Analysis of ECG Data by Energy Efficient Decision Trees on a Reconfigurable ASIC
The goal of the Pilot Innovation Initiative “KI-Sprung” organized by the Federal Ministry of Education and Research in Germany was to develop an energy efficient algorithm that could detect atrial fibrillation episodes in ECG data. For this project the Charité in Germany collected 16000 labelled ECG datasets, with a 50/50 split between records of healthy persons and records, where atrial fibrillation episodes were recorded. The overall goal of this pilot project was to achieve a recall of 90% and a fallout of 20% with minimized energy cost.
To achieve this task, we utilized an approach based on classical machine learning: The decision tree ensemble. We developed an innovative architecture for these types of ensembles to improve classification performance. Further we evaluated a set of features calculated over the ECG data and optimized the set depending on the quality of the features and the cost for the feature calculations. We realized this by estimating the cost of the feature calculation in hardware and used this information to optimize the energy usage and also the classification accuracy of the classification algorithm.
We achieved a recall of 92.7% and a fall-out of 14.7% with a maximum energy consumption of 42nJ for a 2-minute ECG dataset.