University of Cambridge > > Theory - Chemistry Research Interest Group > Machine Learning for Critical Care: Using Neural Networks to Predict Patient Mortality

Machine Learning for Critical Care: Using Neural Networks to Predict Patient Mortality

Add to your list(s) Download to your calendar using vCal

If you have a question about this talk, please contact Lisa Masters.

First Year PhD Report

Intensive care units are constantly under strain, with many vulnerable people needing urgent critical care. Often doctors do not have the capacity to accommodate everyone, and so it is important that patients are discharged when it is safe to do so. However, discharging a patient at the wrong time may result in the patient being readmitted or simply not surviving; both of these cases are highly undesirable. Therefore, it is useful for there to be an understanding of what factors contribute most to the mortality rate of patients in intensive care units.

In the present work, machine learning models are trained to predict whether a given patient with certain measurements will live or die. Neural networks with one hidden layer and two outcomes (alive or deceased) are used in these models, and local minima of the neural network loss functions are found using basin-hopping methods implemented with the open source software GMIN . Area Under the Curve (AUC) values were used to evaluate these models, and AUC values above 0.8 were found for models involving Glasgow Coma Scale (GCS) scores and Blood Urea Nitrogen (BUN) measurements. The effect of using a model trained on one time window and evaluated on different time windows is also investigated, and we find that the AUC value decreases but not substantially. Finally, the effect of label noise is investigated, and it is found that neural networks used on noisy iris flower data are robust, giving AUC values close to 1, even for moderately high proportions of mislabelled training data.

The project is heading in a direction focusing on the thermodynamic properties of the energy landscapes defined by neural network loss functions. By identifying the abstract analogue of ``phase transitions” for this landscape, using some fictitious temperature, it is hoped that complementary neural network solutions on the landscape may be found and combined to yield overall better solutions.

This talk is part of the Theory - Chemistry Research Interest Group series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2024, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity