Artificial intelligence is increasingly being utilized in multiple domains of nephrology which ranges from acute renal failure, chronic kidney disease dialysis as well as detecting rare diseases such as Fabry disease. It is being positioned to play a significant role in addressing healthy inequities specially in the field of organ transplantation Artificial intelligence can also act as a force multiplier in transition to value based care.
The term artificial intelligence was first used at a Conference in 1956 in Dartmouth, NH, USA. Since then, there has been slow progress, partly limited by cloud storage and computing capability, until the development of large learning models in 1992 and subsequent introduction of generative machine learning in 2014. One of the historic events was the defeat of chess champion Gary Kasparov in 1997 by an IBM supercomputer named Deep Blue.
Eventually, the introduction of ChatGPT by OpenAI in 2017 sparked a big public interest in this field, although slowly and steadily, this technology has been refined over the years.
In the last 2 decades, AI has undergone tremendous transformations with the introduction of deep learning and natural language processing (NLP).
Deep learning is increasingly being utilized in kidney imaging, to differentiate between malignant and benign kidney lesions. Additionally, it can also read kidney biopsy images and is used for the diagnosis of kidney transplant rejection.
Natural language processing allows for extracting patient characteristics from clinical notes and has significantly improved in recent years. This has improved the ability to gather relevant data from the vast pool of healthcare notes, which would otherwise have been very time-consuming and labor-intensive.
The diagnosis of acute kidney injury is based on rising serum creatinine and decreased urine output. Both are lagging indicators of renal failure. This gap in care has opened opportunities for utilizing artificial intelligence-based clinical decision support systems for early detection of acute renal failure.
One such example is Deep Mind, which is a Google GI subsidiary. In 2019, Deep Mind developed a machine-learning model that could detect acute kidney injury up to 48 hours in advance. This model utilized a data set of 700,000 individuals from the Veterans Health Administration over 5 years. Machine learning-based dynamic models are more accurate in predicting acute kidney injury than traditional static risk models. This has improved the ability to detect AKI in hospitalized patients, although it still has to be proven that implementing these models leads to clinically meaningful outcomes.
Recently, a machine learning model for predicting hospital-acquired acute kidney injury was deployed at Brigham and Women's Hospital for external validation. This model was able to predict AKI around 22 hours in advance.
Chronic kidney disease is often underrecognized and underreported, partly due to lack of effective screening measures. There is a big unmet need for early and accurate diagnosis of CKD. There is wide variation in referral patterns to nephrology and it ranges across a spectrum of eGFR values. One quality study showed that only 56% of Primary care physicians routinely check kidney function in their patients with Diabetes mellitus. AI/ML driven algorithm integrated with EHR specially in primary care setting can help address these issues, triggering early nephrology referral and improve outcomes in patients with diabetic kidney disease (DKD).
Healthcare data is generally heterogenous and comprises of clinical notes and other patient related details which can be highly variable. In nephrology to some extent some of the data is structured due to the mathematical nature of this specialty, still suffers from inaccuracies and missing entries. If such data is used to train an algorithm, then there may be issues with generalizability as well as interpretability of these results.
Without addressing these issues, these algorithms tend to be limited to a specialized area or hospital, certain demographic rather than a large set of patients. Deep learning has shown remarkable accuracy in image analysis particularly in interpreting kidney biopsy images. It has demonstrated that it is as good as an experience nephrologist in reading kidney biopsy samples. This accuracy could be achieved by training such algorithms with a large set of images which is relatively easier to do in biopsy images due to whole slide image (WSI) technology.
Currently, organ allocation is managed by the United Network for Organ Sharing (UNOS), which is a tier-based system with different priority tiers. These tiers are based on several characteristics such as age, blood type, disease severity, and many others. There is increasing concern that the current organ allocation system has led to inequitable access to transplants for some patients. For example, race-based eGFR calculation has impacted African American patients. AI based allocation system can improve organ allocation, address organ scarcity as well as mismatch issues and it can also predict long-term organ survival. For example, recently, in 2023, an AI-based organ allocation framework has been introduced, named "Continuous distribution”. It is currently being done for only lung transplants, but it may act as a primer for other organ transplants such as kidneys, pancreas, and other organs.
There is a big push towards value-based care model in healthcare. Value-based care models incentivize quality over quantity and focus on improving patient outcomes. Artificial intelligence can act as a force multiplier in value-based care. Predictive artificial intelligence can be used to identify high-risk patients who can be monitored more closely after posthospital discharge and prevent readmissions, leading to lower healthcare costs and improved outcomes.
Recently a large dialysis care organization (LDO) developed a predictive model which could identify dialysis patient for high risk for hospitalization 1 week prior. Through this intervention they were able to reduce hospitalization rate due to " all cause" for their patients. Similarly, predictive AI can be used to detect disease outbreaks, drug dosing, imaging, and much more.
Before clinical implementation, AI algorithms need to undergo a rigorous peer review process and assessment of technology readiness like standards used by NASA
Such algorithms need to undergo trials similar to a randomized control trial which is considered to be the gold standard of generating evidence. However, it is practically not possible to conduct randomized control trials for every clinical scenario. Artificial intelligence can be useful in such scenarios as it can be used to process real-world data to generate real-world evidence. Machine learning models can be used to conduct simulated trials using big data obtained from electronic health records and such simulated trials are much less time-consuming as well as require lower cost.
Majority of hemodialysis in this country is provided in the form of intermittent hemodialysis at freestanding dialysis clinics. It generates a wealth of standardized health data making it an attractive choice for deploying such machine learning models. Artificial intelligence has been used in drug dosing in dialysis. Machine learning models have been shown to decrease darbepoetin alfa as well as intravenous iron consumption.
Intradialytic hypotension can cause muscle cramps, which are uncomfortable for dialysis patients. It is difficult to predict accurately whether a patient will experience intradialytic hypotension. Machine learning models have been used in hemodialysis to predict intradialytic hypotension with higher accuracy as compared to other models
During COVID-19 pandemic a large dialysis organization utilized a predictive model which looked at 80 variables to successfully identify undetectable COVID-19 infection 3 days prior to onset of symptoms
Artificial intelligence is not immune to data privacy laws such as HIPAA. Data protection is one of the key concerns in clinical applications of AI in nephrology.
It is important to ensure that patient data is entered securely and informed consent is Obtained from the patients. Anonymizing the data prior to doing analysis or processing is one of the ways to ensure data privacy. Patients should be informed as to how their data is being used and stored, which in turn will bring more transparency and build trust in such systems.
Despite significant advancements in research on artificial intelligence, its integration in clinical setting remains low. This is due to multiple reasons including data privacy issues, black box nature of algorithms and bias concerns.
One of the challenges is the lack of standards for sharing data between different health systems, which makes it challenging to integrate artificial intelligence algorithms into the routine workflow for nephrologists. Another key factor is that there is a lack of trust among clinicians/healthcare professionals for artificial intelligence which goes back to black box nature of these algorithms. One way of addressing this is to generate high-quality evidence and rigorous validation of these algorithms.
Artificial intelligence-based models can amplify biases and discrimination, particularly if the training data set reflects the healthcare disparities.
One such example is a race-based eGFR calculation, which has been a topic of debate recently, and the growing consensus is that the inclusion of race in these equations has contributed to disparities in kidney transplants among African-American individuals. This led to the development of alternative eGFR measurements that do not include race, such as Cystatin C-based equations. If the AI model incorporates the old race based GFR then the results may be biased.
To minimize bias, several fair practices have been recommended which include selection of representative patients in the model development, careful consideration on incorporating social determinants of health (SDOH), validation of models prior to deployment with tools such as Prediction model Risk Of Bias Assessment Tool (PROBAST). For example, if an AI model learns that low socioeconomic status is associated with poor outcomes even with an effective treatment, then that model may recommend against providing an effective treatment for patients with low socioeconomic status.
Artificial intelligence holds significant promise in transforming the landscape of managing kidney disease in the coming times. It has applications in each and every domain of nephrology, including acute kidney injury, chronic kidney disease progression dialysis kidney transplant as well as kidney pathology.
In this era of precision medicine, incorporating artificial intelligence will allow us to deliver individualized medical care. For example, in the case of dialysis, I envision a future where artificial intelligence models are integrated into routine dialysis delivery, including dialysis machines. A patient comes into the dialysis clinic with slight volume overload and the machine learning model has already alerted the dialysis nurse and physicians with the optimal approach to ultrafiltration by continuous monitoring of blood volume and bioimpedance data as well as lung ultrasonography so as to minimize major hemodynamic shifts during dialysis. This can improve patient experience as well as reduce adverse outcomes.
Artificial intelligence is here to stay, and we will see increasing integration in clinical nephrology in the future.
AI in its current form is neither intended to nor can it replace nephrologist. In coming times, nephrologists who embrace artificial intelligence will not only have a significant advantage, but also will set themselves apart as leaders in their field. We need to familiarize nephrologist with basics of artificial intelligence and need to invest more resources to develop AI competent workforce.
Dr. Prabhat Singh is a Board-certified nephrologist with over 11 years of clinical experience and is currently a Physician Partner at Kidney Specialists of South Texas. He serves as Medical Director for two DaVita hemodialysis units and a home dialysis clinic. He is an Editorial Board Member for BMC Nephrology and the Journal of Medical Case Reports, and also actively teaches at Corpus Christi Commumity, and is Technology Medical Director at Missions of Mercy Charity Clinic.
