Predicting heart beats using co-occurring constrained sequential patterns

The aim of this study is to develop and evaluate a robust method for heart beat detection using a sequential pattern mining framework, based on the multi-modal Physionet 2014 challenge dataset. Each multi-modal patient time series was initially transformed to a symbolic sequence using Symbolic Aggregation Approximation (SAX). A training set was created, by randomly selecting 70% of the data and the rest 30% was used as the test set. Later, all segments of length 100 were extracted, for annotated beat occurrences. Subsequently, an algorithm was used to extract repetitive frequent subsequences, where consecutive symbols are separated by a pre-defined gap range. The patterns for ECG and BP were then ranked based on length and frequency support. For tests, the highest ranked patterns were used to mark beat segments. True beat occurrences were only considered when patterns co-occurred for both ECG and BP within a width of 150 time points. Our results comprise two parts viz. extracted top ranked sequences and gross test statistics. An interpretive highest ranked sequential pattern for ECG looks like [7,7,7,5,5,5,5,5,4,3,10,10,10,2,2,3,3,4,3,4,5,5,5,6,7], for 10 discrete symbols which identify regional signal activity, with a gap range of [2,4] between contiguous elements. As per our test results, the method gives us a sensitivity of 51.66% and a positive predictivity (PPV) of 67.15%. The novelty of mining gap constrained co-occurring frequent sequential patterns lies in its ability to capture approximate co-occurring long clinical episodes across multiple variables, even if the quality of one signal suffers for a certain period of time. A higher PPV indicates that our method did not have a lot of false positives (detecting non-beats). The method is still being improved and will be further tested in the next stages of the Ph