Paper Submission & Registration
8th Dutch Bio-Medical Engineering Conference
12:30   Pregnancy & Neonates
Chair: Rik Vullings
15 mins
Prediction of successful fertilization by machine learning applied to uterine motion estimated by ultrasound speckle tracking
Yizhou Huang, Federica Sammali, Tom Bakkes, Celine Blank, Benedictus Schoot, Massimo Mischi
Abstract: Infertility is emerging as a serious problem in developed countries, where up to 20% of women in reproductive age have difficulty to get pregnant. In-vitro fertilization (IVF) represents the most advanced treatment against infertility. Yet, its success rate remains below 30%. Dysfunction of the uterine activity is believed to affect the success rate of embryo implantation. In this study, we investigate novel uterine motion features and the use of machine learning for probabilistic classification of the uterine activity as either favorable or adverse to embryo transfer (ET). Uterine activity was measured by transvaginal ultrasound (TVUS) in 16 patients undergoing a complete IVF cycle. In this study, we analyzed the measurements taken one hour before ET. Patients were divided in ongoing pregnancy and non-ongoing pregnancy groups based on follow-up TVUS examination. Acquisition of 4-min loops was performed with a WS80A ultrasound scanner (Samsung-Medison) using a transvaginal V5-9 probe. A set of features, such as contraction frequency (CF), standard deviation, and unnormalized first moment, etc. were extracted from the motion signals derived by a dedicated speckle tracking algorithm. Additional features, such as velocity, direction, and coordination of uterine contraction (UC), were extracted from the strain rate by k-space analysis of the spatiotemporal strain distribution measured on both sides of the endometrial cavity. In particular, UC coordination was determined by measuring the similarity in UC direction between both sides of the endometrium cavity using two metrics, namely, mean squared error and cross correlation (CC). Three classifiers, namely, support vector machine, K-nearest neighbors (KNN), and Gaussian mixture model were considered in this study to classify successful and unsuccessful embryo implantation using the extracted features. The proposed classifiers were tested and trained in a nested cross validation loop to avoid overfitting. A full-grid search was adopted for hyperparameter optimization. The classifiers were validated with a leave-one-out approach. Accuracy was used as performance metric. The best classification accuracy (93.8%) was obtained by KNN. CF and coordination of UC based on CC metric were best predictors for embryo implantation. A larger dataset is required in the future to improve the accuracy, robustness, and generalizability of the classifiers.
15 mins
Quantification of dysmature EEG for the assessment of perinatal stress in preterm infants
Laura Smets
Abstract: Introduction Premature infants are at risk to undergo an altered development [1] of which the reason is twofold: (1) They are staying in the Neonatal Intensive Care Unit (NICU) right after birth which is a crucial period for the growth and development of the brain and central nervous system [2]. (2) Preterm infants are experiencing several stressors in the NICU such as maternal separation, noxious stimuli (e.g. bright light and loud noise) and several painful invasive procedures (e.g. heel prick and mechanical ventilation). The accumulation of these early-life experiences leads to an early exposure to stress, perinatal stress [3]. However, no adequate automated method exists to non-invasively quantify and classify perinatal stress in preterm infants by using physiological signal background. Materials and Methods EEG and ECG measurements of 136 preterm infants were recorded for at least three hours at the NICU in the University Hospitals Leuven. The Leuven Pain Scale (LPS) was used to assess their pain levels. Features were extracted from the EEG and HRV signals for quiet sleep (QS) and non-quiet sleep (nQS) segments and were used for the development of a subspace linear discriminant analysis stress classifier. Results and Discussion Different stress classifiers were developed for different age groups and stress intensities which were found to have an area under the curve of [0.72-0.93] for nQS and [0.77-0.96] for QS. The Cohen’s kappa score showed a moderate association between the predicted and true labels which is lower compared to other studies about pain classification. However, these studies elicit pain, whereas this study is based on an unobtrusive approach for the detection of perinatal stress. Further, a lower EEG complexity and a persistence of slow wave activity was found in patients under stress, indicating a dysmature EEG. Also an increased cortical connectivity and brain-heart interaction were associated with stress exposure. Autonomic activity on the other hand did not show an association with stress exposure. These results indicate that an autonomic tool to investigate dysmature EEG could not only be used to assess the brain development, but also to assess stress exposure in preterm infants. REFERENCES [1] J. Trickett, S. Johnson, and D. Wolke, “Behavioural and Educational Outcomes Following Extremely Preterm Birth: Current Controversies and Future Directions,” in Emerging Topics and Controversies in Neonatology, Springer Nature Switzerland, 2020, pp. 367–385. [2] M. Lavanga et al., “A perinatal stress calculator for the neonatal intensive care unit: An unobtrusive approach,” Physiol. Meas., vol. 41, no. 7, p. 75012, Jul. 2020, doi: 10.1088/1361-6579/ab9b66. [3] R. E. Grunau et al., “Neonatal procedural pain exposure predicts lower cortisol and behavioral reactivity in preterm infants in the NICU,” Pain, vol. 113, pp. 293–300, 2005, doi: 10.1016/j.pain.2004.10.020.
15 mins
Happy measure! Stereoscopic vision body length measuring instrument for premature infants inside incubators
Ronald van Gils, Onno Helder, Linda Wauben, Timothy Singowikromo
Abstract: Background Body length and head circumference measurement of premature infants is essential for growth monitoring. But current measuring instruments cause so much stress to premature infants, that length or head circumference measurements of the most vulnerable, unstable infants are simply skipped. Aim Our aim is to develop ‘touchless’, non-disturbing body size measuring instruments that can measure premature infants inside incubators without causing stress to the infant. Methods Multiple device development and validation research tracks have been initiated. One of these tracks is the stereoscopic vision instrument, based on Sokolover e.a. (2014). Through co-design with healthcare professionals of the neonatal intensive care unit (NICU) and medical device developers of the Erasmus MC(-Sophia Children’s) Hospital, students of computer engineering and industrial design engineering developed a functional prototype of the stereoscopic vision instrument. Results The stereoscopic vision instrument captures two digital camera images of the infant from different viewing angles. On both images body points are marked manually: head-end, neck, hip, knee and heel. A touch-screen user-interface was developed, including a screen magnifier to mark body points accurately and fast. The (OpenCV) stereoscopic algorithm calculates the corresponding points in 3D-space and the distance between. The software is developed as an open-source software application. A housing was designed, integrating all hardware in a stand-alone medical device. Conclusions The stereoscopic instrument can measure the total body length without stretching the infant’s legs. The instrument can measure with the required accuracy of 1 mm. Further clinical validation is needed. Next step is adding 2D image recognition to automatically preselect body points.
15 mins
Neurocardiovascular coupling revealing insight in perioperative neonatal physiology
Dries Hendrikx
Abstract: Background. Cerebral blood flow is highly regulated by numerous regulation mechanisms, including cerebral autoregulation and neurovascular coupling [1]. Computational approaches to measure the status of these regulation mechanisms are currently being developed. In this context, signal interaction graphs prove to be a powerful tool, as they allow to create an overview of the coordinated interaction between the brain and the cardiovascular and pulmonary and as such, they allow to study numerous regulation mechanisms at once [2]. Aim. The aim of this study is to apply the innovative integrative brain monitoring approach, based on signal interaction graphs, to study the (patho)physiology of neonates perioperatively. Methods. After exclusion, 37 neonates were enrolled in the study, which were consequently stratified into 5 groups, based on clinical relevance. The signal interaction graph has 7 nodes: heart rate; mean arterial blood pressure; arterial saturation; two cerebral oxygenation channels, measured using near-infrared spectroscopy; and two EEG channels. The interaction between the different signals in the signal interaction graph was quantified using normalized transfer entropy, which allows to capture nonlinear, directed signal coupling. Results. No significant effect of gestational age (p = 0.5478), birth weight (p = 0.5718), sex (p = 0.1187), position of the liver (p = 0.1965), and the size of the defect (p = 0.1383) was found on the connectivity of the graphs. The clinical setting (preoperative, intraoperative, and postoperative) (p < 0.0001) and the clinical group had a significant effect on the graph connectivity (p < 0.0001). Discussion. The physiological differences between the clinical groups, as well as the differences throughout the perioperative period were observed to be reflected in the neurocardiovascular graph. Acute events, such as cardiopulmonary resuscitation and venoarterial extracorporeal membrane oxygenation complications highly affected the structure and connectivity of the neurocardiovascular graph. Before these models can be fully introduced in clinical practice, however, base line data is needed for long term follow-up.

end %-->