Using machine learning to discover correlations in MIMIC high-cadence data sets

Using the MIMIC-IV clinical database and, specifically, the related preliminary MIMIC-IV waveform data release, we endeavored to create a machine learning tool in the form of a gated recurrent unit (GRU) neural network. Our goal was to study the possibility that intrusive measurements of vital signs, such as blood pressure, might be reliably predicted from parallel time series data sets of non-intrusive measurements of other vital signs, such as heart and respiratory rates. Relying on non-intrusive measurements to alert health care professionals of potentially life-threatening conditions in a timely manner can be life-saving in certain situations. To this end we developed and implemented a custom GRU solution in the form of software that can be run in a Web browser to analyze the high cadence numerical data sets of the MIMIC-IV waveform release. Despite the limitations of this data set, such as the small number of records, inconsistencies in what has been recorded, and often, noisy or incomplete data, we were able to obtain potentially promising results.