Multichannel foetal heartbeat detection by combining source cancellation with expectation-weighted estimation of fiducial points

Noninvasive foetal heart rate (fHR) monitoring is important in detecting foetal distress and morbidity. In spite of the improvement achieved in recent years, the accuracy of non-invasive fHR monitors is not satisfactory. In this study a new method is presented with the goal of improving the accuracy of foetal heart beat detection from abdominal recordings of maternal ECG. A dataset of 75 four-channel abdominal ECG recordings (SetA) provided by Physionet Challenge 2013 was used for training. Maternal QRSs were detected and subtracted from notch-filtered (50 Hz) and high-pass filtered (4 Hz) abdominal ECGs. The residual ECG signal was further processed (1st derivative squared and low-pass filtered by moving-average). On the resulting signals, local peaks were searched using an expectation-weighted estimate of the “next” fiducial point (foetal QRS (fQRS)) based on a Gaussian distribution G(μ,σ) with μ indicating the most probable time distance of the “next” fiducial point. A grid search was used to determine (μ,σ) minimizing the standard deviation of the resulting estimated fQRS time series for each recording. The fQRS time series were submitted to Physionet Challenge 2013 Events 4 and 5. In Event 4 (published set of 100 recordings, SetB), a test fHR time series was built by Physionet's Scoring System from fQRS annotations and matched to the reference. Scores were based on test vs. reference mismatches (lower scores indicate better performance). In Event 5 (also using SetB), the RR time series replaced fHR. According to Phase 1 official scores (14 June 2013) our algorithm scored 135.18 (7.11) in Event 4 (Event 5), corresponding to the 10th (9th) position in the ranking. The reference algorithm provided by Physionet scored 3258.56 (102.75). The proposed method substantially improved the foetal heart beat detection accuracy with respect to the reference algorithm.