Artificial Intelligence and the Hunt for Immunological Disorders

Abstract

Purpose of Review:

Artificial intelligence (AI) has pervasively transformed many industries and is beginning to shape medical practice. New use cases are being identified in subspecialty domains of medicine and, in particular, application of AI has found its way to the practice of allergy-immunology. Here, we summarize recent developments, emerging applications and obstacles to realizing full potential.

Recent Findings:

Artificial/augmented intelligence and machine learning are being used to reduce dimensional complexity, understand cellular interactions and advance vaccine work in the basic sciences. In genomics, bioinformatic methods are critical for variant calling and classification. For clinical work, AI is enabling disease detection, risk profiling and decision support. These approaches are just beginning to have impact upon the field of clinical immunology and much opportunity exists for further advancement.

Summary:

This review highlights use of computational methods for analysis of large datasets across the spectrum of research and clinical care for patients with immunological disorders. Here, we discuss how big data methods are presently being used across the field clinical immunology.

Introduction:

Electronic health record (EHR) and scientific data are accumulating at a tremendous rate and patients are living longer with greater complexity.[1] Capitalizing on inferences from these large datasets is an important opportunity, requiring specific methods for gleaning necessary insights.[2] Appropriately constructed computational tools may facilitate interpretation of clinical and scientific information, reduce medical errors and prevent morbidity.[2, 3] For these reasons, artificial/augmented intelligence (AI) and machine learning (ML) have generated interest within the healthcare domain and are advancing use cases spanning diagnostic, quality-care, and research spectra.[4–6]

The terms AI and ML are often used interchangeably; however, AI is an umbrella term which refers to computed decision-making and ML is a subset of AI.[7] Additionally, the term “augmented intelligence” is an important concept relating to enhancement of human effort via complementary digital systems.[8] In ML, a programmed model is built from inputs which enables computer task-learning for making predictions with increasing precision over time.[2]

Recently, the concept of a human-computer collaboration spectrum emerged where AI tools fit along a continuum between fully human and fully computer-driven. [9] This view highlights levels of analytic complexity providing a framework for clarifying forms of ML. For example, a risk calculator may embed classic statistical methods requiring substantial human inference to build; whereas, a deep learning model can automatically classify inputs without much human intervention. Given substantial methodologic diversity, we will use the term “AI” throughout to generally refer to tools across the ML and computational spectrum(Table 1).

Table 1:

A Glossary of Relevant Terms for AI and ML

Term	Definition
Algorithm (Machine Learning)	An equation or set of rules by which a computational task is accomplished.
Artificial Intelligence	See “Machine Learning”(syn.) The process by which a computer can automatically or semi-automatically perform a task and improve upon that performance over time without explicit instruction.
Augmented Intelligence	A model of partnership between humans and AI to enhance learning and task cmopletion.
Big Data	Data characterized by high volume which is accumulated with great velocity having tremendous variety in form.
Black Box	A process whose mechanisms for output are not easily comprehensible.
Classification	A machine learning task performed by separating examples or inputs into a category. Usually done by assessing the probability that a given input data example belongs to a specific output class. This is typically used for discrete data.
Data Mining	The process of discovering patterns in data.
Data, Continuous	A term typically used to imply quantitative data with an infinite number of possibilities. Data which can be counted. Often analyzed via regression.
Data, Discrete	A term typically used to imply qualitative data which can be categorized by classification or into a class.
Deep Learning	A process by which a computer can perform a task by learning complex concepts by first learning simple concepts via a multilayered (i.e. deep) network.
Governance, Data	The process by which strategies, policies and decisions related to data are made. Typically relates to an enterprise process.
Labeled Data	Data examples which have been tagged according to a particular class for training. These enable a compluter to more effectively distinguish features of relevance to the classification task. Used in supervised learrning.
Machine Learning	See “Artificial Intelligence” (syn.) The process by which a computer can automatically or semi-automatically perform a task and improve upon that performance over time without explicit instruction.
Natural Language Processing (NLP)	A sub-discipline of linguistics concerned with how computers and human language interact. A method for enabling computer-based analysis of unstructured text. (Syn. Text Mining)
Predictive Analytics	A collection of methods aimed at ascertaining the probability of a given outcome.
Prescriptive Analytics	An advanced form of analytics which aims to answer the question “what should be done?”.
Regression	A machine learning task, based upon classical statistics, where relationships between variables is represented by a line or curve. This is typically utilized for and to predict continuous, real value, numbers.
Semi-supervised Learning	A machine learning approach where a small amount of labeled data and a large amount of unlabeled data is used to train a model.
Structured Data	Data with a clearly defined category or field prepared as such for ready retrieval and analysis.
Supervised Learning	A machine learning approach where training occurs with labeled data.
Task	A unit of execution or work performed by a computer.
Transparent Box	A process whose mechanisms for output are clearly evident.
Unlabeled Data	Data examples which have no prespecified tags or classification. These do not provide insight towards classification by a computer; thus, the machine must make assessment about inherant relationships among examples. Used in unsupervised learning.
Unstructured Data	Data without a pre-defined category or model. Examples - text, audio data, and videos.
Unsupervised Learning	A machine learning approach where training occurs without labeled data.
Use Case	As applied to systems engineering, a use case defines the series of steps required for a role (i.e. person or agent) to interact with a system to achieve a goal. May also loosely be defined as an application for a particular approach.
	1Witten, I.H. “Data Mining: Practical Machine Learning Tools and Techniques.” 4th ed. Elsevier 2019; Amsterdam.
	2 Goodfellow, I., Bengio, Y., and Courville, A. “Deep Learning”, 1st ed., MIT Press 2016; Cambrige, MA.
	3Shortliffe, E.H., and Cimino, J.J. “Biomedical Informatics: Computer Applications in Health Care and Biomedicine.” 4th ed. Springer-Verlag 2014; London.
	4Burke, J. “Health Analytics: Gaining the Insights to Transform Health Care.” 1st ed. John Wiley & Sons; Hoboken, NJ.

Given the above, how can AI impact the domain of clinical immunology? In facilitating disease diagnosis, computational methods may highlight patients at risk for primary immunodeficiency diseases (PID).[10] For example, a risk-assessment algorithm can be embedded within a health plan data pipeline to identify patients with medical conditions, risks or clinical trajectories to facilitate early intervention. Additionally, rule-based analysis of structured laboratory data can enable population-wide assessment of risk for hypogammaglobulinemia.[11] Yet, these approaches should be seen as the tip of the clinical immunology opportunity iceberg as opportunities range from novel disease gene identification to subject selection for enrollment into clinical trials.[12, 13] In this review, we will provide an overview of AI and then explore its use within clinical immunology for advancing biological understanding, improving diagnosis and driving increased quality of care for patients with immunological disorders.

AI & Healthcare:

How is AI Suited to Manage Big Data?

By every standard of the definition, healthcare data represent ‘Big Data’(BD). In general, BD refers to data which are of tremendous volume, amassed with great velocity and is of substantial variety.[14] This duality of BD warrants consideration for both effectively using it and managing complexity. Unique BD features also require specific tools accomplish data inference.

Traditional ML algorithms are designed to ingest large amounts of data and find patterns. Here, inputs can be derived from routine clinical, laboratory and claims related data to form “examples” which are analyzed by the model after they are “labeled” based upon a pre-defined task.[15] During training, AI algorithms weight certain features more heavily as they relate to predicting a labeled class outcome according to the specific method schema chosen.[2, 7] Weighted features from the model are then retained for analysis of new inputs and may be revised with subsequent training examples.[16] With appropriate structuring, AI algorithms can input large amounts of varied and complex data, analyzed according to feature weights learned in training and output a prediction which can be acted upon.

How Does AI Work?

The task, use case or specific application for AI may be approached in several different ways and can require classification or regression.[7, 16] In all workflows, AI algorithms must be trained, validated and then tested. During testing, algorithmic performance, bias and variance should always be assessed on an “un-seen” dataset.[7, 15] A common approach for AI is supervised learning where training data are labeled (i.e. instances with and without disease) before loading into an algorithm for pattern assessment, followed by validation and performance measurement.[7] (Figure 1: B–D) Some algorithms can ingest unlabeled data to perform “unsupervised” learning where patterns are learned from underlying data structure; however, almost all methods presently use some degree of supervised learning.[17] (Figure 1: A)

Figure 1: Machine Learning Examples.

Classic machine learning approaches include unsupervised methods such as clustering algorithms (A) which find natural relationships between the data. Additionally, supervised learning via linear models(B), decision trees (C) and support vector machines (D) use labeled data to train models for regression or classification. Deep learning methods such as neural networks (E) pass inputs to subsequent layers where learning occurs at each node. Prediction error can be minimized by passing information backwards through a cost function before feeding forward for ultimate output classification. (Network model example constructed with http://alexlenail.me/NN-SVG/index.html)

Despite seeming magical, all AI algorithms are simply mathematical functions which describe input data and map to outputs. Examples of input data in healthcare include structured electronic health record (EHR) elements such as diagnostic codes, vital sign fields or demographic fields as well as unstructured data in the form of text and radiologic image data. Input data represent features of relevance to the desired prediction task and must be selected accordingly. Because inputs may be highly varied without clear inter-relationships, a variety of models must be experimentally applied and subsequently tuned to assess best performance in testing for the given task.[7, 16]

Deep Learning:

Deep learning (DL) is a distinct branch of AI which bears separate mention from traditional ML. While DL theory and application spans several decades, it has only recently become useful in practice with the application of more powerful processing units (i.e. Graphical Processing Units (GPUs)), availability of large datasets and the need to draw inferences from such data.[5, 18] The structure of a DL scheme (Figure 1, E) appears somewhat like a series of interconnected neurons, hence the term ‘neural network’. Here, data flows forward from an input layer to subsequent layers which refine classification at each node, while also feeding backwards to earlier layers (i.e. backpropagation) to minimize error via a cost function.[15]

Deep learning algorithms are particularly well suited for healthcare data.[5] These algorithms can handle disparate data types with ease and their performance often increases in linear fashion with amount of input data.[19] Examples of DL use in healthcare include prediction of mortality in the pediatric ICU [20], assessing the most likely disease course for patients with rheumatoid arthritis [21], mining structured EHR data [6] and extracting clinical concepts from text [22].

Wrangling Healthcare Data:

For AI to become routinely useful in healthcare, validated datasets must be available for analysis.[2] Unfortunately, assembling a useful dataset requires elements to be extracted from EHR data which is known to be unreliable, difficult to obtain and highly variable.[23, 24] One solution for enabling efficient information extraction for healthcare analytics is to partner data science teams, clinicians and laboratory scientists.[25] In this way, the technical knowledge of data scientists can complement clinical and scientific domain expert knowledge to optimize data extraction, dataset construction and analysis along a comprehensive and relevant trajectory.[26]

Using AI to Advance Care of Patients with Immunologic Disorders:

Basic Science and Genomic Applications:

Robust ML applications are evident in the basic sciences and range from basic pattern recognition to modeling complex immune networks, hypothesis generation and analysis efforts. The vast applications across basic science and genomic or bioinformatic approaches are beyond the scope of this review; however, representative examples of relevance to clinical immunology will be highlighted.

At a cellular level, deep convolutional neural networks (CNN) are being used to clarify cognate interactions between human immune cells for understanding cell-cell interactions[27]. Additionally, ML is improving vaccine design, vaccine dose modeling, and prediction of immune responses. [28–30] Within the space of PID research, AI has helped identify candidate genes and understand protein folding interactions in common variable immune deficiency (CVID)[31].

Multiscale ML models, support vector machine (SVM) learning approaches [32] and long-short term memory (LSTM) networks [33] can integrate data about gene expression, protein interactions/pathways, microbiome data and model organism data for disease prediction. Genomic applications include “time-series AI” that may detect DNA sequence elements indicative of gene splicing [34] or methods to assist with variant calling of sequence data. [35, 36]

Data visualization approaches have also been applied to immunologic data sets for dimensionality reduction in the form of principal component analyses, or network diagrams to facilitate reduction of complexity in visualizing high-dimensional data.[37, 38] Graphical approaches to depicting complex data and developing new insights from publicly available data-sets from clinical trials are also possible using a variety of visualization approaches[39].

Disease Diagnosis:

Recent efforts to create computable phenotypes based upon EHR data have proven effective for automatic detection of patients with a variety of clinical conditions.[40, 41] Additionally, structured EHR data can be used to risk stratify patients for PID from large health systems.[10] Recently, a rule-based algorithm called the Software for Primary Immunodeficiency Recognition, Intervention and Tracking (SPIRIT®) Analyzer effectively analyzed claims data (diagnosis codes, pharmacy codes, demographic information) in a study of nearly 200,000 individuals at a large academic medical center.[10] Approximately 1% of the population here had a calculated medium-high risk for PID. One-year follow-up analysis of the risk group revealed that 1% were given a coded PID diagnosis or coded infection concerning for underlying susceptibility. This pilot study provided evidence that SPIRIT® could cull at-risk individuals from a general pediatric population. A longitudinal study is now underway to further assess performance of the SPIRIT® Analyzer and add a learning algorithm to the analysis.

Extending work of digitally encoded phenotypes, the Human Phenotype Ontology (HPO) program recently expanded to include terms relating to immunologic disorders.[42] Use of immune-disease relevant HPO annotation terms may enable linking of genomic variants and patient features to expedite identification of patients and their underlying genomic disease determinants.[43] The combination of HPO terms, unstructured EHR data mined with natural language processing (NLP), and structured laboratory and genomic data are being combined into a large federally funded dataset as part of the “All of Us” program to enable AI driven insights for diagnosis, patient care and research.[44] Other immunobiology, clinical and laboratory networks are being used to refine digital phenotypes and should be expected to facilitate diagnosis and advancement of science in clinical immunology.[45, 46]

Probabilistic models are also useful for disease diagnosis and prognosis [47–49]. Specifically, use of Bayesian networks (BN) and Bayesian classifiers have impacted the field of Allergy-Immunology [50]. These networks combine domain expertise and conditional probabilities which enable calculation of joint probabilities of interest.[51] Recently, a BN was constructed to predict PID risk and provide prescriptive guidance about the most appropriate diagnostic evaluations which showed promising results.[52] This BN was trained on readily-available EHR data and can be accessed by a single end-user clinician via a web interface. Alternatively, the BN can ingest EHR claims data via an application-program interface (API) for automated performance.

Clinical Decision Support:

Clinical decision support (CDS) focuses on providing timely, validated evidence to clinicians, patients and families in a computable format through the EHR.[53] (Figure 2) Formats for CDS range from alerts, best practice reminders, order sets, clinical summaries and AI-based prediction.[54] The backbone of CDS is a rule-driven logic for reminding about routine clinical interventions such as vaccination timing or assessment of cardiovascular risk.[55]

Figure 2: Paradigm for a Data Pipeline Enabling CDS & Patient Decision-making.

An example data pipeline schema originating from the clinician-patient interaction. Data entered into the EHR flows into a database/data warehouse where it can be structured, undergo analysis (asterisk) and transformation to information. This information can be fed back to clinicians to inform subsequent encounters (CDS-clinical decision support) or it may enable patient guidance about personal healthcare decisions as provided through the EHR patient portal. Reused with permission from [10].

A major concern about CDS relates to alert-fatigue and best practice notifications without appropriate outcomes-based automation[55, 56]. Balancing the number of alerts with patient safety is an ongoing challenge. Evidence for this includes a recent study which noted 80% of ignored alerts in an ICU setting were appropriate. [57] However, inappropriately ignored alerts in the same study were six times more likely to be associated with an adverse event. This report demonstrated that a CDS system for improving primary-care based asthma management was accessed in less than 20% of appropriate cases suggesting that usability is still a major hurdle [58, 59]. In every domain, CDS will require expert knowledge and deep clinical workflow understanding in order to drive efficacy and minimize intrusion.[53]

Yet the potential to improve care with CDS is real. For example, patient-facing, internet-of-things (IoT) technologies like wearables can reach patients dynamically and longitudinally to suggest improved health behaviors in real-time. One such application of immunology focused IoT technology is the Allergic Rhinitis and its Impact on Asthma (ARIA) Phase 3. This mobile application guides appropriate self-management of rhinitis, conjunctivitis and asthma via an innovative Mobile Airways Sentinel Network (MASK) [60]. This and other AI-based CDS have potential to merge different data sets by incorporating clinical and biomedical research into patient care, augmenting hypothesis-based research[61].

Current Limitations & Concerns about AI:

The promise of AI in healthcare is evident, yet several challenges remain before potency of these methodologies is fully realized. First, hype surrounding AI ability exceeds the science at present.[62] While AI is a powerful tool, we cannot expect technology to supersede insights of a well-trained clinician. This sentiment is beautifully conveyed by Cerrato and Halamka who state “…our enthusiastic take on digital innovation should not give readers the impression that AI will ever replace a competent physician. That said, there is little doubt that a competent physician who uses all the tools that AI has to offer will soon replace the competent physician who ignores these tools.”.[54] Algorithms are often considered to be a “black box” where mechanisms of output are not clearly evident raising suspicion and often leading to mistrust.[63] Concern about AI is warranted given previous indiscretions as in the situation with IBM’s Watson and its erroneous recommendations for cancer patients.[64]

Implicit bias is another major issue within the field of AI.[65] Thus, algorithms must be trained on a diverse population with respect to geography, disease state, gender, socioeconomic status and race or ethnicity to minimize performance gaps. Bias is already an appropriate concern within healthcare at-large and not isolated to the use of AI, but AI may further worsen such inequities.[65, 66] For example, healthcare quality is known to differ across socioeconomic strata which can affect performance of CDS systems.[67] Even high quality studies have inadvertently excluded underrepresented minorities thus exacerbating bias.[68] The field of AI science will need to overcome these and other challenges in order for the technology to have full utility in allergy-immunology and beyond.

Conclusions:

The fields of AI and medicine, specifically allergy-immunology, can be advanced together in exciting ways to improve patient care, research and clinical operations. Here we present some current applications and background information of relevance to AI’s potential for impacting clinical immunology. To advance the field of clinical immunology, health systems must be refined to facilitate implementation of AI, Allergist-Immunologists must seek to integrate AI/ML appropriately and look for collaboration with data scientists.

Key Points:

• Artificial intelligence, augmented intelligence and machine learning represent a subset of biomedical informatics focused on analyzing large datasets to make inferences.

• Analysis of large datasets with AI is already improving clinical work and research efforts in the field of clinical immunology.

• Use of AI and ML within the clinical immunology domain will expand knowledge and provide new insights.

• Clinical immunologists should become familiar with the principles of AI and data science or partner with data scientists to tackle new problems and advance our speciality.