A clinically applicable approach to continuous prediction of future acute kidney injury - Tomašev et al., Nature, July 2019

Original article: https://www.nature.com/articles/s41586-019-1390-1

One-sentence summary

In this paper, a research group from Deep Mind trained a recurrent neural network (RNN) to detect acute kidney injury (AKI) up to 48 hours in advance of its onset, at a significantly higher accuracy than previous models.

What did they do?

They identified key features for predicting AKI using both clinician input as well as features identified by the deep learning model. Using these features, they trained an RNN which continually outputs the probability of an AKI occurring in the next 48 hours, with a sensitivity of 55.8%.

What does this mean?

As well as supporting the prediction of AKI with a reasonable time window for clinical intervention, the algorithm helps us to better understanding the important predictive factors for AKIs.

This algorithm is the best performing AKI prediction algorithm to-date. However, it still has a relatively high false positive rate (with a ratio of 2 false alerts: 1 true alert).

Future work may look to increase the positive predictive value and to extend this time window. The algorithm will also require prospective validation.

Figures from original paper

Example of successful AKI prediction
Structure of the RNN algorithm developed