Abstract
With the advances in computational technology, artificial intelligence (AI) systems have been growing exponentially and promise to become tools that are able to overcome some of the most difficult issues of medical research and patient care. Current progress in AI systems offers significant advantages in healthcare, with the potential to minimize the gap between data, knowledge and patient care. The purpose of this chapter is to examine how AI methods might affect data analysis in biomedicine and more specifically in anaesthesia. By the time this book has been published, the anaesthesia and critical care literature will be abound with manuscripts that use AI methods. It will therefore become crucial for the clinician to understand what AI is all about. A detailed understanding of AI requires an extensive knowledge of computational science and complex mathematical concepts. This chapter will provide the reader with the main insights needed to understand the basic concepts of the underlying modelling framework used by AI and will briefly review the different AI methods, their applications and limitations.
Introduction
With the advances in computational technology, artificial intelligence (AI) systems have been growing exponentially and promise to become tools that are able to overcome some of the most difficult issues of medical research and patient care. Current progress in AI systems offers significant advantages in healthcare, with the potential to minimize the gap between data, knowledge and patient care. The purpose of this chapter is to examine how AI methods might affect data analysis in biomedicine and more specifically in anaesthesia. By the time this book has been published, the anaesthesia and critical care literature will be abound with manuscripts that use AI methods. It will therefore become crucial for the clinician to understand what AI is all about. A detailed understanding of AI requires an extensive knowledge of computational science and complex mathematical concepts. This chapter will provide the reader with the main insights needed to understand the basic concepts of the underlying modelling framework used by AI and will briefly review the different AI methods, their applications and limitations.
Basic Concepts
Because AI covers such a wide spectrum, it always needs to be defined in the context of what it is going to be used for. Still, the definition of AI can be reduced to ‘a system that mimics human cognition’. This definition might not represent the actual computational basis of AI, but it summarize its current applications in healthcare well. Examples of medical applications of AI include image recognition on chest x-rays or CT Scan, ECG interpretation, surgery assistance, clinical decision making and differential diagnosis support. These are all considered applications derived from human cognition. However, AI techniques also carry out tasks that cannot be performed by ‘human cognition’, such as big data processing, pattern recognition and other types of data analyses.
To better understand the concept of AI, it is necessary to know how AI works: a system consisting of an input, a black-box, and a certain output (Figure 20.1). The input involves any type of information relevant to what the AI is going to be used for (e.g. images, patient data or drug concentration). The black-box refers to the underlying AI model, which consists of statistical procedures and computational algorithms that process the information input(s) and link them to the output. AI has used two major approaches, the rule-based approach, and the machine learning (ML) approach and mainly differs in the underlying black-box structure.
Rule-based AI
Rule-based AI is basically an extensive computational algorithm, built from several ‘knowledge-based rules’ (Figure 20.2). In this approach, the black-box is characterized by many inference rules formulated a priori, which specify the ‘direction’ towards an output given the input conditions. In other words, rule-based AI models are constructed by many ‘IF–THEN’ rules which link the inputs to the outputs. The main difference with other AI approaches is that ‘rule-based’ AI models do not learn from data because the outputs are the direct result from applying the static rules defined by the developer. Therefore, no data are required to develop the model.
This approach was the first AI system to be developed. Examples of this type of AI system are differential diagnosis support systems. In such systems, the inputs are defined by the patient characteristics, symptoms and other clinical data, which are processed by algorithms (e.g. ‘is the patient male?’ or ‘does the patient have anaemia?’) to lead to a conclusion (output), such as a disease or condition.
Machine Learning aI
Frequently mixed up with the AI concept, machine learning (ML) models refer to any type of AI model that is able to learn associations from data. In contrast to the rule-based models, ML models are designed to develop their own ‘rules’ based on the data provided to the system.
ML models are subdivided according to the learning method they employ: supervised learning, unsupervised learning and reinforced learning. Additionally, ML models can be grouped according to the similarity of the algorithms they use for the specific task they are being used for.
Supervised Learning
In supervised learning, the objective of the ML model is to predict a known output (Figure 20.3). The term ‘supervised’ means that the output or target is defined a priori, and therefore the training process of the model aims to get the most accurate output.
Fig. 20.3 Representation of the development process of a machine learning model using supervised learning. Data are usually split into a training and a validation dataset. After the extraction of features from raw data, the ML system processes the input/output pairs in order to find the most accurate model. Afterwards, the model is validated by using data without outputs in it (unlabelled data), and the model is evaluated.
Supervised learning is a two-step process. The first step is the training process, in which the system is ‘fed’ labelled data (a dataset which contains the outputs). In this step of the process, the system evaluates possible associations between the inputs and the outputs, and then derives a model that links them better. Next, the model is exposed to unlabelled data (dataset not containing outputs), and the estimated outputs (‘predicted data’) are evaluated to test the accuracy of the model.
Supervised learning algorithms and models are classified into two groups:
– Regression models: Regression models are quite similar to classic regression approaches. In this type of model, the target is to obtain the most accurate prediction of a predefined output. An example of this type of ML are the risk models such as the Framingham Risk Score and the guide for antithrombotic therapy in atrial fibrillation, in which the output is well defined (e.g. mortality due to cardiovascular disease), and used in the derivation of the model. Other types of regression models include regression trees, ensemble methods and many other types of models, all of them designed to predict or estimate an outcome.
– Classification models: With these types of models, the target is to obtain a precise classification or identification of an output. Similar to the regression models, the classification models are also trained with a specific classification output. However, in contrast to regression models, classification models aim to recognize patterns in data inputs to link them to predefined output. Examples of this type of model are the automated EKG interpretation systems and image recognition in chest x-rays, in which the recognition of certain patterns (e.g. ST abnormalities) is linked to a predefined set of diagnoses (e.g. myocardial infarction).
Both types of models, and especially classification models, usually need preprocessing of the raw data before they can be properly examined by the model. This process, commonly known as feature extraction, is mostly used in image recognition, where the image is processed into a large set of pixels or vectors, and each one of them is handed to the model as a unique input. However, more complex types of ML models, such as neural networks, are designed to also extract features from raw data.
Related posts:
Full access? Get Clinical Tree
