AI Model Predicts Patients at Most Risk of Complication During Treatment for Advanced Kidney Failure

Millions of individuals with chronic kidney disease (CKD) undergo hemodialysis, a treatment that circulates their blood through a machine to eliminate toxins. A prevalent complication associated with this procedure is a sudden decrease in blood pressure, referred to as intradialytic hypotension (IDH). IDH is linked to higher rates of mortality and increased hospitalizations, and until now, there has been no reliable method to predict its occurrence. Now, a new artificial intelligence (AI) model can predict which patients are at greater risk of experiencing a drop in blood pressure.

The idea for this model originated from a previous study conducted by the University of Portsmouth (Hampshire, UK). Two years ago, the researchers has developed an algorithm capable of estimating the duration of a patient’s hospital stay upon being diagnosed with bowel cancer. By utilizing AI and data analytics, they could predict the length of hospitalization, the likelihood of readmission after surgery, and the chances of mortality within one or three months. Building upon this work, the researchers have now developed a machine learning tool, having gathered pre-dialysis and real-time data from 10 treatment centers over a span of two decades (2000-2020), involving a total of 3,944 patients. The dataset comprised 73,323 dialysis sessions, during which 36,662 IDH events were recorded.

Image: The machine learning tool prevents people suffering from painful and sometimes fatal low blood pressure during hemodialysis (Photo courtesy of 123RF)

From this information, the researchers identified 33 variables to determine which individuals were most at risk. These variables included observations routinely collected during clinical care, such as weight, temperature, age, blood pressure, medication, and treatment details. Machine learning algorithms were employed to construct a predictor aimed at preventing IDH events. Among the five different algorithms assessed, the Random Forest model exhibited the highest overall predictive accuracy at 75.5%, while the Bidirectional Long Short-Term Memory model achieved the best sensitivity at 78.5%. Additionally, the analysis highlighted that both systolic and diastolic blood pressures are crucial predictor variables. The study also evaluated the algorithm using solely pre-dialysis data inputs to simulate conditions at the beginning of a dialysis session. Although the prediction performance decreased in this scenario, it remained clinically relevant. Future efforts by the researchers will focus on developing a decision-support system for clinicians and conducting a clinical trial.

“This research highlights the value of using machine learning in healthcare, particularly in complex situations like hemodialysis,” said project lead, Dr Shamsul Masum from the University’s School of Electrical and Mechanical Engineering. “Predicting hypotension not only helps clinicians intervene early but also opens the door to personalized care. As we continue to develop and refine these models, the goal is to create a practical decision-support system that could enhance dialysis management, patient safety and quality of care.”