Understanding deep learning models for Length of Stay prediction on critically ill patients through latent space visualization

Abstract

Background and Objective:

Continuous, real-time monitoring of Length of Stay (LoS) for critically ill patients in Intensive Care Units (ICUs) is essential for anticipating patient needs, reduce the risk of adverse events, optimize resource allocation, plan incoming patients, and improve overall care. While previous research has focused primarily on predicting LoS, less attention has been given to how these prediction systems can be used by non-machine learning experts in real hospital environments for capacity planning.

Methods:

In this work, we predict remaining ICU LoS using data from the Amsterdam University Medical Center dataset, which is the first freely available European critical care dataset and shows higher variability in LoS compared to U.S. datasets. We applied state-of-the-art sequence-to-sequence deep learning models – Long Short-Term Memory Networks, Gated Recurrent Units, Temporal Convolutional Networks with and without attention, and Transformer models – on 271 input features extracted from 20,481 ICU stays. Additionally, the latent spaces of these models were extracted, projected onto a 2D space and explored interactively to intuitively understand how the models learn patterns from the clinical data over time.

Results:

The TCN model with attention (TCN-att) returned the best performance, reducing the prediction error by 2.24 days from a MAE of 6.94 days to a MAE of 4.70 days. Latent space analysis revealed that with just 5–6 h of patient data, the models could clearly differentiate between short ($0<\text{LoS}\le 3$ days) and long ($7\le \text{LoS}<+\infty$ days) ICU stays, identify a patient cluster at risk of in-hospital mortality, and more.

Conclusion:

Beyond reporting prediction error metrics, this study shows how an interactive dashboard can be used to gain insights into the inner workings of complex algorithms and how these can be integrated into clinical decision support systems. This approach allows for real-time intuitive understanding of how models learn and how patient predictions evolve, facilitating their use in real hospital environments.