Machine Learning and Artificial Intelligence Algorithms to Optimize the Performance and Delivery of Acute Dialysis

Study Background \& Rationale:

Acute dialysis is required in approximately 10-15% of all patients admitted to intensive care units (ICUs). Acute dialysis may take the form of continuous renal replacement therapy (CRRT), intermittent hemodialysis (IHD) or slow-low efficiency dialysis (SLED). Worldwide, CRRT remains the predominant form of acute dialysis, with over 75% of acute dialysis being CRRT.

The application of acute dialysis in the ICU is associated with poor patient outcomes. Despite advances in medical technology and care, mortality remains between 40-60%, which is similar to outcomes observed with severe acute respiratory distress syndrome (ARDS). Additionally, even when patients survive their critical illness, up to 10% of patients require ongoing long-term chronic dialysis therapy. This has a significant effect on the quality of life of survivors of critical illness, as well as important effects on their families, often requiring changes in work and housing status, as well as relocation to sites in closer proximity to dialysis centres. This results in not only significant healthcare and social costs (approximately $100,000/patient/year in Alberta, Canada), but also very important reductions in the health-related quality of lives for these patients. Currently, while evidence exist regarding the optimal initiation of acute dialysis, there is a paucity of evidence to predict timing of modality transition or liberation from this therapy. Using a completely integrated electronic Clinical Information System (eCIS) such as Connect Care (EPIC, Wisconsin, USA) in Alberta, the investigators can develop predictive algorithms that may anticipate patient and acute dialysis needs.

Once, acute dialysis is initiated, several factors may affect kidney recovery following acute dialysis. Intra-dialytic hypotension has been identified as a leading modifiable factor, but unfortunately, one where clinicians have limited capacity to accurately predict. This is an important knowledge gap that must be addressed. It has also been previously identified by our study team as one of the most important key performance indicators (KPI) for acute dialysis, especially in IHD and SLED. For CRRT, filter life has been identified as the most important and studied KPI. Both of these KPIs are currently being utilized by the ongoing QUALITY CRRT and DIALYZING WISELY programs to improve the performance and delivery of acute dialysis to critically ill patients. These two programs have been successfully implemented across Alberta and have established the infrastructure necessary to initiate the next steps in these Continuous Quality Initiatives for acute dialysis, the SMART DIALYSIS program.

Recently, advances in computer and machine processing have led to the 4th industrial revolution featuring the development of smart machines, devices, and learning algorithms that can aid humans in the management of patients and optimize the delivery of healthcare. Machine Learning (ML) and Artificial Intelligence (AI) algorithms have been previously used in medicine, but have only begun their implementation into critical care nephrology.

Current initiatives are primarily focused on pattern recognition and risk prediction. This program will contain 4 distinct phases.

In Phase 1, the investigators will continue work from our QUALITY CRRT and DIALYZING WISELY program and will aim to better understand the landscape surrounding decisions around transitions between acute dialysis modalities, termination, and attempts at liberation from acute dialysis, the incidence of intra-dialytic hypotension, and timing of filter clotting in the ICU.

In Phase 2, the investigators will develop models and subsequent Clinical Decision Support (CDS) tools and/or Best Practice Advisories (BPAs) for clinicians to better predict and manage 1) transitions between acute dialysis therapies, 2) management of intra-dialytic hypotension, 3) prediction of filter life and 4) liberation from acute dialysis. Concurrent with this work, we will work to develop an AI/ML Augmented Acute Dialysis Dashboard (i.e., SMART DIALYSIS Digital Dashboard) embedded within our electronic Clinical Information System (eCIS) to present these KPIs to clinicians.

Phase 3 will look into implementing and evaluating the performance and acceptability of these ML/AI algorithms in clinical practice.

Finally, Phase 4 will take our Alberta-derived results and look to implement and benchmark these across other large healthcare authorities to ensure that the algorithms have been developed and validated appropriately and ethically. These will include partners across Canada, the US, Europe and Australia and New Zealand.