Machine Learning Models Can Improve Efforts to Offer Dynamic, Precise Kidney Care

Artificial intelligence (AI) and deep and machine learning can become implanted in many areas of nephrology, according to Azra Bihorac, MD, MS, FCCM, FASN, the senior associate dean of the Office of Research with the University of Florida College of Medicine, Gainsville, Florida, at the American Society of Nephrology’s (ASN) 2023 Kidney Week Annual Meeting, taking place November 2 to 5 in Philadelphia, Pennsylvania. Bihorac explained further that AI literacy is becoming crucial for health care professionals.

“I think that a huge, huge transformation is coming,” Bihorac said during her presentation, which outlined the different ways in which AI can influence kidney care. “I would say the number 1 task for you is to become AI literate…[because] we do need to dominate this conversation and we can't have others bringing this evolution to us.”

In particular, supervised machine learning may be beneficial for health care professionals to be familiar with, according to Bihorac. Supervised machine learning is a type of AI and machine learning that can train data and algorithms to classify and label data to predict outcomes based on trends in the data. Bihorac explained that research has identified the benefits of supervised machine learning in the perioperative surgical setting, for example. In this setting, supervised machine learning can be used to assess risk in real time and identify potential major complications.

Supervised machine learning is a type of AI and machine learning that can train data and algorithms to classify and label data to predict outcomes based on trends in the data. Image Credit: © nobeastsofierce - stock.adobe.com

On the other hand, deep learning is better for larger data sets, “allow[ing] you to use the dense data in a time series framework,” Bihorac said during the session. In the surgery setting, for example, deep learning can help health care professionals interpret data, augment dynamic risk prediction for surgery, and, using a network/cloud, deliver data to physicians or other health care professionals within a network, Bihorac explained.

There is a laundry list of other capabilities that AI could have in kidney care, according to Bihorac. The original Sequential Organ Failure Assessment (SOFA) is a system that assesses organ function to predict acuity and mortality outcomes at a hospital or intensive care unit (ICU). However, Bihorac noted that deepSOFA, which uses deep learning to assess similar inputs, can take SOFA further by improving understanding of physiological patterns, assessing illness severity, and predicting when patients are not doing well.

These different ways to input data and monitoring of organs are “urgently needed in nephrology,” Bihorac said. It also opens the limited door of testing in nephrology. “We need to have more diagnostic tests beyond just creatinine or urine,” Bihorac said.

AI can also be used for autonomous patient monitoring, which can allow pharmacists to monitor patients better. According to Bihorac, a collaborative team of researchers at the Universities of Florida and Pittsburg are currently testing how pharmacists can use data from AI models to inform better drug use, specifically factoring in patient risk and disease severity, allowing pharmacists to alert patients and manage their drug use.

Bihorac noted further that the power of machine learning and deep learning models in health care may come from their ability to work for the individual patient. The technology may be able to achieve this because it is dynamic and can incorporate vast amounts of data in real time.

“[Patients] are different [people] every day, Bihorac emphasizes. “Every day something is done to you, or something is happening in your body.”

On the personalization front, AI might be able to recognize individual faces, gauge risk of delirium, and help with the management of delirium, helping to even detect early signs of dementia, Bihorac explained. Facial recognition offers practitioners a visual cue to understand more about patient risks and disease severity and could lead to better treatment. Further, managing delirium is important for older adults because they are at a much higher risk.

But AI models, such as large language models (LLM), contain biases that could impact patients, Bihorac warns. Since the people who input data into these models are biased themselves, these AI machines could become “bias optimizers” that learn and grow from the biases already integrated into their programing. Once they are trained with these biases, they cannot be untrained, Bihorac noted.

“[While] we [as people] are born with biases, we fix them,” Bihorac explains. “Machines really can't do that...they cannot fix those biases.”

So ideal algorithms should be explainable, dynamic, precise, autonomous, fair, and reproduceable, according to Bihorac; this could allow for optimized deep learning models and more equitable health care. Further, with “billions” of parameters to secure data, some machine learn models with LLM will be ready for large-scale implementation, according to Bihorac.

“Just in the last 2 years, everyone who wants to have power will have their own foundational model because that's what [the world is] going towards,” Bihorac concludes. “We need to be part of that.”

Reference

Biharoc A. Beyond AKI: How is AI Improving Critical Care in Illness? Session. ASN Kidney Week Annual Meeting. November 2 to 5, 2023. Philadelphia, PA.