Machine Learning Based Classification of Myocardial Infarction Conditions Using Smartphone-derived Seismo- and Gyrocardiography

Saeed Mehrang1, Mojtaba Jafari Tadi2, Olli Lahdenoja2, Matti Kaisti2, Tuija Vasankari3, Tuomas Kiviniemi3, Juhani Airaksinen3, Mikko Pankaala2, Tero Koivisto2
1Dep. of Future Tech., University of Turku, 2University of Turku, 3Turku Heart Center


Abstract

Acute myocardial infarction (AMI) is a life-threatening heart condition caused by sudden rupture of atherosclerotic plaque inside one or more of the coronary arteries. AMI must be treated shortly after its emergence due to the potential shortage of blood flow to heart muscle. Untreated AMI may result in permanent damage to cardiac muscle cells within hours. Goals: This paper aims at presenting an AMI detection framework that works based on seismocardiogram (SCG) and gyrocardiogram (GCG) signals recorded by built-in inertial sensors of a smartphone. Advanced machine learning and signal processing methods are deployed to detect indications of AMI. Methods: From 22 AMI patients, before and after receiving percutaneous coronary intervention (PCI), tri-axial SCG and GCG signals were recorded by a smartphone. The signals were bandpass-filtered [4-50Hz], cleaned from artefacts using median absolute deviation of Shannon entropy, and subsequently features were extracted. A variety of features (151 in total) from the time domain, Fourier domain, and wavelet transform were extracted. The formed feature matrix was subsequently fed into two different classifiers namely random forest (RF) and support vector machines (SVM). Leave-one-out cross-validation was used to assess the performance of the method. Results: Results show that AMI condition can be discriminated from the treated heart (after PCI) with accuracy rates of 0.80 and 0.82 for RF and SVM, respectively. Sensitivity and specificity rates were 0.73 and 0.86 for RF as well as 0.77 and 0.86 for SVM. Conclusion: The presented approach can be deployed as a remote monitoring tool for the detection of AMI. Currently only the sensor logging is implemented in an app while the data analysis algorithms operate offline. In the future, the presented data analysis framework could be integrated in the app and served as a simple tool for the detection of AMI.