Analysing irregularity and complexity in short-length heart rate variability signals

Abstract

Heart rate variability (HRV) analysis is a powerful non-invasive means to help diagnose several cardiac ailments. The non-linear and non-stationary nature of HRV necessitates the use of non-linear statistics such as 'irregularity' or 'complexity' measurement to understand the complex nature of this physiological signal. But, most of such measures need long-term recordings of HRV to extract any relevant signal information. To do so from short-term (2-15 minutes of data) HRV data has been a long-standing challenge in medical diagnosis. Using short-term HRV data in place of the long-term ones significantly reduces the non-stationarity involved, in addition to providing economizing benefits.

The most popularly used non-linear tools to measure signal irregularity from short-term HRV data are the entropy measures like Approximate entropy (ApEn) and Sample entropy (SampEn). These methods become unreliable and inaccurate most times, due to their extreme sensitivity and dependence on parameters like the tolerance of inter-vector distance r. For a given length (N) of signal, one must choose the most appropriate value of r to obtain an accurate estimate of entropy. Different inputs of r give different entropy outcomes for the same signal and hence an incorrect parameter selection may give completely wrong estimates of signal regularity. For biological signals, the recommended choice of r is 0.1-0.2 times the standard deviation (SD) of the signal, found to be inappropriate in many cases of study. Current research focuses on addressing this issue of irresolute r-selection by either (a) introducing a criterion for appropriate selection of the parameter or (b) reducing the impact of the parameter on entropy calculations.

This work aims at handling the issue differently; by proposing to eliminate the need for using r as an input parameter in entropy estimations. We use a data driven, non-parametric approach that would automatically generate all appropriate r values for a given signal and eventually generate a complete profile (instead of a single estimate) of entropy values. Here, we focus on achieving enhanced information retrieval by generating the complete entropy profile of a given signal in contrast to a single entropy value that comes from traditional algorithms. Having a full profile of irregularity based evidence is extremely beneficial when it comes to extracting information from short-length (less than 2 minutes of data) data. This work also focuses on signal classification of short-length HRV data based on factors like disease and age. From short-length segments of heart rate recordings, we extract maximum regularity based information using our non-parametric, data driven non-linear approach, and thereby classify signals with high accuracy and consistency.