National audience

Monitoring sleep of premature infants is a vital aspect of clinical care, as it can reveal potential future pathologies and health issues. This study presents a novel approach to automatically estimate and track Quiet Sleep (QS) in 33 newborns using ECG, respiration, and video motion features. Using an annotated dataset from 15 neonates (10 preterm, 5 full-term) encompassing 127.2 hours, a comprehensive feature extraction and selection process was employed. Three classifiers (Random Forest, Logistic Regression, K-Nearest Neighbors) were evaluated to develop a QS estimation model. A compact and interpretable model was selected, achieving a balanced accuracy of 84.6 $\pm$ 7.5% . The robustness of the model was further enhanced by incorporating a switching mechanism between models using only ECG and respiration when video data was unavailable. The study further explored the evolution of QS during hospitalization using a large dataset with 18 newborns (16 preterm and 2 term) and 1396.6 hours of data. It highlighted an increase in QS duration and mean interval duration with post-menstrual age. The results offer valuable insights into the developmental progress of healthy preterm infants and underscore the potential of continuous, non-invasive monitoring in neonatal intensive care units.

In image processing, wavelet transform (WT) offers multiscale image decomposition, generating a blend of low-resolution approximation images and high-resolution detail components. Drawing parallels to this concept, we view feature maps in convolutional neural networks (CNNs) as a similar mix, but uniquely within the channel domain. Inspired by multitask learning (MTL) principles, we propose a wavelet-based dual-task (WDT) framework. This novel framework employs WT in the channel domain to split a single task into two parallel tasks, thereby reforming traditional single-task CNNs into dynamic dual-task networks. Our WDT framework integrates seamlessly with various popular network architectures, enhancing their versatility and efficiency. It offers a more rational approach to resource allocation in CNNs, balancing between low-frequency and high-frequency information. Rigorous experiments on Cifar10, ImageNet, HMDB51, and UCF101 validate our approach's effectiveness. Results reveal significant improvements in the performance of traditional CNNs on classification tasks, and notably, these enhancements are achieved with fewer parameters and computations. In summary, our work presents a pioneering step toward redefining the performance and efficiency of CNN-based tasks through WT.

Deep learning and Convolutional Neural Networks (CNNs) have driven major transformations in diverse research areas. However, their limitations in handling low-frequency in-formation present obstacles in certain tasks like interpreting global structures or managing smooth transition images. Despite the promising performance of transformer struc-tures in numerous tasks, their intricate optimization com-plexities highlight the persistent need for refined CNN en-hancements using limited resources. Responding to these complexities, we introduce a novel framework, the Mul-tiscale Low-Frequency Memory (MLFM) Network, with the goal to harness the full potential of CNNs while keep-ing their complexity unchanged. The MLFM efficiently preserves low-frequency information, enhancing perfor-mance in targeted computer vision tasks. Central to our MLFM is the Low-Frequency Memory Unit (LFMU), which stores various low-frequency data and forms a parallel channel to the core network. A key advantage of MLFM is its seamless compatibility with various prevalent networks, requiring no alterations to their original core structure. Testing on ImageNet demonstrated substantial accuracy improvements in multiple 2D CNNs, including ResNet, MobileNet, EfficientNet, and ConvNeXt. Furthermore, we showcase MLFM's versatility beyond traditional image classification by successfully integrating it into image-to-image translation tasks, specifically in semantic segmenta-tion networks like FCN and U-Net. In conclusion, our work signifies a pivotal stride in the journey of optimizing the ef-ficacy and efficiency of CNNs with limited resources. This research builds upon the existing CNN foundations and paves the way for future advancements in computer vision. Our codes are available at https://github.com/AlphaWuSeu/MLFM.

By seamlessly integrating wavelet transforms into the image patching stage of ViT, we leverage the power of multi-level wavelet transforms to decompose images into a diverse array of frequency-domain features. These features, integrated with spatial characteristics at equivalent scales, enrich image details, enhancing ViT's proficiency in delineating intricate textures and distinct edges. Consequently, we registered a notable 2.7% accuracy enhancement on the ImageNet100 dataset in ViT. Our wavelet patching module, designed for versatility, seamlessly fits into various ViT derivatives without necessitating architecture modifications. This advancement has uplifted the performance of several leading vision transformers by 0.46–4.3%, preserving parameter efficiency without notable FLOPs increment.

The follow-up of the development of the premature baby is a major component of its clinical care since it has been shown that it can reveal a pathology. However, no method allowing an automated and continuous monitoring of this development has been proposed. Within the framework of the Digi-NewB European project, our team wishes to offer new clinical indices qualifying the maturation of newborns. In this study, we propose a new method to characterize motor activity from video recordings. For this purpose, we have chosen to characterize the motion temporal organization by drawing inspiration from sleep organization. Thus, we propose a fully automatic process allowing to extract motion features and to combine them to estimate a functional age. By investigating two datasets, one of 28.5 hours (manually annotated) from 33 newborns and one of 4,920 hours from 46 newborns, we show that the proposed approach is relevant for monitoring in clinical routine and that the extracted features reflect the maturation of preterm newborns. Indeed, a compact and interpretable model using gestational age and three motion features (mean duration of intervals with motion, total percentage of time spent in motion and number of intervals without motion) was designed to predict post-menstrual age of newborns and showed an admissible mean absolute error of 1.3 weeks. While the temporal organization of motion was not studied clinically due to a lack of technological means, these results open the door to new developments, new investigations and new knowledge on the evolution of motion in newborns.

This study was designed to test if heart rate variability (HRV) data from preterm and full-term infants could be used to estimate their functional maturational age (FMA), using a machine learning model. We propose that the FMA, and its deviation from the postmenstrual age (PMA) of the infants could inform physicians about the progress of the maturation of the infants. The HRV data was acquired from 50 healthy infants, born between 25 and 41 weeks of gestational age, who did not present any signs of abnormal maturation relative to their age group during the period of observation. The HRV features were used as input for a machine learning model that uses filtering and genetic algorithms for feature selection, and an ensemble machine learning (EML) algorithm, which combines linear and random forest regressions, to produce as output a FMA. Using HRV data, the FMA had a mean absolute error of 0.93 weeks, 95% CI [0.78, 1.08], compared to the PMA. These results demonstrate that HRV features of newborn infants can be used by an EML model to estimate their FMA. This method was also generalized using respiration rate variability (RRV) and bradycardia data, obtaining similar results. The FMA, predicted either by HRV, RRV or bradycardia, and its deviation from the true PMA of the infants, could be used as a surrogate measure of the maturational age of the infants, which could potentially be monitored non-invasively and in real-time in the setting of neonatal intensive care units.

OBJECTIVE: This study was designed to test the diagnostic value of visibility graph features derived from the heart rate time series to predict late onset sepsis (LOS) in preterm infants using machine learning. METHODS: The heart rate variability (HRV) data was acquired from 49 premature newborns hospitalized in neonatal intensive care units (NICU). The LOS group consisted of patients who received more than five days of antibiotics, at least 72 hours after birth. The control group consisted of infants who did not receive antibiotics. HRV features in the days prior to the start of antibiotics (LOS group) or in a randomly selected period (control group) were compared against a baseline value calculated during a calibration period. After automatic feature selection, four machine learning algorithms were trained. All the tests were done using two variants of the feature set: one only included traditional HRV features, and the other additionally included visibility graph features. Performance was studied using area under the receiver operating characteristics curve (AUROC). RESULTS: The best performance for detecting LOS was obtained with logistic regression, using the feature set including visibility graph features, with AUROC of 87.7% during the six hours preceding the start of antibiotics, and with predictive potential (AUROC above 70%) as early as 42 h before start of antibiotics. CONCLUSION: These results demonstrate the usefulness of introducing visibility graph indexes in HRV analysis for sepsis prediction in newborns. SIGNIFICANCE: The method proposed the possibility of non-invasive, real-time monitoring of risk of LOS in a NICU setting.

Video-based motion analysis recently appeared to be a promising approach in neonatal intensive care units for monitoring the state of preterm newborns since it is contact-less and noninvasive. However it is important to remove periods when the newborn is absent or an adult is present from the analysis. In this paper, we propose a method for automatic detection of preterm newborn presence in incubator and open bed. We learn a specific model for each bed type as the camera placement differs a lot and the encountered situations are different between both. We break the problem down into two binary classifications based on deep transfer learning that are fused afterwards: newborn presence detection on the one hand and adult presence detection on the other hand. Moreover, we adopt a strategy of decision intervals fusion in order to take advantage of temporal consistency. We test three deep neural network that were pre-trained on ImageNet: VGG16, MobileNetV2 and InceptionV3. Two classifiers are compared: support vector machine and a small neural network. Our experiments are conducted on a database of 120 newborns. The whole method is evaluated on a subset of 25 newborns including 66 days of video recordings. In incubator, we reach a balanced accuracy of 86%. In open bed, the performance is lower because of a much wider variety of situations whereas less data are available.

Different approaches have been proposed in the literature to detect the frailty of an elderly person. In this paper, we propose a solution for detecting the frailty of older adults based on the monitoring of activities of daily living (ADL). The elderly's daily routine, is characterized by indexes determined by depth sensors such as the percentage of time in the lying position, the percentage of time in a sitting position during the day, the number of falls, the number of visits, the number of outing, and the walking speed. These indexes are intended to be an indication of frailty. Measuring frailty is difficult and requires data collection over several months. In this communication, we hypothesize that the elderly person organizes the daily life around their environment, behavior or social relations and has a well-defined routine life and we use a model to simulate the routine (normal) or non-routine (abnormal) day, according to the variance of frailty indexes over a six-month period. The classification of the type of the days (normal/ abnormal) for two different databases to lead to an accuracy of 99% and 100%. A patient is considered frail when the weekly percentage of maintaining routine decreases steadily.

We reuse photoplethysmography (PPG) datasets obtained during the clinical testing of a wrist-worn reflectance pulse oximeter and build a deep neural network (DNN) with the goal of achieving RMSE <; 3.5% for SpO2 levels ranging from 70% to 100%, which is the required clinical accuracy. We show that, with supervised learning alone, the DNN does not achieve the required performance (RMSE = 4.4%). When pretraining the DNN in an unsupervised manner with a method based on contrastive representation learning to take advantage of a set of unlabelled PPG signals, the DNN achieves clinical-grade accuracy (RMSE = 2.91%). This work therefore highlights the importance of semi-supervised learning for the development of wearable medical devices, for which unlabelled data is abundant but labelled data is scarce.

In this paper, we perform complex network analysis on a connectivity dataset retrieved from a monitoring system in order to classify simple daily activities. The monitoring system is composed of a set of wearable sensing modules positioned on the subject's body and the connectivity data consists of the correlation between each pair of modules. A number of network measures are then computed followed by the application of statistical significance and feature selection methods. These methods were implemented for the purpose of reducing the total number of modules in the monitoring system required to provide accurate activity classification. The obtained results show that an overall accuracy of 84.6% for activity classification is achieved, using a random forest classifier, and when considering a monitoring system composed of only two modules positioned at the neck and thigh of the subject's body.

This paper proposes an online Bayesian method to detect change points in auto-regressive (AR) processes with unknown model orders. The AR model is frequently used in the spectral analysis of RR series extracted from electrocardio-graphic signals (ECG) [1, 2]. By relaxing the model order constraint, we aim to detect apnea-bradycardia (AB) episodes from abrupt changes in the model space. An efficient recursive algorithm inspired from the work of Godsil et al. [3, 4] is proposed to update with fixed complexity the joint posterior distribution of the AR coefficients and model orders. Simulation results show fast convergence of the estimated distribution, thus making it an efficient tool to detect underlying AR model changes in time series. For AB detections with annotated ECG data, the detection sensitivity (TP/(TP + FN)) reaches 98% over a total of 50 episodes with 92% specificity (TN/(TN + FP)). We also discovered an interesting property in terms of detection delay (-3.64s ± 4.34), compared with the experts' off-line annotations. The negative mean in detection delay suggests that AR model changes might occur before the onset of AB episodes while from the clinical point of view, it is essential to achieve reliable early stage detection of AB episodes to enable the initiation of quick nursing actions [5].

In neonatal electroencephalography (EEG) heart activity is a major source of artifacts which can lead to misleading results in automated analysis if they are not properly eliminated. In this work we propose a combination of empirical mode decomposition (EMD) and adaptive filtering (AF) to cancel electrocardiogram (ECG) noise in a simplified EEG montage for preterm infants. The introduction of EMD prior to AF allows to selectively remove ECG preserving at maximum the original characteristics of EEG. Cleaned signals improved up to 17% the correlation coefficient with original datasets in comparison with signals denoised solely with AF.

We analyzed the performance of five common estimators of the Hurst parameter (H) in simulated data, i.e. surrogate and synthetic through a fractional Gaussian noise model. Inter-breath signals of 30 preterm infants were used to generate surrogates as well as the synthetic data, which was produced according 3 typical patterns: erratic (EB), periodic (PB) and regular breathing (RB). The discrete wavelet transform was the most efficient for PB, with a concordance correlation coefficient (CCC) of 0.92. The iterative fGn-based estimator was the most efficient for EB and RB, with a CCC of 0.92 and 0.97 respectively, and performed better in the surrogates test with an estimation error of 0.008.

The aim of this study is to facilitate the home follow-up of patients treated with cardiac implantable devices. A new procedure to synthesize 12-lead ECG from intracardiac EGM is proposed. It is based on the estimation of: (i) a 3D representation of the cardiac electrical activity both for ECG (VCG) and EGM (VGM), and (ii) the transfer function between the VGM and the VCG. The extraction of VCG and VGMis performed by comparing four different algorithms based on PCA and ICA, whereas the non-linear transfer function between VCG and VGM is estimated using a specific neural network. Results demonstrate the effectiveness of the proposed method in comparison with our previous work. Indeed, the correlation coefficients, between the real ECG and the synthesized ECG, lie between 0.78 and 0.99, whereas correlation coefficients of the previous method (combining PCA and linear Wiener filter) lie between 0.6 and 0.94.

In this paper, the objective was to investigate the heart rate variability in two selected groups of premature infants (sepsis vs non-sepsis). We studied the RR interval series not only by linear methods - time domain and frequency domain, but also by non-linear methods - chaos theory and information theory, in order to find the optimal parameters to distinguish sepsis premature infants from non-sepsis ones. The results show that indexes of information theory are useful parameters for the diagnosis of late neonatal infection in premature infants with recurrent apnea-bradycardia.

Monitoring patients in intensive care units is a critical task. Simple condition detection is generally insufficient to diagnose a patient and may generate many false alarms to the clinician operator. Deeper knowledge is needed to discriminate among alarms those that necessitate urgent therapeutic action. We propose an intelligent monitoring system that makes use of many artificial intelligence techniques: artificial neural networks for temporal abstraction, temporal reasoning, model based diagnosis, decision rule based system for adaptivity and machine learning for knowledge acquisition. To tackle the difficulty of taking context change into account, we introduce a pilot aiming at adapting the system behavior by reconfiguring or tuning the parameters of the system modules. A prototype has been implemented and is currently experimented and evaluated. Some results, showing the benefits of the approach, are given.

This paper deals with the conception of a new system for sleep staging in ambulatory conditions. Sleep recording is performed by means of five electrodes: two temporal, two frontal and a reference. This configuration enables to avoid the chin area to enhance the quality of the muscular signal and the hair region for patient convenience. The electroencephalopgram (EEG), eletromyogram (EMG), and electrooculogram (EOG) signals are separated using the Independent Component Analysis approach. The system is compared to a standard sleep analysis system using polysomnographic recordings of 14 patients. The overall concordance of 67.2% is achieved between the two systems. Based on the validation results and the computational efficiency we recommend the clinical use of the proposed system in a commercial sleep analysis platform.

Telemedicine is producing a great impact in the monitoring of patients located in remote nonclinical environments such as homes, elder communities, gymnasiums, schools, remote military bases, ships, and the like. A number of applications, ranging from data collection, to chronic patient surveillance, and even to the control of therapeutic procedures, are being implemented in many parts of the world. As part of this growing trend, this paper discusses the problems in electrocardiogram (ECG) real-time data acquisition, transmission, and visualization over the Internet. ECG signals are transmitted in real time from a patient in a remote nonclinical environment to the specialist in a hospital or clinic using the current capabilities and availability of the Internet. A prototype system is composed of a portable data acquisition and preprocessing module connected to the computer in the remote site via its RS-232 port, a Java-based client-server platform, and software modules to handle communication protocols between data acquisition module and the patient's personal computer, and to handle client-server communication. The purpose of the system is the provision of extended monitoring for patients under drug therapy after infarction, data collection in some particular cases, remote consultation, and low-cost ECG monitoring for the elderly.