Assessing the signal quality of electrocardiograms from varied acquisition sources: A generic machine learning pipeline for model generation.

BACKGROUND AND OBJECTIVE: Long-term electrocardiogram monitoring comes at the expense of signal quality. During unconstrained movements, the electrocardiogram is often corrupted by motion artefacts, which can lead to inaccurate physiological information. In this situation, automated quality assessment methods are useful to increase the reliability of the measurements. A generic machine learning pipeline that generates classification models for electrocardiogram quality assessment is presented in this article. The presented pipeline is tested on signals from varied acquisition sources, towards selecting segments that can be used for heart rate analysis in lifestyle applications. METHODS: Electrocardiogram recordings from traditional, wearable and ubiquitous devices, are segmented in 10 s windows and manually labeled by experienced researchers into two quality classes. To capture the electrocardiogram dynamics, a comprehensive set of 43 features is extracted from each segment, based on the time-domain signal, its Fast Fourier Transform, the Autocorrelation function and the Stationary Wavelet Transform. To select the most relevant features for each acquisition source we employ both a customized hybrid approach and the state-of-the-art Neighborhood Component Analysis method and compare them. Support Vector Machines (SVM), Decision Trees, K-Nearest-Neighbors and supervised ensemble methods are tested as possible binary classifiers. RESULTS: The results for the best performing models on traditional, wearable and ubiquitous electrocardiogram datasets are, respectively: balanced-accuracy: 89%, F1-score: 93% with the Fine Gaussian SVM model and 10 features; balanced-accuracy: 93%, F1-score: 93% with the Fine Gaussian SVM model and 11 features; balanced-accuracy: 95%, F1-score: 86%, with the Fine Gaussian SVM model and 8 features. CONCLUSIONS: According to the results, our generic pipeline can generate classification models tailored to individual acquisition sources, provided that a standard Lead I or Lead II is available. Such models accurately establish whether the electrocardiogram quality is good or bad for heart rate analysis. Furthermore, removing bad quality segments decreases errors in heart rate calculation.

Reveal Temporal Patterns of Smoking Behavior in Real Life Using Data Acquired through Automatic Tracking Systems.

Accurately monitoring and modeling smoking behavior in real life settings is critical for designing and delivering appropriate smoking-cessation interventions through mHealth applications. In this paper, we inspect smoking patterns based on data collected from 52 volunteers during a 4-week period of their everyday lives. These data are acquired by an automatic data acquisition system comprising an electric lighter, two wearable sensors and one mobile phone, which together can automatically track smoking events, collect concurrent context and physiology, and trigger pop-up questionnaires. We visualize temporal patterns of smoking at the level of the week, day and time of the day. Statistical analysis on all subjects has demonstrated significant differences at the levels evaluated. Distinct emotions during smoking at individual level are also found. Quantified smoking patterns can upgrade our understanding of individual behaviors and contribute to optimizing intervention plans.

Toward Quantifying the Psychopathology of Eating Disorders From the Autonomic Nervous System Perspective: A Methodological Approach.

The phenomenology of Eating Disorders (ED) relates with altered functioning of the Autonomic Nervous System (ANS). The lack of agreement in what comes to the direction and significance of such alterations is possibly due to the variability in the ED spectrum. As the stress response system is an integral part of the ANS, we propose to investigate ANS tonic variations and phasic activations in response to stressors. We hypothesize that, while using stress as a test probe, characteristic ANS dysregulations in ED may be found when considering several physiological signals measured over time, and weighted by the individual psychological profiles. In this article we describe a novel methodological approach to investigate this hypothesis with the aim of providing further clarification on the ED spectrum conceptualization. The proposed methodology has been designed to be easily integrated in clinical practice and, eventually, in daily life. The population under observation includes both patients in treatment for ED, and matched controls. The study session has the duration of 1 day, including: (1) the administration of a stress task in a controlled environment and (2) naturalistic data collection. The stress task is designed to elicit both mentally and physically driven ANS activation. The naturalistic component intends to illustrate the psychophysiology in everyday life. We use wearable devices to continuously and non-invasively measure bio-signals related to ANS functioning. This information is complemented with psychometric information from validated stress and ED scales and ecological momentary assessments. The protocol has received ethical approval and has been implemented in practice, currently accounting for 37 patients (out of 120) and 16 controls (out of 60). Ongoing work focus on the definition and implementation of a data processing pipeline to quantitatively test our hypothesis, both standard statistical methods and more exploratory machine learning approaches will be considered.

Motion Artifact Reduction for Wrist-Worn Photoplethysmograph Sensors Based on Different Wavelengths.

Long-term heart rate (HR) monitoring by wrist-worn photoplethysmograph (PPG) sensors enables the assessment of health conditions during daily life with high user comfort. However, PPG signals are vulnerable to motion artifacts (MAs), which significantly affect the accuracy of estimated physiological parameters such as HR. This paper proposes a novel modular algorithm framework for MA removal based on different wavelengths for wrist-worn PPG sensors. The framework uses a green PPG signal for HR monitoring and an infrared PPG signal as the motion reference. The proposed framework includes four main steps: motion detection, motion removal using continuous wavelet transform, approximate HR estimation and signal reconstruction. The proposed algorithm is evaluated against an electrocardiogram (ECG) in terms of HR error for a dataset of 6 healthy subjects performing 21 types of motion. The proposed MA removal method reduced the average error in HR estimation from 4.3, 3.0 and 3.8 bpm to 0.6, 1.0 and 2.1 bpm in periodic, random, and continuous non-periodic motion situations, respectively.