A Focus on the ICU’s Mortality Prediction Using a CNN-LSTM Model

1Sakshi Hooda, Suman Mann


In healthcare, the ability to predict mortality accurately is an important aspect because it allows for empirical risk estimations, allowing further for prognostic hospital benchmarking, patient stratification, and decision-making. Currently, most of the techniques for prediction come in the form of scoring systems seeking to determine disease severity. Also, the models involve certain, rigid summarized physiological data and admission attributes. The implication is that there tends to be some degree of biasness in the systems, especially because they involve manual effort that also demands regular updating to address shortcomings that are reported frequently. Therefore, deep learning algorithms have evolved in a quest to allow for the automatic extraction of features, as well as their selection independent of human intervention. In this study, an approach combining notes for mortality prediction and a deep learning algorithm was proposed. The role of the CNN-LSTM entailed the mapping of various notes in relation to possibilities of mortality outcomes. In the custom architecture, there is a combination of recurrent and convolutional layers with previous capturing semantic associations in certain notes on an independent basis. Therefore, this experimental study strived to determine the performance of the CNN-LSTM approach towards mortality prediction via the use of notes for the initial 48, 24, and 12 hours of a given patient’s hospital stay. In the findings, the study established that when CNN-LSTM is implemented, it exhibits superior performance when compared to the baseline. Thus, there was evidence of a proof-of-concept in relation to the efficacy of combining deep learning and notes towards improvements in outcome prediction.


ICU’s, Mortality Prediction, CNN-LSTM

Paper Details
IssueIssue 6