Is it possible that one day, a computer could take over the position of a doctor? Scientific research and development in artificial intelligence are discovering new ways that these human-like machines can aid medical professionals in doing their jobs more accurately and efficiently. Artificial intelligence, also known as machine intelligence, is a type of knowledge that allows machines to think like humans in order to complete difficult tasks and calculations. These machines not only sort through tons of data and are able to organize it in ways that are easily accessible, but they also translate languages, perceive visually, and make decisions about what steps to take when treating a patient. In today’s society, artificial intelligence machines are necessary for the management of tremendous amounts of medical data that is saved for each patient. AIs are beneficial because of the digital ways to manage and collect medical data, their ability to spot patterns in a set of data, and their contribution to medication discovery and genomics.
Artificial intelligence is a concept that has been discussed for far longer than most realize and it traces back to myths from the Greeks about robots that could think and act like humans. Hesiod and Homer were Greek poets who told the story of Talos around 700 BCE. Talos was described as a bronze man who was created by the god of invention. Not only was Talos artificial, but he had a cord that stemmed from within him that held a specific type of power, similar to the power that is within machinery today. Pandora, another ancient figure, was created by Zeus as an artificial woman who would punish civilization by opening an evil box. In both stories, the communities end up in shambles and it is as if the gods were trying to tell us that the interaction between artificially created beings and humans will be inevitably destructive. Although AI was thought to have been destructive to society back then, research and development in artificial intelligence has proven that this is not the case and that AIs make our society better off.
The first true design for artificial intelligence was developed by classical philosophers who attempted to describe the process of human thinking using symbols that machines were able to understand and analyze. The first human-like computer was produced in the 1940s. This computer was based on mathematical reasoning and encouraged researchers to build the first mechanical “brain.” John McCarthy is the scientist who is given credit for the first AI machinery. John McCarthy was an assistant mathematics professor at Dartmouth College. Himself, his students, and other research scientists worked for six to eight weeks to brainstorm ideas about a new machine that would think and process as humans do. This task proved to be harder than the team thought. Although they were not successful in producing the first analytical machine at that time, the term ‘artificial intelligence’ was developed. John McCarthy performed more research at the Massachusetts Institute of Technology where he finally developed LISP, a programmable language that is used in many AI’s today.
In 1960, the Dendral experiments were performed by Edward Feigenbaum. These experiments were performed originally to allow organic chemists to analyze molecules, but instead, a lot of information about AI’s analytical processes were obtained. It was uncovered that AI could be programmed to do cognitive tasks just like humans that could extend into fields of medicine, biology, engineering, etc. Edward Feigenbaum was also the winner of the A.M. Turing Award in 1944 for his developments in AI. The Turing Test was created by Alan Turing and put machines to the test as they had to accurately hold conversations with humans without being recognized as machines. If an AI was able to hold a believable conversation with humans, they were considered to be intelligent.
Some of the first artificial intelligence machines were based on the WABOT-1 robot that was created in Japan in 1972. It was the first computer with intelligence similar to human beings. The robot consists of a vision system, conversational system, and a limb system that allowed it to react and move similarly to a human being. Eyes, ears, and mouth were all present and contained receptors to sense the environment. It was reported that the WABOT-1 had a mental capacity of a 1-year-old so more development took place to increase this capacity even farther over a significant amount of time. Sophia, the first robot to ever receive citizenship in Saudi Arabia, was created by David Hanson who is the founder of Hanson-Robotics in Hong Kong. Sophia has facial expressions, a sense of humor and the motility and mental capacity of an adult human. The developments of AI have increased significantly since the development of WABOT-1 to the point where robots are now able to gain citizenship.
A very important function of AI that is intriguing to many computer scientists is its ability to process and store large amounts of data in short amounts of time. Computers are currently far superior to humans in their ability to quickly sort through, collate, measure and analyze large amounts of data. AI specializes in the field of what is known as “big data.” Big data is extremely large or copious amounts of data that are organized typically by computers to uncover certain patterns, trends, or associations. Big data is characterized by IBM as any collection of information that is at high volume, velocity, and variety. Volume refers to copious amounts of data that are generated from hundreds of different available sources. Variety refers to using multiple different types of data to make informed decisions. Velocity refers to the rapid increase of data all the time and the need for instant decision making.
When AI’s receive data, they are able to sort through the data in seconds, pinpointing similarities between different data sets that were learned before. The machines can take these similarities and process them to create very specific algorithms that allow the machines to make accurate predictions. AI’s ability to make rapid predictions and learn from other sets of data is what makes it the most efficient tool to manage over half of the world’s stock trades. AI can express itself by applying rules that have been programmed into it or by looking at specific examples to come up with similar conclusions. A typical AI can be knowledge-based or machine learning-based which aids in reducing misdiagnosis and improving patient care. Although AI has excellent analytical skills, they are not as good at understanding context. This makes them very efficient tools for doctors or physicians in medicine because doctors are able to apply this missing context that AI does not have the ability to comprehend. Since they can analyze data and create predictions very quickly, medical professionals can use an AI’s judgment as another opinion to help them correctly diagnose a patient.
A very common application of AI medicine is the computer-based interpretation of medical images by using segmentation. Image segmentation is when the computer breaks down an image into individual parts that are easier to read and analyze. To do this, the AI must examine each pixel of an image and assign a designation to it to create groups of pixels based on color, intensity, shape, or texture. AI’s function based on input, processing, storage, and outputting information to doctors. An example of AI functionality for segmentation would be the input of a digital photograph and the output would be a classification of sorts. A photograph of a patient’s skin lesion can be inputted by a Dermatologist and AI would segment it in order to determine if the lesion is benign or malignant. Besides analyzing an individual photo, AIs can compare imaging from other databases to their own images to find similarities or differences between the two. If the images are similar, the AI can diagnosis two patients with similar illnesses. The process of analyzing photos allows AI to learn from each case and store that information for future use.
Health records are compilations of a patients health history which includes diagnoses, treatments, medications, etc. An AI can use each patient’s medical history to guide patient management software. By accessing these medical records to view all past medical information that could pertain to a new discovery, AI will save time and improve efficiency within medical offices. Analyzing past medical information for a patient could lead AI to a conclusion of a new diagnosis, even for a disease or illness that seems to be unrelated. AI machines can work at a rapid pace and learn from each individual case very quickly. They can sort through the data of tons of patients at a time in much less time than any doctor or clinician. Nurses, doctors, physicians, and pharmacists see multiple patients every day. Each patient needs a specific type of care and this takes up a lot of time. Instead of medical professionals having to perform a consultation on their own, the most prevalent risks that are within a patient’s medical record could automatically be generated by AI. This would allow professionals to avoid lengthy consultations since the AI would generate what knowledge is necessary for diagnosing patients or providing them with medications. Consultations could be written into a summary letter by the AI and the doctor would amend or approve this letter after close inspection.
Usually, traditional methods of pattern spotting in a data sequence use linear regression models to create an equation that represents a certain pattern. Today, AI allows doctors and medical professionals to uncover complex patterns that cannot be formed into a simple equation. The evidence presented could be interpreted by the AI in the same way and with the same knowledge that a doctor would. Today, AIs can process limitless amounts of inputs and this is why AI outputs are being trusted with certain tasks where experts disagree. Pulmonary tuberculosis on chest radiographs is difficult for humans to understand, but yet AI is able to correctly diagnosis TB with a specificity of 100%.
Not only is AI able to segment pictures to make conclusions, but it is also able to segment pieces of DNA and make predictions about certain genetic related diseases. A segmentation procedure provides a conceptually simple approach to finding patterns in genomic data. Segmentation is the process of splitting each DNA piece into separate categories while also assigning labels to the segments. Each genome is divided by researchers into nonoverlapping segments, assigning each one multiple different types of labels. Each region that has the same segment label is then considered to have characteristics in common and can be grouped together in observed data. For example, ChromHMM is a type of software that annotates parts of the genome like histones, open chromatin, transcription factor binding, etc. The presence or absence of annotation can represent the different structures, locations, or spacing.
A category of genomic sequencing is SAGA, semi-automated genome annotation. These algorithms take as input a collection of genomics data sets from a particular cell type. They output a set of integer labels that correspond to the genomic activity of that section. The types of activity include an active promoter, active transcription, active chromosome repair, etc. Once the whole genome is labeled, it can be examined by a researcher and assigned an interpretation term that indicates its function. The Segway model is part of the SAGA category. It allows the entire human genome to be analyzed even when pieces of DNA are missing or damaged. It uses the Dynamic Bayesian Network, also known as DBN, which is a way to calculate the value of a variable based on the value beforehand even if data is missing from the genome. It related the genomes together based on each step as it makes conclusions about each piece of the genome. Segway has a specific label for each defined base pair in the genome and this variable/label is accompanied by an observation variable for each part. Each data value is weighted equally so that each data value has the same contribution to the model. Segway then uses each one of these data points and formulates an inverse sine function to represent that entire data set. Unlike the ChromHMM model discussed earlier, Segway analyzes a full data set and has a more precise resolution to ensure accuracy.
In an experiment led by Michael M. Hoffman, open chromatin that was derived from myeloid leukemia cells within humans was analyzed by Segway. Using this AI machinery, researchers were able to rediscover protein-coding genes and different chromatin states depending on the location of the genome. It also determined that transcription factor binding always occurs near a transcription start site and this was expected based on patterns and past research. Specific labels were associated with locations on the gene where enhancers and insulators. Patterns occurred in areas of the genome where there was little biochemical activity or in regions where the gene was not expressed. Thomas Chittenden, a geneticist at Wuxi NextCODE said, “AI is going to lead to the full understanding of human biology and give us the means to fully address human disease.”
AI has the ability to aid researchers in developing new medications that could save millions of lives. IBM Watson, a computer system capable of answering questions, is being used by Pfizer, a pharmaceutical company, to develop new drugs that use the strength of the immune system to fight certain cancers. Sanofi, a drug developing company located in Paris, is working to develop an AI platform in the United Kingdom for research about therapies for patients with metabolic diseases. In Massachusetts, AI systems are being used to search for new cancer treatments by Berg, a biotech company. A model was created using 1,000 cancerous and healthy human cells. The cells were all exposed to levels of oxygen and sugar that were similar to that of cancer cells. Each cell was then analyzed to view any changes in lipids, metabolism, enzymes, or proteins. Major differences between cancerous and healthy human cells were discovered during this experiment. By using certain amounts of sugar and oxygen, Berg was able to identify the true biological cause of disease and efficient ways to treat these problems. Many different pharmaceutical companies, including Berg, believe that AI is beneficial to drug discovery because instead of continuing to implement clinical trials that often end in unapproved medications, AI can identify what the issue is by using true human biology. This issue can then be correctly directly instead of hoping that a medication will be safe enough to use in the market. Reoccurring molecules in cancer cells were discover by the co-founder of Berg. BPM31510, a drug that Dr. Narain believes would be helpful in cancer treatment, is currently stuck in stage two of a clinical trial where it is being used on patients with pancreatic cancer.
AI is most efficient when used to sort through big data from multiple different sources that can all be used to make informed predictions. BenevolentBio has a new system of AI machinery that collects data from research papers, clinical trials, patient records, and many other sources. All this data is saved within the cloud where AI is able to access it at any time and form billions of biological relationships between topics like genes, disease and its symptoms, proteins, tissues, drugs, cells, DNA, etc. Graphs can then be created based on the different relationships found. For example, a disease and its symptoms along with the genes that cause it and the tissues/proteins that are affected can all be graphed together.