Waveform Estimation Research Articles

Time–Frequency Learning Framework for rPPG Signal Estimation Using Scalogram-Based Feature Map of Facial Video Data

Remote photoplethysmography (rPPG) is a non-contact camera based method that seems to be the most promising technology for remote cardiac health assessment. The conventional rPPG methods have played a significant role in the growth of low-cost camera based cardiac health monitoring system. However, they are based on certain assumptions, and their performance may degrade with real-time dynamic interferences. Recently, deep neural network (DNN) based methods have been introduced in rPPG domain and they proves to perform well in noisy environment. In this work, we proposed a novel 2D-convolution neural network (CNN) based time-frequency learning (TFL) framework to estimate rPPG signal using a scalogram based feature map of facial video. The TFL network estimates the periodic rPPG signal from the feature map that is obtained using the raw RGB traces. The feature map generation involves the enhancement of pulsatile profile of the raw RGB traces with the help of wavelet transformation. The enhanced pulsatile profile in the feature map helps the proposed network to characterize a wide-range of frequencies associated with the cardiac activity. The estimated rPPG signal from the network is used for heart rate (HR) and heart rate variability (HRV) measurements. The proposed network is tested and validated on three publicly available UBFC_rPPG, COHFACE, and ECG-fitness datasets and outperforms the state-of-the-art methods. The proposed TFL framework achieves 0.569 bpm mean absolute error (MAE) and 0.997 Pearson correlation coefficient (ρ) for HR estimation on the UBFC_rPPG dataset. On the COHFACE and ECG-fitness datasets, the MAE and (ρ) pair are found to be (0.987, 0.958) and (4.267, 0.937), respectively. The performance results show the effectiveness of the designed feature map along with the TFL network in the estimation of rPPG waveform for non-contact vitals measurement.

Total loss torque waveform estimation for a turbocharged diesel engine

The indicated (gas) torque is a key parameter for reciprocating engine. It reports on various aspects of the engine, such as the quality of the combustion process, the fuel combustion rate and the average effective pressure. This is a good indicator of the stability of the engine at work. The form of the instantaneous indicated torque is precious information for the control and diagnosis of the engine. The study of the simulation of crankshaft dynamics illustrates the need to know the total loss torque (i.e. the addition of friction torque, auxiliary torque and load torque) to achieve the indicated torque. To date, it has proven difficult to obtain an accurate predictive expression of the instantaneous total loss torque. This expression includes adequate simulation of friction of all moving parts of the engine (piston ring and skirt, crankshaft bearings, valve gear...), auxiliaries and load. Based on experimental results conducted on a four-cylinder turbocharged diesel engine, this work offers a simple and practical solution for estimating the total torque loss. The proposed method makes it possible to successfully estimate total torque losses under a wide range of engine operating conditions. The result prediction has been confronted to experimental data.

ECG Analysis-Based Cardiac Disease Prediction Using Signal Feature Selection with Extraction Based on AI Techniques

ECG (Electrocardiogram) performs classification using a machine learning model for processing different features in the ECG signal. The electrical activity of the heart is computed with the ECG signal with machine learning library. The key issue in the handling of ECG signals is an estimation of irregularities to evaluate the health status of patients. The ECG signal evaluate the impulse waveform for the specialized tissues in the cardiac heart diseases. However, the ECG signal comprises of the different difficulties associated with waveform estimation to derive certain features. Through machine learning (ML) model the input features are computed with input ECG signals. In this paper, proposed a Noise QRS Feature to evaluate the features in the ECG signals for the effective classification. The Noise QRS Feature model computes the ECG signal features of the waveform sequences. Initially, the signal is pre-processed with the Finite Impulse response (FIR) filter for the analysis of ECG signal. The features in the ECG signal are processed and computed with the QRS signal responses in the ECG signal. The Noise QRS Feature evaluate the ECG signal with the kNN for the estimation and classification of features in the ECG signals. The performance of the proposed Noise QRS Feature features are comparatively examined with the Discrete Wavelet Transform (DWT), Dual-Tree Complex Wavelet Transforms (DTCWT) and Discrete Orthonormal Stockwell Transform (DOST) and the machine learning model Cascade Feed Forward Neural Network (CFNN), Feed Forward Neural Network (FFNN). Simulation analysis expressed that the proposed Noise QRS Feature exhibits a higher classification accuracy of 99% which is ~6 – 7% higher than the conventional classifier model.

Machine-Learning Classification of Pulse Waveform Quality.

Pulse measurements made using wearable devices can aid the monitoring of human physiological condition. Accurate estimation of waveforms is often difficult for nonexperts; motion artifacts may occur during tonometry measurements when the skin-sensor contact pressure is insufficient. An alternative approach is to extract only high-quality pulses for use in index calculations. The present study aimed to determine the effectiveness of using machine-learning analysis in discriminating between high-quality and low-quality pulse waveforms induced by applying different contact pressures. Radial blood pressure waveform (BPW) signals were measured noninvasively in healthy young subjects using a strain-gauge transducer. One-minute-long trains of pulse data were measured when applying the appropriate contact pressure (67.80 ± 1.55 mmHg) and a higher contact pressure (151.80 ± 3.19 mmHg). Eight machine-learning algorithms were employed to evaluate the following 40 harmonic pulse indices: amplitude proportions and their coefficients of variation and phase angles and their standard deviations. Significant differences were noted in BPW indices between applying appropriate and higher skin-surface contact pressures. The present appropriate contact pressure could not only provide a suitable holding force for the wearable device but also helped to maintain the physiological stability of the underlying tissues. Machine-learning analysis provides an effective method for distinguishing between the high-quality and low-quality pulses with excellent discrimination performance (leave-one-subject-out test: random-forest AUC = 0.96). This approach will aid the development of an automatic screening method for waveform quality and thereby improve the noninvasive acquisition reliability. Other possible interfering factors in practical applications can also be systematically studied using a similar procedure.

A noninvasive method of estimating patient-specific left ventricular pressure waveform

Background and objectiveThe left ventricular pressure waveform is indispensable for the construction of the pressure strain loop when investigating coronary artery disease (CAD) patients. In previous studies by others, exclusion of CAD patients has not allowed a reliable estimation of the left ventricular pressure waveform from the pressure strain loop of these patients. To remedy this, we propose a patient-specific noninvasive method for the estimation of left ventricular pressure. MethodsA simplified systemic circulation model consisting primarily of a single fiber model and a 1D simulation of the arterial tree was used. Sensitivity analysis based on the Morris method was performed to select a subset of the important parameters. Following this, the important parameter subset and the set of all the parameters were identified in the model using the pressure waveform of a peripheral artery as input, in a two-step process. In addition, the left ventricular pressure waveform was estimated using the set of all parameters. ResultsReducing the size of the parameter subset significantly decreases the computational cost of parameter optimization in the first step and greatly simplifies the identification of the full parameter set in the second step. Comparison with the reference left ventricular pressure waveform from CAD patients, showed that the proposed method provides a good estimate of the reference left ventricular pressure waveform. The correlation coefficients between the estimated and reference were r = 0.907, r = 0.904 and r = 0.780 for systolic blood pressure, pulse pressure and mean blood pressure, respectively. ConclusionsThis work may provide a convenient surrogate for the estimation of the left ventricular pressure waveform.

NABNet: A Nested Attention-guided BiConvLSTM network for a robust prediction of Blood Pressure components from reconstructed Arterial Blood Pressure waveforms using PPG and ECG signals

Backgroundand Motivations: Continuous Blood Pressure (BP) monitoring is crucial for real-time health tracking, especially for people with hypertension and cardiovascular diseases (CVDs). The current cuff-based BP monitoring methods are non-invasive but discontinuous while continuous BP monitoring methods are mostly invasive and can only be applied in a clinical setup to patients being monitored by advanced equipment and medical experts. Several studies have reported different techniques for predicting BP values from non-invasive Photoplethysmogram (PPG) and Electrocardiogram (ECG) signals. Apart from BP readings, estimating ABP waveforms from non-invasive signals can provide vital body parameters such as Mean Arterial Pressure (MAP) which can be used to determine poor organ perfusion, nutrient supply to organs, and cardiovascular diseases (CVDs), etc. MethodsIt is challenging to estimate ABP waveforms while maintaining a high BP prediction performance and ABP waveform pattern. In this work, we propose a novel approach for ABP waveform estimation by separating the task into BP prediction and a normalized ABP waveform estimation through segmentation from PPG, PPG derivatives, and ECG signals, and combining afterward. We propose the Nested Attention-guided BiConvLSTM Network or NABNet which uses LSTM blocks during segmentation for better handling of the existing phase shifts between PPG, ECG, and ABP signals. Several experiments were performed to improve the ABP reconstruction performance, which was combined with an existing BP prediction pipeline for the non-invasive estimation of ABP waveforms. ResultsThe proposed framework can robustly estimate ABP waveforms from PPG and ECG signals by reaching a high MAP performance and low construction error while maintaining the overall Grade A performance of the BP prediction pipeline. ConclusionLinearly translating the range-normalized, synthesized ABP segments by corresponding SBP and DBP predictions from the BP prediction pipeline managed to robustly estimate ABP waveforms from PPG and ECG signals.

Demonstration of static atomic gravimetry using Kalman filter

The measurement precision of the static atomic gravimetry is limited by white Gaussian noise in short term, which costs previous works an inevitable integration to reach the precision demanded. Here, we propose a statistical model based on the quantum projection noise and apply the associated Kalman filter with the waveform estimation in static atomic gravimetry. With the white Gaussian noise significantly removed by the Kalman-filter formalism, the measurement noise of the gravimetry is reshaped in short term and shows τ1/2 feature that corresponds to a random walk. During 200 h of static measurement of gravity, the atomic gravimeter using Kalman filter demonstrates a sensitivity as good as 0.6 nms2/s and highlights a precision of 1.7 nms2 at the measuring time of a single sample. The measurement noise achieved is also lower than the quantum projection limit below ∼30 s.

Respiratory Modulation of Sternal Motion in the Context of Seismocardiography

Wearable sensing has enabled physiological monitoring in everyday life. Clinically, cardiovascular and respiratory monitoring commonly requires separate devices. However, for feasibility in wearable systems, they are integrated into a single device. Since seismocardiography (SCG) records the motion of the chest wall, it has the potential to monitor both vital signs. The primary application of SCG is cardiovascular monitoring. As a result, its ability to detect respiration has been less defined. This study characterizes how SCG is affected by respiration and consequently, how respiratory metrics can be derived from these effects. The sternal motion of 40 subjects was recorded using an inertial measurement unit with a spirometer for reference. Three scenarios were examined with the subject at rest, holding their breath at peak inhalation, and holding their breath when exhaled. Three main aspects of respiratory modulation were explored: amplitude modulation, frequency modulation, and baseline wandering. The SCG amplitude was observed to be dependent on respiratory volume whereas baseline and frequency modulation were dependent on inhalation phase. All three effects were employed separately in the calculation of respiratory rate. Estimations from baseline wandering, amplitude modulation, and heart rate modulation produced r-squared values of 0.71, 0.44, and 0.66, respectively. The accuracy of tidal volume measurement was limited by a high inter-subject variability. Estimation of the respiratory waveform produced an average r-squared of 0.76 using multivariate linear regression. This demonstrates the potential of SCG as a tool for respiratory monitoring within an integrated, non-invasive cardiorespiratory monitoring system from a single point of contact.

Estimation of blood pressure waveform from facial video using a deep U-shaped network and the wavelet representation of imaging photoplethysmographic signals

Background:The remote measurement of physiological signals from video has gained a particular attention over the last past years. Estimating cardiovascular parameters like oxygen saturation and arterial blood pressure (BP) is covered by a limited volume of studies and remain a very challenging issue. Recent attempts demonstrated that BP can be estimated from facial video but under very controlled scenarios or with moderate performances. The data used in these works have not been publicly released or were gathered in a clinical setting. Methods:We, in contrast, propose a framework for estimating BP from publicly available data in order to allow replication and to facilitate fair comparison. We developed and trained a deep U-shaped neural network to recover the blood pressure waveform from its imaging photoplethysmographic (iPPG) signal counterpart. The model predicts the continuous wavelet transform (CWT) representation of a BP signal from the CWT of an iPPG signal. Inverse CWT transform is ultimately computed to recover the BP time series. Results:The proposed framework has been evaluated on 57 participants using international standards developed by the AAMI and the BHS. Results exhibit close agreement with ground truth BP values. The method satisfies all standards in the estimation of mean and diastolic BP (grade A) and nearly all standards in the estimation of systolic BP (grade B). Conclusions:This is, to the best of our knowledge, the first demonstration of a deep learning-oriented framework that manages to predict the continuous blood pressure waveform from facial video analysis. Codes developed during the study are publicly available (https://github.com/frederic-bousefsaf/ippg2bp).

A Review of Noninvasive Methodologies to Estimate the Blood Pressure Waveform.

Accurate estimation of blood pressure (BP) waveforms is critical for ensuring the safety and proper care of patients in intensive care units (ICUs) and for intraoperative hemodynamic monitoring. Normal cuff-based BP measurements can only provide systolic blood pressure (SBP) and diastolic blood pressure (DBP). Alternatively, the BP waveform can be used to estimate a variety of other physiological parameters and provides additional information about the patient’s health. As a result, various techniques are being proposed for accurately estimating the BP waveforms. The purpose of this review is to summarize the current state of knowledge regarding the BP waveform, three methodologies (pressure-based, ultrasound-based, and deep-learning-based) used in noninvasive BP waveform estimation research and the feasibility of employing these strategies at home as well as in ICUs. Additionally, this article will discuss the physical concepts underlying both invasive and noninvasive BP waveform measurements. We will review historical BP waveform measurements, standard clinical procedures, and more recent innovations in noninvasive BP waveform monitoring. Although the technique has not been validated, it is expected that precise, noninvasive BP waveform estimation will be available in the near future due to its enormous potential.

A periodic adaptive controller for the torque loop of variable speed brushless DC motor drives with non-ideal back-electromotive force

In this paper, an adaptive current controller is proposed for variable speed brushless direct current (BLDC) motor drives to minimize the output torque ripples caused by parametric and periodically varying uncertainties. Phase-to-phase non-ideal back-electromotive force (back-EMF) in BLDC motor changes periodically with respect to the shaft angle, and hence the period of these signals alters depending on the rotor frequency. To address these problems, the uncertain current dynamics of the BLDC motor is reformulated by transforming the time variable, then the periodic adaptive controller employing the instantaneous estimation values of the unknown periodic signal is developed to achieve the torque ripple reduction. The periodic estimation of the non-ideal back-EMF waveform is achieved considering the switching between conduction and commutation periods. Also, the update rules based on direct adaptation for parametric uncertainties are derived, and thus, hybrid differential-periodic adaptation rules are obtained considering the switching phenomenon. Asymptotic convergence of the phase currents to the reference values is proven by an appropriate Lyapunov–Krasovskii function depending on the angular position. Comprehensive numerical simulation studies have been successfully carried out to verify the performance and the effectiveness of the proposed controller for variable speed applications.

A strategy to personalize a 1D pulse wave propagation model for estimating subject-specific central aortic pressure waveform

The central aortic pressure (CAP) provides insights into the prediction, prevention, diagnosis, and treatment of cardiovascular disease, but can't be directly measured non-invasively. Therefore, the development of a noninvasive CAP estimation method based on the non-invasively measured peripheral pressure waveform is critical for clinical decisions based on the CAP. Some existing widely applied methods, such as the generalized transfer function (GTF) method relating measured peripheral pressure to the CAP, do not or only partly account for inter-subject or intra-subject variability of the cardiovascular system. To overcome this pitfall, we propose a subject-specific central aortic pressure estimation method in this paper. The novel method presented can derive an accurate aortic pressure from the peripheral pressure based on an individualized pulse wave propagation model using a GTF method as a first guess. To develop a strategy to personalize a pulse wave propagation model one usually needs to optimize many input parameters. Therefore, we present a two-step approach with the screening method of Morris and the adaptive sparse generalized polynomial chaos expansion (agPCE) algorithm for the sensitivity analysis of the wave propagation model. First, for a-priori defined output of the model, a subset of important parameters is identified using the screening method of Morris. Next, a quantitative variance-based sensitivity analysis is performed using agPCE. This approach is applied to a 1D pulse wave propagation model to get the personalized parameters of the pulse wave propagation model for the estimation of a subject-specific central aortic pressure waveform and is validated with 26 patients. Compared with the GTF method, the proposed method showed better performance in estimating the central aortic pulse wave and predicting the parameters.

Acoustic array signal processing with the Stockwell transform

The Stockwell transform is a wavelet-like transformation based on a Fourier kernel weighted by a symmetric, frequency-scaled, time-shifted window function. Thus, it is suitable for analysis of non-stationary waveforms and transients in particular. In its discrete-time, orthogonal basis realization known as the discrete orthogonal Stockwell transform (DOST), it is possible to transform an N-point sequence in O(N log N) operations. In this presentation, we show how to perform direction-of-arrival and transient waveform estimation in the Stockwell basis in a manner that is similar to traditional multi-channel frequency-domain (discrete Fourier transform) techniques. This enhances detection of multiple transients within the same data frame sequence.

Noninvasive estimation of aortic pressure waveform based on simplified Kalman filter and dual peripheral artery pressure waveforms

Background and ObjectiveAortic pressure (Pa) is important for the diagnosis of cardiovascular disease. However, its direct measurement is invasive, not risk-free, and relatively costly. In this paper, a new simplified Kalman filter (SKF) algorithm is employed for the reconstruction of the Pa waveform using dual peripheral artery pressure waveforms. MethodsPa waveforms obtained in a previous study were collected from 25 patients. Simultaneously, radial and femoral pressure waveforms were generated from two simulation experiments, using transfer functions. In the first, the transfer function is a known finite impulse response; and in the second, it is derived from a tube-load model. To analyze the performance of the proposed SKF algorithm, variable amounts of noise were added to the observed output signal, to give a range of signal-to-noise ratios (SNRs). Additionally, central aortic, brachial and femoral pressure waveforms were simultaneously collected from 2 Sprague-Dawley rats and the measured and reconstructed Pa waveforms were compared. ResultsThe proposed SKF algorithm outperforms canonical correlation analysis (CCA), which is the current state-of-the-art blind system identification method for the non-invasive estimation of central aortic blood pressure. It is also shown that the proposed SKF algorithm is more noise-tolerant than the CCA algorithm over a wide range of SNRs. ConclusionThe simulations and animal experiments illustrate that the proposed SKF algorithm is accurate and stable in the face of low SNRs. Improved methods for estimating central blood pressure as a measure of cardiac load adds to their value as a prognostic and diagnostic tool.

Impact of Respiratory Fluctuation on Hemodynamics in Human Cardiovascular System: A 0-1D Multiscale Model

To explore hemodynamic interaction between the human respiratory system (RS) and cardiovascular system (CVS), here we propose an integrated computational model to predict the CVS hemodynamics with consideration of the respiratory fluctuation (RF). A submodule of the intrathoracic pressure (ITP) adjustment is developed and incorporated in a 0-1D multiscale hemodynamic model of the CVS specified for infant, adolescent, and adult individuals. The model is verified to enable reasonable estimation of the blood pressure waveforms accounting for the RF-induced pressure fluctuations in comparison with clinical data. The results show that the negative ITP caused by respiration increases the blood flow rates in superior and inferior vena cavae; the deep breathing improves the venous return in adolescents but has less influence on infants. It is found that a marked reduction in ITP under pathological conditions can excessively increase the flow rates in cavae independent of the individual ages, which may cause the hemodynamic instability and hence increase the risk of heart failure. Our results indicate that the present 0-1D multiscale CVS model incorporated with the RF effect is capable of providing a useful and effective tool to explore the physiological and pathological mechanisms in association with cardiopulmonary interactions and their clinical applications.

Estimation of Blood Pressure Waveform from Facial Video Using a Deep U-Shaped Network and the Wavelet Representation of Imaging Photoplethysmographic Signals

Estimation of Blood Pressure Waveform from Facial Video Using a Deep U-Shaped Network and the Wavelet Representation of Imaging Photoplethysmographic Signals

Simultaneous adaption of the gain and phase of a generalized transfer function for aortic pressure waveform estimation

GoalThis paper proposes and validates a completely adaptive transfer function (CATF) based on an autoregressive exogenous (ARX) model which adjusts the gain and phase of a generalized transfer function (GTF) simultaneously to estimate the aortic pressure waveform from a brachial pressure waveform. MethodsInvasive aortic and brachial pressure waveforms were recorded from 34 subjects for the validation of the proposed method. Individual transfer functions (ITFs) were trained based on the pressure waveforms using an ARX model. The GTF was derived by averaging the ITFs. CATF was then obtained by adjusting both the gain and phase of the GTF using regression formulas calculated from the ITFs and brachial hemodynamic parameters. Meanwhile the quantitative contributions of the adaption of gain and phase of the GTF were investigated respectively. The root-mean-square-error of the total waveform and absolute errors of common hemodynamic indices including systolic and diastolic blood pressures (SBP and DBP, respectively), pulse pressure (PP) and augmentation index were used to evaluate the performance of the proposed method in the data divided into low, middle and high PP amplification groups. ResultsThe CATF achieved lower errors for DBP and PP in the low PP amplification group (1.79 versus 2.10 mmHg and 5.08 versus 6.23 mmHg, respectively, both P < 0.05) and PP in the middle amplification group (1.43 versus 1.92 mmHg, P < 0.05) compared with the GTF. SignificanceThe proposed method provides a step towards the development of an improved and clinically useful non-invasive approach for estimating the aortic pressure waveform from a peripheral pressure waveform.

Attack on PPG Biometrics: Presentation Attack by Stealth Recording and Waveform Estimation.

To develop a photoplethysmogram (PPG)-based authentication system with countermeasures, we investigate a "presentation attack" against the authentication. The attack uses the PPG for performing measurements on various sites on each subject's body. It records PPG on a nongenuine measurement site stealthily, generates a spoofing signal based on the recorded PPG, and transmits the signal to the authentication device. To investigate the feasibility of the attack, we developed a PPG-based authentication system. We recorded the PPGs of the subjects' bodies using the developed system and investigated the feasibility of attack in the experiment. The results indicated that an attack can occur with a probability of more than 80 % under ideal conditions.

IPPG 2 cPPG: Reconstructing contact from imaging photoplethysmographic signals using U-Net architectures

Imaging photoplethysmography (iPPG) is an optical technique dedicated to the assessment of several vital functions using a simple camera. Significant efforts have been made to reliably estimate heart and respiratory rates. Currently, research is focusing on the remote estimation of oxygen saturation and blood pressure (BP). The limited number of publicly available data tends to restrict the advancements related to BP estimation. To overcome this limit, we propose to split the problem in a two-stage processing chain: (i) converting iPPG to contact PPG (cPPG) signals using available video dataset and (ii) estimate BP from converted cPPG signals by exploiting large existing databases (e.g. MIMIC). This article presents the first developments where a method for converting iPPG signals measured using a camera into cPPG signals measured by contact sensors is proposed. Real and imaginary parts of the continuous wavelet transform (CWT) of cPPG and iPPG signals are passed to various deep pre-trained U-shaped architectures. Conventional metrics and specific waveform estimators have been implemented to validate the relevance of the predictions. The results exhibit good agreements towards a large portion of metrics, showing that the neural architectures properly estimated cPPG from iPPG signals through their CWT representations. The performance indicates that BP estimation from iPPG signals converted to cPPG signals can now be envisaged. Consequently, future work will focus on the integration of models dedicated to BP estimation trained on MIMIC. This is the first demonstration of a method for accurate reconstruction of cPPG from iPPG signals satisfying pulse waveform criteria.

Unifying frame rate and temporal dilations for improved remote pulse detection

Remote photoplethysmography (rPPG) is the monitoring of blood volume pulse from a camera at a distance. 3-Dimensional Convolutional Neural Networks (3DCNNs) have shown promising performance on the rPPG task, although it is critical that we understand the impact of both video and model parameters. In this paper, we explore the effect of frame rate, temporal kernel width, and – more generally – temporal receptive field on the reliability of heart rate and waveform estimation carried out by 3DCNNs. We train and evaluate 32 3DCNNs with different temporal parameters on a new large-scale database for physiological monitoring in an interview scenario. We show that previous studies reporting null effects of frame rate changes on pulse estimators may no longer be valid when using CNNs, and decreasing the frame rate may actually improve performance. In particular, we found that models trained on videos with frame rates as low as 12.9 frames per second (fps) perform better than those trained on videos recorded at a full 90 fps, perhaps due to the temporal receptive fields becoming larger in time dimension when the fps decreases. Using this insight, we propose RemotePulseNet, a novel 3DCNN architecture that exploits temporally dilated convolutions with increasing dilation rate to drastically increase the receptive field. We compare its performance with that of recent state-of-the-art pulse estimation methods, and show that both RemotePulseNet and the low frame rate 3DCNNs produce high-quality pulse signals from faces captured under a challenging interview scenario. The source code and instructions for obtaining a copy of the test data are made available with this paper.