[Paper Review] Predicting Blood Pressure Response to Fluid Bolus Therapy Using Attention-Based Neural Networks for Clinical Interpretability

Uma Girkar, Ryo Uchimido+4 more|arXiv (Cornell University)|Dec 3, 2018

Hemodynamic Monitoring and Therapy15 references9 citations

TL;DR

This study develops an attention-based deep learning model using time-series clinical data from the MIMIC-III database to predict blood pressure response to fluid bolus therapy (FBT) in ICU patients. The stacked LSTM with attention mechanism achieved the highest performance (accuracy: 0.852, AUC: 0.925), enabling both high predictive accuracy and clinical interpretability by identifying key physiological timepoints and features influencing FBT success.

ABSTRACT

Determining whether hypotensive patients in intensive care units (ICUs) should receive fluid bolus therapy (FBT) has been an extremely challenging task for intensive care physicians as the corresponding increase in blood pressure has been hard to predict. Our study utilized regression models and attention-based recurrent neural network (RNN) algorithms and a multi-clinical information system large-scale database to build models that can predict the successful response to FBT among hypotensive patients in ICUs. We investigated both time-aggregated modeling using logistic regression algorithms with regularization and time-series modeling using the long short term memory network (LSTM) and the gated recurrent units network (GRU) with the attention mechanism for clinical interpretability. Among all modeling strategies, the stacked LSTM with the attention mechanism yielded the most predictable model with the highest accuracy of 0.852 and area under the curve (AUC) value of 0.925. The study results may help identify hypotensive patients in ICUs who will have sufficient blood pressure recovery after FBT.

Motivation & Objective

• To improve prediction of blood pressure response to fluid bolus therapy (FBT) in ICU patients, who often do not respond to FBT despite its widespread use.
• To address the clinical challenge of avoiding unnecessary fluid administration by identifying patients likely to benefit from FBT.
• To develop a machine learning model that combines high predictive performance with clinical interpretability for real-world ICU use.
• To investigate whether time-series modeling with recurrent neural networks (RNNs) outperforms time-aggregated models in predicting FBT response.
• To identify clinically meaningful features and critical timepoints influencing FBT success using attention mechanisms.

Proposed method

• Utilized a large-scale ICU database (MIMIC-III) with 17,977 hypotensive patients receiving FBT within 24 hours of ICU admission.
• Extracted 29 clinical features including demographics, comorbidities, vital signs, lab values, and vasopressor doses, both time-static and time-varying.
• Applied regularized logistic regression, multilayer perceptrons, and recurrent neural networks (LSTM and GRU) with and without attention mechanisms.
• Used temporal segmentation into 12, 36, and 72 timesteps to evaluate the impact of time granularity on model performance.
• Employed attention mechanisms to identify the most influential timesteps before FBT for model decision-making.
• Used autoencoders to derive compact, informative feature representations from raw clinical data.

Experimental results

Research questions

• RQ1 Can recurrent neural networks with attention mechanisms improve prediction of blood pressure response to FBT compared to traditional time-aggregated models?
• RQ2 Which clinical features are most predictive of successful FBT response, and are they clinically plausible?
• RQ3 Can attention mechanisms in RNNs identify key timepoints in the clinical trajectory that influence FBT outcome?
• RQ4 Does using distributed representations (autoencoder outputs) improve model performance compared to raw features?
• RQ5 What is the optimal temporal granularity for modeling FBT response using time-series data?

Key findings

• The stacked LSTM with attention mechanism achieved the highest performance with an accuracy of 0.852 and an area under the ROC curve (AUC) of 0.925.
• The model using 72 timesteps (approximately 20-minute intervals) outperformed coarser granularities, indicating that fine-grained temporal modeling enhances prediction.
• The LASSO-regularized logistic regression identified respiratory rate, diastolic blood pressure, temperature, bicarbonate, and base excess as top predictive features, which are underappreciated in prior clinical studies.
• Attention mechanisms highlighted timepoints closest to FBT administration as most influential, aligning with clinical intuition and enhancing model interpretability.
• The use of autoencoder-derived feature representations yielded comparable performance to raw features, suggesting utility in high-dimensional clinical data settings.
• The GRU and LSTM models with attention outperformed their non-attention counterparts, demonstrating the value of attention in identifying salient temporal patterns.