Introduction: Artificial intelligence is a branch of computer science that enables the analysis of complex medical data. Their potential to exploit meaningful relationships in data sets can be used in many clinical scenarios for diagnosis, treatment, and prediction of outcomes.

Methods: Medline and Internet searches were conducted using the keywords “artificial intelligence” and neural networks (computers). Additional references were obtained by cross-referencing key articles. An overview of different artificial intelligence techniques was presented and important clinical applications were reviewed.

Results: Artificial intelligence techniques have been explored in relatively great depth in almost every field of medicine. Artificial neural networks are the most commonly used analytical tool, while other artificial intelligence techniques such as fuzzy expert systems, evolutionary computation, and hybrid intelligent systems have been used in different clinical settings.

Discussion: Artificial intelligence techniques have the potential to be applied in almost all areas of medicine. Further clinical trials with appropriate designs are necessary before these emerging technologies can be used in real clinical settings.

Keywords: Artificial intelligence, neural networks (computers)

Artificial intelligence (AI) is defined as "a field of science and engineering concerned with the computational understanding of what is often called intelligent behavior, and the creation of artifacts that exhibit such behavior". Aristotle attempted to formalize "correct thinking" (logic) through his syllogism (a three-part deductive reasoning). Many modern works were inspired by this, and early studies of the workings of the mind helped to establish contemporary logical thinking. Programs that make computers behave in a way that makes them appear intelligent are called artificial intelligence systems. British mathematician Alan Turing (1950) is one of the founders of modern computer science and artificial intelligence. He defined intelligent behavior in computers as the ability to achieve human-level performance in cognitive tasks, which later became the popular "Turing test". Since the middle of the last century, researchers have explored the potential applications of intelligent technology in various medical fields. Gunn first studied the application of artificial intelligence technology in the field of surgery in 1976, when he explored the possibility of using computer analysis to diagnose acute abdominal pain. In the past two decades, there has been a surge of interest in medical artificial intelligence.

Modern medicine faces the challenge of acquiring, analyzing and applying the large amount of knowledge necessary to solve complex clinical problems. The development of medical artificial intelligence is related to the development of artificial intelligence projects aimed at helping clinicians formulate diagnoses, make treatment decisions, and predict outcomes. They are designed to support healthcare workers in performing their daily tasks, assisting with tasks that rely on data and knowledge operations. These systems include artificial neural networks (ANNs), fuzzy expert systems, evolutionary computing, and hybrid intelligent systems.

Artificial Neural Networks

Artificial neural networks are the most popular AI technique in medicine, judging by the number of publications over the past 20 years. Artificial neural networks are computational analysis tools inspired by biological nervous systems. They consist of a network of highly interconnected computer processors, called "neurons," that are capable of parallel computation for data processing and knowledge representation. Their ability to learn from historical examples, analyze nonlinear data, handle imprecise information, and apply models to independent data makes them very attractive analytical tools in medicine.

The first artificial neuron using a simple binary threshold function was invented by McCulloch and Pitts (1943). The next major milestone was the introduction of the perceptron as a practical model by psychologist Frank Rosenblatt in 1958. Many variations of the basic perceptron network have been proposed, but the most popular model is the multilayer feedforward perceptron. These networks consist of layers of neurons, typically an input layer, one or more intermediate or hidden layers, and an output layer, each fully connected to another layer. These neurons are connected by links, each of which has a numerical weight associated with it. A neural network "learns" by repeatedly adjusting these weights. An important feature of artificial neural networks is that they can learn from the experience they have in the training environment. The use of multilayer feedforward perceptrons was limited due to the lack of suitable learning algorithms until Paul Werbos (1974), a PhD student, introduced "back-propagation" learning. Some other popular network designs include Hopfield networks, radial basis functions, and self-organizing feature maps.

Artificial neural networks have found a variety of applications in the real world. Their ability to accurately classify and recognize patterns has attracted researchers to apply them to solve many clinical problems. As we realize that diagnosis, treatment, and predictive outcomes in many clinical situations rely on the complex interaction of many clinical, biological, and pathological variables, there is a growing need for analytical tools such as artificial neural networks that can exploit the complex relationships between these variables. Baxt was one of the first researchers to explore the clinical potential of artificial neural networks. He developed a neural network model that could accurately diagnose acute myocardial infarction, which was prospectively validated by the latter with similar accuracy. Since then, artificial neural networks have been applied in almost all fields of medicine. Artificial neural networks have been used in clinical diagnosis, image analysis in radiology and histopathology, data interpretation, and waveform analysis in critical care settings.

Stamey et al. developed a neural network-derived classification algorithm that can classify prostate glands as benign or malignant. The model was subsequently validated in prospective studies with a diagnostic accuracy of 90%, a sensitivity of 81%, and a specificity of 92%. Other surgically related applications of anal network diagnosis include abdominal pain and appendicitis, retained common bile duct stones, glaucoma, and back pain. Artificial neural networks have also been used to diagnose cytology and histology specimens. PAPNET, a neural network-based computer automated screening system, has been developed to assist cytologists in cervical screening and is one of the few commercially available artificial neural network models. Breast, gastric, thyroid, oral epithelial, urothelial, pleural, and peritoneal effusion cytology have all been analyzed by neural networks with varying degrees of success. In radiology, human observation and directly digitized images can be used as input to the network. Artificial neural networks have been used to interpret plain radiographs, ultrasound, CT, MRI, and radioisotope scans.

Neural pattern recognition capabilities have been used to analyze a variety of waveforms, including electrocardiogram interpretation to diagnose myocardial infarction, atrial fibrillation, and arrhythmias. Analysis of the electroelectrograph (EEG) by neural networks has led to its application in the diagnosis of epilepsy and sleep disorders. They were also trained to analyze electromyography and Doppler ultrasound waveforms and hemodynamic patterns in critical care patients.

Prognosis is extremely important in planning appropriate treatment strategies and follow-up. Accurate identification of high-risk patients may lead to targeted aggressive adjuvant therapy, which can help cure the disease and prolong survival. The ability of neural networks to exploit nonlinear relationships between variables is particularly suitable for analyzing complex cancer data. Studies have shown that neural networks can predict survival in patients with breast and colorectal cancer. They also performed better than colorectal surgeons in predicting the prognosis of colorectal cancer patients. The authors of this article have demonstrated the generalizability of neural networks, and once trained on specific data, the network was able to predict the outcomes of patients from independent institutions without retraining. Artificial neural networks have also been used to predict the prognosis of lung and prostate cancer. They have been applied to predict prognosis in intensive care units and have outperformed the APACHE II disease severity scoring system.