Challenges of Developing Robust AI for Intrapartum Fetal Heart Rate Monitoring.

Background: CTG remains the only non-invasive tool available to the maternity team for continuous monitoring of fetal well-being during labour. Despite widespread use and investment in staff training, difficulty with CTG interpretation continues to be identified as a problem in cases of fetal hypoxia, which often results in permanent brain injury. Given the recent advances in AI, it is hoped that its application to CTG will offer a better, less subjective and more reliable method of CTG interpretation. Objectives: This mini-review examines the literature and discusses the impediments to the success of AI application to CTG thus far. Prior randomised control trials (RCTs) of CTG decision support systems are reviewed from technical and clinical perspectives. A selection of novel engineering approaches, not yet validated in RCTs, are also reviewed. The review presents the key challenges that need to be addressed in order to develop a robust AI tool to identify fetal distress in a timely manner so that appropriate intervention can be made. Results: The decision support systems used in three RCTs were reviewed, summarising the algorithms, the outcomes of the trials and the limitations. Preliminary work suggests that the inclusion of clinical data can improve the performance of AI-assisted CTG. Combined with newer approaches to the classification of traces, this offers promise for rewarding future development.

Bag of Patterns for Nearest Neighbour Neonatal EEG Recall.

Clinical neurophysiologists often find it difficult to recall rare EEG patterns despite the fact that this information could be diagnostic and help with treatment intervention. Traditional search methods may take time to retrieve the archived EEGs that could provide the meaning or cause of the specific pattern, which is undesirable as time can be critical for sick neonates. If neurophysiologists had the ability to quickly recall similar patterns, the prior occurrence of the pattern may help make an earlier diagnosis. This paper presents a system that may be used to assist a clinical neurophysiologist in the recall of neonatal EEG patterns. This paper compares two brute force approaches for the task of neonatal EEG recall and looks at the performance accuracy, speed and memory requirements. This system was tested on six different neonatal EEG pattern types with 430 events in total and the results are presented in this paper.

An Approximate Nearest Neighbour System For Neonatal EEG Recall.

Clinical neurophysiologists often find it difficult to recall rare EEG patterns despite the fact that this information could be diagnostic and help with treatment intervention. Traditional search methods may take time to retrieve the archived EEGs that could provide the meaning or cause of the specific pattern which is not acceptable as time can be critical for sick neonates. If neurophysiologists had the ability to quickly recall similar patterns, the prior occurrence of the pattern may help make an earlier diagnosis. This paper presents a system that may be used to assist a clinical neurophysiologist in the recall of neonatal EEG patterns. The proposed system consists of an alignment technique followed by an approximate nearest neighbour search algorithm called locality sensitive hashing. The system was tested on six different neonatal EEG pattern types with 430 events in total and the results are presented in this paper.

Application of audio fingerprinting to neonatal EEG.

A clinical neurophysiologist must recognize patterns in EEG signals to evaluate the health of a patient's brain activity. Rare or unusual patterns may take time to correctly identify. The ability to automatically assist this recall would be beneficial in ensuring that appropriate measures could be taken in a timely fashion. Audio fingerprinting is a method used to identify songs using only a snippet of the song. Fingerprints are extracted from a sub-section of the song and matched against a database of previously computed fingerprints. In this paper, a fingerprint quantization technique is implemented on neonatal EEG data to attempt to identify sections of EEG data when only seeing a sub-section of the data. The impact of signal distortions is investigated and results from a database of one hour recordings from 40 newborns are presented.

Stream computing for biomedical signal processing: A QRS complex detection case-study.

Recent developments in "Big Data" have brought significant gains in the ability to process large amounts of data on commodity server hardware. Stream computing is a relatively new paradigm in this area, addressing the need to process data in real time with very low latency. While this approach has been developed for dealing with large scale data from the world of business, security and finance, there is a natural overlap with clinical needs for physiological signal processing. In this work we present a case study of streams processing applied to a typical physiological signal processing problem: QRS detection from ECG data.

SoC-based architecture for biomedical signal processing.

Over the last decades, many algorithms have been proposed for processing biomedical signals. Most of these algorithms have been focused on the elimination of noise and artifacts existing in these signals, so they can be used for automatic monitoring and/or diagnosis applications. With regard to remote monitoring, the use of portable devices often requires a reduced number of resources and power consumption, being necessary to reach a trade-off between the accuracy of algorithms and their computational complexity. This paper presents a SoC (System-on-Chip) architecture, based on a FPGA (Field-Programmable Gate Array) device, suitable for the implementation of biomedical signal processing. The proposal has been successfully validated by implementing an efficient QRS complex detector. The results show that, using a reduced amount of resources, values of sensitivity and positive predictive value above 99.49% are achieved, which make the proposed approach suitable for telemedicine applications.

Automated detection of perturbed cardiac physiology during oral food allergen challenge in children.

This paper investigates the fully automated computer-based detection of allergic reaction in oral food challenges using pediatric ECG signals. Nonallergic background is modeled using a mixture of Gaussians during oral food challenges, and the model likelihoods are used to determine whether a subject is allergic to a food type. The system performance is assessed on the dataset of 24 children (15 allergic and 9 nonallergic) totaling 34 h of data. The proposed detector correctly classified all nonallergic subjects (100% specificity) and 12 allergic subjects (80% sensitivity) and is capable of detecting allergy on average 17 min earlier than trained clinicians during oral food challenges, the gold standard of allergy diagnosis. Inclusion of the developed allergy classification platform during oral food challenges recorded would result in a 30% reduction of doses administered to allergic subjects. The results of study introduce the possibility to halt challenges earlier which can safely advance the state of clinical art of allergy diagnosis by reducing the overall exposure to the allergens.

Robustness of time frequency distribution based features for automated neonatal EEG seizure detection.

In this paper we examined the robustness of a feature-set based on time-frequency distributions (TFDs) for neonatal EEG seizure detection. This feature-set was originally proposed in literature for neonatal seizure detection using a support vector machine (SVM). We tested the performance of this feature-set with a smoothed Wigner-Ville distribution and modified B distribution as the underlying TFDs. The seizure detection system using time-frequency signal and image processing features from the TFD of the EEG signal using modified B distribution was able to achieve a median receiver operator characteristic area of 0.96 (IQR 0.91-0.98) tested on a large clinical dataset of 826 h of EEG data from 18 full-term newborns with 1389 seizures. The mean AUC was 0.93.

An automated system for grading EEG abnormality in term neonates with hypoxic-ischaemic encephalopathy.

Automated analysis of the neonatal EEG has the potential to assist clinical decision making for neonates with hypoxic-ischaemic encephalopathy. This paper proposes a method of automatically grading the degree of abnormality in an hour long epoch of neonatal EEG. The automated grading system (AGS) was based on a multi-class linear classifier grading of short-term epochs of EEG which were converted into a long-term grading of EEG using a majority vote operation. The features used in the AGS were summary measurements of two sub-signals extracted from a quadratic time-frequency distribution: the amplitude modulation and instantaneous frequency. These sub-signals were based on a model of EEG as a multiplication of a coloured random process with a slowly varying pseudo-periodic waveform and may be related to macroscopic neurophysiological function. The 4 grade AGS had a classification accuracy of 83% compared to human annotation of the EEG (level of agreement, κ = 0.76). Features estimated on the developed sub-signals proved more effective at grading the EEG than measures based solely on the EEG and the incorporation of additional sub-grades based on EEG states into the AGS also improved performance.

The effects of lossy compression on diagnostically relevant seizure information in EEG signals.

This paper examines the effects of compression on EEG signals, in the context of automated detection of epileptic seizures. Specifically, it examines the use of lossy compression on EEG signals in order to reduce the amount of data which has to be transmitted or stored, while having as little impact as possible on the information in the signal relevant to diagnosing epileptic seizures. Two popular compression methods, JPEG2000 and SPIHT, were used. A range of compression levels was selected for both algorithms in order to compress the signals with varying degrees of loss. This compression was applied to the database of epileptiform data provided by the University of Freiburg, Germany. The real-time EEG analysis for event detection automated seizure detection system was used in place of a trained clinician for scoring the reconstructed data. Results demonstrate that compression by a factor of up to 120:1 can be achieved, with minimal loss in seizure detection performance as measured by the area under the receiver operating characteristic curve of the seizure detection system.

EEG-based neonatal seizure detection with Support Vector Machines.

OBJECTIVE: The study presents a multi-channel patient-independent neonatal seizure detection system based on the Support Vector Machine (SVM) classifier. METHODS: A machine learning algorithm (SVM) is used as a classifier to discriminate between seizure and non-seizure EEG epochs. Two post-processing steps are proposed to increase both the temporal precision and the robustness of the system. The resulting system is validated on a large clinical dataset of 267 h of EEG data from 17 full-term newborns with seizures. RESULTS: The performance of the system using event-based metrics is reported. The system showed the best up-to-date performance of a neonatal seizure detection system. The system was able to achieve an average good detection rate of ~89% with one false seizure detection per hour, ~96% with two false detections per hour, or ~100% with four false detections per hour. An analysis of errors revealed sources of misclassification in terms of both missed seizures and false detections. CONCLUSIONS: The results obtained with the proposed SVM-based seizure detection system allow for its practical application in neonatal intensive care units. SIGNIFICANCE: The proposed SVM-based seizure detection system can greatly assist clinical staff, in a neonatal intensive care unit, to interpret the EEG. The system allows control of the final decision by choosing different confidence levels which makes it flexible for clinical needs. The obtained results may provide a reference for future seizure detection systems.

Performance assessment for EEG-based neonatal seizure detectors.

OBJECTIVE: This study discusses an appropriate framework to measure system performance for the task of neonatal seizure detection using EEG. The framework is used to present an extended overview of a multi-channel patient-independent neonatal seizure detection system based on the Support Vector Machine (SVM) classifier. METHODS: The appropriate framework for performance assessment of neonatal seizure detectors is discussed in terms of metrics, experimental setups, and testing protocols. The neonatal seizure detection system is evaluated in this framework. Several epoch-based and event-based metrics are calculated and curves of performance are reported. A new metric to measure the average duration of a false detection is proposed to accompany the event-based metrics. A machine learning algorithm (SVM) is used as a classifier to discriminate between seizure and non-seizure EEG epochs. Two post-processing steps proposed to increase temporal precision and robustness of the system are investigated and their influence on various metrics is shown. The resulting system is validated on a large clinical dataset of 267h. RESULTS: In this paper, it is shown how a complete set of metrics and a specific testing protocol are necessary to extensively describe neonatal seizure detection systems, objectively assess their performance and enable comparison with existing alternatives. The developed system currently represents the best published performance to date with an ROC area of 96.3%. The sensitivity and specificity were ~90% at the equal error rate point. The system was able to achieve an average good detection rate of ~89% at a cost of 1 false detection per hour with an average false detection duration of 2.7 min. CONCLUSIONS: It is shown that to accurately assess the performance of EEG-based neonatal seizure detectors and to facilitate comparison with existing alternatives, several metrics should be reported and a specific testing protocol should be followed. It is also shown that reporting only event-based metrics can be misleading as they do not always reflect the true performance of the system. SIGNIFICANCE: This is the first study to present a thorough method for performance assessment of EEG-based seizure detection systems. The evaluated SVM-based seizure detection system can greatly assist clinical staff, in a neonatal intensive care unit, to interpret the EEG.

Parallel artefact rejection for epileptiform activity detection in routine EEG.

The EEG signal is very often contaminated by electrical activity external to the brain. These artefacts make the accurate detection of epileptiform activity more difficult. A scheme developed to improve the detection of these artefacts (and hence epileptiform event detection) is introduced. A structure of parallel Support Vector Machine classifiers is assembled, one classifier tuned to perform the identification of epileptiform activity, the remainder trained for the detection of ocular and movement-related artefacts. This strategy enables an absolute reduction in false detection rate of 21.6% with the constraint of ensuring all epileptic events are recognized. Such a scheme is desirable given that sections of data which are heavily contaminated with artefact need not be excluded from analysis.

Predicting the neurodevelopmental outcome in newborns with hypoxic-ischaemic injury.

The prediction of outcome in newborns with hypoxic ischemic encephalopathy (HIE) is a problematic task. Here, the ability of a combination of clinical, heart rate and EEG measures to predict outcome at 2 years is investigated. One hour of EEG and ECG recordings were obtained from newborns 24 hours after birth. Each newborn was reassessed at 24 months to investigate their neurodevelopmental outcome. From the EEG and ECG recordings, a set of 12 features was extracted. To classify each baby's outcome this data, along with clinical information was fed to a support vector machine. On a per patient basis an ROC area of 0.768 was achieved with 73.68% of newborns being assigned the correct outcome. Overall, this system presents a promising step towards the use of multimodal data for the prediction of neurodevelopmental outcome in newborns with HIE.

Energy efficient on-sensor processing in Body Sensor Networks.

Body Sensor Networks (BSNs) have tremendous potential in facilitating the real-time monitoring of the health of an individual in their own environment. However to truly exploit this potential, the powerful signal processing and analysis techniques available in the hospital environment must also be deployed in BSNs. In this paper, techniques in algorithm development, communications, hardware architecture and circuit design are described that will achieve the necessary power savings to make intelligent BSNs a reality.

The effect of lossy ECG compression on QRS and HRV feature extraction.

This paper describes the performance of beat detection and heart rate variability (HRV) feature extraction on electrocardiogram signals which have been compressed and reconstructed with a lossy compression algorithm. The set partitioning in hierarchical trees (SPIHT) compression algorithm was used with sixteen compression ratios (CR) between 2 and 50 over the records of the MIT/BIH arrhythmia database. Sensitivities and specificities between 99% and 85% were computed for each CR utilised. The extracted HRV features were between 99% and 82% similar to the features extracted from the annotated records. A notable accuracy drop over all features extracted was noted beyond a CR of 30, with falls of 10% accuracy beyond this compression ratio.

Speech recognition features for EEG signal description in detection of neonatal seizures.

In this work, features which are usually employed in automatic speech recognition (ASR) are used for the detection of neonatal seizures in newborn EEG. Three conventional ASR feature sets are compared to the feature set which has been previously developed for this task. The results indicate that the thoroughly-studied spectral envelope based ASR features perform reasonably well on their own. Additionally, the SVM Recursive Feature Elimination routine is applied to all extracted features pooled together. It is shown that ASR features consistently appear among the top-rank features.

Heart rate based automatic seizure detection in the newborn.

This work investigates the efficacy of heart rate (HR) based measures for patient-independent, automatic detection of seizures in newborns. Sixty-two time-domain and frequency-domain features were extracted from the neonatal heart rate signal. These features were classified using a sophisticated support vector machine (SVM) scheme. The performance was evaluated on a large dataset of 208 h from 14 newborn infants. It was shown that the HR can be useful for the detection of neonatal seizures for certain patients yielding an area under the receiver operating characteristic (ROC) curve of up to 82%. On evaluating the system using multiple patients an average ROC area of 0.59 with sensitivity of 60% and specificity of 60%, were obtained. Feature selection was performed and in the majority of patients the performance was degraded. Further analysis of the feature weights found significant variability in feature ranking across all patients. Overall, the patient-independent system presented here was seen to perform well in some patients (2 out of 14) but performed poorly when tested on the entire group.

Gaussian mixture models for classification of neonatal seizures using EEG.

A real-time neonatal seizure detection system is proposed based on a Gaussian mixture model classifier. The system includes feature transformation techniques and classifier output postprocessing. The detector was evaluated on a database of 20 patients with 330 h of recordings. A detailed analysis of the choice of parameters for the detector is provided. A mean good detection rate of 79% was obtained with only 0.5 false detections per hour. A thorough review of all misclassified events was performed, from which a number of patterns causing false detections were identified.

Heart rate variability during sleep in healthy term newborns in the early postnatal period.

Normative time- and frequency-domain heart rate variability (HRV) measures were extracted during quiet sleep (QS) and active sleep (AS) periods in 30 healthy babies. All newborn infants studied were less than 12 h old and the sleep state was classified using multi-channel video EEG. Three bands were extracted from the heart rate (HR) spectrum: very low frequency (VLF), 0.01-0.04 Hz; low frequency (LF), 0.04-0.2 Hz, and high frequency (HF), >0.2 Hz. All metrics were averaged across all patients and per sleep state to produce a table of normative values. A noticeable peak corresponding to activity in the RSA band was found in 80% patients during QS and 0% of patients during AS, although some broadband activity was observed. The majority of HRV metrics showed a statistically significant separation between QS and AS. It can be concluded that (i) activity in the RSA band is present during QS in the healthy newborn, in the first 12 h of life, (ii) HRV measures are affected by sleep state and (iii) the averaged HRV metrics reported here could assist the interpretation of HRV data from newborns with neonatal illnesses.