Algorithmic Medicine

1.1 Introduction

For long health services have faced several challenges, chief among them being rising expenditure and workforce shortages without clear solutions in sight (Topol, 2019). At the same time, there has been an unprecedented generation of medical data ranging from sources such as electronic health records, medical imaging and laboratory units (Sidey-Gibbons & Sidey-Gibbons, 2019). Clinicians have for long relied on computers to analyse such data as the analysis of such complex, and large datasets exceed their human capacity. In this context, the emergence of artificial intelligence (AI) with its ability to significantly enhance the data analysis process has presented an opportunity for clinicians and healthcare administrators to gain better

insights (Reddy, 2018). An opportunity to optimise care delivery, reduce healthcare delivery costs and support a stretched workforce.

Of the various AI approaches, the most pertinent to analysing data is machine learning (ML), which comprises aspects of mathematics, statistics and computational science (Sidey-Gibbons & Sidey-Gibbons, 2019). ML is the core of changes occurring in medicine because of AI. Unlike non-AI methods and software, which rely mainly on traditional statistical approaches, ML software utilises pattern detection and probabilistic approaches to predict medical outcomes (Reddy, 2018). This utilisation of ML algorithms and other AI approaches to deliver medical care is what can be termed as algorithmic medicine. The ability to predict crucial medical outcomes through AI algorithms can make healthcare more precise and efficient. Beyond medical care, AI can also support healthcare administration, drug discovery, population health screening and social assistance (Reddy, 2018), thus expanding the scope of algorithmic medicine beyond the confines of clinical care, i.e. direct clinician to patient care. This ability and promise have ignited the interest of governments and other healthcare stakeholders to consider incorporation of AI in healthcare administration and delivery seriously. This chapter outlines what would be involved for this to occur and what the impact will be.

1.2 AI in Medicine – A History

Before we define AI and describe its techniques, it will be pertinent to review the history of AI in healthcare. The concept of intelligent machines is not new and in fact can be traced to Ramon Llull’s theory of reason-ing machine in the 14th century (Reddy, 2018). However, modern AI can be tracked back to the past 70 years with the term originating from the workshop organised by John McCarthy at Dartmouth College in 1956 (AAIH, 2019). In the following decade, the availability of faster and cheaper computers allowed experimentation with AI models particularly in the areas of problem-solving and interpretation of spoken language (Anyoha, 2017). However, as work progressed in these areas, the lack of requisite computational power and the limitations of the then algorithmic models came to fore. In the 1980s, there was a revival of interest in AI particularly so in expert systems, which were modelled to mimic the decision-making pro-cess of a human expert . However, again these types of models fell short of expectations, and interest in AI in both academia and industry

waned. Commencing in the mid-2000s, the availability of suitable technical hardware and emergence of neural networks, an advanced form of ML, cou-pled with their demonstrable performance in image and speech recognition once again brought AI back to the limelight. Since then, significant funding and interest has led to further advances in algorithms, hardware, infrastructure, and research.

Paralleling the general history of AI, its use in medicine formally commenced with the DENDRAL project in the 1960s, which was an early expert system with an objective to define organic compound structures by investigating their mass spectra (AAIH, 2019). The development of this system required new theories and programming. This was followed by MYCIN in the 1970s, which was aimed at identifying infections and recommending appropriate treatment. The learning from MYCIN was extrapolated to develop the CADUSEUS system in the 1980s. This system was then hyped as the most knowledgeable medical expert system in existence. In line with the general history of expert system, the application of expert systems in medicine fell short of expectations. The sophistication of neural networks and availability of hardware to run these algorithms presented a new opportunity for the use of AI in medicine (Naylor, 2018; Reddy, 2018). Since then, increasing evidence has been detailed of what AI models can do in terms of medical imaging interpretation, support for clinical diagnosis, drug discovery and clinical natural language processing.

1.3 AI Types and Applications

Before we discuss the different types of AI and its applications, it is important to define what AI is? There are numerous definitions of AI in the literature, but this one derived from computer science describes AI as “the study of intelligent agents and systems, exhibiting the ability to accomplish complex goals” (AAIH, 2019). However, this definition is oriented to an academic perspective. From an application and industry perspective, AI can be best described as “machines assuming intelligence”. Now that we have defined AI, it is pertinent to mention here two levels of AI: General and Narrow AI. General AI, also referred to as Artificial General Intelligence, is when AI exhibits “a full range of cognitive abilities or general intelligence actions by an intelligent agent or system” (AAIH, 2019), while Narrow AI, also referred to as Weak AI, is where AI is specified to address a singular or limited task.

The predominant approach of AI, currently, is ML (Figure 1.2). This approach involves performing tasks without explicit instructions relying mainly on patterns and relationships in the training data and environment (AAIH, 2019). To develop ML models, you will need to define the necessary features, i.e. dependent and independent or input and target variables, and develop datasets including the features. Further to this, you split up the data-set into training and test datasets to allow for internal validation. Following this, the datasets are trained or tested with relevant ML algorithms. If the training dataset contains the input data and the appropriate output/target variable, then it is termed supervised learning (El Morr & Ali-Hassan, 2019). However, if there is no known output and the algorithm is left to detect hidden patterns or structures within the dataset, then this is unsupervised learning. In recent years, a hybrid form where the training set has a mix of labelled and unlabelled data and the expectation is that a function predict-ing the target variable is arrived at, which is termed semi-supervised learning (El Morr & Ali-Hassan, 2019). ML algorithm development does not necessarily have to adopt the training approach described above. Reinforcement learning, a relatively newer form of ML, involves a process of maximising reward function based on the actions taken by the agent (AAIH, 2019). A trial-and-error approach is adopted to eventually arrive at optimal decision-making by the agent. In generative learning, the model development involves creating new examples from the same distribution as the training set and in certain instances with a particular label. The evolutionary algorithm model builds on this approach where initially developed algorithms are tested for their fitness, similar to an evolutionary process, until peak performing algorithms are identified and no more progress in fitness of the group can be derived (AAIH, 2019).

While there are numerous ML algorithms in use, a couple of commonly used algorithms in medicine are linear regression, logistic regression, decision trees, random forest and support vector machines (SVMs). An advanced form of ML that excels at analysing complex patterns between variables in datasets is deep learning (DL) (Topol, 2019). This approach is inspired by the architecture and ability of human brains whereby learning and complex analysis is achieved through interconnected neurons and their synapses. This is computationally simulated through many layers of artificial neurons between the input and output variables. These artificial neurons through a hierarchical and interconnected process are programmed to detect complex features and the model depending on complexity of data adds necessary number of layers (auto-didactic quality) (Topol, 2019). Sandwiched between the input and output layers are the hidden layers (see Figure 1.3), which adds to the feature optimisation and model performance but also creates opacity about the decision-making process of the model.

While there are myriad ways as to how neural networks and AI are in use in healthcare, three applications where they are mostly used or have most promise are profiled: computer vision, natural language processing and robotics.

1.3.1 Computer Vision Computer Vision (CV) is where computers assist in image and video recognition and interpretation (Howarth & Jaokar, 2019). Increasingly DL has become central to the operation of CV.

This is due to DL’s many layers useful for identifying and modelling the different aspects of an image. In particular, convolutional neural networks (CNNs), a form of DL, involve a series of convolutions and max-pooling layers (see Figure 1.4) as its under-lying architecture has been found to be very useful in image classification (AAIH, 2019; Erickson, 2019). CNNs are credited for reviving interest in neural networks in recent years. The way the CNNs work is by commenc-ing with low-level features in the image and progress to higher-level features that represent the more complex components of the image. For example, the first layers will identify points, lines and edges, and the latter layers will combine these to identify the target class. An early example of CNN was AlexNet, an image classification model (AAIH, 2019). More recent versions are CNNs with specialised layers including ResNet, ResNeXt and region-based CNN (Erickson, 2019).

CNNs are increasingly being applied in medical image interpretation (Erickson, 2019): for example, to classify chest X-rays that have malignant nodules and those that haven’t. Here, a set of labelled or annotated chest X-rays are used to train the neural networks to compute features that are reliable indicators of malignancy or lack. CNNs can be used for segmentation too where the class of interest is delineated from the remaining area of non-interest. However, CNNs are not restricted to analysing chest X-rays and have also been used to interpret CT, MRI, fundoscopy, histopathology and other images (Erickson, 2019; Reddy, 2018; Reddy, Fox, & Purohit, 2019).

1.3.2 Natural Language Processing

Natural language processing (NLP) is a process of computationally represent-ing, transforming and utilising different forms of human language, i.e. text 8 ◾or speech (Wu et al., 2020). Unlike other data, computing human language is not straightforward as there is a lot of imprecision in human language (Chen, 2020). Also, the unit component of language is not necessarily con-ducive to computation. To address this natural language must be initially reencoded into a logical construct before it can be administered for informa-tion extraction or translation. For many years, NLP reliant on traditional ML approaches like SVM and logistic regression, which were trained on very high dimensional and sparse features, yielded shallow models (Friedman, Rindflesch, & Corn, 2013). However, the advent of DL and its use in NLP has resulted in better performing models. This is because DL enables multi-level automatic feature representational learning.

An important reason for the success of DL in NLP is because of distrib-uted representation, which describes the similar data features athwart multiple scalable and interdependent layers (Young, Hazarika, Poria, & Cambria, 2018). Examples of distributed representation include word embeddings, word2vec and character embeddings. These examples follow the distributional hypothesis, where it is assumed that words with similar meanings tend to occur in a similar context. Thus, the models aim to capture the characteristics of the neighbours of a word to predict meaning. DL has also been useful in Automatic Speech Recognition (ASR), sometimes referred to as speech-to-text (Chen, 2020). Recurrent neural networks have been demonstrated to work well for ASR by lending the algorithm tolerance to complex language conditions such as accents, speed and background noise.

Clinical use of NLP has extended to the vector representation of clinical documents such as clinical guidelines, extracting clinical concepts from electronic medical records or discharge summaries through named entity recognition, mapping clinical ideas and diagnoses with codified guidelines, and developing human-to-machine instructions (Rangasamy, Nadenichek, Rayasam, & Sozdatelev, 2018). NLP can also be potentially used for non- clinical healthcare purposes such as efficient billing and accurate prior authorisation approval through the extraction of information from unstructured physician notes or electronic health records. Further uses of NLP include transcription and chatbot services (Reddy, Fox, et al., 2019).

1.3.3 Robotics

Robots are machines that can carry out complex action and can be programmed by computers (Ben-Ari & Mondada, 2017). Not all robots are programmed by computers and are purely mechanical in nature However, for this chapter, we will review those robots that are programmable by a robot. Robots can be of two categories: fixed and mobile, depending on the environment they operate. Fixed robots like industrial robots operate in a well-defined environment, while mobile robots move and perform activities in poorly defined and uncertain environments. Algorithms work in robots through embedded computers that run on pseudocode utilising a mix of natural language, mathematics and programming structures.

In healthcare, robots are used in various ways, including in surgery, hospitals and aged care (Pee, Pan, & Cui, 2019; Reddy, 2018). One such application that has become popular in recent years is robotic-assisted surgery (Svoboda, 2019).

In this format, surgeons control multiple robotic arms through a hand-operated console (Figure 1.5). This application enables surgeons’ greater vision and dexterity to operate in hard-to-reach areas.

Yet, this is not AI robotics which is about robots operating in an auto-mated or semi-automated fashion. In this regard, trials are being held to allow for independent operation of surgical procedures (Svoboda, 2019). More straightforward or repetitive tasks like suturing and valve repair lend themselves to surgical automation, while complex surgical tasks may take many more years to be automated. Elsewhere, robotic assistants have been used either to support the elderly as social companions or to guide them with medications, appointments and in unfamiliar environments. As AI-enabled robots attain more autonomous functionality through intelligent algorithms, their use in various areas of healthcare is only to increase (Reddy, 2018; Reddy, Fox, et al., 2019).

As AI algorithms and models evolve, there will be broader applicability of them in healthcare to drive efficiency and improved patient outcomes. Demonstrable evidence in the areas of CV, NLP, AI robotics and predictor models will enable adoption and broader use of AI within healthcare broadly and specifically within clinical care models.

1.4 Challenges and Solutions

While AI has enabled unprecedented sophistication and performance in medicine that very few technologies can match, it has also presented significant challenges in its implementation (Reddy, Allan, Coghlan, & Cooper, 2019). While medical data are abundant, they all are not necessarily structured or standardised to train AI models (Wang, Casalino, & Khullar, 2018). While the human brain is capable of inferring patterns from heterogeneous and noisy data, AI models are less so. Utilisation of incorrect and non- representative data can have several implications in the context of healthcare delivery, including the introduction or affirmation of biases and exacerbation of health disparities. Also, in a clinical setting, reliance on a model trained on inaccurate data can have medico-legal repercussions (Reddy, Allan, et al., 2019). Another issue that has emerged specifically with the use of DL models is the opacity of decision-making that is intrinsic to these models. When trained on large datasets, DL models use their many layers to simulate complicated regularities in the data. However, the layered non-linear feature learning makes it impractical to interpret the learning process (Hinton, 2018). The inability to clearly explain the DL model’s conclusion basis presents an obstacle to its use in clinical medicine. For example, if a DL model were to make a clinical recommendation or diagnosis without a clear rationale, it will find little acceptance amongst clinicians. Further to this, the training of ML models involves several parameters (rules) (Beam, Manrai, & Ghassemi, 2019). Because of the use of randomness in training many ML models, there are different possibility parameters arrived at each time the model is retrained, thus limiting reproducibility of the models. Finally, the mathematical accuracy of AI models means nothing if there is no impact on patient outcomes. Currently, very few studies have presented evidence of the down-stream benefit of AI models in medicine.

While these are relatively significant challenges for the adoption and applicability of AI in medicine, they are not without solutions. Most medical DL models are relatively small and focused on medical image interpretation which has fewer issues in terms of structure and reproducibility (Beam et al., 2019). Increasingly, medical researchers are utilising shared or open-source datasets to train their models and providing open access to the code used for the training. These measures allow for transparency and reproducibility of AI models. Also, academic and transdisciplinary collaborations present an opportunity to test and embed AI models in routine clinical care (Sendak, Gao, Nichols, Lin, & Balu, 2019). To address bias or safety and quality issues that may arise from the use of AI models, a governance model that incorporates fairness, transparency, trustworthiness and accountability has been proposed (Figure 1.6) (Reddy, Allan, et al., 2019).

Fairness requires representation from the community at which the AI medical application is aimed at in determining how the software developer uses data (Reddy, Allan, et al., 2019). The representation could be at a data governance panel that reviews datasets used for training such AI medical applications. While it is not feasible for all software developers to constitute such panels, they could potentially draw advice from a government-instituted committee. Information from the group can contribute to less discriminatory or less biased AI models being developed. Transparency stresses the explainability of medical AI models. Where possible, algorithms that lend themselves to explainability are to be utilised, and when DL types of algorithms are necessary, functional understanding of the model conveyed through interpretable frameworks. Also, in clinical practice, informed consent is obtained from patients before use of AI medical applications in the treatment and management of their medical conditions. These initiatives are also required to ensuring trustworthiness of AI medical applications in addition to educating clinicians and the general community about AI and its use and limitations. Through this education and subsequent understanding, AI stands a better chance of being accepted by the medical and patient populations. Finally, accountability is about ensuring the safety and quality of AI medical application through appropriate regulatory and clinical governance processes. This requires input and involvement from a range of governmental and non-governmental bodies. Further to this, legal frame-works and guidance need to be constituted as to who becomes responsible if there were AI-related medical errors or mishaps. In essence, accountability is extending beyond the AI medical application to cover a range of players (Reddy, Allan, et al., 2019). This is necessary to ensure the appropriate and safe use of AI in medicine.

1.5 The Future

As costs of running healthcare, the volume of medical data, the time required to train and deploy work-ready medical workforce and complexity of medical delivery increase, it is inevitable for stakeholders to explore an increased role for AI. The rate and extent at which AI gets adopted in routine clinical care delivery are not guaranteed. However, based on current evidence, one can speculate where AI can contribute to and benefit clinical care. AI can replace some of the mundane or repetitive tasks that clinicians engage with leaving them more time to engage with patients in a meaningful manner. Also, areas which require analysis of complex or voluminous data may benefit from AI’s ability to infer patterns from the data contributing to an augmented medicine model. Further, the progression of research and trials in AI robotic systems can eventuate in the automation of certain aspects of surgery, aged care and hospital logistics (Pee et al., 2019; Svoboda, 2019). All these developments herald an era of algorithmic medicine.