PhysioMiner : a scalable cloud based framework for physiological waveform mining

Abstract

This work presents PhysioMiner, a large scale machine learning and analytics framework for physiological waveform mining. It is a scalable and flexible solution for researchers and practitioners to build predictive models from physiological time series data. It allows users to specify arbitrary features and conditions to train the model, computing everything in parallel in the cloud. PhysioMiner is tested on a large dataset of electrocardiography (ECG) from 6000 patients in the MIMIC database. Signals are cleaned and processed, and features are extracted per period. A total of 1.2 billion heart beats were processed and 26 billion features were extracted resulting in half a terabyte database. These features were aggregated for windows corresponding to patient events. These aggregated features were fed into DELPHI, a multi algorithm multi parameter cloud based system to build a predictive model. An area under the curve of 0.693 was achieved for an acute hypotensive event prediction from the ECG waveform alone. The results demonstrate the scalability and flexibility of PhysioMiner on real world data. PhysioMiner will be an important tool for researchers to spend less time building systems, and more time building predictive models.