Introduction to Artificial Intelligence and Machine Learning in Nephrology

The terms artificial intelligence (AI), deep learning, and digital health have successfully transcended their niche origins and now become part of our public lexicon. AI plays a major role in our day-to-day life, whether it be internet searches or shopping recommendations. However, the successes of AI in medicine have been more muted. There is even controversy about the nomenclature, with the American Medical Association (AMA) using the term augmented intelligence.1 The robust use of mathematics in the clinical decision-making process in nephrology suggests that our field is ideal for clinical adoption of AI. To adopt a new technology, clinicians and scientists need to be conversant with the basics, be fluent in its vocabularies and nomenclature, and be inspired by the potential clinical applications to improve outcomes for patients with or at risk of kidney disease. Recognizing this need, the American Society of Nephrology (ASN) has created the Augmented Intelligence and Digital Health (AIDH) task force to develop a roadmap for kidney health in this new era. However, for these tools to be incorporated into our clinical practice, it is important to understand the fundamentals and the potential clinical applications of some of these newer approaches. Over the next few issues of CJASN, we present a series of articles authored by leaders in the field to acquaint the readership with the basics and the nomenclature in this field. First, Park and Hu delve deep into bias in clinical AI. This topic is of pertinence in nephrology, in view of the recent acknowledgment and replacement of the biased eGFR estimates and the Kidney Donor Risk Index (KDRI). One hazard of AI is magnification of bias through the inclusion of flawed or incomplete databases and development of decision making on the basis of these data. The authors provide a framework for illustrating the types of bias that can arise throughout the AI lifecycle with the goal of preventing and mitigating bias. Next, Heilbroner and Miotto explain deep learning, with a particular emphasis on medicine. Deep learning mimics a human neuron, where neural network layers are typically arranged sequentially and composed of many basic, nonlinear operations, such that one layer is fed into the next layer. Deep learning is the bedrock of most modern machine learning (ML) operations and enables inferences based on multimodal data including images, waveforms, and videos. Finally, Arivazhagan and Van Vleck provide an overview of natural language processing (NLP). NLP is an area of AI dealing with the ability of a computer program to understand human language as it is spoken and written—referred to as natural language. This technology powers some of the tools we use in our day-to-day activities, including chatbots and predictive text completion. In the next issue, Khezeli, Siegel, Shickel, Ozrazgat-Baslanti, Bihorac, and Rashidi introduce the concept of reinforcement learning (RL). RL focuses on learning optimal course of action from interactions with the environment to maximize a reward. This has been of utility in the nonclinical domain, with the well-publicized examples of beyond human proficiency in complex games, including those requiring the skills of intuition, persuasion, and strategy.2–4 This could be useful in clinical settings for decision making, with the ultimate reward function being better patient outcomes. Then, Shickel, Loftus, Ren, Rashidi, Bihorac, and Ozrazgat-Baslanti introduce the concept of the transformer model, one of the more significant breakthroughs in AI in recent times. A transformer model is a deep learning approach that learns context and thus meaning by tracking relationships in sequential data. They have already resulted in remarkable advancements in computer vision and NLP (ChatGPT) and are poised to enhance both the research and clinical enterprise. Finally, Balczewski, Cao, and Singh introduce the fundamentals of risk prediction, which is germane to day-to-day clinical practice. Anchored to a clinical case, the authors progress through several concepts including data quality, model form, model performance/metrics, and ethical considerations. From fundamentals, we then move to potential clinical applications. Bajaj and Koyner discuss the various applications of AI in diagnosis, prognosis, treatment, and subphenotyping for AKI. They present a comprehensive review of the current literature, limitations, and potential future directions for using AI in the setting of clinical AKI. Liu, Takeuchi, Chen, and Neyra made the case for how AI could help clinical decision making to improve delivery, resource allocation, and outcomes in patients who require continuous KRT. Kotanko, Zhang, and Wang discuss the utility of multimodal data and AI applications for patients requiring maintenance dialysis and point toward the need for prospective studies, clinical trials, and real-world routine use to power broad translation. Farrell and Chan discuss the utility of NLP in nephrology research and discuss how it could be used in improving clinical trial recruitment by identifying patients with specific diseases or symptoms and facilitating patient-centered outcomes research. Finally, in a pair of linked articles, leading computational kidney pathology researchers discuss how to develop and implement clinical grade kidney pathology into clinical care including data quality/availability, independent validation, and complementary qualitative and quantitative methodologies while adhering to privacy, safety, and ethical standards. In summary, we present one of the first comprehensive article collections in nephrology, addressing both sides of the coin, both fundamentals and clinical applications. Acknowledging the potential for transforming health care, we should remember that the provider-patient relationship is and always will be at the heart of medicine. AI and digital health will augment our ability to improve care and maintain health. Thus, at least for this physician-scientist, augmented intelligence is the preferred term over artificial intelligence, since, if used correctly and in an unbiased and ethical manner, it will augment our ability, not replace it. Disclosures G.N. Nadkarni reports consultancy agreements with Daiichi Sankyo, GLG consulting, Reata, Renalytix, Siemens Healthineers, Qiming Capital, and Variant Bio; ownership interest in Doximity, Data2Wisdom LLC, Pensieve Health, Renalytix, Nexus iConnect, and Verici; research funding from Renalytix; honoraria from Daiichi Sankyo; patents or royalties from Renalytix; an advisory or leadership role for Renalytix; and speakers bureau for Daiichi Sankyo. Funding This work was supported by research grants R01DK127139, R01DK12713902S1, R01HL155915, and U01DK133093 from the National Institute of Diabetes and Digestive and Kidney Diseases (G. Nadkarni).