ICU False alarm detection
The main objective of this project was to address the high false alarm rates in the ICU when trying to predict an arrhytmia. The data set comprises monitor signals from ICU with labeled arrhythmia alarms. We used Naive Bayes as a classifier (as was specified) and analyzed very different features to get the best possible metrics. This project was developed on a team to participate on a Kaggle challenge as part of our learning process of Physiological Signal Processing. We ended top 5 at our class out of 34 groups (each of them being composed of 3/4 people). It was modified after the challenge, mainly to embellish it and make it more readable and reusable.
Several features were obtained of the dataset to enhance the performance of our classifier, mainly by statistical methods based on entropy or gaussian modelling. We ended up using Heart Rate Mean and Standard Deviation over the second channel and several heart rate variability statistics over the second and the fifth channels (such as the mean, the standard deviation or the entropy), obtaining a 68 percent of accuracy on the public leaderboard and a 74 percent of accuracy on the private one.
To see how the project was deployed, you can see the following Jupyter Notebook.