Improving Prediction of Need for Mechanical Ventilation using Cross-Attention

TL;DR

This paper introduces a deep learning model with multi-head attention to improve the accuracy and reduce false positives in predicting the need for mechanical ventilation in ICU patients, demonstrating significant performance gains on a public dataset.

Contribution

The paper presents a novel FFNN-MHA model applying multi-head attention for personalized, accurate MV prediction, outperforming baseline models on the MIMIC-IV dataset.

Findings

0.0379 increase in AUC over baseline models

17.8% reduction in false positives

Effective for critical care prediction tasks

Abstract

In the intensive care unit, the capability to predict the need for mechanical ventilation (MV) facilitates more timely interventions to improve patient outcomes. Recent works have demonstrated good performance in this task utilizing machine learning models. This paper explores the novel application of a deep learning model with multi-head attention (FFNN-MHA) to make more accurate MV predictions and reduce false positives by learning personalized contextual information of individual patients. Utilizing the publicly available MIMIC-IV dataset, FFNN-MHA demonstrates an improvement of 0.0379 in AUC and a 17.8\% decrease in false positives compared to baseline models such as feed-forward neural networks. Our results highlight the potential of the FFNN-MHA model as an effective tool for accurate prediction of the need for mechanical ventilation in critical care settings.