AI and Home Dialysis: From Prediction to Personalization

VIENNA — “We are going to change anemia treatment. We are going to change fistula control. The future is going to be AI (artificial intelligence),” declared Adrian C.M. Covic, MD, PhD, at the opening of a session dedicated to AI and home dialysis at 62nd European Renal Association (ERA) Congress 2025.

Covic, vice-rector for scientific research at Grigore T. Popa University of Medicine and Pharmacy, Iași, Romania, emphasized that dialysis has seen limited improvement in mortality over the past 30-40 years. “We haven’t changed much,” he said, “but AI may be the game changer.”

“Home dialysis is a highly standardized process that generates a large volume of medical data,” he explained in his presentation.

Different machine learning models are already used in a cycle of AI-based clinical decision support — a cycle from problem identification through algorithm development and validation to real-world deployment. These models can process data from up to half a million patients and uncover variables that clinicians might overlook.

And AI, he reported, has outperformed traditional models like the Kidney Failure Risk Equation in predicting when dialysis should begin and in forecasting mortality. In a head-to-head study, only 2 of 10 experienced nephrologists outperformed an AI model in predicting survival at 30 and 90 days. And “only one was better than AI at 1 year,” said Covic.

Beyond survival, AI models are also accurately predicting major adverse cardiac event and gastrointestinal bleeding hospitalization risks with hemodialysis.

Diagnostic capabilities are also improving, Covic added. He described how natural language processing applied to clinical notes identified symptoms with far greater sensitivity than International Classification of Diseases, 10th Revision, coding. In simulated clinical cases, the large language model-based AI system AMIE, which has been optimized for diagnostic dialogue, outperformed physicians in diagnostic accuracy. “AMIE was far better than the doctors in picking the right symptoms,” he said.

On the treatment side, AI models are helping clinicians anticipate intradialytic hypotension 15-75 minutes in advance, with an area under the receiver operating characteristic curve (AUC) of 0.89, and manage vascular access by classifying aneurysms and predicting stenosis. In one example, a model achieved 86% classification accuracy. Another AI-enabled device, the DeepVAQ system, used photoplethysmography data to detect stenosis with an AUC of 0.86.

Covic also presented evidence that AI improved anemia control. A model trained on 170,000 clinical records achieved a mean absolute prediction error in hemoglobin concentration of 0.59 g/dL and led to improved hemoglobin levels during implementation.

“We are close, for the very first time, to a meta-analysis for anemia management in dialysis through AI,” he said.

Even device integration is advancing, with AI-enabled tools like phonangiography and robotic tomographic ultrasound improving vascular assessments. You won’t need a technician — the device itself will standardize the analysis, he said.

Despite his optimism, Covic issued a strong caveat: AI models are increasingly opaque, and “we are not going to learn the algorithms,” he said. “This process is close to beliefs in divinity.”

Expanding Access Through Policy

Shifting the focus to access, Rajnish Mehrotra, MD, MS, argued for policy solutions that leverage home dialysis with peritoneal dialysis (PD) to close the global dialysis gap.

“Nearly 2 million people die with kidney failure every year because they have limited access to kidney replacement therapy care,” said Mehrotra, Belding H. Scribner Endowed Chair in and professor of medicine and head of the Division of Nephrology at the University of Washington School of Medicine, Seattle.

PD is ideal for low-resource settings, he said. “Hemodialysis is personnel-intensive. PD is personnel-efficient.” PD also has a markedly lower carbon footprint, less than a quarter that of in-center hemodialysis. “It substantially reduces the transportation burden on patients,” he added.

Mehrotra cited Thailand’s 2008 PD-first mandate, which rapidly expanded access. “People who previously would have died because they did not get dialysis lived,” he said. Even though the policy was later relaxed, PD rates in Thailand remain high.

In the US, financial reform drove a doubling of PD use after 2011. Reimbursement parity and bundled payments neutralized the incentives that had favored hemodialysis. “This public policy has safely grown the use of PD and has been a spectacular success,” said Mehrotra.

Telemedicine in Daily Practice

Finally, Sabrina Milan Manani, MD, Department of Nephrology, Dialysis and Transplant, San Bortolo Hospital, Vicenza, Italy, turned to practical innovations that bring AI into patients’ homes. “Home dialysis is a corner[stone] of patient-centered care in nephrology,” she said.

Her team uses a hybrid model of home visits and telemonitoring for home dialysis. With cloud-connected cyclers, clinicians review overnight dialysis sessions each morning and make remote prescription changes. “It’s a two-way communication system,” she noted.

Manani presented outcomes from her center: Patients with remote monitoring had fewer urgent visits and fewer hospitalizations for disease-specific events. “We observed that the patients felt more confidence with the care team and with the treatment,” she said.

She also described a real-world protocol for integrating telemedicine, with frequent reviews of new patients’ data during the first 15 days of treatment, tapering to weekly reviews for stable patients.

We only invite the patient to come to the hospital if there is a clinical problem, she added.

However, she acknowledged that not all patients benefit equally. Older individuals may struggle with connectivity or tech literacy, and the evidence for complication reduction remains mixed. “The significant results were about disease-specific hospitalization and technical failure, but the evidence was low or very low,” said Manani.

Still, she sees future promise in wearable devices and AI-enhanced alerts. “Flexibility and innovation are essential for promoting home therapies,” she concluded. “Remote monitoring could integrate the traditional follow-up, allowing safe, high-quality care.”

The Road Ahead

Key challenges in the integration of AI into kidney care remain. As Covic put it, “Prediction is currently the most widely used application of AI and machine learning in dialysis,” but rigorous studies are needed. “We should now move to this sort of randomized control trials,” he urged.

The issues of trust and interpretability loom large with AI. “It is really difficult, or even impossible, to know how they produce results,” Covic said of AI models. But the results, increasingly, are hard to ignore. “Optimized machine learning models could enhance risk identification and drive preemptive interventions,” he concluded.

No funding or relevant financial relationships were declared.

Liam Andrew Davenport, MA (Hons), is a UK-based medical journalist and writer with more than 20 years’ experience. He studied medical sciences and anthropology at Emmanuel College, Cambridge, England.