Enhancing Dialysis Care using AI: Predicting Intradialytic Hypotension

A collaborative effort between the University of Portsmouth (UK) and Portsmouth Hospitals University NHS Trust (PHUT) (UK) has led to the development of a machine learning tool that predicts which patients are most at risk of their blood pressure dropping during dialysis; a condition known as intradialytic hypotension (IDH). This innovation addresses a critical gap in clinical care for chronic kidney disease (CKD) patients.

The idea for this new machine learning tool was inspired by a previous study led by the University and the Trust. In 2022, the researchers developed an algorithm that anticipated the length of hospital stay for patients who were diagnosed with bowel cancer. They found that with the algorithm, they were able to accurately predict this, as well as the rate of readmittance after surgery, and mortality rate over a one or three-month period. 

The Need for Innovation in Dialysis Care

CKD affects over three million people in the UK, with 31,000 relying on hemodialysis to manage their condition. Despite recent advancements in dialysis, end-stage kidney failure patients face persistent challenges, including a significant risk of mortality. Among these challenges, IDH remains one of the most common and severe complications, characterized by a sudden and potentially fatal drop in blood pressure during dialysis.

IDH is linked to higher rates of mortality and increased hospitalization. Up until now, predicting IDH has posed a significant challenge for clinicians. Traditional assessment methods rely on subjective clinical judgment, which can be time-consuming and inconsistent—not all patients can receive this level of care.

However, a new machine learning tool promises to revolutionize this process by providing a reliable, data-driven approach to identifying at-risk patients before complications arise. Identifying at-risk patients not only helps clinicians intervene early but also helps them to decide on a treatment pathway far quicker and with more accuracy. 

“This model offers great promise that could pave the way to a future where AI/ML can be used to personalize treatments for individuals on dialysis and significantly reduce the risk of IDH and other complications.”  

Dr Nicholas Sangala, Consultant Nephrologist, Wessex Kidney Centre, Portsmouth. 

How the AI Tool Works

Using only routinely measured clinical observations, the researchers identified key variables and developed a machine learning tool to be used as an IDH predictor.  The AI model was trained using data from 73,323 dialysis sessions with 36,662 IDH events, gathered from 10 treatment centres and 3,944 patients. Researchers analyzed pre-dialysis and real-time observations, focusing on variables such as:

• Patient weight before and after dialysis
• Blood pressure before and after dialysis
• The amount of fluid removed during dialysis
• The total volume of blood processed

Using five machine learning algorithms, including Random Forest (RF) and Bidirectional Long Short-Term Memory (Bi-LSTM), the team identified 33 critical variables that influence IDH risk. Both systolic and diastolic blood pressures were highlighted as key predictors.  

Among the five different IDH prediction models assessed, the RF algorithm demonstrated the highest overall predictive accuracy at 75.5%, while the Bidirectional Long Short-Term Memory model excelled in sensitivity, achieving 78.5%. Notably, when the IDH prediction model relied solely on pre-dialysis data—a scenario mimicking current clinical practice—its accuracy slightly decreased, however, it still retained clinical significance.

Impact on Clinical Practice

The integration of this tool into dialysis workflows could drastically improve patient care by:

1. Early risk detection: By identifying high-risk patients at the beginning of a dialysis session, clinicians can proactively adjust treatment plans to mitigate complications.
2. Personalized treatment: The AI model could potentially allow for tailored interventions, optimizing fluid removal rates and blood pressure management based on individual risk profiles.
3. Real-time monitoring: Future iterations of this tool aim to incorporate real-time data inputs, enhancing its predictive capabilities during ongoing dialysis sessions.
4. Resource Optimization: By reducing IDH incidence, the tool could decrease hospital readmissions and associated healthcare costs, easing the burden on healthcare systems. This is especially useful for healthcare systems like the National Health Service (NHS) in the UK, where there are many patients waiting in line for treatment.

Challenges & Future Directions  

While the study demonstrates significant potential, several limitations and practical challenges remain. Seamless adoption of the tool requires compatibility with existing electronic health record systems and dialysis machines, thus integration with existing clinical workflow might be difficult. This is especially true for hospitals that might not have as much funding, as an upgrade to a more current computational system might be needed.

Moreover, rigorous testing in clinical settings is essential to validate the model’s performance and assess its impact on patient outcomes. Training data for this IDH predictor also needs to be ensured that it is representative of CKD patients worldwide, ensuring that all races and genders are included. Clinical trial outcomes would give hospitals information to decide whether adopting this tool, which would be costly, is worth it.

Furthermore, the study compared predictive machine learning models that only use pre-dialysis data with those utilizing both pre- and post-dialysis data, revealing that the latter tends to be more reliable. Thus, it is not an entirely fair comparison.  

Conclusion

The AI-based IDH predictor represents a major step forward in nephrology. By transforming complex datasets into actionable insights, it empowers clinicians to deliver safer and more effective care. The development of this tool highlights the transformative role of AI in healthcare—not as a replacement for clinicians, but rather as an invaluable ally in tackling intricate medical challenges that were previously lacking solutions.