This repository contains supplementary materials from a research project for determining whether an ECG arrhythmia alarm in the ICM is a true alarm or a false alarm. This is done by classifying the ECG segment immediately following an alarm into either that of a true alarm or a false alarm, as accurately as possible and as early as possible. This research resulted in an advancement of the state of the art, mostly resulting from 2015 PhysioNet/CinC Challenge (https://www.physionet.org/content/challenge-2015/1.0.0/).
Contents in this repository:
- Deep learning model: ResNet + BiLSTM. Model architecture (figure below) and source code python (link)
- Prequential evaluation: Growing window version. Source code python (link)
- Data preparation: WFDB ECG Segment splitting and data preparation. Source code python (link)
- Data sets: 2015 PhysioNet Challenge data sets (download)
The result datasets are those used to generate the figures and tables in the paper. (The corresponding figure/table numbers are specified for each dataset.)
- Classification time for varying interval: one row for each ECG segment; one column for each batch-interval (4 msec, 0.5 sec, 1 sec, 2 sec) (download) (see Figure 3 in text).
- Model's output probability over time: one row for each ECG segment; one row for each sample within the sample interval (4 msec) (download true alarm) (download false alarm)(see Figure 4 in text).
- Classification times (with the 4 msec interval) for all ECG segments, comparing approaches. (download boxplots) (see Figure 5 in text)
- Classification accuracy among the three model structures (ResNet+BiLSTM, ResNet only, BiLSTM only) (download table) (see Table 1 in text).
