Master thesis presentation Adam Söderholm, MPCAS

Abstract: Machine learning models interpreting ECG signals enable scalability to increase the number of examined patients and hence democratize healthcare. This thesis develops and evaluates a machine learning model for detecting various types of heartbeat arrhythmias. The initial aim was to interpret signals from mobile ECG devices but due to lack of data, the model was developed and evaluated on the MIT-BIH Arrhythmia Database. The evaluation of the model followed an inter-patient evaluation scheme to mimic the performance of the model in a real-world clinical setting. The evaluation was performed following requirements to ensure that regulations for medical devices agreed with the result obtained. The final model design included ensemble learning between a convolutional neural network and a multilayer perceptron network. The developed model was able to detect non-ectopic, supraventricular ectopic and ventricular ectopic heartbeats with a recall of 97 %, 65 % and 82 %, respectively. 

Password: 437889

Supervisor: Gustav Karlsson
Examiner: Mats Granath
Opponent: Viking Zandhoff