Bloque II. La Revelación: Dios interviene en la historia Unidad 2: DIOS SE DA A
ESTÁNDARES DE APRENDIZAJE EVALUABLES INSTRUMENTOS DE EVALUACIÓN
detection and classification
With an aim to recognise the abnormal patterns in long-term ECG monitoring, the chapter initially described the experiments that were performed using data analysis pipelines for V,A,L,R classification. A preliminary feature extraction algorithm for the MITDB MIT-BIH arrhythmia database has been described giving consideration to the digital format of the records along with the annotations related to the normal and abnormal heart beats that were used in ECG analysis to detect cardiac arrhythmia. The PhysioNet Waveform Database (WFDB) library which provided software routines to query and analyse the MITDB records using these annotations, was used for feature extraction. Currently, there exists no software library in literature that could take an ECG recording in real-time, extract features and annotate the beat samples as belonging to a certain arrhythmia class types. The methods of feature extraction using WFDB routines is presented to determine the effectiveness of using k-Nearest
Neighbours (k-NN) and RandomForestClassifier models and pattern recognition
for arrhythmia detection. Initial exploratory analysis was performed on the MITDB records using a reduced feature set (Age, Gender, RR-Interval, signal value (mV)) and class labels containing only the abnormal beat rhythms V,A,L,R representing PVCs, PACs, Left Branch Bundle Blocks (LBBB) and Right Branch Bundle Blocks (RBBB) as adequate number of samples for these arrhythmia types were found in MITDB dataset. The data analysis for V,A,L,R classification is presented in the subsection 4.2.2 Once the effectiveness of using machine learning classification models was established a detailed feature extraction algorithm was developed by extracting features from physiological parameters of the ECG waveform. As was argued that an ECG signal is a non-stationary signal, the statistical properties of the signal could vary over time. Also, no two normal heartbeats or abnormal heartbeats could belonging to the same class type can
4.9. Summary: Early warning arrhythmia detection and classification 149
have the same waveform, so it was difficult to obtain feature values that were consistent in describing and classifying the waveform into a specific arrhythmia type 4.2.3. This problem was resolved by performing wavelet and spectral analysis over the sub-wave portions of the ECG waveform. The methods of spectral analysis of an ECG wave belonging to a class type and it’s comparative analysis to differentiate between class types are presented in section on spectral analysis section 4.3. The data analysis section 4.3.3 describes in detail the pipelines used for features transformation (standardisation, normalisation, scaling and dataset imbalance removal) used in data analysis preprocessing tasks prior to generating and fine-tuning the classification models. The rationale behind using a particular supervised learning classification model has also been discussed to explain the choice of a particular model. The results obtained from the data analysis using the classification models are presented in detail in results section using classification report containing accuracy, precision, recall and f1-scores. The classification models were serialised and deployed on a wearable target device, so that the classification of arrhythmia types could be done in real-time in-situ. In the ECG signal acquisition and processing section 4.5 the method of signal acquisition is described, followed by detailed signal processing techniques for signal denoising, filtering and conditioning. The freshly acquired ECG signal is noisy due to electrical interference and motion artefacts, several filtering techniques and wavelet transforms are described along with the rationale of using these techniques. The signal filters were developed in MATLAB signal processing and wavelet toolboxes initially, to model the filter parameters, though since the filter was deployed on the wearable target device a SciPy model was developed using the parameters similar to the MATLAB models. The detailed specification of the digital filter is presented in the filter design subsections 4.5.1 and 4.5.2 Once the ECG samples were filtered, they were digitised according to MITDB compatible data format so that the classification models trained on MITDB dataset could be used on the ECG samples captured in real-time. In order to achieve this task, the WFDB routines were used to transform fresh ECG samples to digitised MITDB format, subsection 4.5.3. A digitisation pipeline along with the feature extraction pipeline provided a unique method of achieving the real-time ECG
signal acquisition, transformation and extraction objectives in real-time signal acquisition phase, presented in 4.6 subsection. An extended feature algorithm was developed and is presented as Algorithm 4.2 to achieve the real-time feature extraction pipeline.
Chapter 5
Trauma Analysis
5.1
Introduction
The fatal cardiac arrhythmia can lead to emergencies and trauma conditions, and at any time and location and whilst the individuals remain engaged in their day-to-day activities. As the objective of the research study was to predict early signs of arrhythmia and to produce early warning signs and to predict survival, a reliable trauma scoring measure or measures were required to ascertain patient’s health status when an emergency occurred. The vital signs (Holcomb et al. 2005; Lockwood, Conroy-Hiller, and Page 2004) and certain physiological parameters could help in calculating trauma scores to indicate the trauma condition of the patient. In order to perform trauma scoring a health monitoring kit was required to perform arrhythmia classification and trauma analysis simultaneously in real-time. The challenges in performing trauma scoring tasks were that bedside monitors and equipment in hospitalised settings are normally used in trauma scoring and some vital signs such as the respiratory rate and the blood pressure, are normally obtained using clinical instruments such as the Spirometer and the Sphygmomanometer, which are not wearable. Also, trained staff is normally required to attend triage emergencies to interpret the vital signs and to manually calculate and interpret the trauma scores. In the absence of direct measurements, these measures had to be calculated and approximated so that all the vital signs became available for calculations. A solution has been proposed in this chapter to acquire and calculate intermediate vital signs and
intermediate trauma scores from ECG and Photoplethysmogram (PPG) signals (Dinh, Luu, and Cao 2017) using the MITDB WFDB libraries (Ary L Goldberger et al. 2000) and signal processing techniques on a wearable resource constrained device. A PPG is an optically obtained plethysmogram that can be used to detect blood volume changes in the micro-vascular bed of tissue. Furthermore, the arrhythmia and trauma related information had to be integrated with Electronics Health Records (EHR) if a trauma event occurred. In order to make this provision the challenge was to encapsulate the trauma related scores, the vital signs profile and the cardiac arrhythmia related information, along with location information, in a standard format and according to clinical terminology acceptable by EHR globally. The trauma and arrhythmia related information could be transmitted in real time and could be shared across multiple EHR repositories and Decision Support Systems (DSS) worldwide for further consultation with medics having diverse skill-sets and for research.
In this chapter operation of the Composite Health Monitoring (CHM) kit (Walinjkar 2018a) explained in section 3.1 is presented in a form of a unique pipeline for ECG and PPG signal acquisition and processing tasks, followed by machine learning based prediction and classification along with the encapsulation of trauma scores, vital signs profile and location information to be transmitted to EHR servers using standard telemetry protocols. A novel trauma scoring algorithm is also presented in this chapter that could aggregate a combination of trauma scores that could determine the prediction of survival of an individual under trauma condition. To determine the efficacy of this algorithm, it was tested on the MIMIC Numerics dataset, which is essentially a vital signs dataset of patients admitted to the ICU. The National Early Warning Signs (NEWS), Revised Trauma Scores (RTS), and TRauma Injury Severity score (TRISS) and Prediction of Survival (Ps) scores were calculated using the vital signs extracted from the dataset (Champion et al. 1989). The vital signs and physiological parameters are usually obtained in a hospitalised environment or from ambulatory equipment, however, the CHM kit presented in this chapter, could calculate these scores in real-time and could provide trauma and prediction survival scores to the critical care team ahead of emergencies. For trauma scoring, the proposed