A GRU network: Analyzing ER data

Research & experiment

Medical data analysis using a RNN

While they may dominate the news cycle, the multihead transformer models that are at the heart of Large Language Models are not the only leading edge neural network architectures in existence.

Recurrent Neural Networks (RNNs) represent a conceptually different class of neural networks, specifically suited to analyze, and predict, time series data.

In our recent study, we made use of just such a neural network in an attempt to discover hidden correlations in medical data collected in hospital emergency rooms in the United States. Our work was specifically motivated by the possibility that the results of life-critical but intrusive measurements, such as blood pressure, may be reliably predicted by other, readily available, but less intrusively obtained signals.

While our study was preliminary, the results seemed promising. They also demonstrated the value of small neural networks: While large networks may take considerable time (days, weeks, or more) to train and powerful hardware, small networks can often be trained in a matter of minutes. This can prove especially useful if the goal is to construct a bespoke, patient-specific model while a patient is being treated for a medical condition.

Gated recurrent units

In our study, we made use of a large database of emergency room visits: version IV of MIMIC, the Medical Information Mart for Intensive Care, a deintified database of tens of thousands of emergency room patients. Specifically, we studied the MIMIC waveform database and the extracted high-cadence data sets that include blood pressure, electro-cardiogram, and other vital sign measurements.

These data sets are in the form of parallel time series. Our goal was not, however, temporal prediction. We did not intend to predict the future behavior of the signal from present data. Rather, our goal was to study if the present-behavior of one data series can be predicted from the present behavior of another, parallel measurement. For instance, is it possible to predict reliably a patient's blood pressure from their heart rate, respiratory rate, or cardiogram values?

To this end, we utilized a specific type of a neural network, a Gated Recurrent Unit (GRU) solution. GRU networks are especially well adapted to model time series data, but they can also be used to analyze the cross correlation between parallel time series efficiently.

Our interactive solution was implemented in browser-based JavaScript, integrating patient and parameter selection with the actual training (backpropagation) of the neural network. Our goal was to address a simple conceptual question: Is it possible to train a GRU network using the early portion of a patient's stay, so that later on, the trained network can reliably indicate a sudden, dangerous change in the patient's condition?

Our implementation of a GRU network has been promising: at least in a few select cases, a properly trained model indeed accurately indicated a rapid, life-threatening drop in the patient's blood pressure on the basis of other available data. Though these are very preliminary results, they nonetheless indicate that a solution of this type may, one day, serve as a "sentinel", watching critical care patients in real time and alerting medical staff in a timely manner if intervention is required.

An rPPG demo... »

© 2023-2025, Viktor T. Toth., All Rights Reserved.