Authors:
Giovanni Rosa
1
;
Marco Russodivito
1
;
Gennaro Laudato
1
;
Angela Rita Colavita
2
;
Simone Scalabrino
1
and
Rocco Oliveto
1
Affiliations:
1
STAKE Lab, University of Molise, Pesche (IS), Italy
;
2
ASREM – Regione Molise, Italy
Keyword(s):
ST Anomalies, Temporal Window, Decision Support System, Machine Learning.
Abstract:
Nowadays, Computerized Decision Support Systems (CDSS) play an important role in medical support and preventative care. In those scenarios, the monitoring of biomedical data, such as the ECG signal, is fundamental. The ECG signal may reveal a variety of abnormalities or pathological conditions. Some examples are Ischemia and Myocardial Infarction (MI), with a significant impact on the world’s population. Both these conditions can be diagnosed by observing changes in specific sections of the ECG, such as the ST segment and/or T-wave of heartbeats. Much effort was devoted by the scientific community to aim at automatically identifying ST anomalies. The main drawback of such approaches is often a trade-off between the accuracy in the classification, the robustness to noise, and the real-time responsiveness. In this work, we present RAST, a robust approach for a Real-time Accurate screening of ST segment anomalies. RAST takes as input a sequence of 10 successive heartbeats extracted from
an ECG recording and provides as output the classification of the ST segment trend. We evaluated two versions of RAST, namely RAST-BINARY, and RAST-TERNARY: the first capable of distinguishing only between an ST anomaly and Normal Sinus Rhythm and the second able to distinguishing between ST elevation, ST depression, and normal rhythm. Moreover, we conducted an extensive study by experiment also (i) the validation within the intra- and inter-patient strategies and (ii) the ideal number of successive heartbeats in which to observe an anomalous episode of change in the ST segment. As a result, both RAST-BINARY and RAST-TERNARY can achieve an F1 score of 0.94 with a window of 4 heartbeats in the inter-patient validation. For the intra-patient validation, both versions achieve an F1 score of 0.73 using a longer observation window.
(More)