M. I. Bogachev

The paper discussed two general approaches to estimate the probability that a physiological process consisting of consecutive heartbeat intervals exceeds a certain threshold. The first approach is based on searching for characteristic precursors of such events during the learning procedure, and later giving an alarm whenever a typical precursor appears during the online testing procedure. The learning procedure can be performed either using the available realizations of relevant physiological records, or on the mathematical model representing this process. As a learning model, both available realizations of heartbeat intervals and a well-known multiplicative random cascade model with 1/f power spectrum recently suggested for representation of physiological rhythms are used. This first approach is usually called the precursory pattern recognition technique and exploits solely short-term memory. The second approach is based on the return interval statistics between large events and thus is able to exploit long-term memory. The paper shows that the second approach yields at least of the same efficiency being tested on the simulated records, obtained using the multiplicative random cascade model. In the real 24-hour heartbeat interval records, the second approach is clearly superior to the first approach. It is shown that the same result can be obtained in the simulated records in the presence of measurement noise, due to the better noise robustness of the second approach. Another advantage of the second approach is that it does not require extensive learning or tuning procedures and thus is easier to implement. The paper suggests possible applications of the same techniques to other physiological rhythms, such as blood pressure and some parameters of the ECG like S-T segment level and slope. Furthermore, combining information obtained using both approaches in order to improve the overall predictability is suggested.