InfarctWatch: A Mobile Acute Myocardial Infarction Detection Service Using Commercial Smartwatch Electrocardiogram
16.13 MB
avoin
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
Pysyvä osoite
Verkkojulkaisu
DOI
Tiivistelmä
Many acute myocardial infarction patients do not seek medical attention early enough after they start to feel symptoms. Significant percentage of patients seek medical attention many hours after the infarction symptoms, worsening their prognosis. The reasons for the delay are complex in nature, but studies have found older age, first-time MI, lower income, symptom onset at night, underestimating symptom severity, fear of bothering others in a false alarm, and lower education level correlates with longer delay. This study implemented a system called InfarctWatch that enables patients to evaluate their commercial smartwatch ECG report for signs of myocardial infarction. Smartwatch is a common non-invasive piece of clothing accessory and is normal to wear in everyday life. This is why InfarctWatch works with smartwatches, to offer an inconspicuous service. InfarctWatch app will send the ECG report to server-hosted software called InfarctAPI. It extracts the ECG signal from the report with image processing techniques, preprocesses the extracted signal, and finally calculates features. These features are used to predict if the ECG contains signs of myocardial infarction with a trained machine learning model. InfarctWatch app displays the result and encourages patient to make immediate actions when needed. The model is trained using Lead I ECG signal from STAFF III database, consisting ECGs from 104 patients that are recorded before, after and during a balloon inflation procedure. This procedure effectively simulates ischemia and infarction, making this database essential for training MI detecting models. InfarctWatch is expected to function with any smartwatch model allowing its user to download the ECG report. For signal extraction, the drawn signal is expected to locate inside a grid. No integration development is done to any smartwatch, which requires InfarctWatch app to provide a robust, but in the end not very user-friendly solution for the user to upload their ECG file. Random forest model performed best with the test split resulting 0.912 accuracy, 0.947 precision, 0.9 recall, and 0.923 F1-score. The small scaled testing with 4 different smartwatches and 12 subjects resulted many false positives, due to how different in nature the training and extracted smartwatch ECG input data actually is. InfarctWatch did succeed well in extracting and processing the signal for ML prediction from the reports with a distinct visual design. The concept this mobile AMI detection system prototype demonstrates is considered impactful in making infarction patients seek medical attention earlier.