“Everybody has certain red flags they want to keep an eye on, and big data tools enable you to do that in fairly close to real time.”
Brian Loughman, Financial Reporting Investigations Team
This report from the Journal of Medical Internet Research focuses on the AI methodology of recurrent neural network in deploying a continuous prediction model for in-hospital acute kidney injury (AKI) as part of clinical decision support. This paper is an example of more diverse AI in clinical medicine projects that we are observing especially in the past few years after a period during which we saw predominantly convolutional neural network (CNN) in medical imaging in the realms of radiology, dermatology, pathology, and ophthalmology (and now cardiology and gastroenterology).
The study involved close to 70,000 adult patients who were admitted to two different tertiary hospitals in Seoul. The investigators developed 2-stage hierarchical prediction models using RNN algorithms for occurrence of AKI within 7 days as well as trajectory of creatinine values up to 72 hours; both of these models were evaluated using internal and external validation data sets.
The authors added that explainability was brought forth by model-agnostic interpretation methods such as SHapley Additive exPlanations (SHAP), partial dependence plots, individual conditional expectations, and accumulated local effects plots (but did not use Local Interpretable Model-agnostic Explanations, or LIME).
The authors further explained that cases with missing data as well as patients with an estimated glomerular filtration rate less than 15 mL/min/1.73 m2 or end stage kidney disease were excluded. The group showed that a continuous AKI prediction model using RNN algorithms was able to provide real-time assessment of future AKI occurrences as well as individualized risk factors for these inpatients.
While this prediction model using RNN and time series data outperformed an XGBoost-based model, deep reinforcement learning may be the ultimate AI tool to help the clinician in the ICU, OR, or ED setting for real-time decision support. This methodology, however, will require many samples of the scenarios which may be very difficult to collect for this tool to learn from. As in CNN and RNN models, perhaps a generative AI methodology will need to help with this DRL methodology.
Read the full report here