Discover the scientific fundamentals behind OMsignal technology, as well as numerous situations where our solutions have been applied.
The purpose of this work was to develop a personalized method of categorizing heart rate variability (HRV) from recordings taken during daily life while participants wore the OMbra. Low HRV has been associated with increased stress levels and a wide array of adverse health outcomes. Twenty women wore the OMbra several times a week during their daily life. HRV was calculated on 5 minute windows when participants were in a still, seated position. Personalized HRV zones were calculated based on the relationship between HRV and heart rate (HR) during the first 100 HRV segments that were recorded for each participant. Participants had on average 23.1% (+/-9.8) high HRV, 55.3% (+/-13.2) average HRV, 16.1% (+/-7.1) low HRV and 5.6% (+/-6.9) very low HRV periods during their daily life recordings. This work presents a novel method of obtaining high quality HRV metrics as people go about their daily lives using the OMbra. A novel method is proposed to provide personalized HRV ratings that take into account changes in HRV due to increased or decreased HR.
Heart rate, heart rate variability, electrocardiogram, bio-sensing textile.
In this work, a deep convolutional neural network (CNN) is proposed to detect atrial fibrillation (AF) among the normal, noisy and other categories of cardiac arrhythmias electrocardiogram (ECG) recordings. The proposed CNN is trained by stochastic gradient descent with the categorical cross-entropy loss function. The network performance is evaluated on training (75%) and validation (25%) data sets that are obtained from 2017 Physionet/CinC challenge database. The proposed CNN model respectively achieves the average accuracy and F1 score of 87% and 0.84 on validation data set.
One of the main advantages of this work besides high accuracy and reliability, is to simplify the feature extraction process and to remove the need for detecting ECG signal fiducial points and extracting hand-crafted features unlike conventional methods available in the literature. Moreover, it provides an opportunity for ECG screening in a large population, especially for atrial fibrillation screening, using wearable devices such as OM apparel that records high-quality single channel ECG signal.
The work described in this report is concerned with verifying the accuracy of the breathing rate algorithm (based on the inhale peak detections in the OMsignal box firmware). For this purpose, nine raw data files collected from men wearing OMshirts and thirty raw data files collected from women wearing OMbras had a portion of their useable recording manually scored by an engineer to identify true inhale peaks. The performance of the event detection algorithm is assessed based on sensitivity and positive predictive value (PPV). The accuracy of breathing rate was assessed using the 95th percentile of the absolute difference between the rate computed from the manual scores and the automatically detected IN-IN (inhale to inhale) intervals. The overall accuracy is 2.1 (+2.7) BPM which is within the desired tolerance of 5 BPM. Additionally, the per file average sensitivity and PPV is 94% and 94% respectively. The breathing rate algorithm is assessed to work well for both men and women in the context of running.
The purpose of this report is to assess the accuracy of the breathing depth and ventilation metrics as computed by OMshirts and OMbras. 5 men and 5 women participated in an incremental exercise test where their breathing was monitored by a metabolic cart, at a lab that will be referred to as the Peak Center. At the same time, their biometrics were recorded with OMsignal technology. There was a technical failure in the OMsignal recording of 2 of the men. One of the women had difficulty adhering to the protocol, which resulted in unreliable data. In the end, 3 men and 4 women were analyzed in detail. The subjects were scored based on their Pearson correlation coefficients between their metrics as measured by a metabolic cart, and their analogous metrics as measured by OMsignal technology. Additionally a measure of worst case bias based on comparison of a linear regression model with a kNN model is used in order to quantify non linearities.
In the end, all three men and two of the four women were within tolerance. One of the women who failed (W1) had a significant bias, but this occurred in a region of the test that was not important for identifying AT or VT. The final woman who failed (W5) had an issue with excess breath detections in the OMsignal data. The subject reported that she believes this pattern is consistent with how she breathes. However, these detections caused significant issues for AT/VT identification, which was impossible for this subject from the OMsignal data. In conclusion, the ventilation measure from OMsignal technology was shown to be sufficiently correlated and linear with true minute ventilation to be useful for AT/VT identification in 6 out of 7 analyzed subjects. The 7th subject was accurately identified in the separate AT/VT validation as having no discernible inflection points. Given this, it is recommended to always allow an option of ’not available’ in applications that require assessing AT/VT.
In this work, a fast, reliable and accurate ECGbased human identification strategy is presented. A neural network is trained to identify individuals from a pool of 33 participants, given a window of 5 heart beats. The participants are drawn from the OMsignal MyHeart project. The windows are extracted from ECG recordings captured by OMsignal apparel while the participants go about their daily activities. The neural network is augmented with an uncertainty measure based on Monte Carlo dropout, allowing predictions to be made only on high quality data. On a testing dataset created from different recordings of the same participants and unobserved until development had finished, a window accuracy of 99.7% is achieved with a window rejection rate of 18.4%. Using majority voting classification across all collected windows, 31 out of 33 of the participants are correctly identified. The individuals who are not correctly identified belong to a subset of users who had less than 5 recordings on 5 separate days. This is hypothesized to be the lower limit on the amount of data necessary to observe enough variation to make accurate identifications of an individual.
Identifying an individual from their ECG requires the presence and measurement of subtle features in the signal. The performance of this system provides encouraging evidence that OMsignal apparel, combined with machine learning algorithms may be able to find subtle patterns associated with developing medical conditions.
A system is presented for the automated detection of the heart rate that corresponds to a persons ventilatory and anaerobic thresholds (VT and AT). The system is based on an analysis of heart rate, breathing depth and breathing rate measurements from OMsignal apparel while users run. In most cases, the system is capable of providing a reasonable estimate of the AT and VT after 5 free form runs, ie. without a pre-specified protocol. The system automatically determines the AT and VT from this data, closely approximating the values determined by the previously used manual process relying on a human annotator. Due to the difficulty of analyzing free form running data, a sufficiently accurate algorithm has thus far been elusive. Using a sequence of filters and algorithms, automatic AT and VT determination is cast as a computer vision problem, and finally solved with a random forest. On a test set of 40 users, the AT and VT determined by the system were both within 10 BPM of the manually determined values in 95% of the cases, and were both within 7.5 BPM in 85% of the cases. The largest differences occurred on users that had ventilation-HR curves that showed ambiguous inflection points or low variation. Additionally, the system’s AT and VT assessments were compared against ground truth values determined by a metabolic cart system on 10 users while they underwent an incremental exercise protocol. In all but one case, the difference was less than 7.5 BPM, and the mean absolute error of VT and AT were 3.56 BPM and 4.75 BPM respectively. The remaining errors are likely due to inherent uncertainty of using free form data as opposed to a regimented protocol. Use cases requiring more accuracy or precision should use such a protocol.
Because of the limited amount of resources of the OMsignal hardware box, the only possibility to record and send the raw electrocardiogram (ECG) and raw breathing signals is to compress them. But compression creates a loss of signal quality that needs to be minimized. We present results of our compression algorithm where the percent root mean square error of the resulting decompressed signal compared to the original raw signal is less than 0.1 percent.
ECG, breathing, signal compression.
In this work, a deep convolutional neural network (CNN) is developed to focus on human identification problem using electrocardiogram (ECG) signals that are collected by OMsignal apparel from 33 women while doing their daily activities. The signature windows including 10 consecutive heart beats are extracted from the filtered ECG signal to be applied to the CNN model. The CNN is trained by stochastic gradient descent with a categorical cross-entropy loss function. The network performance is evaluated on validation and testing data sets. On validation and testing data sets created from different recordings of the same participants, an overall window accuracy of 95.25% and 95.95% are respectively achieved. Using majority voting classification across all collected windows, 100% of the participants with more than five ECG daily recordings are correctly identified.
One of the main advantages of this work besides high accuracy and reliability, is to simplify the feature extraction process and to remove the need for detecting ECG signal fiducial points and extracting hand-crafted features unlike conventional methods available in the literature.
OM Running Dynamics are based on utilizing the accelerometer signal from the OMsignal box while users run. The OMsignal box is securely attached to the OMbra or OMshirt and measures electrocardiogram, breathing and acceleration. This document describes how the acceleration signal is used to obtain detailed information about a user’s running technique. First, steady state running is identified and then the bio-mechanical metrics are calculated; temporal gait features, asymmetry, braking impulse, impact and vertical oscillation on each step. This document summarizes the metrics obtained via OMsignal Running Dynamics.
Biomechanics, Accelerometer, Bio-sensing Textile.
The OMsignal heart rate measurement algorithms were demonstrated to be accurate within 5 beats per minute during running and jogging. This accuracy was verified for 95% of the recording time with 95% statistical certainty. Data for the verification was taken from 20 recordings by women and men wearing the OMsignal bio-sensing garments. In this work, an accurate ECG-based daily mental stress level prediction strategy is presented. Multiple support vector machines (SVM) with linear kernel functions are individually trained to predict daily stress levels of women who participated in the OM- signal MyHeart project. In this study, participants are asked to answer a daily survey to determine the quality of their sleep, exercise, valence, control and rumination during the last 24-hour. Using the aforementioned items, a daily stress score was defined to be used as the target value for constructing the stress prediction model. The model is designed by the use of heart rate variability (HRV) metrics calculated from a 5-minute data window moving over daily ECG recordings. A 30 dimensional feature vector, including the first five minimum and maximum values of SDNN and RMSSD (two popular HRV metrics) as well as heart rate is extracted to represent each individual daily ECG record. The leave-one-out cross- validation method is used to train and validate our user-dependent SVM model. On validation data, an average accuracy of 85.26% is achieved for predicting daily stress scores of the users with sufficient number of daily survey data.
The OMsignal heart rate measurement algorithms were demonstrated to be accurate within 5 beats per minute during running and jogging. This accuracy was verified for 95% of the recording time with 95% statistical certainty. The overall heart rate accuracy was measured to be 4 bpm (CI+ 4.71 bpm). The average sensitivity and PPV of the algorithm were measured to be 90.95% and 99.6% Data for the verification was taken from 20 recordings by women and men wearing the OMsignal bio-sensing garments.
Heart Rate, Electrocardiogram, Bio-sensing Textile.