Philip de Chazal

Automatic classification of heartbeats using ECG morphology and heartbeat interval features

A method for the automatic processing of the electrocardiogram (ECG) for the classification of heartbeats is presented. The method allocates manually detected heartbeats to one of the five beat classes recommended by ANSI/AAMI EC57:1998 standard, i.e., normal beat, ventricular ectopic beat (VEB), supraventricular ectopic beat (SVEB), fusion of a normal and a VEB, or unknown beat type. Data was obtained from the 44 nonpacemaker recordings of the MIT-BIH arrhythmia database. The data was split into two datasets with each dataset containing approximately 50 000 beats from 22 recordings. The first dataset was used to select a classifier configuration from candidate configurations. Twelve configurations processing feature sets derived from two ECG leads were compared. Feature sets were based on ECG morphology, heartbeat intervals, and RR-intervals. All configurations adopted a statistical classifier model utilizing supervised learning. The second dataset was used to provide an independent performance assessment of the selected configuration. This assessment resulted in a sensitivity of 75.9%, a positive predictivity of 38.5%, and a false positive rate of 4.7% for the SVEB class. For the VEB class, the sensitivity was 77.7%, the positive predictivity was 81.9%, and the false positive rate was 1.2%. These results are an improvement on previously reported results for automated heartbeat classification systems.

Combination of EEG and ECG for improved automatic neonatal seizure detection.

OBJECTIVE Neonatal seizures are the most common central nervous system disorder in newborn infants. A system that could automatically detect the presence of seizures in neonates would be a significant advance facilitating timely medical intervention. METHODS A novel method is proposed for the robust detection of neonatal seizures through the combination of simultaneously-recorded electroencephalogram (EEG) and electrocardiogram (ECG). A patient-specific and a patient-independent system are considered, employing statistical classifier models. RESULTS Results for the signals combined are compared to results for each signal individually. For the patient-specific system, 617 of 633 (97.52%) expert-labelled seizures were correctly detected with a false detection rate of 13.18%. For the patient-independent system, 516 of 633 (81.44%) expert-labelled seizures were correctly detected with a false detection rate of 28.57%. CONCLUSIONS A novel algorithm for neonatal seizure detection is proposed. The combination of an ECG-based classifier system with a novel multi-channel EEG-based classifier system has led to improved seizure detection performance. The algorithm was evaluated using a large data-set containing ECG and multi-channel EEG of realistic duration and quality. SIGNIFICANCE Analysis of simultaneously-recorded EEG and ECG represents a new approach in seizure detection research and the detection performance of the proposed system is a significant improvement on previous reported results for automated neonatal seizure detection.

Sleep/wake measurement using a non-contact biomotion sensor

We studied a novel non‐contact biomotion sensor, which has been developed for identifying sleep/wake patterns in adult humans. The biomotion sensor uses ultra low‐power reflected radiofrequency waves to determine the movement of a subject during sleep. An automated classification algorithm has been developed to recognize sleep/wake states on a 30‐s epoch basis based on the measured movement signal. The sensor and software were evaluated against gold‐standard polysomnography on a database of 113 subjects [94 male, 19 female, age 53 ± 13 years, apnoea–hypopnea index (AHI) 22 ± 24] being assessed for sleep‐disordered breathing at a hospital‐based sleep laboratory. The overall per‐subject accuracy was 78%, with a Cohen’s kappa of 0.38. Lower accuracy was seen in a high AHI group (AHI >15, 63 subjects) than in a low AHI group (74.8% versus 81.3%); however, most of the change in accuracy can be explained by the lower sleep efficiency of the high AHI group. Averaged across subjects, the overall sleep sensitivity was 87.3% and the wake sensitivity was 50.1%. The automated algorithm slightly overestimated sleep efficiency (bias of +4.8%) and total sleep time (TST; bias of +19 min on an average TST of 288 min). We conclude that the non‐contact biomotion sensor can provide a valid means of measuring sleep–wake patterns in this patient population, and also allows direct visualization of respiratory movement signals.

Registration of digital retinal images using landmark correspondence by expectation maximization

A method for registering pairs of digital images of the retina is presented, using a small set of intrinsic control points whose matching is not known. Control point matching is then achieved by calculating similarity transformation (ST) coefficients for all possible combinations of control point pairs. The cluster of coefficients associated with the matched control point pairs is identified by calculating the Euclidean distance between each set of ST coefficients and its Rth nearest neighbour, followed by use of the Expectation ‐ Maximization (EM) algorithm. Registration is then achieved using linear regression to optimize similarity, bilinear or second order polynomial transformations for the matching control point pairs. Results are presented of (a) the cross-modal image registration of an optical image and a fluorescein angiogram, (b) temporal registration of two images of an infant eye, and (c) mono-modal registration of a set of seven standard field optical photographs. For cross-modal registration, using a set of independent matched control points, points are mapped with an estimated accuracy of 2.9 pixels for 575 £ 480 pixel images. Bilinear and second-order polynomial transformation models all prove to be appropriate for the final registration transform. q 2004 Elsevier B.V. All rights reserved.

Person identification using automatic integration of speech, lip, and face experts

This paper presents a multi-expert person identification system based on the integration of three separate systems employing audio features, static face images and lip motion features respectively. Audio person identification was carried out using a text dependent Hidden Markov Model methodology. Modeling of the lip motion was carried out using Gaussian probability density functions. The static image based identification was carried out using the FaceIt system. Experiments were conducted with 251 subjects from the XM2VTS audio-visual database. Late integration using automatic weights was employed to combine the three experts. The integration strategy adapts automatically to the audio noise conditions. It was found that the integration of the three experts improved the person identification accuracies for both clean and noisy audio conditions compared with the audio only case. For audio, FaceIt, lip motion, and tri-expert identification, maximum accuracies achieved were 98%, 93.22%, 86.37% and 100% respectively. Maximum bi-expert integration of the two visual experts achieved an identification accuracy of 96.8% which is comparable to the best audio accuracy of 98%.

SleepMinder: An innovative contact-free device for the estimation of the apnoea-hypopnoea index

We describe an innovative sensor technology (SleepMinder™) for contact-less and convenient measurement of sleep and breathing in the home. The system is based on a novel non-contact biomotion sensor and proprietary automated analysis software. The biomotion sensor uses an ultra low-power radio-frequency transceiver to sense the movement and respiration of a subject. Proprietary software performs a variety of signal analysis tasks including respiration analysis, sleep quality measurement and sleep apnea assessment. This paper measures the performance of SleepMinder as a device for the monitoring of sleep-disordered breathing (SDB) and the provision of an estimate of the apnoea-hypopnoea index (AHI). The SleepMinder was tested against expert manually scored PSG data of patients gathered in an accredited sleep laboratory. The comparison of SleepMinder to this gold standard was performed across overnight recordings of 129 subjects with suspected SDB. The dataset had a wide demographic profile with the age ranging between 20 and 81 years. Body weight included subjects with normal weight through to the very obese (Body Mass Index: 21-44 kg/m2). SDB severity ranged from subjects free of SDB to those with severe SDB (AHI: 0.8-96 events/hours). SleepMinders AHI estimation has a correlation of 91% and can detect clinically significant SDB (AHI>15) with a sensitivity of 89% and a specificity of 92%.

Assessment of sleep/wake patterns using a non-contact biomotion sensor

We evaluate a contact-less continuous measuring system measuring respiration and activity patterns system for identifying sleep/wake patterns in adult humans. The system is based on the use of a novel non-contact biomotion sensor, and an automated signal analysis and classification system. The sleep/wake detection algorithm combines information from respiratory frequency, magnitude, and movement to assign 30 s epochs to either wake or sleep. Comparison to a standard polysomnogram system utilizing manual sleep stage classification indicates excellent results. It has been validated on overnight studies from 12 subjects. Wake state was correctly identified 69% and sleep with 88%. Due to its ease-of-use and good performance, the device is an excellent tool for long term monitoring of sleep patterns in the home environment in an ultraconvenient fashion.

Multimodal detection of sleep apnoea using electrocardiogram and oximetry signals

A method for the detection of sleep apnoea, suitable for use in the home environment, is presented. The method automatically analyses night-time electrocardiogram (ECG) and oximetry recordings and identifies periods of normal and sleep-disordered breathing (SDB). The SDB is classified into one of six classes: obstructive, mixed and central apnoeas, and obstructive, mixed and central hypopnoeas. It also provides an estimated apnoea, hypopnoea and apnoea–hypopnoea index. The basis of the method is a pattern recognition system that identifies episodes of apnoea by analysing the heart variability, an ECG-derived respiration signal and blood oximetry values. The method has been tested on 183 subjects with a range of apnoea severities who have undergone a full overnight polysomnogram study. The results show that the method separates control subjects from subjects with clinically significant sleep apnoea with a specificity of 83 per cent and sensitivity of 95 per cent. These results demonstrate that home-based screening for sleep apnoea is a viable alternative to hospital-based tests with the added benefit of low cost and minimal waiting times.

Assessment of sleep-disordered breathing using a non-contact bio-motion sensor

Obstructive sleep apnoea is a highly prevalent but under‐diagnosed disorder. The gold standard for diagnosis of obstructive sleep apnoea is inpatient polysomnography. This is resource intensive and inconvenient for the patient, and the development of ambulatory diagnostic modalities has been identified as a key research priority. SleepMinder (BiancaMed, NovaUCD, Ireland) is a novel, non‐contact, bedside sensor, which uses radio‐waves to measure respiration and movement. Previous studies have shown it to be effective in measuring sleep and respiration. We sought to assess its utility in the diagnosis of obstructive sleep apnoea. SleepMinder and polysomnographic assessment of sleep‐disordered breathing were performed simultaneously on consecutive subjects recruited prospectively from our sleep clinic. We assessed the diagnostic accuracy of SleepMinder in identifying obstructive sleep apnoea, and how SleepMinder assessment of the apnoea–hypopnoea index correlated with polysomnography. Seventy‐four subjects were recruited. The apnoea–hypopnoea index as measured by SleepMinder correlated strongly with polysomnographic measurement (r = 0.90; P ≤ 0.0001). When a diagnostic threshold of moderate–severe (apnoea–hypopnoea index ≥15 events h−1) obstructive sleep apnoea was used, SleepMinder displayed a sensitivity of 90%, a specificity of 92% and an accuracy of 91% in the diagnosis of sleep‐disordered breathing. The area under the curve for the receiver operator characteristic was 0.97. SleepMinder correctly classified obstructive sleep apnoea severity in the majority of cases, with only one case different from equivalent polysomnography by more than one diagnostic class. We conclude that in an unselected clinical population undergoing investigation for suspected obstructive sleep apnoea, SleepMinder measurement of sleep‐disordered breathing correlates significantly with polysomnography.

Non-episode-dependent assessment of paroxysmal atrial fibrillation through measurement of RR interval dynamics and atrial premature contractions.

A method is presented for classifying a single lead surface electrocardiogram recording from a Holter monitor as being from a subject with paroxysmal atrial fibrillation (PAF) or not. The technique is based on first assessing the likelihood of 30-min segments of electrocardiogram (ECG) being from a subject with PAF, and then combining these per-segment likelihoods to form a per-subject classification. The per-segment assessment is based on the output of a supervised linear discriminant classifier (LDC) which has been trained using known data from the Physionet Atrial Fibrillation Prediction Database (which consists of two hundred 30-min segments of Holter ECG, taken from 53 subjects with PAF, and 47 without). One of two LDCs is used depending on whether there is a significant correlation between observed low-frequency and high-frequency spectral power in the RR power spectral density over the 30-min segment. If there is high correlation, then the LDC uses spectral features calculated over a 10-min window; in the low-correlation case, both spectral features and atrial premature contractions are used as features. The classifier was tested for its ability to distinguish PAF and non-PAF segments using three independent data sets (representing a total of 1370 segments from 50 subjects). The cumulative sensitivity, specificity, and accuracy on a per-segment basis were 43.0, 99.3, and 80.5%, respectively on these independent test sets. By combining the results of segment classification, a per-subject classification into PAF and non-PAF classes was performed. For the 50 subjects in the independent data sets, the sensitivity and specificity of the per-subject classifier were 100%.