Deep learning for health and medicine

Abstract

Lately, a data-centered era has emerged due to the rapidly growing health and medical fields that have yielded big data such as patient-related examinations, texts, speeches, and figures; hence, the need for technology upgrades in the current systems and processes. Machine learning (ML), a subset of artificial intelligence (AI), applies models used by computer systems for learning instructions based on the weightage assigned to the parameters and devoid of any direct instructions, via increasingly advancing algorithms, especially those for deep learning (DL). DL can automatically develop models of several processing layers for learning data representations at several abstraction levels. Research suggests that DL algorithms are getting more attention for being superior to those of ML in advanced predictive power; they are promising in learning models and in retrieving parameters from compound data. With the many representation levels and outcomes, which have outpaced human efficiency, DL has been adopted widely in health and medical informatics, for example, in molecular diagnostics for pharmacogenomics and in recognizing variants of pathogens; in explaining experimental data for sequencing DNA and splicing genes; in classifying protein structures; in predictions for medical imaging and developing drugs, among others. This chapter discusses the novel applications of DL in improving health and medicine; it first introduces ML and DL and, next, highlights their applications in health, and finally, links their relevance to the future perspectives of modern health.

1 Introduction

Lately, there have been dramatic advances in data acquisition techniques in life sciences, along with developments in computational biology and automatic storing technologies that have transformed today’s biology into a data-driven discipline. Hence, nowadays, research depends on data, and there are many promising interventions to biological issues. Bioinformatics aims at evaluating such big data in several domains, by maintaining, extracting, or exploring it. Computational biology algorithms could also enable managing data extraction. Such techniques, known as data mining, could reveal useful relationships, trends, and performances in biological data [1].

Lately, a subfield of computational science has gained more attention, namely, soft computing, which reflects all the techniques that smoothly process data while managing uncertain real-world circumstances. While hard computing targets precision, soft computing functions in the realm of incomplete facts, uncertainty, inadequacy, and estimation for solving issues [2]. Previously, only primitive systems could be accurately modeled and analyzed via computing methodologies; yet, systems of health, biology, managerial research, humanities, and similar complicated disciplines were hard to manage via traditional mathematical and analytical approaches. Soft computational technologies supplement one another; hence, biological procedures are more nearly mimicked through soft computational technologies than conventional ones, which mainly apply logic, like sentential or predicate logic. One of the key components of soft computing is neural networks, which have a broad reach in categorizing and reflecting computationally biological data. Neural networks are solid and demonstrate great learning and abstracting capability in data-driven contexts where they apply machine learning (ML) algorithms [3].

2 Machine learning for health

ML is the study of processing strategies that are automated by experience [4], as a diagrammatic representation of an algorithm with the help of artificial intelligence (AI) [5]. Quantitative approaches create a blueprint for sample results to be repeated and determined, without being expressly programmed, and are called “training data.” In a variety of applications [6], ML algorithms are used to perform the necessary functions, including in healthcare [7], medicine [8], email filtering [9], and computer vision [10]. One ML category is strongly connected to predictive analytics, which focuses on the use of computers to make forecasts, but mathematical learning is not all ML. Study of mathematical optimization introduces techniques, theory, and implementation fields in ML. Data mining is a related field of research for the analysis of exploratory data, which relies on unattended learning. ML is also referred to as mathematical wisdom in its approach to market problems.

The twenty-first century is two decades old, and AI will be one of the most important tools for transformation and enabling human life during this century. AI and related services and networks are well established in their architecture to change global competitiveness, work habits, and lifestyles and generate enormous wealth. No mystery that this change is mostly driven by powerful technologies (ML) and innovations such as deep convolution networks, opposing networks from generations, gradient-driven tree models (GBM), and deep reinforcement training (DRL). But, these not the only AI areas affected by conventional industrial and technology industries. Health is an area that is considered ideal for implementing AI tools and technology.

Compulsory approaches such as electronic medical records (EMRs) now have next-generation healthcare programs equipped to evaluate massive data resources. AI/ML tools are designed to make this flow more meaningful. The quality of automation and intelligent decision-making in primary and tertiary care and public health care is expected to improve. AI tools can have the greatest impact because they can improve the quality of life for thousands of people around the world.

Machine learning in healthcare (MLH) is usually intended to predict clinical effects based on different predictors. MLH shows enormous promise, and ML-based instruments can reach human-level diagnostic and prognostic functions defined in almost all clinical areas [11]. However, the number of MLH resources used in clinical applications represents only a fraction of the investment in the entire field, showing that most MLH applications do not surpass the first edition [12]. On closer inspection, ML researchers seem to be inclined to conclude that a reliable prognosis (and thus, news) has been shown if experts from domain stakeholders can translate it into clinical practice [13]. With a high error rate of 6.8, the translation is complicated, but some potential problems can be easily handled.

ML in pharmacy has recently made significant advancement. An ML algorithm has been developed by Google to identify mammographic cancer tumors. A robust skin cancer recognition learning algorithm is being used by Stanford. In a recent article by JAMA, the results of a deep-tech algorithm that diagnosed retinopathy in retinal imaging were published. In the decision-making process, ML adds another arrow. It is an easy one here. For some systems, ML is even more attractive than others. Disciplines can benefit directly from algorithms in reproducible or systematic systems. Many of those with broad picture datasets, such as radiology, cardiology, and pathology are also promising candidates. For all these procedures, ML should be able to look at the images, recognize patterns, and know the focus areas that need to be more accurate. ML will help the family practitioner or bedside trainee in the long term, by offering impartial guidance that improves reliability, results, and accuracy.

2.1 Supervised learning

Supervised learning begins with the goal of predicting a known outcome or a known goal. In ML competitions, where individual participants are judged based on their results on standard datasets, recurring supervised learning problems include handwriting recognition (such as handwritten number recognition), classification of photographs of objects, for example, if it is a cat or a dog, and text classification. Another example is a clinical trial of heart disease or a financial report. All these are what an intelligent individual would do well, and so the machine also tries to estimate human output. Supervised learning focuses on grouping, which means choosing to explain and predict a new data body among subgroups, which could include predicting an uncertain parameter such as the temperature in San Francisco tomorrow afternoon.

What are...