Artificial Intelligence in Nephrology: Pioneering Precision with Multimodal Intelligence

Nephrology education

GPT-4 had shown remarkable potential, outperforming USMLE scores typically earned by early-stage medical students, with valid insights for >84% of responses.¹⁷ These advancements highlight AI’s potential to enhance medical education and clinical reasoning, with possible applications in nephrology. Virtual AI-powered environments can provide learning experiences for trainees, simulating complex clinical scenarios like AKI or dialysis complications. Mixed-reality applications such as the CyranoHealth app offer interactive, case-based training tailored to healthcare professionals.^18,19

Early Diagnosis and Prediction

AI has shown potential in predicting and managing AKI and CKD.^20,21 Various ML models predict adverse outcomes in hospitalized patients at risk of these diseases, offering valuable tools for early intervention and resource allocation.^20-25 During the COVID-19 pandemic, the AKI risk in hospitalized patients became a significant concern. To address this, Ponce and colleagues developed a prognostic score using ML techniques.²⁰ They applied a 10-fold cross-validation method to fit models and validated their accuracy using the Area Under the Receiver Operating Characteristic Curve (AUC ROC). They studied 870 patients from the Latin America AKI COVID-19 Registry and reported a 0.82 AUC ROC, indicating the model’s strong predictive ability. Similarly, Vaid and colleagues developed several ML models, including logistic regression, Least Absolute Shrinkage and Selection Operator (LASSO), random forest (RF), and eXtreme Gradient Boosting (XGBoost), to predict the dialysis need or death at various time points after admission of patients with COVID-19 and AKI.²¹ LASSO is a regression technique combining variable selection and regularization by shrinking less significant feature coefficients to zero, enhancing model simplicity and interpretability. RF, an ensemble learning method, constructs multiple decision trees (DTs) and aggregates their predictions to improve accuracy and reduce overfitting, making it effective for classification and regression tasks. XGBoost is a highly efficient and powerful gradient boosting algorithm that iteratively improves model accuracy by optimizing weak learners while handling missing data and large datasets with exceptional speed and performance.^13-15 In their multicenter cohort study involving 6093 patients, the XGBoost model without imputation outperformed other models, achieving a 0.85-0.87 AUC ROC and demonstrating AI’s effectiveness in predicting critical outcomes in patients with AKI. Another study used DL to identify sub-phenotypes of sepsis-associated AKI in 4001 ICU patients.²² By analyzing 188 variables, including vital signs and lab results, they identified three unique AKI sub-phenotypes with distinct comorbidities and outcomes. These findings highlight AI’s potential in refining AKI classification and guiding personalized treatment strategies. A systematic review on predicting mortality in patients with AKI using AI showed the highest and lowest pooled AUC for the broad learning system model, elastic net final, and the proposed clinical model, respectively.²³ AI can also contribute to the early sepsis diagnosis. A meta-analysis of seven studies demonstrated a 0.89 (95% CI: 0.86–0.92) pooled AUC for ML models predicting sepsis onset 3 to 4 hours in advance, with a 0.81 (95% CI: 0.80–0.81) sensitivity and 0.72 (95% CI: 0.72–0.72) specificity. In comparison, the pooled AUROCs for existing sepsis scoring systems, including the Modified Early Warning System (MEWS), Systemic Inflammatory Response Syndrome (SIRS), and Sequential Organ Failure Assessment (SOFA), were 0.70, 0.50, and 0.78, respectively.²⁴ These findings suggested that ML models might outperform traditional sepsis scoring systems in predicting sepsis. AI’s application in sepsis prediction is still being researched, and the best model type is yet to be formulated. AI has also been applied to CKD management, with several studies backing its potential to enhance early detection and prognosis. For instance, Debal et al., explored the capability of different models to predict CKD stages using binary and multi-category approaches.²⁵ They tested three models: RF, Support Vector Machine (SVM), and DT. SVM is a supervised ML algorithm that identifies an optimal hyperplane to separate data into distinct classes, maximizing the margin for effective classification or regression. DT is a tree-structured algorithm that splits data into branches based on feature values, creating a decision hierarchy to classify data or predict outcomes in an interpretable manner.^5-9 To choose the most important features for these models, they used methods like analysis of variance and recursive feature elimination, combined with cross-validation to improve accuracy. After testing these models with tenfold cross-validation, they found he RF model to outperform SVM and DT, especially when using recursive feature elimination. Tangri and colleagues developed models and equations to predict kidney failure risk by integrating extensive datasets.²⁶ The meta-analysis included data from 31 cohorts, comprising 721,357 participants with CKD stages 3 to 5 spanning four continents. Using risk factors from the original risk equations, new pooled kidney failure risk equations were developed by calculating and combining cohort-specific hazard ratios. The original equations demonstrated excellent discrimination, effectively distinguishing individuals who developed kidney failure with consistent performance across subgroups by age, race, and diabetes status. Another study showed that an automated, laboratory-based clinical decision support system could enhance physician compliance with guidelines for timely CKD monitoring.²⁷ An e-technology-based program was designed to identify patients at CKD risk and automatically order relevant screening tests.²⁸

AI and Creatinine Clearance

In critically ill patients, creatinine clearance (CrCl) is a key glomerular filtration rate indicator and can fluctuate daily. Huang et al. developed and validated predictive models using a gradient-boosting ML algorithm to forecast CrCl one day in advance, leveraging data from 2,825 patients in the EPaNIC multicenter database.²⁹ Three models were created: the “Core” model, incorporating demographic, admission diagnosis, and daily lab results; the “Core + BGA” model, adding blood gas analysis (BGA) data; and the “Core + BGA + Monitoring” model including high-resolution monitoring data. Model performance was evaluated against the actual CrCl by mean absolute error and root-mean-square error. All models demonstrated higher accuracy and lower prediction errors. Prediction models utilizing routinely collected ICU clinical data accurately predicted next-day CrCL. These models might have applications in adjusting drug dosages and identifying at-risk patients.

Dialysis

Dialysis is characterized by its structured and routine nature and produces lots of patient-specific data, including dialysis prescriptions (e.g., treatment duration, ultrafiltration volumes, and flow rates) and intradialytic biosignals (e.g., blood pressure, heart rate). These datasets, often stored in electronic health records (EHR), are ripe for AI/ML analysis, offering enhanced diagnosis, prognosis, and treatment recommendations. Chan et al. employed NLP to extract symptoms from EHRs, achieving higher sensitivity in identifying hemodialysis-related symptoms than traditional coding methods.³⁰ Zhang et al. utilized AI/ML-driven image analysis to classify vascular access aneurysms with impressive accuracy, demonstrating AI’s potential in image-based diagnostics.³¹ Prognostically, Lee et al. developed an RNN model for predicting intradialytic hypotension with high accuracy, showcasing AI’s capacity to foresee critical events during dialysis sessions.³² Accurate dry weight estimation is challenging but essential to reduce morbidity and mortality. Some studies on neural network development showed using bio-impedancemetry, blood volume monitoring, and blood pressure as inputs to predict AI-based dry weight showed that the neural network’s predictions outperformed those of experienced nephrologists in most cases, highlighting its potential.^33,34

Despite these promising developments, routine AI/ML implementation in dialysis remains limited, with only a few such as ML-driven anemia management models, reaching clinical practice.³⁵

Vascular access management

AI is now being utilized in vascular access (VA) management through applications in preoperative planning, monitoring, and predictive modeling. The Vexev Ultrasound Imaging System exemplifies advancements in preoperative mapping by employing robotic tomographic ultrasound to provide comprehensive 3-D vascular data comparable to CT or MRI scans without radiation or nephrotoxic contrast agents risks. This hands-free system integrates cloud-based platforms like vxView for remote expert stenosis, aneurysms, and thrombosis analysis.³⁶ ML can also aid surgical planning, as shown by Doneda et al., whose models achieved 96.8% accuracy in predicting arteriovenous fistula (AVF) maturation and postoperative blood flow velocities, optimizing resource utilization, and reducing failure rates.³⁷ In monitoring and risk prediction, AI-powered tools like DeepVAQ use photoplethysmography (PPG) sensors and DL to assess VA quality with over 92% accuracy, aiding early complication detection.³⁸ Mel spectrogram analysis, as utilized by Chung et al., transforms audio data into visual formats for DL models, successfully predicting AVF stenosis and malfunction with high accuracy.³⁹ Risk models like PREDICT-AVF and AVF-FM incorporate patient-specific data to forecast complications such as thrombosis and recurrent interventions. The PREDICT-AVF web application, developed using prospective cohort data, achieved a 0.75 AUROC for one-year intervention prediction.⁴⁰

Peritoneal dialysis

AI has now been integrated into the PD care. One example is a chatbot system designed to support PD patients. This system, offers a range of functionalities, including instructional videos, clinical reminders, dietary guidelines, and automatic PD guidance.⁴¹ For clinical predictions, Noh et al. utilized ML algorithms to assess mortality risk among 1,730 PD patients, finding that survival-tree models outperformed traditional methods like Cox regression, with a 0.769 concordance index compared to 0.745.⁴² Additionally, AI has been applied to predict and manage peritonitis. Zhang et al. used a systematic ML approach to analyze immune responses to microbiologically well-defined infections in 83 PD patients presenting with acute peritonitis.⁴³ Many biomarkers were used to characterize pathogen-specific local immune responses to genera such as Streptococcus and coagulase-negative Staphylococcus showing value of nonlinear approaches for analyzing complex biomedical datasets where traditional statistical methods fall short and single biomarkers lack sensitivity and specificity. The nature of the immune signatures varied depending on the mathematical model applied. By directly comparing three ML approaches—RF, SVMx, and ANNs—the study identified RF as the most effective model for microbiological and clinical outcome prediction.

Personalized Medicine

The idea “one therapy fits all” is outdated. Precision medicine focuses on optimizing treatment outcomes and minimizing adverse effects for each patient. AI has propelled personalized medicine development. Destere et al. demonstrated a forward-backwards ML approach to optimize the early rituximab regimen intensification in patients with membranous nephropathy with high underdosing risk.⁴⁴ This method identified the best combination of variables to predict rituximab underdosing using training data and validated these predictions on a test set. CURATE.AI predicts optimal dosages and treatment outcomes based on individual patient data. It continuously adjusts patient profiles as the disease status changes, optimizing dosages for single drugs and drug combinations.⁴⁵ IBM Watson for renal oncology uses AI to predict personalized treatment plans and responses.⁴⁶

Kidney imaging

Radiomics involves mathematically extracting quantitative features from medical images.⁴⁷ Recent studies use AI, including radiomics-based ML and DL, to derive obscure diagnostic insights through visual inspection of digital images.^48-50 Total kidney volume (TKV) is a crucial imaging marker evaluating Autosomal Dominant Polycystic Kidney Disease (ADPKD) severity and progression. DL networks distinguish renal parenchyma from pathological cysts without manual tracking, estimating TKV as effectively as automated kidney segmentation. Recent studies apply DL algorithms in ultrasound, CT, and MRI to compute TKV in ADPKD patients, matching manual method accuracy.^49-51 Goel et al. created a CNN model using 3D-US images to segment PKD regions for TKV computation, achieving a 0.80 Dice score in the test set.⁵⁰ (The Dice coefficient is a statistical similarity measure, commonly used to evaluate the overlap between two sets or binary classifications). Li et al. conducted a radiomics analysis using diffusion-weighted imaging from fMRI data. The authors developed a logistic regression model to differentiate between individuals with CKD and healthy subjects, achieving 93% sensitivity and 70% specificity.⁵²

Kidney Transplantation

AI’s application has been researched for pre-transplant evaluation, organ allocation systems, and long-term post-transplant care. Kidney Donor Risk Index, Kidney Donor Profile Index (KDPI), and Estimated Post-Transplant Survival Score tools have enhanced organ allocation by combining donor-recipient data to predict graft survival and optimize matches. Bae et al. reported challenges like unchanged discard rates post-KDPI implementation, but newer AI-driven models like neural networks have demonstrated superior predictive accuracy in post-transplant outcomes, including graft survival and delayed graft function (DGF).⁵³ For instance, Kawakita et al. highlighted ML model’s advanced performance in predicting DGF, which can improve decision-making and resource allocation.⁵⁴ Advanced DL systems, such as those developed by Marsh et al., have reduced unnecessary organ discards by enhancing pre-operative graft biopsy analysis.⁵⁵ AI has also been used for waitlist management, patient education, and post-transplant monitoring. Tools like ML-based waitlist prediction models and risk stratification algorithms, as described by Pineda et al., might help refinine patient care strategies.⁵⁶ Pineda et al.⁵⁶ identified gene signatures associated with rejection utilizing the ML model. In post-transplant settings, AI helped predict complications like pneumonia and cytomegalovirus infection, as shown by Luo et al. and Sheppard et al., respectively.^57,58 AI-powered surgical systems can also enhance precision through augmented reality and robotic assistance, while personalized immunosuppressive protocols optimize Tacrolimus and Mycophenolate Mofetil (MMF) dosing, demonstrated in studies leveraging genetic polymorphisms and neural networks.⁵⁹ As the field advances, AI integration with regenerative medicine and machine perfusion may redefine KT processes, ensuring equitable organ allocation and improving global transplant success rates.

Kidney histopathology

Kidney biopsy pathology is the gold standard in diagnosing kidney parenchymal diseases. However, more standardization in the diagnostic approach is needed. Variability in specimen preparation and pathological examination among observers pose challenges for reproducibility, affecting diagnosis accuracy. AI has seen notable advancements in this field. Whole-slide images (WSIs) are produced through high-throughput slide digitization, enabling efficient utilization of computer-assisted histopathological analysis.⁶⁰ Pathology image segmentation with DL involves data preparation, image preprocessing (including normalization and augmentation), model selection and construction (software choice, model training), post-processing techniques application, feature extraction, and correlating these features with diseases.^60-64 Recently, DL algorithms have shown promise in pathology image analysis, including tumor region identification and metastasis detection.⁶¹ Jayapandian et al. developed a DL method for segmenting renal cortex tissue structures across various stains (HE, PAS, silver, trichrome). They indicated PAS-stained whole slide images (WSIs) as optimal for achieving consistent segmentation of diverse structures, including glomerular tufts, Bowman’s capsules, tubules, peritubular capillaries, arteries, and arterioles.⁴ Zeng et al. developed DLV algorithms to detect glomerular lesions and classify and quantify various intrinsic glomerular cell types. Additionally, they devised a network-driven mesangial hypercellularity score using PAS-stained slides in 400 Chinese patients diagnosed with immunoglobulin A nephropathy (IgAN).⁶² Weis et al. showed that DL algorithms, particularly CNN, can accurately identify subtle and overlapping glomerular morphological changes from conventional microscopy. By categorizing nine glomerular alteration classes and training CNNs on datasets curated by nephropathology experts (23,395 images), the models achieved excellent classification performance.⁶³

Clinical trials

AI is increasingly integrating into research and development, fueled by advancements in computational technology. Its utility includes analyzing biometric data, imaging, and facilitating trial design, recruitment, retention, and outcome analysis.^64,65 AI-powered tools can streamline protocol drafting using large language models (LLMs) like retrieval-augmented generation (RAG). These tools integrate external knowledge sources, such as KDIGO guidelines, to generate contextually relevant trial documents. However, human oversight remains essential to ensure logical accuracy and precision. AI can potentially enhance patient recruitment by analyzing EHRs and leveraging tools like Criteria2Query for clinical trial eligibility query generation.⁶⁶ ML models can further optimize recruitment by managing missing data and identifying high-risk patients, especially in trials on progressive kidney diseases. In pretrial stages, digital twins, such as TWIN-GPT, can enable virtual simulations of patient trajectories.⁶⁷ This approach can help in predicting outcomes and adverse events, advancing drug development and hypothesis testing without direct patient involvement. During the consent process, AI-driven chatbots can be used to simplify documentation and improve patient comprehension, though concerns about dehumanization persist. Additionally, real-time AI-driven data analysis can enable dynamic trial adjustments, ensuring timely outcomes. Table 1 and Figure 2 show the overview of AI applications in Nephrology.

Close

Figure 2:

Application of AI in Nephrology. AI: Artificial intelligence.

Export to PPT

Addressing Bias in AI Models

Bias in AI can severely affect the accuracy and fairness of clinical decisions.^70,72,73 Types of bias include:

• Selection Bias: When training data does not represent the general patient population.

• Measurement Bias: Due to inconsistencies in data collection methods across facilities.

• Algorithmic Bias: When the training data or algorithm processing introduces bias.

• Confirmation Bias: When data confirming pre-existing beliefs is prioritized.

• Cultural Bias: When models do not account for cultural differences.

Diversifying data collection, standardizing measurements, using bias detection and mitigation tools, and continuously monitoring model performance can help tackle data bias. Effective AI implementation includes continuous learning from new data, active bias evaluation, model interpretability and explainability workflows, clinician input feedback loops, and fostering multidisciplinary collaboration to ensure clinical relevance and technical reliability.

AI’s integration in nephrology stands to transform the field by boosting diagnostic precision, enabling tailored treatment plans, and optimizing clinical workflows. With the capacity to analyze complex datasets—such as EHRS, imaging, and genetic information—AI might empower earlier detection and kidney disease intervention, leading to better patient outcomes. Ensuring patient trust through transparency, explainability, and ethical AI practices remains essential. As AI technologies continue to evolve, robust clinical validation and interdisciplinary collaboration will play pivotal roles in unlocking its full potential in nephrology. Addressing existing challenges and creating a supportive regulatory framework will be key to establishing AI as a more efficient, equitable, and advanced healthcare system cornerstone.