Artificial Intelligence in Medical

 

Medical Artificial Intelligence

Artificial intelligence (AI) was coined by John McCarthy at the Dartmouth conference in 1956,1 and the field of artificial intelligence originated from thereon. Kaplan and Haenlein defined AI as “the ability to process external data systematically and learn from it to achieve specific goals and tasks”. 2 AI involves using machines to simulate human thinking processes and intelligent behaviors, such as thinking, learning, and reasoning, and aims to solve complex problems that can only be solved by experts.3 As a branch of computer science, the field of AI mainly studies the following contents: machine learning, intelligent robot, natural language understanding, neural network, language recognition, image recognition, and expert system.4 The concept of artificial intelligence in medicine (AIM) originated in the early 1970s.5 It aimed to increase the efficiency of medical diagnosis and treatment with the aid of AI systems. After the 1980s, the development of AIM could be roughly divided into four stages: (1) infancy (1980s): the “decision tree” algorithm was proposed, and artificial neural networks continued to develop; (2) adolescence (1990s): “expert systems” continued to mature due to the emergence of support vector machines; (3) coming-ofage (2000s): the concept of “deep learning” was proposed, and machine learning became a prominent theme of AIM; and (4) currently, we are in the “maturation period” (2010s): as the technologies are relatively advanced. However, the ability to communicate with people still needs to be improved. Therefore, we are still in the stage of “weak” AI.

2. Application of AIM:-

2.1. Machine learning

In 1959, Arthur Samuel coined the term “machine learning” to denote a category of algorithms and construction of classifiers. The algorithm automatically learns through the input data and builds a model based on the input data to accurately predict new data. Subsequently, the algorithms of machine learning experienced many major breakthroughs: the backpropagation algorithm was proposed in the early 1960s. In 1982, Paul applied the automatic differentiation method in neural networks. In 1986, Ross Quinlan proposed a well-known machine learning algorithm, called the “decision tree”, which involves classifying data according to the set rules. Tin Kam Ho created an important algorithm, called the “random forest” (double spatial feature extraction algorithm), based on random subspace method, which was built using the decision tree. In 1995, Vladimir invented the support vector machine (SVM) model. In 2006, Geoffrey Hinton, a leader in the field of deep learning, proposed the deep learning algorithm. Deep learning is actually based on machine learning, and convolutional neural network (CNN) is one of its representative algorithms. In 2008, according to the standards of Association for the Advancement of Medical Instrumentation and the British and Irish Hypertension Society, a novel measurement model of blood pressure based on CNN— convolutional recurrent neural network-blood pressure (CRNN-BP) was constructed to solve the problem related to the extraction of pulse waveform feature points and low robustness in traditional medicine, and to increase the precision of the model. With continuous development of assisted diagnosis technology, a large data repository is generated during screening, diagnosis, and treatment of diseases. Organizing and analyzing these data in a short duration can be a challenge for the doctors. Therefore, machine learning is increasingly used in medicine to help doctors predict diseases and treatment outcomes. Random forest is one of the most efficient algorithms in machine learning. In recent times, random forest plays an important role in medicine, particularly in predicting diseases. Patients with a history of idiopathic hemorrhagic ulcers may demonstrate a higher incidence of ulcer recurrence. The safety of the patient is endangered, if a serious complication, such as ulcer rupture, occurs. In 2018, machine learning was used to build a model with high accuracy to predict idiopathic peptic ulcer rebleeding, which was called IPU-ML. In another instance, severe hand-foot-mouth disease caused by enterovirus can lead to serious complications, such as pulmonary edema and myocarditis, in very few children. In 2019, to predict severe hand-foot-mouth disease, Cat Boost model was build, which demonstrated higher specificity and sensitivity than other models, such as decision tree and SVM. In addition, machine learning can predict the efficacy of radiotherapy. For example, patients usually undergo radiation therapy, when they are suffering from lung cancer, especially small cell lung cancer. However, long term radiotherapy can lead to serious complications, such as radiation pneumonitis, which can result in respiratory failure and death. Using artificial neural networks, researchers deduced a method for predicting radiation pneumonitis. They also established a network with an extensive training in memory and data, which could exhibit higher accuracy in predicting the complications. The “black box” problem of machine learning needs to be resolved. “Black box” is a neural network comprising CNNs for feature extraction and recurrent neural networks with long short-term memory. Generally, a neural network comprises a series of neural layers, including input, processing, and output. The intermediate processing in the neural network is called the “black box”. It is the system that hides its internal structure from the user. By resolving the problem of black box, the accuracy and computing power of machine learning can be improved and its application range can be expanded, which in turn can result in more contributions to the medical field. 

2.2. Intelligent robots:-

 In 1979, the Robot Institute of America defined robot as “a reprogrammable, multi-functional manipulator designed to move materials, parts, tools, or other specialized devices through various programmed motions for the performance of a variety of tasks”. 20 Intelligent robots were used for surgery in 1980s.21 For example, PUMA 560 was employed in neurosurgical biopsy in 1985 and prostate surgery in 1988. ROBODOC, which was developed in 1992, was the first intelligent robot authorized by the U.S. Food and Drug Administration (FDA). It was mainly used for hip replacement procedures in orthopedic surgery. Currently, there are three kinds of robotic surgical systems approved by the FDA, including the ZUES, Da Vinci, and automated endoscopic system for optimal positioning robots. Owing to their minimally invasive, precise, and intelligent features, intelligent robots are widely used in orthopedics, urology, orthopedics, stomatology, and other fields. According to the type of orthopedic surgery, robots can be classified into three categories, including joint surgery robots, spinal orthopedic robots, and traumatic orthopedic robots. Femoral neck fractures can possibly occur in elderly patients with symptoms of hip deformity, pain, and dysfunction. These fractures can lead to complications, such as fracture nonunion and avascular necrosis of the femoral head, and the best treatment for these fractures is surgery. In 2018, a research explored methods to reduce the amount of haemorrhagia during the femoral neck fracture surgeries. They compared the two surgical methods of orthopedic surgery robot and manual nailing and concluded that with the assistance of surgical robots, the surgeon could locate the operative site accurately and reduce punctures, thus reducing the amount of haemorrhagia during the surgery. Intelligent robots are also widely used in gynecological surgery. For example, during the early stages of ovarian cancer, patients might suffer from abdominal mass, ovarian tumor pedicle torsion, and tumor rupture. Therefore, performing surgical treatments during the early stages of this cancer is important. A meta-analysis demonstrated that the Da Vinci System had many advantages during surgery, as it enabled the removal of numerous lymph nodes and a low blood transfusion rate in the patients. Therefore, this kind of surgical approach was even safer than the laparoscopic surgery. The robots that assist the surgeons in clinical practice are mostly discrete robots with limited mobility. However, in recent years, continuous robots (a novel bionic robot with “invertebrate” flexible structure) have been emerging, and they will gradually replace the discrete ones with their flexible bending characteristics and good adaptability to the environment. Continuous robots are expected to become the main force of surgery in the future. Although intelligent robots are widely used in the field of orthopedics, they still demonstrate drawbacks, such as high costs, large size, and limited application range. With constant advancement of medical and AI technologies, intelligent robots will be made to gradually adapt to the development direction of surgeries in the future.

2.3. Image recognition technology:-

Image recognition is the technology of processing and analyzing images through computers. It is an important technology in the field of AI, which is based on deep learning. The development of image recognition technology includes three stages—text recognition, digital image recognition, and object recognition. The recognition process can be divided into five steps, including processing the input, image preprocessing, image extraction, classifier construction, and producing the output. This technology can process image data rapidly and efficiently. For example, a study found out that the map projection technology could identify the bones that were most prone to fractures between femoral and trochanter more efficiently and accurately than traditional heat map technology. They established that the distribution area of the fracture line was relevant to the age and gender of the patients. Image recognition technology has been highly significant in the diagnosis and treatment of intertrochanteric fractures. In addition, it is also widely used in disease prediction, diagnosis, and lesion identification.

Currently, image recognition technology is applied in many clinical disciplines. Cervical cancer is one of the four major causes of death in women, and it is caused mostly due to infection with human papillomavirus. Patients do not experience any obvious symptoms in the early stages. However, symptoms, such as anemia and cachexia, may be experienced later. Although many treatments are available for patients with cervical cancer, such as surgery, radiotherapy, and chemotherapy, the prognosis of patients depends highly on whether the cancer is diagnosed at an early stage. Based on deep learning, intelligent image recognition of the cervix could assist doctors in the early diagnosis of cervical cancer with an accuracy rate of approximately 90%.

In 2017, a study explored whether the accuracy of diagnosing fungal keratitis could be enhanced using image recognition technology. The researchers adopted a method, called slit lamp microscopy, analyzed the experimental data, and finally concluded that the detection method based on image recognition demonstrated higher specificity and sensitivity in the diagnosis of fungal keratitis than cornea smear. In addition, the image recognition method could assist doctors with inadequate knowledge in diagnosing the disease more accurately.

Deep learning plays an important role in the application of image recognition technology to identify lesions. In 2017, a study was using CNNs to identify malignant breast lesions. This method performed better than the most advanced system in computer-aided detection, as the recognition accuracy of CNN was higher.

Although image recognition technology can aid doctors in diagnosing diseases in clinical practice, it cannot completely replace the role of doctors. Due to the different equipment used in varied hospitals, the resolution of the acquired images can also be different, which can affect the final diagnosis to some extent. Several difficulties are encountered during the application of image recognition technology. For example in multi-layer neural convolution, the training model requires a large amount of data, and the computational efficiency needs to be further improved. Besides, high performance supercomputers have not yet been popularized. Therefore, further research in the future is necessary to resolve the problems related to hardware devices, optimization algorithms, and technology integration.

2.4. Expert system:-

The expert system is a computer system that simulates the decision making ability of human experts. As one of the earliest successful AI software, it uses the existing knowledge system to reason and solve a series of complex problems. The development phase of the expert system can be roughly divided into three periods: the initiation period (1965–1971), maturation period (1972–1977), and the development period (1978–). In the early 1960s, the first expert system—Dendral system was designed. In 1972, the University of Leeds developed the AAPHelp system to assist in the diagnosis of acute abdomen. In 1974, the University of Pittsburgh excogitated the INTERNIST-I system, which was mainly used for the diagnosis of complex diseases in internal medicine. In 1976, Stanford University developed the intelligent diagnosis system MYCIN, which could aid in diagnosing infectious diseases. However, it was not used in clinical practice for various reasons, such as ethics. A study stated that the use of personal digital assistants (PDA) to provide expert knowledge to untrained helpers can significantly improve the quality of first aid. The evaluation of results concluded that the application of expert systems could improve the quality of bystanders' first aid, meanwhile strengthen the weakest link of the survival chain.

In another study, the expert system was used to diagnose different types of headaches, such as tension headaches, migraine headaches, and drugdependent headaches.42 The computerized headache assessment tool (CHAT) accurately diagnosed 94.4% of migraine headaches and 93% of daily syndromes. The average accurate rate of diagnosis was 98%. Therefore, the introduction of the CHAT expert system can aid doctors in identifying the type of headache diagnosed, which is of significant value in health care.42 In an analysis of the MIT-BIH arrhythmia database, a fuzzy expert system was adopted to distinguish arrhythmias and ischemic heart beats. The results demonstrated an average sensitivity of 96% and an average specificity of 99%.

The expert system exhibits a strong ability of clinical decision-making and shows advantages in terms of identification and diagnosis of diseases. However, it is also necessary to enhance the accuracy of the system, combine the system with the patient's medical history, and simultaneously integrate the doctor's clinical experience. In addition, the knowledge and findings in medicine must be constantly updated to provide doctors with frontier diagnosis and treatment plans.

3. Conclusion:-

AI is expected to encounter bigger challenges in future development: in data mining and machine learning, researchers are required to construct a model of black box predictor to resolve the “black box” problem and invent the 5th generation wireless technology (5G) and Internet of Things (IoT) integrated continuous robots; in image recognition technology, a more efficient training model needs to be created, and the expert system should continuously expand the knowledge base to provide more information to medical personnel.

In the past 10 years, significant breakthroughs have been experienced in the field of AI. Research institutions in many countries around the world have demonstrated extensive cooperation in realizing these breakthroughs. The amount of literature related to AI has developed rapidly at this stage, as researchers in China, Europe, and the United States have realized significant achievements in this field. Particularly, China is gradually becoming a leader in the field of AI.

With the high-speed transmission of 5G network, real-time remote technical guidance for remote collaborative surgery can guarantee the stability, reliability, and safety of the surgery. It enables experts to understand the progress of the procedure and the patient's condition in real time, thereby effectively reducing the risk of the surgery.

AI has completely changed the traditional model of medicine, significantly improved the level of medical services, and guaranteed human health in various aspects. A broader development prospect for medical AI is highly expected in the future.