What is the study of how computer programs can improve their performance without explicit programming?

Artificial Intelligence, Machine Learning, and Deep Learning

Artificial intelligence (AI) is a broad term that can be construed to mean a focus in computer programming and development designed to train machines to perform tasks as well as, or better than, humans.Machine learning is a related application of artificial intelligence, in which a computer is provided programming and large amounts of data to learn its own patterns, rather than the patterns and limits set by a human programmer, and thereby improve from experience.Deep learning is a class of machine learning techniques that uses multilayered neural networks. The application of AI to medical tasks is not new317 but has exploded in recent years owing to advances in computing power, hardware capabilities, and software advances. In essence, AI algorithms attempt to provide classifications.317 As applied to medical imaging, cardiopulmonary imaging in particular, a given AI algorithm may attempt to determine if a lung nodule is present or not, or determine the amount of reticular abnormality in a patient with fibrotic lung disease compared to normal lung. One prominent application of AI in pulmonary imaging is lung nodule detection and characterization of lung malignancies called radiomic analysis (i.e., the quantification of the phenotypic features of a lesion from medical imaging),318 particularly for genomic prediction. Additional cardiopulmonary applications for AI include pneumonia detection, detection and quantification of obstructive pulmonary disease, and the CT determination of fractional flow reserve for atherosclerotic coronary disease, among others. Artificial intelligence methods have been applied to the assessment of ILDs at CT and have shown the ability to predict pulmonary function in patients with IPF better than visual scoring,319 potentially identifying nonvisual CT parameters that may predict the severity of functional impairment.

The applications of AI are expanding rapidly and growing immensely in sophistication. The translation of AI methods into clinical practice has been somewhat slow; issues regarding integration into workflow, cost, and reimbursement for applications must be addressed. Nevertheless, it seems clear that AI will play a large role in medicine, and imaging in particular, in the years to come.

XIII Machine Learning

Machine learning is a type of artificial intelligence (AI) that provides computers with the ability to learn without being explicitly programed. Machine learning focuses on the development of computer programs that can change when exposed to new data.

One study has shown that machine learning is up to 93% accurate in correctly classifying a suicidal person and 85% accurate in identifying a person who is suicidal and has a mental illness but is not suicidal or neither. These results provide evidence for using advanced technology as a decision-support tool to help clinicians and caregivers identify and prevent suicidal behavior, says John Pestian, PhD, professor in the Divisions of Biomedical Informatics and Psychiatry at Cincinnati Children's Hospital Medical Center and the study's lead author (Neurosciencenews, n.d.).

Imaging Biomarkers and Deep Machine Learning

Radiographic features, including nodule size, contour, and growth, have traditionally been the main differentiators between benign and malignant nodules and are inputs into risk prediction calculators.15,25,28 Nodule size, one of the strongest predictors of malignancy,61 was traditionally assessed via measurement of the largest transverse diameter using a manual 2D caliper. More recently, however, screening trials and nodule management guidelines have recommended measurement of volume rather than diameter because it is more representative of a nodule’s true size, has less interobserver and intraobserver variability, and is more sensitive to changes comparedto 2D measurements.62 For volumetric measurements to be reliable, nodule segmentation (i.e., the delineation of a nodule boundary) must be accurate.63 There are software packages that provide manual, semiautomated, and automated volumetrics. Although these provide repeat measurements that are reliable, there is variation between software packages that can make comparisons between images done at differing institutions difficult.38 Volumetric measurements, while gaining traction, have not been widely adopted in the United States as standard practice for nodule evaluation.

There have previously been attempts to develop CAD algorithms to help radiologists distinguish benign from malignant nodules and improve diagnostic accuracy; however, these did not achieve widespread acceptance.64 A new paradigm shift in computer-based diagnosis was seen after the introduction ofconvolutional neural networks (CNNs). This set of techniques allows computers to detect patterns beyond human perception, to classify imaging findings further. Research using CNNs to classify pulmonary nodules has demonstrated superiority to standard CAD techniques by reducing the number of false positives.65–68 This form of deep machine learning has the ability to learn previously unknown features but requires large amounts of validated data. Unfortunately, there are few large validated imaging datasets with pathology-confirmed cancer available to allow training, tuning, and testing.69 Efforts are underway to advance this technology for analysis of both incidental and screen-detected nodules. Ideally, much like clinician intuition or risk prediction calculators, deep machine learning would incorporate all available imaging, clinical, and biochemical data into determining likelihood for malignancy.

12.1 Introduction

Machine learning is a collection of methods that enable computers to automate data-driven model building and programming through a systematic discovery of statistically significant patterns in the available data. While machine learning methods are gaining popularity, the first attempt to develop a machine that mimics the behavior of a living creature was conducted by Thomas Ross in 1930s [1]. In, 1959 Arthur Samuel defined machine learning as a “Field of study that gives computers the ability to learn without being explicitly programmed” [2]. While the demonstration by Thomas Ross, then a student at the University of Washington and his professor Stevenson Smith, included a Robot Rat that can find a way through artificial maze [1], the study presented by Arthur Samuel included methods to program a computer “to behave in a way which, if done by human beings or animals, would be described as involving the process of learning.” With the evolution of computing and communication technologies, it became possible to utilize these machine learning algorithms to identify increasingly complex and hidden patterns in the data. Furthermore, it is now possible to develop models that can automatically adapt to bigger and complex data sets and help decision makers to estimate impacts of multiple plausible scenarios in a real time.

The transportation system is evolving from a technology-driven independent system to a data-driven integrated system of systems. For example, researchers are focusing on improving existing Intelligent Transportation Systems (ITS) applications and developing new ITS applications that rely on quality and size of the data [3]. With the increased availability of data, it is now possible to identify patterns such as flow of traffic in real time and behavior of an individual driver in various traffic flow conditions to significantly improve efficiency of existing transportation system operations and predict future trends. For example, providing real-time decision support for incident management can help emergency responders in saving lives as well as reducing incident recovery time. Various algorithms for self-driving cars are another example of machine learning that already begins to significantly affect the transportation system. In this case, the car (a machine) collects data through various sensors and takes driving decisions to provide safe and efficient travel experience to passengers. In both cases, machine learning methods search through several data sets and utilize complex algorithms to identify patterns, take decisions, and/or predict future trends.

Machine learning includes several methods and algorithms, some of them were developed before the term “machine learning” was defined and even today researchers are improving existing methods and developing innovative and efficient methods. It is beyond the scope of this book to provide in-depth review of these techniques. This chapter provides brief overview of selected data preprocessing and machine learning methods for ITS applications.

Machine Learning and Image Analysis

The design and development of a retinal image analysis system involves the combination of some of the processing steps as explained above, with specific sizes of features and specific operations used to map the input image into the desired interpretation output. Instead of being programmed, the steps can also be learnt, throughmachine learning, so that for example the features used for detection61 as in pixel classification (below), or how to combine the output of feature detectors into an output can be learned.62

This term, machine learning, is used when an algorithm is improved incrementally by changing parameters so that it is slightly improved every step. During training, the correct interpretation, or reference standard, also called ground truth, is required, which is typically created by retinal specialists or ophthalmologists.63 A theoretical disadvantage of using a supervised system with a training set is that the provenance of the different settings is implicit and may not be clear – resulting in a black box. However, because all retinal image analysis algorithms undergo some optimization of parameters based on their initial performance, this is only a relative, not absolute difference.

As mentioned, two distinct stages are required for a supervised learning/classification algorithm to function: A training stage, in which the algorithm “statistically learns” to correctly classify images, regions of images, or even pixels from the reference standard, and a deployment, testing or classification stage in which the algorithm classifies previously unseen images keeping the algorithm settings constant as established during learning. For proper assessment of supervised classification method functionality, training data and performance testing data sets must be completely separately.57

Until recently, retinal image analysis used machine learning in modular fashion, i.e., one or more of the processing steps are implemented using machine learning. Recent studies are showing remarkable performance improvements using convolutional neural networks, a machine learning approach where all steps are learnt, as will be explained below.

10.02.2.1 The Steps of Machine Learning

Before outlining all the main ML algorithms, it could be helpful for the non-expert readers to follow through a basic ML method (the “how” question posed in the paragraph above).

ML workflow starts with three questions that the researchers should answer:

What is the type of data I am working with?

What do I want from the data?

How and where will be this information applied to?

A classic methodology (Kuhn and Johnson, 2013; Dwyer et al., 2018; Iniesta et al., 2016) of ML applied to our example will require the following steps:

Loading the data: data, extracted by sensors or present in our dataset, can be loaded into one of the available software for ML, such as R, MATLAB or Python (Ozgur et al., 2017).

Processing the data: pre-processing is a critical part of ML and includes centering and scaling (when data are measured from different scales), transformations (e.g., log, square root, power) or Box-Cox transformation to resolve skewness of a distribution, or to treat outliers in our dataset. Another way to deal with large dataset or multicollinearity is to apply data reduction techniques such as Principal Component Analysis (PCA) or Partial Least Squares (PLS). It is crucial to resolve missing values, as any of these will reduce our sample; a way to predict missing values is to use imputation. Consider removing (highly correlated predictors) or adding (dummy variable) predictors prior to modeling. Finally, for ML it is critical to divide the sample in a dataset for testing (training set) the ML and another (test or validation set) to retest the ML. Thus, prediction performances can be measured with in-sample (in the same dataset used to fit the model) and out-of-sample (performance is measured with new data) estimates.

Features (independent variables) selection: this work can be part of pre-processing or it can be done with the ML itself. PCA, Least Absolute Shrinkage and Selection Operator (LASSO) regression or correlations can effectively reduce redundant features; this part is critical to enhance accuracy and generalizability of the model.

Building and training the model: using the training set is possible to estimate model performance; a part of the sample is used to fit a model and the other part is for efficacy testing, this procedure is repeated multiple times. Resampling is used to estimate and optimize generalizability of our ML. Methods of resampling include: k-fold cross-validation, leave-one-out cross-validation, generalized cross-validation, leave-group-out cross-validation or bootstrap. Generally, cross-validation is the most used, robust and preferred strategy as it applies to both small and large samples. Additionally, several models with different ML algorithms can be trained to evaluate their performances.

Selecting model fit and between models: it is suggested to select the simplest model (PCA or correlation matrix could help at this point) and when it comes to select among ML models, the one with the best accuracy estimate is preferred.

Testing the model on the validation set: this will assess model accuracy on unused data.

Introducing the knowledge into a specific environment

In summary, given these features, ML is defined as an intelligent agent that learns from and act on the environment and improves from experience on a predetermined task with a performance measure (Das et al., 2015). It can be divided in supervised learning, unsupervised learning, reinforcement learning and recommender systems, which will be discussed in the following paragraphs.

Machine Learning and Artificial Neural Networks

Machine learning (ML) is a core branch of AI that aims to give computers the ability to learn without being explicitly programmed (Samuel, 2000). ML has many subfields and applications, including statistical learning methods, neural networks, instance-based learning, genetic algorithms, data mining, image recognition, natural language processing (NLP), computational learning theory, inductive logic programming, and reinforcement learning (for a review see Mitchell, 1997).

Essentially, ML is the capability of software or a machine to improve the performance of tasks through exposure to data and experience. A typical ML model first learns the knowledge from the data it is exposed to and then applies this knowledge to provide predictions about emerging (future) data. Supervised ML is when the program is “trained” on a pre-defined set of “training examples” or “training sets.” Unsupervised ML is when the program is provided with data but must discover patterns and relationships in that data.

The ability to search and identify patterns in large quantities of data and in some applications without a priori knowledge is a particular benefit of ML approaches. For example, ML software can be used to detect patterns in large electronic health record datasets by identifying subsets of data records and attributes that are atypical (e.g., indicate risks) or that reveal factors associated with patient outcomes (McFowland, Speakman, & Neill, 2013; Neill, 2012). ML techniques can also be used to automatically predict future patterns in data (e.g., predictive analytics or predictive modeling) or to help perform decision-making tasks under uncertainty. ML methods are also applied to Internet websites to enable them to learn the patterns of care seekers, adapt to their preferences, and customize information and content that is presented. ML is also the underlying technique that allows robots to learn new skills and adapt to their environment.

Artificial neural networks (ANNs) are a type of ML technique that simulates the structure and function of neuronal networks in the brain. With traditional digital computing, the computational steps are sequential and follow linear modeling techniques. In contrast, modern neural networks use nonlinear statistical data modeling techniques that respond in parallel to the pattern of inputs presented to them. As with biological neurons, connections are made and strengthened with repeated use (also known as Hebbian learning; Hebb, 1949). Modern examples of ANN applications include handwriting recognition, computer vision, and speech recognition (Haykin & Network, 2004; Jain, Mao, & Mohiuddin, 1996). ANNs are also used in theoretical and computational neuroscience to create models of biological neural systems in order to study the mechanisms of neural processing and learning (Alonso & Mondragón, 2011). ANNs have also been tested as a statistical method for accomplishing practical tasks in mental health care, such as for predicting lengths of psychiatric hospital stay (Lowell & Davis, 1994), determining the costs of psychiatric medication (Mirabzadeh et al., 2013), and for predicting obsessive compulsive disorder (OCD) treatment response (Salomoni et al., 2009).

ML algorithms and neural networks also provide useful methods for modern expert systems (see Chapter 2). Expert systems are a form of AI program that simulates the knowledge and analytical skills of human experts. Clinical decision support systems (CDSSs) are a subtype of expert system that is specifically designed to aid in the process of clinical decision-making (Finlay, 1994). Traditional CDSSs rely on preprogrammed facts and rules to provide decision options. However, incorporating modern ML and ANN methods allows CDSSs to provide recommendations without preprogrammed knowledge. Fuzzy modeling and fuzzy-genetic algorithms are specific ancillary techniques used to assist with the optimization of rules and membership classification (see Jagielska, Matthews, & Whitfort, 1999). These techniques are based on the concept of fuzzy logic (Zadeh, 1965), a method of reasoning that involves approximate values (e.g., some degree of “true”) rather than fixed and exact values (e.g., “true” or “false”). These methods provide a useful qualitative computational approach for working with uncertainties that can help mental healthcare professionals make more optimal decisions that improve patient outcomes.

1.1 Shallow and deep learning

Machine learning refers to algorithms that successively reprocess data into more abstract and compact forms (LeCun, Bengio, & Hinton, 2015). Deep learning has led to a revolution in the flexibility and accuracy of machine learning. In conventional or shallow learning, algorithms are provided with selected features of sensor data (fundamental pitch contours from an audio signal, for example). In deep learning, multiple layered neural networks iteratively abstract information more directly from sensor data (e.g., the audio signal itself). A historical example is illustrative. In 2009, Messinger et al. predicted infant smiling using a one-dimensional support vector machines from the output of manifolds of face image data (shallow learning) (Messinger, Mahoor, Chow, & Cohn, 2009). By contrast, in 2021, Ertugrul et al. reliably detected infant smiling using a deep neural network directly from a video model of the infant's face (Ertugrul et al., 2021).

1.1 Introduction

Machine Learning is concerned with algorithmic induction. Its aim is to develop algorithms that are able to generalize from a given set of examples. This is quite a general description, and Machine Learning is a wide field. Here we will confine ourselves to two exemplary settings, viz. concept learning and sequence prediction.

In concept learning, the learner observes examples taken from some instance space X together with a label that indicates for each example whether it has a certain property. The learner's task then is to generalize from the given examples to new, previously unseen examples or to the whole instance space X. As each property of objects in X can be identified with the subset C ⊆ X of objects that have the property in question, this concept C can be considered as a target concept to be learned.

EXAMPLE 1. Consider an e-mail program that allows the user to classify incoming e-mails into various (not necessarily distinct) categories (e.g. spam, personal, about a certain topic, etc.). After the user has done this for a certain number of e-mails, the program shall be able to do this classification automatically.

Sequence prediction works without labels. The learner observes a finite sequence over an instance set (alphabet) X and has to predict its next member.

EXAMPLE 2. A stock broker has complete information about the price of a certain company share in the past. Her task is to predict the development of the price in the future.

In the following, we will consider each of the two mentioned settings in detail. Concerning concept learning we also would like to refer to the chapter on Statistical Learning Theory of von Luxburg and Schölkopf in this volume, which deals with similar questions in a slightly different setting.

Unsupervised machine learning

Unsupervised machine learning is a method that captures unknown patterns or features in a dataset and then produces the most suitable representation associated with the dataset (Goodfellow et al., 2016). Unsupervised machine learning is the chain rule of probability from inputs X = { x1, x2, …, xi } .

To distinguish between supervised and unsupervised machine learning, the former requires both a feature and a label, while the latter may require only a feature. Unsupervised learning algorithms include principal components analysis used in studies (Burnap et al., 2015; De Choudhury, Counts, & Horvitz, 2013a; De Choudhury, Counts, & Horvitz, 2013b; De Choudhury, Gamon, et al., 2013; Maxim et al., 2020; Schwartz et al., 2014; Stankevich et al., 2020), and k-means clustering in (Fodeh et al., 2019; Maxim et al., 2020; Prieto et al., 2014; Wang et al., 2017; Yang et al., 2020).