatively new support vector machines.
Neural Networks
Artificial neural networks work much like the human brain and have the ability to learn from training data and store knowledge in the network. In the learning phase, it maps relation between inputs and the corresponding expected outputs. During the learning phase, knowledge is acquired and stored in the network in the form of synap-tic weights and biases. This knowledge is used to make future predictions in response to new data or inputs during the testing phase. Usually, the network has one input and one output layer, and one or more hidden layers depending on the complexity of the problem. Learning can be supervised; that is, the network is provided with both the inputs and their desired outputs during the leaning process, or it can be unsupervised or self-organizing learning. There are a number of learning algorithms available (Haykin, 1999), and among them back-propagation learning algorithm is the most widely used. In this method, an error signal based on the difference between network-generated output (g) and desired output (di) is propagated in the backward direction to adjust the synaptic weights according to the error signal. During the learning process, the aim is to minimize an objective function such as the mean-squared error (E),
Neural networks are frequently used as diagnostics, and therefore it is important to have good generalization ability, that is, good performance in predicting results in response to unseen data. One limitation of neural networks is the possibility of being stuck in local minima during training rather than converging to the global minimum. To overcome this the network is usually trained several times with random initial weights to avoid converging to the local minima. Neural networks have found the majority of their applications in pattern recognition, time-series prediction, signal processing and financial forecasting.
Fuzzy Logic
Fuzzy sets were introduced by Zadeh (1965), and they deal with imprecise and uncertain information or data. Naturally, this has been found suitable for many medical and health-related problems, as it relates to the way humans think. Since the early work of Zadeh, there has been an exponential rise in the number of scientific papers applying fuzzy sets in biology, medicine and psychology areas (Teodorescu et al., 1998).
support Vector Machines
Support vector machines are a relatively new machine learning tool and have emerged as a powerful technique for learning from data and solving classification and regression problems. This has been particularly effective for binary classification applications. SVMs originate from Vapnik’s statistical learning theory (Vapnik, 1995). SVMs perform by nonlinearly mapping the input data into a high dimensional feature space (by means of a kernel function) and then constructing a linear optimal separating hyper-plane by maximizing the margin between the two classes in the feature space.
For m training data with input-output pairs (y1 ,x1),^,(ym,xm) where each input data belongs to a class yye{-1,+1}/=1 m, the decision function for a new data (x.) can be given by the sign of the following function (Gunn, 1998):
where, at is a nonnegative Lagrange multiplier corresponding to xt, K(.) is a kernel function and b is the bias. The Lagrange multipliers are obtained as the solution of a convex quadratic programming problem. The data points xis corresponding to a> >0 are called support vectors. Such xt s are the only data points in the training set relevant to classification since the decision surface is expressed in terms of these points alone (support vectors, SV). For linearly separable problems, the number of SVs and the hyperplane are determined by a subset of the training set only. For nonlinearly separable problems, ai in SVs are constrained by an upper bound C, which is regarded as a regularization parameter. This parameter makes a trade-off between margin maximization and minimization of classification errors in the training data set (Gunn, 1998).
Hybrid Systems
Recently, researchers have started looking into ways of combining various AI tools in order to maximize performance of the AI system. The main idea behind this is to offset limitation of one system by cascading with another AI tool. As a result, hybrid systems like Neuro-Fuzzy (neural networks and fuzzy logic), Neuro-SVM (neural networks and support vector machines) and Fuzzy-SVM (fuzzy logic and support vector machines) systems have evolved. Hybrid systems have been applied in many applications, including some biomedical areas (Teodorescu et al., 1998).
applications in healthcare AND medicine
In addition to applications in medical diagnostic systems, AI techniques have been applied in many biomedical signal-processing tasks, including analysis of ECG, EEG and human movement data (Nazeran & Behbehani, 2001). Neural network models have played a dominant role in a majority of these AI-related applications in health and medicine. Many of these applications are for pattern recognition or classification. A typical classification application usually has a number of steps or procedures as shown by the flow diagram (see Figure 1). This involves feature extraction from the input data before feeding these features to the classifier for designing and developing automated classification models, and finally testing the models for generalization.
Medical Decision Support Systems
Medical decision support systems (MDSS) are designed to construct a knowledge database by way of receiving a list of symptoms as input features and their corresponding disease type(s) as the output. Such a developed symptom-to-disease mapping system then facilitates the diagnostic process by generating new responses due to a new set of symptoms. Neural networks have been used to aid MDSS. Silva and Silva (1998) developed such a neural network-based MDSS system for a relatively small set of 20 randomly selected diseases and reported encouraging results. Disease symptoms in this study were represented as sinusoidal damped waveforms. Hybridization has been shown to improve diagnostic accuracy. For example, Dumitrache et al. (1998) reported an improvement in accuracy (by 28%) for medical diagnosis in cardiac disease using a hybrid decision-making system compared to a classical expert system.
Figure 1. Stages of a typical pattern recognition task
Cardiology
Several studies have applied neural networks in the diagnosis of cardiovascular disease, primarily in the detection and classification of at-risk people from their ECG waveforms (Nazeran & Behbehani, 2001). Celler and Chazal (1998) have applied neural networks to classify normal and abnormal (pathological) ECG waveforms: 500 ECG recordings (155 normal and 345 abnormal) were used to extract features from the QRS complex for training and testing the classifier. The abnormal ECG recordings had six different disease conditions. The classifier was able to recognize these waveforms with 70.9% accuracy.
Electroencephalography
AI tools, including neural networks, fuzzy clustering and SVMs, have been shown to be useful for analyzing electrical activity of the brain, the electroencephalogram (EEG) signals. Features extracted from EEG recordings of the brain have been used with AI tools for improving communication between humans and computers and also for effective diagnosis of brain states and epileptic seizures (Garrett et al., 2003; Geva & Kerem, 1998; Nazeran & Behbehani, 2001).
Opthalmology
Neural networks have been shown to be an effective diagnostic tool to identify glaucoma disease.
Glaucoma is more prevalent in older age and can cause loss of vision. Papadourokis et al. (1998) applied backpropagation neural network to classify normal patients and patients with glaucomatic optic nerve damage from perimeter examination. Several neural network models were tested using 715 cases, including 518 glaucoma cases, and they reported 90% recognition accuracy with two hidden layer networks and training with 80% of the input data. In an effort to compare effectiveness of different AI techniques in recognizing glaucoma diagnosis, Chan et al. (2002) used standard automated perimetry data to compare classification performance of several classifiers including multiplayer perceptron and support vector machines (SVM). In-depth analysis of performance of these classifiers was carried out using areas under the receiver operating characteristic (ROC) curves and also sensitivity (true positive rates) and specificity (1 – false positive rates) measures. Machine classifiers were found to perform superiorly in the classification tasks, whereas SVM showed significantly improved performance compared to a multiplayer percep-tron. A self-organizing fuzzy structure has also been developed and applied to predict the onset of hypoglycemia for diabetic patients (Hastings et al., 1998).
Gait Analysis and Rehabilitation
Gait is the systematic analysis of human walking. Various instrumentations are available to analyze different aspects of gait (cf. Begg et al., 1989). Among its many applications, gait analysis is being increasingly used to diagnose abnormality in lower limb functions, and also to assess the progress of improvement as a result of treatments and interventions. Recently, neural networks and fuzzy logic techniques have been applied for gait pattern recognition and clustering gait types. Barton and Lees (1997) classified gait patterns based on hip-knee angle diagrams and Holzreiter and Kohle (1993) used neural network models to identify normal and pathological gait patterns from measurements of foot-to-ground reaction forces using force platforms. Wu et al. (1998) applied a back propagation neural network to classify gait patterns of patients with ankle arthrodesis and normal subjects, and reported a superior classification by the neural networks compared to statistical technique (linear discriminant analysis) (98.7% vs. 91.5%). Gait analysis is being increasingly used in rehabilitation settings, and also combining with AI techniques to improve gait control and functionality. Tong and Grant (1998) applied neural networks to optimize sensor sets for control of FES system in people with spinal cord injury and showed improvements in accuracy as a result of neural network aided control. Fuzzy logic has also been recently applied with great success in: clustering children gait with and without neurological disorder (O’Malley, 1997) and also detection of gait events such as the foot contact and take-off during walking in the analysis of paraplegic gait (Skelly & Chizeck, 2001).
Support vector machine (SVM) has recently been applied to classify young and elderly gait patterns (Begg et al., 2003). Gait changes with aging, with potential risks of loss of balance and falls. Recognizing gait patterns with potential falls risks would help to detect at-risk people so that rehabilitation programs could be undertaken to minimize the risk of falls. AI techniques such as neural networks and SVMs have demonstrated their potentials for detecting gait degeneration due to aging and appear to have potential applications in falls prevention in the elderly.
SUMMARY AND FUTURE TRENDS
There are plenty of future challenges for AI to be routinely used in medicine and health. The use of automated medical decision support system in routine use in the clinic would make a significant impact on our healthcare system.
One possibility is that healthcare in the future will be built on knowledge networks (Erlandson & Holmer, 2003). Applications of telemedicine and informatics in healthcare can help to provide support to patients in remote areas, and to share expert knowledge or limited resources. Furthermore, effective networking between informatics and biomedical engineering can also help to complement each other’s knowledge and to fight in partnership the various challenges faced by the medical and healthcare systems.
One major aim of a classification task is to improve its recognition accuracy or generalization performance. Selecting features that offer the most discriminatory information between the classes or medical groups could help to improve classification accuracy. At the same time, removing redundant features might also improve the medical recognition task. It has been demonstrated in a number of studies that selecting a small number of good features in fact improves the recognition accuracy (Begg et al., 2003; Chan et al., 2002; Yom-Tov & Inbar, 2002), in addition to the added advantages of classifier simplicity and fast processing time. Therefore, a pattern recognition task may be improved by decomposing it into the following stages (see Figure 2).
Recently hybrid systems are emerging, combining various AI tools with improved performance in medical diagnosis and rehabilitation.
Figure 2. Stages of a pattern recognition task involving”feature selection” sub-task
Application of hybrid systems in medical diagnosis has already provided increased efficiency in diagnosis of cardiac disease by as much as 28% compared to classical expert system. This and other applications provide much hope and encouragement for more future applications based on hybridization of AI systems in helping to solve problems in medicine and healthcare.
conclusion
Artificial intelligence, particularly neural networks, fuzzy logic and the recently introduced support vector machines, played a key role over the years for many important developments in medicine and healthcare. Despite such developments there are many future challenges and currently only a few AI-based systems are routinely used in the clinical setting. Continued developments in AI fields are providing much impetus that is needed to tackle the many problems of the healthcare system.
KEY TERMS
ECG: The electrical activity of the heart recorded from the body surface using electrodes as electrical potential is known as electrocardiogram or ECG.
EEG: The electrical activity of the brain recorded from the scalp as electrical potential is known as electroencephalogram or EEG.
Fuzzy Logic: The concept of fuzzy logic is that many classes in the natural environment are fuzzy rather than crisp. It deals with imprecise and uncertain data.
Gait Analysis: Analysis of human walking patterns. It is used to analyze abnormality in lower limb problems and assess treatment or intervention outcomes.
Hybrid Systems: Integration of two or more artificial intelligence tools to improve efficiency or system performance.
Neural Networks: Neural networks resemble the human brain and able to store knowledge during training and use this for decision making during testing phase.
Support Vector Machines: Introduced by Vapnik and capable of learning from data for solving classification and regression problems.
