Written by Clay Smith
This machine learning system accurately predicted which patients would develop acute kidney injury or need dialysis. And it did so 41 hours before the patients actually developed it! Imagine the ramifications not just for AKI but for predicting rapid response, need for ICU care, and more.
Why does this matter?
As computers get smarter, they will be better able to predict outcomes based on machine learning and artificial intelligence (AI). This is one example, and we will cover another one tomorrow. This will impact staffing models, resource utilization, and clinical care both now and even more in the future.
How did the computer know I was going to order nachos fundidos?
Using demographics, vital signs, vital sign trends, laboratory values, laboratory value trends, interventions, medications, transfusions, nurse documentation, diagnostics, location, length of stay, prior cardiac arrest, urinary bag, and LVAD status, a machine learning system was able to predict acute kidney injury (AKI) and need for dialysis in 48 hours with very high accuracy (AUC 0.9). It analyzed over 121,000 inpatients using 60% for derivation and 40% for validation. It predicted stage 2 AKI 41 hours before it actually occurred. The clinical implications are clear. This computer can tell with a high degree of accuracy which patients will develop AKI almost 2 days before it actually happens, which would allow us to take corrective, protective, and preventive measures to ward this off. All the variables are readily available in the electronic health record. This is applicable to the ED, ward, and ICU settings. Predicting AKI is the tip of the iceberg. Soon, using readily available data in the chart, AI will be able to predict multiple outcomes while the patient is still waiting in the ED for a bed. Imagine knowing which patients will decompensate and need a rapid response or ICU-level care hours or days before it happens.
The Development of a Machine Learning Inpatient Acute Kidney Injury Prediction Model. Crit Care Med. 2018 Mar 28. doi: 10.1097/CCM.0000000000003123. [Epub ahead of print]
Peer reviewed by Thomas Davis