See the Full Issue
January 2024
Artificial Intelligence in Vet Med
The rapid progression of AI has been facilitated by the enormous increase in data capture and availability, computer storage and processing capabilities, and neural network architectures. As governments, businesses, and organizations are struggling to keep up with the pace of developments, it begs the questions: What is on the horizon for veterinary medicine, and what will be the impact of AI?
What’s on the Horizon?
Artificial intelligence (AI) is no longer science fiction—it’s science fact. It’s the realization of a concept that has captured human imagination since the ancient Greeks first proposed that reasoning was not limited to humankind. Since AI’s humble roots in the 1950s, it has progressed to generate a vast array of valid and practical outputs by solving defined problems through data analysis.
More recently, rapid progression of AI has been facilitated by the enormous increase in data capture and availability, computer storage and processing capabilities, and neural network architectures. As governments, businesses, and organizations are struggling to keep up with the pace of developments, it begs the questions: What is on the horizon for veterinary medicine, and what will be the impact of AI?
Broadening Narrow AI
Current AI applications all fall within the realm of “narrow” or “weak” AI; that is, algorithms that take in data/inputs and produce a limited, specified output. While the nomenclature belies the complexity of some of these tools, they are far removed from the “general” or “strong” AI that would be more equivalent to broad human intelligence—where multimodal data feeds in to produce a variety of unspecified outputs as the algorithm learns and adapts. The move from narrow to general AI represents a singularity point beyond which the capabilities of AI will be equivalent to, or greater than, human intelligence. Beyond that, the dawn of superintelligence would leave humans in its wake (at least in theory, we hope).
Narrow AI tools can be thought of as discrete islands of solutions in the ocean of problems, each answering a defined question within its given boundaries. As the number of islands increases, they fill more gaps in the ocean. This “island generation” is where the field of AI currently sits, as individual companies and organizations work to develop individual tools to meet specific needs in the environment—like a computational microevolutionary process. This island-hopping potential already presents a paradigm shift in the way we practice veterinary medicine, using these tools across a broadening array of both clinical and nonclinical applications.
General AI—when it comes—presents more of a coalesced land mass, breaking down boundaries between individual AI tools and functioning intelligently across multiple areas in a similar way to human beings. Similar to macroevolution, what the outcomes of this will look like and when it is likely to happen nobody really knows, but it’s likely to be vastly different. Hence, there are some loud calls to delay its emergence and introduce regulations and controls across a number of key industries and applications. This article will therefore focus on the visible horizon of narrow AI.
The AI Toolbox
Machine Learning (ML): Machine learning is a subset of AI that involves the development of models that enable computers to learn from, and make predictions or decisions based on data. Neural networks are an effective and frequently used ML technology that mimic the plasticity of the human brain in creating dynamic neural connections, which are either reinforced or weakened dependent on feedback during the learning process.
The fundamental idea behind ML is to train algorithms to recognize patterns in data with known outcomes and use these patterns to make predictions or decisions without being explicitly programmed. Data used can include numerical data (such as clinical parameters), categories or labels (such as species and patterns), text and images. It can also process temporal data, where data is collected over a period of time, typically at regular intervals.
Most types of AI utilized in veterinary medicine involve ML due to the breadth of data types these algorithms can use. These tools will go through a number of iterations of improvement during development, upon which a model is selected and typically deployed as a static, unlearning model.
ML can be deployed to predict risk of acute occurrences, such as seizure activity or anaphylactic reactions through continuous remote monitoring devices. Longer term, it can be used for epidemiological modeling of disease patterns in populations and, at the individual level, the risk of developing disease. For example, a tool has been developed to predict likelihood of chronic renal insufficiency in cats up to two years in advance, taking into account multimodal data including age and blood and urine parameters. ML tools have also been developed to aid diagnosis of challenging conditions, including Addison’s disease and leptospirosis in dogs.
Machine learning can be used for epidemiological modeling of disease patterns in populations and, at the individual level, the risk of developing disease.
Deep Learning (DL): Deep learning is a subset of ML that involves neural networks with many layers, also known as deep neural networks. These networks are capable of learning and representing patterns in data. Unlike traditional ML, DL algorithms can learn to extract features from raw data, eliminating the need for manual labeling.
DL tools excel at learning intricate patterns from large datasets—such as diagnostic images, histopathology, and photographs—making them invaluable in analyzing visual data. Outputs may include direct interpretation, such as fecal parasite detection, reticulocyte counts, and mitotic figure detection. They may also be used to segment, categorize, and label images, for example, highlighting areas of abnormality in radiology. Automated error detection and reorientation of images are also useful to improve human efficiency.
With the almost universal access to high-quality images and video through smartphones, there is a huge and growing data pool to inform the development of image-based AI tools. Photographic analysis is being used directly to assess skin and ocular lesions in dogs and horses, respectively, and video footage can be used to detect lameness in horses and livestock. Intra- and interspecies mapping of images has potential to improve animal welfare in many areas: facial pain recognition in prey species, comparative imaging, mapping skull anatomy from MRI to head photographs of dogs to aid the diagnosis of Chiari-like malformation.
The ability of computers to analyze images beyond the limitations of human vision increases the potential capabilities. The field of radiomics converts medical images into mineable high-dimensional data, extracting huge amounts of quantitative features beyond human ability. This field is rapidly expanding in both human and veterinary medicine, helping with earlier diagnosis and greater predictive accuracy of the progression of diseases such as cancer.
Surgical robots equipped with AI can assist surgeons in performing intricate procedures with precision and minimal invasiveness, assisting the human with prediction and precision guidance. Applications in surgery are advancing rapidly, from teaching aids to autonomous robots able to perform end-to-end anastomosis on pig intestines in a research environment. With this technology, there is potential for procedures to be more widely available beyond geographical and human resource constraints.
Applications in surgery are advancing rapidly, from teaching aids to autonomous robots able to perform end-to-end anastomosis on pig intestines in a research environment.
AI-powered robotic systems can also automate repetitive tasks in healthcare facilities, such as medication dispensing and sample analysis, enhancing efficiency, reducing human errors, and freeing up human resources to perform more interesting and skilled tasks.
Reinforcement Learning (RL): These algorithms can learn how to make sequences of decisions by interacting with their environment and with positive and negative feedback simultaneously. Originally created for gaming, RL tools can analyze multiple input data and potential outcomes sequentially to determine optimum pathway; a classic example being the AlphaGo system, which defeated the human Go world champion.
In healthcare research, RL is being explored for personalized and dynamic treatment regimens to optimize therapy and for resource scheduling and allocation, taking into account seasonal trends, staffing, and inpatient levels. Additional data points could also include clinician experience and preferences.
In the field of drug discovery and development, in silico trials are increasingly used as a precursor to in vitro trials, providing a much more cost-effective and rapid method to narrow down potential drug targets and therapeutic molecules. These trials may also reduce the need for in vivo testing on animals.
Natural Language Processing (NLP): This refers to the branch of DL concerned with the ability to contextualize text and spoken words in much the same way human beings can. NLP can automate transcription of clinical notes and provide additional context and references, automatically inserting them into the text to enhance the clinical history and inform ongoing case management. Automated translation also opens the door to increasing accessibility of veterinary care beyond language barriers, helping to democratize access to care.
In the field of education, NLP can help to optimize learning experience for the individual, whether that be learning styles, communication preferences, or adaptations to support neurodiverse individuals.
Large Language Models (LLMs): These neural networks are specific models that excel in the field of NLP as they can be trained with large amounts of data to contextualize, summarize, generate, and predict new content. We’re increasingly familiar with open access LLMs, such as ChatGPT and Google Bard. These tools can help to write practice blogs, formulate social media content, and even craft challenging emails to owners or communications to staff. Interestingly, the learning process is two-way, with the human learning how to optimize the prompts provided to the LLM in order to optimize the outputs generated.
In education, LLMs are already being used by students to generate assignments. Rather than penalizing use of technology, it may be wiser to learn to integrate its use into education systems as it integrates more in the real world.
Generative Adversarial Networks (GANs): These algorithms consist of two neural networks, a generator and a discriminator, which work together to produce realistic data. In medicine, GANs are employed for generating synthetic medical images, which can augment scarce datasets. In veterinary medicine, where progress has often been hampered by paucity of standardized data sets, GANs could help to fill data gaps.
By generating diverse and realistic data, GANs aid in training AI models effectively, leading to improved diagnostic accuracy and possibly a revolution in the field of evidence-based veterinary medicine.
Explainable AI (XAI): Key to adoption and deployment of AI is trust and understanding of the technology. XAI focuses on making AI models more interpretable and transparent. Decloaking the black box and understanding the decisions made by AI systems is crucial, especially in critical areas like diagnosis and treatment and where the ultimate outcome could be euthanasia of a patient. XAI techniques foster trust and acceptance of AI technologies in medical practice.
In the field of drug discovery and development, in silico trials are increasingly used as a precursor to in vitro trials, providing a much more cost-effective and rapid method to narrow down potential drug targets and therapeutic molecules.
Looking Ahead
A horizon is a fitting concept for a technology that is apparently broad, limitless, and at a distance we can’t define. While we can scan the horizon of what lies ahead, we cannot know how fast it is approaching or the full extent of how it will impact animal health and our working lives. We also miss some of the dips and bumps in the road on the journey to get there, and with AI, there are many unknown and unpredictable outputs that necessitate ongoing quality assessment and improvement.
By looking at current technology and the vehicles delivering AI tools though, we can make predictions about what the near future will hold and consequently, how we need to adapt to the change. Understandably, there may be some hesitation around how AI will impact working life in veterinary practice, but early indicators show immense promise to improve diagnostics, treatments, and outcomes for patients, while also improving communications, assisting workflow efficiency, and optimizing the scarce human resources across the veterinary workforce.
Photo credits: your_photo/Collection via Getty Images Plus, gorodenkoff/Collection via Getty Images Plus, sompong_tom/Collection via Getty Images Plus, amgun/Collection via Getty Images Plus, Userba011d64_201/Collection via Getty Images Plus