◊ ◊ ◊

“200 years of solicitude: Past successes and the future potential for fetal monitoring” – Gari Clifford, PhD

Although the notion that we can hear the sound of a fetus’ heartbeat, and that this is connected to its health has probably been known for much longer than the 400 years of written history on the subject, the origins of modern fetal monitoring could arguably be tied to a set of guidelines published in 1833 by Evory Kennedy, an English physician, for identifying fetal distress together with recommendations on how to auscultate the fetal heart rate to assist with intrapartum monitoring.
(Of course, others, both known and unknown, had suggested the notion before but did not publish guidelines – or at least any that survive to this day). Subsequent developments in the 19th Century around the Pinard horn and fetoscope increased wider adoption of these techniques and were followed by Cremer’s description of the abdominal and invasive fetal electrocardiogram in 1906 and Henry’s description of electronic fetal phonocardiography in 1931. Although Hon first described a system for capturing continuous fetal ECG in 1958, in 1964, electronic fetal monitoring took off with Doppler-based systems around the designs of Callagan.
This technology still forms the standard of care in medical practice today. However, despite the efforts by several organizations in the 1990s to standardize interpretation of fetal heart rate tracings and uterine contractions, there is scant evidence that continuous fetal heart rate monitoring does much more than increase C-section rates.
Nevertheless, there have been many developments over the last decade that provide a basis for optimism. In this talk, I will briefly frame this history, indicating what I think works, and is flawed in this field, and then make the classic mistake of all academics at their peak by prognosticating about the future of this field.

 

◊ ◊ ◊

“Principles of fetal electrocardiography” – Rik Vullings, PhD

Non-invasive fetal electrocardiography is studied to ultimately support gynaecologists in assessing the fetal condition. This method entails the measuring of signals that are a mixture of fetal electrocardiogram (ECG), maternal ECG, and other interferences and noise. To extract the fetal ECG, or alternatively suppress the interferences, various signal processing techniques have been proposed, some of which will be discussed during the lecture.
In many situations, these techniques enable the detection of fetal R-peaks and subsequent calculation of the fetal heart rate (FHR). Yet, we know that FHR alone often is not good enough to accurately assess the fetal condition and further supporting information is necessary. This information might come from analysis of the fetal ECG waveform. Before the fetal ECG waveform can be used, challenges relating to remaining noise and movement of the fetus need to be properly addressed. Some solutions to these challenges will be proposed, leading to an example usage of the fetal ECG for the detection of congenital heart disease.

 

◊ ◊ ◊

“Kalman filters basics”, “Kalman filters for fetal ECG processing” and LAB – Reza Sameni, PhD

The Kalman filter (KF) and its variants have been used for optimal state estimation and filtering for more than six decades. The essence of Kalman filtering is to “optimally” combine prior information about a system governed by linear/nonlinear dynamics driven by stochastic inputs, with posterior information obtained from noisy measurements of the system. The primary applications of the KF were mainly for navigational and aerospace systems with well-defined dynamics supported by laws of physics. However, as mathematical modeling of biological signals/systems became popular, KFs were also applied for filtering and estimation in biological systems. More recently, the notion of optimal combination of priors and posteriors has been found to be far more general than state estimation and optimal filtering; to the extent that in the modern Machine Learning era, many data fusion and learning algorithms are derived and interpreted from a Kalman filtering perspective.
In this lecture, we start by an overview of the “art of biological systems modeling.” We will see that mathematical models of biological systems are not unique and any model is only a simplification of the system, at a certain level of abstraction. Therefore, multiple mathematical models can coexist for a given biological system, each resulting in a different KF formulation.
The case study will be for the challenging problem of fetal electrocardiogram (fECG) extraction from noninvasive maternal abdominal recordings. We will show how one can mathematically model the signals recorded from the maternal body surface and design an Extended Kalman filter (EKF) for extracting the fECG from background noise and interferences, such as the maternal ECG. It is later shown how the combination of EKFs and blind source separation can be used to design a unified spatio-temporal filtering framework for fECG extraction and analysis. In the lab session, we will cover the implementation details and engineering aspects of KFs for biomedical applications in Matlab, including the model formation, KF parameter selection, sanity checking and temporal adaptation of the model/filter parameters.

 

◊ ◊ ◊

“Physiological basis and processing principles of heart rate variability” – Maria G. Signorini, PhD

The lesson presents the main characteristics of the heart rate signal (HRV).
From the extraction from the ECG signal by recognition of the R peak of the QRS complex the time series of variability is obtained. There are different ways of representing the time series in samples or in time.
The importance of the measurement of the HRV signal lies in the relationship between the variability of the heartbeat and the physiological systems that control its duration. The autonomic nervous system with its sympathetic and parasympathetic branches act dynamically modifying the duration of the heartbeat. This regulation is altered in the presence of pathological states.
The presentation shows how time and frequency domain analysis is able to quantify the contribution of ANS in the regulation of HRV. Examples of spectral estimation Non-parametric and Autoregressive show how it is possible to quantify these mechanisms and evaluate their dynamics in healthy and pathological subjects.
The lesson shows how it is possible to obtain the foetal heart rate signal from cardiotocographic recordings and shows some elaborations in the time domain, in the frequency domain and by extracting complexity parameters in normal fetuses and with pathology.
The availability of measurements in different domains allows to classify the HRV signal highlighting the physiological contributions of control. Reference will be made to machine learning approaches for classification

◊ ◊ ◊

“Fetal cardiac electrophysiology and pathology” – Tammo Delhaas, MD, PhD

There are not only significant developmental or age-related changes in the ionic currents that are responsible for the generation of the cardiac action potential, but also in the microscopic and macroscopic anatomic and neural substrates that govern the physiology of cardiac depolarization and repolarization. In this lecture, the physiology of both impulse formation and conduction within the developing heart is discussed.

 

◊ ◊ ◊

“Medico-legal aspects of perinatal asphyxia: a conundrum waiting for clinical and technical solutions” – Ernesto D’Aloja, MD, PhD

Care for newborn infants at risk for hypoxia ischemia is a priority in health care and understanding the pathophysiology of hypoxic ischemic brain injury is quite essential to the design of effective interventions. Before the advent of neonatal hypothermia, clinicians were not able to provide much care to neonates suffering from HIE besides systemic supportive care. Up to now, preventive perinatal measures to reduce the risk of a such devastating event seem to be the only way to pursue.
Neonatal HIE is an injury that occurs in the immature brain, resulting in delayed cell death via excitotoxicity, inflammation and oxidative stress. These adverse events in the developing brain often lead later on in life to long lasting detrimental neurological defects such as mental retardation, epilepsy, cerebral palsy, learning disabilities, and other neurophysiological handicaps.
Although major concerns have been expressed regarding the possibility of a misuse of the diagnostic label of HIE in a legal context, there is a real need to have an objective tool to discriminate between the hypoxic cause of neonatal encephalopathies and all the other possible
ones. As well known in forensic medicine, even a flawed instrument may be useful when you have no clue to decide if a damage is a direct/indirect consequence of a human behaviour. In this case, the results of the analysis should not be used to confirm the true asphyctic nature of the newborn neurological damage, but to rule out the possibility of such an eventuality. This approach – considering a ‘putative’ signature of asphyxia in the perinatal period reachable in a near future – may be of some help in denying the causal relationship between damage and professional’s (mis)conduct, even when a clear cause of damage has not been identified.
A robust negative prediction value of the analysis may help in the court not to wrongly convict health professional, while a positive prediction value – whenever achievable – has to satisfy a strict medico-legal exclusion criterion, being it only one evidence among the others.
The possibility to have a metabolomics snapshot at birth, before, during, and after the TH, will help all the expert witness examined in the court to ground their opinion on laboratory data, relying on the synergic effects of several metabolic pathways and not on a single metabolite or value (e.g., lactate and/or blood pH value at birth).

 

◊ ◊ ◊

“2020 Guidelines on Neonatal Resuscitation: what may be of interest to an
engineer?” –
Daniele Trevisanuto, MD, PhD

Worldwide, asphyxia accounts for about a quarter of 2.5 million neonatal deaths. In Europe, there are about 5,000-20,000 asphyxiated neonates per year, and approximately one-half of them die or suffer neurodevelopmental impairment. Providing appropriate care at birth remains a crucial intervention for reducing neonatal mortality and morbidity.
A new edition of the guidelines on neonatal resuscitation is published every 5 years. The objective of this talk is to present the 2020 edition of the American Heart Association guidelines, to revise the evidence behind each recommendation and to highlight the gaps of knowledge. New technologies, such as pulse-oxymetry, ECG, CO2 detector, are now standard of care for the management of the newborn infants at birth. Others (i.e. respiratory function monitoring, near infrared spectrography, EEG) have been assessed and their impact on the relevant neonatal outcomes needs to be confirmed. New technologies (i.e. high fidelity manikins) have been also successful developed with the purpose to improve education of health care staff.
A strict collaboration between clinical engineers and healthcare providers may significantly improve antenatal, perinatal and postnatal monitoring of the fetus/neonate; technical and not technical skills of healthcare staff can also benefit from this collaboration.

 

◊ ◊ ◊

“Fetal growth monitoring and issues” – Silvia Visentin, MD, PhD

Fetal growth monitoring during pregnancy has been an important practice amongst obstetricians usually done to verify the health status of a fetus.
The monitoring possibilities come from the use of ultrasound, cardiotocography, and still experimental surveillance systems. One of the main purposes of their use is the identification of some fetal problems that could put their health at risk, both during pregnancy and in labor. They can also be used to verify the better timing of delivery in presence of pathological conditions, such as intrauterine growth retardation. This is a problem that is characterized, in addition to growth below the fetal potential, by an impairment of the cardiovascular and neurological system, which can be evidenced by a cardiac performance that is subclinically lower than expected and a suboptimal yield during the school age.
The in utero study of cardiac strain and neural networks can select a category of infants and therefore children who can benefit from early medical interventions, aimed at reducing the long-term effects of what is believed to be fetal programming.

 

◊ ◊ ◊

“Nonlinear-type time-frequency analysis and its application to fetal ECG” – Hau-Tieng Wu, MD, PhD

I will discuss a signal processing tool for the mission of extracting fetal ECG from a few channels of trans-abdominal maternal ECGs. The mission could be understood as the single channel blind source separation (scBSS). The algorithm for scBSS combines de-shape algorithm, random matrix and diffusion geometry. The developed tool could be applied to other biomedical time series other than fetal ECG, particularly when the single (or few) channel time series is composed of multiple oscillatory components with complicated statistical features, like time-varying amplitude, frequency and non-sinusoidal pattern, and the signal quality is compromised by nonstationary noise and artifact. If time permits, I will go through the theoretical support of the algorithm based on harmonic analysis, differential geometry and random matrix theory, and discuss the recently developed Ramanujan-based approach and its application to other biomedical signals.

 

◊ ◊ ◊

“Fundamentals of antepartum and intrapartum fetal monitoring” – Salvatore Tagliaferri, MD, PhD

The electronic Fetal Heart Rate (FHR) monitoring is one of the most widespread not invasive method to evaluate the fetal well-being during the antepartum and intrapartum period, especially in pregnancies complicated by Fetal Growth Restriction. The antepartum monitoring is performed starting from the 38th week of gestation once a week in healthy pregnancies, but it can be started before and repeated more frequently in case of high-risk pregnancies. The cardiotocograph allows obtaining two tracings that can be recorded simultaneously on the same strip of paper. A centimeter scale is shown on the horizontal axis of the trace. The FHR values are highlighted on the upper portion of the trace (range of 50-210 beats per minute). On the lower portion, uterine activity is recorded, measured in Relative Units (0-100). The presence of significant beat to beat variation suggests intact baroreflex, sympathetic/parasympathetic tone and central control indicating normal central nervous system (CNS) responsiveness and normal local CNS metabolic environment reflecting fetal health. The Short Term Variability is the most significant indicator of fetal homeostasis, especially when it is compared to long and medium term variability: normal STV values reflect a healthy ANS, normal activity of chemoreceptors, baroreceptors and cardiac responsiveness, while low STV values are associated with impending deterioration of fetal oxygen supply and therefore fetal distress.
The relationship between cardiotocographic parameters and fetal pH resembles the observations made for fetal Doppler vessels, as each individual component (tone, gross body movement, breathing movement, amniotic fluid volume, and heart rate reactivity) is independently altered by hypoxemia but their combined consideration in a composite score best predicts pH and fetal outcome at birth.

 

 

◊ ◊ ◊

“Gaussian Processes”, “Deep learning” and LAB – Petar M. Djurić, PhD

Gaussian processes – During labor, fetal heart rate (FHR) and uterine activity (UA) are continuously monitored by obstetricians for assessing the fetal well-being. There are also a number of machine learning methods that are applied to these signals with the objective of extracting information from the signals that would be of help to obstetricians in making timely and good decisions. In this lecture, we describe a class of machine learning methods known as Gaussian processes. They are nonparametric in nature and represent a flexible Bayesian machinery for learning functions or mappings from data. They are data driven, data-efficient, and can handle uncertainties in a principled way. In the lecture, we also explain how they can be used for computerized analysis of FHR and UA signals.

Deep Gaussian processes – Deep learning methods are typically based on artificial neural networks whose architectures comprise multiple layers. Similar to deep neural networks, one can also create deep learning methods based on sequences of Gaussian processes (GPs). These deep learning methods are known as deep Gaussian processes (DGPs). DGPs have many attractive features. They can be viewed as multi-layer hierarchical generalizations of GPs, and they are formally equivalent to neural networks with multiple, infinitely wide hidden layers. In this lecture, we introduce DGPs and explain supervised and unsupervised learning by DGPs for classification of fetal heart rate tracings.

 

◊ ◊ ◊

“Non-invasive fetal ECG source separation methods” – Joachim Behar, PhD

In this presentation we will be talking about blind source separation (BSS): separation of a set of source signals from a set of mixed signals, without the aid of information (or with very little information) about the source signals or the mixing process.
Such methods were used in the field of non-invasive fetal ECG (NI-FECG) analysis using principal component analysis (Kanjilal et al 1997) and independent component analysis (ICA) (De Lathauwer et al 2000). In essence, these approaches aim to separate the underlying statistically independent sources into three categories: maternal ECG, fetal ECG and noise. In this lecture we will cover the mathematical basis of PCA and ICA, how these algorithms may be used in the context of NI-FECG analysis and methods for assessing the relative performance of these algorithms.

 

◊ ◊ ◊

“Big data for clinical decision-support in labour: challenges and future directions”Antoniya Georgieva, PhD

How can we use routinely collected cardiotocography (CTG) data during labour for research? What outcome should we use to train our algorithms? What is the role of clinical risk factors? Do healthcare professionals want new technology in delivery wards? How do we ensure they want and use well new technologies?
This lecture will attempt to address the above questions, based on our extensive experience working with all routinely collected CTGs at Oxford since 1993.

 

 

◊ ◊ ◊

” Intrapartum cardiotocography (CTG) feature detection and fetal state estimation using signal processing and machine learning” – Philip Warrick, PhD

This talk will focus on automated approaches to interpreting intrapartum cardiotocography (CTG). The goal is to assist clinical staff in making better obstetrical decisions from the information contained in the two acquired CTG signals: uterine pressure (UP) and fetal heart rate (FHR). The methods and rationale for acquiring these signals will be discussed as well as the accumulated clinical consensus on their most informative aspects. Key signal preprocessing steps will be introduced that account for such artifacts as missing signal and maternal heart rate interference. Two interpretation problems will be addressed within a machine learning context. The first is the automated detection of the key clinical features such as FHR baseline, acceleration and deceleration. The second addresses the key rationale for intrapartum CTG and its most challenging aspect, that is, inference on the fetal state that is timely and accurate. Relevant machine learning challenges will be discussed including signal representation; learning strategies, with a focus on neural network approaches; and databases and the associated curation of ground truths. Finally, we will assess the state of the art, the limitations of such interpretation and consider future directions.

 

◊ ◊ ◊

“Advantages and limitations of non-linear HRV indexes” – Roberto Sassi, PhD

During labor, the fetus faces repetitive hypoxic insults due to maternal uterine contractions, possibly leading to acidemia (lack of oxygen) and brain injury. Acidemia at birth can be quantified via pH measurements from the umbilical cord of the newborn. However, such post-partum measurements cannot be used to help clinicians identifying compromised fetuses. Fetal heart rate (FHR) is one of the few available measurements (sometimes the only one) to monitor the fetal well-being during labor. The quantification of FHR variability (FHRV) has been widely studied for the prompt identification of acidemia intra-partum.
During the lecture, we will review the use of non-linear time-series metrics employed to quantify FHRV.
After a general introduction, listing the possible approaches, we will discuss a few techniques commonly employed for FHRV quantification. In particular, we will start dealing with the definition of the entropy of a time series. Then we will review Phase Rectified Signal Averaging (PRSA), a technique which is robust to phase desynchronizations, which are common in FHR series due to the low signal quality. Finally, we will discuss applications of entropy and PRSA-based indexes to the identification of fetal acidemia during labor and to the monitoring of the growth of the fetus during pregnancy.