Session S31.5

Detection of Myocardial Ischemia with Hidden Semi Markovian Models

J Dumont*, G Carrault, P Gomis, GS Wagner, AI Hernandez

Université de Rennes
Rennes, France

Myocardial ischemia is still a major public health problem. A reliable detection of acute or transient ischemia episodes may contribute to the reduction in the delay between the appearance of ischemia and the application of adapted reperfusion procedures; and to improve the risk-stratification of the patients with respect to a future acute ischemic stroke. Among electrocardiographic data, ST segment analysis remains the most common method used in monitoring systems. It is, however, admitted that to overcome its low specificity, other indicators have to be jointly analyzed. Although ischemia is a dynamic process, the final decision (ischemia is absent/present) often relies on static measurements and the comparison to a predefined threshold. To improve this detection, it appears important to take into consideration the temporal evolution of the retained indicators. In this work, a methodology to characterize the joint evolution of a number of clinical parameters affected by ischemia is proposed and evaluated. It is based upon Hidden Semi Markovian Models (HSMM). Unlike standard HMM, each state of the model is constituted by a multivariate Normal Probability Distribution (NPD) to characterize a part of the feature space. Also, the time spent in each state is modeled by a NPD. The Viterbi algorithm has been adapted to the HSMM case to train the models. The performances of this new methodology were evaluated on the STAFF3 database (annotated ischemic episodes induced by balloon inflation). Two models, an ischemic and a reference model, were defined. The detection of an ischemic event is decided when the likelihood of the ischemic model largely exceeds the likelihood of the reference model. Several models have been tested according to the variables introduced. It appears that the best compromise between sensitivity (SE) and specificity (SP) is achieved by considering only the RT interval (SE = 89.1%; SP = 87.5%). When the amplitude of the R wave is introduced, these performances are observed (SE = 94.7; SP = 71.9%). The classical method using only the ST segment provides lower performance values (SE = 79.1; SP = 56.1%). Thus, the HSMM technique seems to be a potential method for non-invasive detection of myocardial ischemia.

(Abstract Control Number: 21)