Automated triage parameters estimation from ECG

1. Technical note

Deriving patients’ vital signs parameters, and especially the time dynamics of those parameters, is a constant challenge in the cases of mass casualties as: floods, earthquakes, terrorist attacks, battle fields, and other natural, man-made, or hybrid disasters [1]. In the case of a mass casualty incident, when the number of injured reaches hundreds of people, the triage process must be performed time efficiently, preferably in less than 30 seconds [2]. According to the triage demands [3], the heart rate (HR) and respiratory rate (RR) vital signs are needed to be obtained immediately in order to instantly determine the medical status of an injured person. An optimization of the procedure can be achieved if biosensors are used to derive the vital signs parameters.

To select the most suitable HR and RR derivation methods, recently published methodologies have been properly analyzed in terms of their low computation characteristics. Accordingly, HR extraction from ECG signals is possible only after at least two methods are applied. The first type of methods are used to eliminate the noise in the signals and the second type of methods refer to QRS-complex and R-peak detection.

OPEN IN VIEWER

Following the literature, we propose methodology that will solve the issues regarding the processing of large amounts of data received in real-time and the large power consumption due to the constant Bluetooth connection with the biosensors.

The raw ECG signal obtained from a particular biosensor is the input in the methodology whose steps are depicted in Fig. 1. The use of 30-seconds-long single sensor signals are supported in [4]. R-peak detection of the ECG signal is a common initial step for both the HR and RR extraction. In this stage, the well-known Pan-Tompkins algorithm [5] is chosen as the most appropriate since it has been shown to offer low-complexity characteristics during QRS analysis examination [6], making it suitable for portable devices.

Pan-Tompkins relies on two learning stages and one detection stage. The first learning stage initializes detection thresholds based upon signal and noise peaks detected during the learning process. In the second learning stage, RR-interval average and limit values are initialized. The detection stage does the recognition process and produces a pulse for each QRS-complex. The Pan-Tompkins algorithm is applied in the preprocessing stage of the methodology, where a bandpass and derivative filter are first applied to the ECG signal and therefrom, a decision rule checks whether the waveform represents a QRS-complex. The R-waves obtained as an output of the algorithm are later used for HR and RR estimation.

Having detected the number of R-peaks (NRP) for a known signal length (SL), the number of beats per second can be easily calculated and hereupon the number of heart beats per minute (HR):

𝐻𝑅 = ( 𝑆𝐿 𝑁𝑅𝑃 ) - 1 * 60(1)

In parallel, the R-peaks are used to obtain the RR. The procedure for the RR computation relies on the idea presented in [7], where the authors provide a low-complexity methodology for breathing rate computation. Several modifications of the above approach are proposed, and they are explained in the sequel.

In order to enable the accurate identification of the locations of the local maxima, a kurtosis computation technique is applied to measure the peakedness of the signal’s distribution. Local regression with weighted linear least squares and a 1st degree polynomial model is chosen as the most appropriate for data smoothing which is the second step. Once the ECG signal is smoothed, its local maxima (peaks) are found by applying a peak finder method. The number of peaks are the number of respirations (NR) in the signal, and are used to calculate the number of respirations per minute (RR):

𝑅𝑅 = ( 𝑆𝐿 𝑁𝑅 ) - 1 * 60(2)

Once the parameters are determined, they are used to perform the triage.

Table 1 presents the ranges and the consequential descriptive classes of injuries.

To test the methodology accuracy, Massachusetts General Hospital/Marquette Foundation (MGH/MF) Waveform [8] Physionet database is used. The estimation robustness is measured via sensitivity (SE) and positive predictivity (PPV) metrics that are chosen as appropriate for this type of methodology evaluation because they provide the real insight into the methodology’s ability not to miss the real positive subjects, or the most severely injured individuals in a mass casualty incident. In the current case, SE reflects the number of true detected heart beats (or respirations) over the real number of heart beats (or respirations), whereas the PPV refers to the number of true detected heart beats (or respirations) over the number of detected heart beats (or respirations). Secondly, the triage robustness is measured where the SE and PPV are defined as the methodology’s ability to correctly identify the high-priority injuries among all the injuries and its ability to correctly identify high-priority injuries among the truly high-priority injuries, correspondingly. Eventually, the time needed to analyze the ECG signal is also considered as most important to achieve real-time triage calculation.

The proposed methodology provides the best results for both HR and RR estimation, achieving SE of 0.87 for HR, 0.74 for RR, and a PPV of 0.94; for the triage prediction those metrics are 0.76 for SE and 0.87 for PPV. All of the calculations are finished in a very favorable average-case execution time of 0.02 seconds. The methodology is developed for multiple platforms and together with the testing database is publicly available at https://github.com/bojanakoteska/.