Corrigendum: Building an open source classifier for the neonatal EEG background: a systematic feature-based approach from expert scoring to clinical visualization.

[This corrects the article DOI: 10.3389/fnhum.2021.675154.].

Neuronal coupling modes show differential development in the early cortical activity networks of human newborns.

The third trimester is a critical period for the development of functional networks that support the lifelong neurocognitive performance; yet the emergence of neuronal coupling in these networks is poorly understood. Here, we used longitudinal high-density electroencephalographic (EEG) recordings from preterm infants during the period from 33 to 45 weeks of conceptional age to characterize early spatiotemporal patterns in the development of local cortical function and the intrinsic coupling modes (phase-phase, amplitude-amplitude, and phase-amplitude correlations). Absolute local power showed a robust increase with conceptional age across the full frequency spectrum, while local phase-amplitude correlations showed sleep state -specific, biphasic development that peaked a few weeks before normal birth. Amplitude-amplitude and distant phase-amplitude correlations decreased globally at nearly all frequencies. In contrast, the phase-phase correlations showed frequency- and region-selective development, with an increase of coupling strength with conceptional age between frontal, central, and occipital regions at low-delta and alpha frequencies together with a wider-spread decrease at other frequencies. Our findings together present the spectrally and spatially differential development of the distinct intrinsic coupling modes during the neonatal period and provide their developmental templates for future basic and clinical research.Significance statement Neuronal activity coupling in cortical networks is a fundamental mechanism underlying higher brain functions. However, emergence of these functional networks within the structural connectome at early maturation is poorly understood. Here, we study the human development of cortical function and distinct neuronal coupling modes (amplitude-amplitude, phase-phase, and phase-amplitude coupling) by investigating longitudinal high-density electroencephalographic recordings in preterm-born infants and discussing their potential links to established neurophysiological processes. Our findings disclose robust, spatially and spectrally specific developmental trajectories for all coupling modes to be used as developmental baselines for future research. The findings together indicate that neuronal coupling modes develop independently during the early neonatal period, supporting the notion that these modes reflect different coupling mechanisms and should be considered separately.

Quantitative EEG features during the first day correlate to clinical outcome in perinatal asphyxia.

OBJECTIVE: To assess whether computational electroencephalogram (EEG) measures during the first day of life correlate to clinical outcomes in infants with perinatal asphyxia with or without hypoxic-ischemic encephalopathy (HIE). METHODS: We analyzed four-channel EEG monitoring data from 91 newborn infants after perinatal asphyxia. Altogether 42 automatically computed amplitude- and synchrony-related EEG features were extracted as 2-hourly average at very early (6 h) and early (24 h) postnatal age; they were correlated to the severity of HIE in all infants, and to four clinical outcomes available in a subcohort of 40 newborns: time to full oral feeding (nasogastric tube NGT), neonatal brain MRI, Hammersmith Infant Neurological Examination (HINE) at three months, and Griffiths Scales at two years. RESULTS: At 6 h, altogether 14 (33%) EEG features correlated significantly to the HIE grade ([r]= 0.39-0.61, p < 0.05), and one feature correlated to NGT ([r]= 0.50). At 24 h, altogether 13 (31%) EEG features correlated significantly to the HIE grade ([r]= 0.39-0.56), six features correlated to NGT ([r]= 0.36-0.49) and HINE ([r]= 0.39-0.61), while no features correlated to MRI or Griffiths Scales. CONCLUSIONS: Our results show that the automatically computed measures of early cortical activity may provide outcome biomarkers for clinical and research purposes. IMPACT: The early EEG background and its recovery after perinatal asphyxia reflect initial severity of encephalopathy and its clinical recovery, respectively. Computational EEG features from the early hours of life show robust correlations to HIE grades and to early clinical outcomes. Computational EEG features may have potential to be used as cortical activity biomarkers in early hours after perinatal asphyxia.

Clinical outcome prediction with an automated EEG trend, Brain State of the Newborn, after perinatal asphyxia.

OBJECTIVE: To evaluate the utility of a fully automated deep learning -based quantitative measure of EEG background, Brain State of the Newborn (BSN), for early prediction of clinical outcome at four years of age. METHODS: The EEG monitoring data from eighty consecutive newborns was analyzed using the automatically computed BSN trend. BSN levels during the first days of life (a of total 5427 hours) were compared to four clinical outcome categories: favorable, cerebral palsy (CP), CP with epilepsy, and death. The time dependent changes in BSN-based prediction for different outcomes were assessed by positive/negative predictive value (PPV/NPV) and by estimating the area under the receiver operating characteristic curve (AUC). RESULTS: The BSN values were closely aligned with four visually determined EEG categories (p < 0·001), as well as with respect to clinical milestones of EEG recovery in perinatal Hypoxic Ischemic Encephalopathy (HIE; p < 0·003). Favorable outcome was related to a rapid recovery of the BSN trend, while worse outcomes related to a slow BSN recovery. Outcome predictions with BSN were accurate from 6 to 48 hours of age: For the favorable outcome, the AUC ranged from 95 to 99% (peak at 12 hours), and for the poor outcome the AUC ranged from 96 to 99% (peak at 12 hours). The optimal BSN levels for each PPV/NPV estimate changed substantially during the first 48 hours, ranging from 20 to 80. CONCLUSIONS: We show that the BSN provides an automated, objective, and continuous measure of brain activity in newborns. SIGNIFICANCE: The BSN trend discloses the dynamic nature that exists in both cerebral recovery and outcome prediction, supports individualized patient care, rapid stratification and early prognosis.

A growth chart of brain function from infancy to adolescence based on EEG.

BACKGROUND: In children, objective, quantitative tools that determine functional neurodevelopment are scarce and rarely scalable for clinical use. Direct recordings of cortical activity using routinely acquired electroencephalography (EEG) offer reliable measures of brain function. METHODS: We developed and validated a measure of functional brain age (FBA) using a residual neural network-based interpretation of the paediatric EEG. In this cross-sectional study, we included 1056 children with typical development ranging in age from 1 month to 18 years. We analysed a 10- to 15-min segment of 18-channel EEG recorded during light sleep (N1 and N2 states). FINDINGS: The FBA had a weighted mean absolute error (wMAE) of 0.85 years (95% CI: 0.69-1.02; n = 1056). A two-channel version of the FBA had a wMAE of 1.51 years (95% CI: 1.30-1.73; n = 1056) and was validated on an independent set of EEG recordings (wMAE = 2.27 years, 95% CI: 1.90-2.65; n = 723). Group-level maturational delays were also detected in a small cohort of children with Trisomy 21 (Cohen's d = 0.36, p = 0.028). INTERPRETATION: A FBA, based on EEG, is an accurate, practical and scalable automated tool to track brain function maturation throughout childhood with accuracy comparable to widely used physical growth charts. FUNDING: This research was supported by the National Health and Medical Research Council, Australia, Helsinki University Diagnostic Center Research Funds, Finnish Academy, Finnish Paediatric Foundation, and Sigrid Juselius Foundation.

Automatic assessment of infant carrying and holding using at-home wearable recordings.

Assessing infant carrying and holding (C/H), or physical infant-caregiver interaction, is important for a wide range of contexts in development research. An automated detection and quantification of infant C/H is particularly needed in long term at-home studies where development of infants' neurobehavior is measured using wearable devices. Here, we first developed a phenomenological categorization for physical infant-caregiver interactions to support five different definitions of C/H behaviors. Then, we trained and assessed deep learning-based classifiers for their automatic detection from multi-sensor wearable recordings that were originally used for mobile assessment of infants' motor development. Our results show that an automated C/H detection is feasible at few-second temporal accuracy. With the best C/H definition, the automated detector shows 96% accuracy and 0.56 kappa, which is slightly less than the video-based inter-rater agreement between trained human experts (98% accuracy, 0.77 kappa). The classifier performance varies with C/H definition reflecting the extent to which infants' movements are present in each C/H variant. A systematic benchmarking experiment shows that the widely used actigraphy-based method ignores the normally occurring C/H behaviors. Finally, we show proof-of-concept for the utility of the novel classifier in studying C/H behavior across infant development. Particularly, we show that matching the C/H detections to individuals' gross motor ability discloses novel insights to infant-parent interaction.

Multiplex dynamic networks in the newborn brain disclose latent links with neurobehavioral phenotypes.

The higher brain functions arise from coordinated neural activity between distinct brain regions, but the spatial, temporal, and spectral complexity of these functional connectivity networks (FCNs) has challenged the identification of correlates with neurobehavioral phenotypes. Characterizing behavioral correlates of early life FCNs is important to understand the activity dependent emergence of neurodevelopmental performance and for improving health outcomes. Here, we develop an analysis pipeline for identifying multiplex dynamic FCNs that combine spectral and spatiotemporal characteristics of the newborn cortical activity. This data-driven approach automatically uncovers latent networks that show robust neurobehavioral correlations and consistent effects by in utero drug exposure. Altogether, the proposed pipeline provides a robust end-to-end solution for an objective assessment and quantitation of neurobehaviorally meaningful network constellations in the highly dynamic cortical functions.

Detailed statistical analysis plan for ALBINO: effect of Allopurinol in addition to hypothermia for hypoxic-ischemic Brain Injury on Neurocognitive Outcome - a blinded randomized placebo-controlled parallel group multicenter trial for superiority (phase III).

BACKGROUND: Despite therapeutic hypothermia (TH) and neonatal intensive care, 45-50% of children affected by moderate-to-severe neonatal hypoxic-ischemic encephalopathy (HIE) die or suffer from long-term neurodevelopmental impairment. Additional neuroprotective therapies are sought, besides TH, to further improve the outcome of affected infants. Allopurinol - a xanthine oxidase inhibitor - reduced the production of oxygen radicals and subsequent brain damage in pre-clinical and preliminary human studies of cerebral ischemia and reperfusion, if administered before or early after the insult. This ALBINO trial aims to evaluate the efficacy and safety of allopurinol administered immediately after birth to (near-)term infants with early signs of HIE. METHODS/DESIGN: The ALBINO trial is an investigator-initiated, randomized, placebo-controlled, double-blinded, multi-national parallel group comparison for superiority investigating the effect of allopurinol in (near-)term infants with neonatal HIE. Primary endpoint is long-term outcome determined as survival with neurodevelopmental impairment versus death versus non-impaired survival at 2 years. RESULTS: The primary analysis with three mutually exclusive responses (healthy, death, composite outcome for impairment) will be on the intention-to-treat (ITT) population by a generalized logits model according to Bishop, Fienberg, Holland (Bishop YF, Discrete Multivariate Analysis: Therory and Practice, 1975) and ."will be stratified for the two treatment groups. DISCUSSION: The statistical analysis for the ALBINO study was defined in detail in the study protocol and implemented in this statistical analysis plan published prior to any data analysis. This is in accordance with the Declaration of Helsinki and the International Conference on Harmonization Good Clinical Practice guidelines. TRIAL REGISTRATION: ClinicalTrials.gov NCT03162653. Registered on 22 May 2017.

Early EEG-burst sharpness and 2-year disability in extremely preterm infants.

BACKGROUND: Automated computational measures of EEG have the potential for large-scale application. We hypothesised that a predefined measure of early EEG-burst shape (increased burst sharpness) could predict neurodevelopmental impairment (NDI) and mental developmental index (MDI) at 2 years of age over-and-above that of brain ultrasound. METHODS: We carried out a secondary analysis of data from extremely preterm infants collected for an RCT (SafeBoosC-II). Two hours of single-channel cross-brain EEG was used to analyse burst sharpness with an automated algorithm. The co-primary outcomes were moderate-or-severe NDI and MDI. Complete data were available from 58 infants. A predefined statistical analysis was adjusted for GA, sex and no, mild-moderate, and severe brain injury as detected by cranial ultrasound. RESULTS: Nine infants had moderate-or-severe NDI and the mean MDI was 87 ± 17.3 SD. The typical burst sharpness was low (negative values) and varied relatively little (mean -0.81 ± 0.11 SD), but the odds ratio for NDI was increased by 3.8 (p = 0.008) and the MDI was reduced by -3.2 points (p = 0.14) per 0.1 burst sharpness units increase (+1 SD) in the adjusted analysis. CONCLUSION: This study confirms the association between EEG-burst measures in preterm infants and neurodevelopment in childhood. Importantly, this was by a priori defined analysis. IMPACT: A fully automated, computational measure of EEG in the first week of life was predictive of neurodevelopmental impairment at 2 years of age. This confirms many previous studies using expert reading of EEG. Only single-channel EEG data were used, adding to the applicability. EEG was recorded by several different devices thus this measure appears to be robust to differences in electrodes, amplifiers and filters. The likelihood ratio of a positive EEG test, however, was only about 2, suggesting little immediate clinical value.

EEG response to a high volume (1.5 mL/kg) caudal block in infants less than 3 months.

INTRODUCTION: The substantial compression of the dural sac and the subsequent cranial shift of cerebrospinal fluid caused by a high-volume caudal block has been shown to significantly but transiently reduce cerebral blood flow. The aim of the present study was to determine whether this reduction in cerebral perfusion is significant enough to alter brain function, as assessed by electroencephalography (EEG). METHODS: Following ethics approval and parental informed consent, 11 infants (0-3 months) scheduled to undergo inguinal hernia repair were included in the study. EEG electrodes (using nine electrodes according to the 10-20 standard) were applied following anesthesia induction. Following a 5 min baseline period, a caudal block was performed (1.5 mL/kg), whereafter the EEG, hemodynamic, and cerebral near-infrared spectroscopy responses were followed during a 20 min observation period that was divided into four 5 min segments. Special attention was given to alterations in delta power activity since this may indicate cerebral ischemia. RESULTS: All 11 infants displayed transient EEG changes, mainly represented by increased relative delta power, during the initial 5-10 min postinjection. The observed changes had returned close to baseline values 15 min postinjection. Heart rate and blood pressure remained stable throughout the study. CONCLUSION: A high-volume caudal block appears to increase intracranial pressure, thereby reducing cerebral blood flow, to the extent that it transiently will affect cerebral function as assessed by EEG (increased delta power activity) in approximately 90% of small infants. TRIAL REGISTRATION NUMBER: ACTRN12620000420943.

Evaluation of self-supervised pre-training for automatic infant movement classification using wearable movement sensors.

The recently-developed infant wearable MAIJU provides a means to automatically evaluate infants' motor performance in an objective and scalable manner in out-of-hospital settings. This information could be used for developmental research and to support clinical decision-making, such as detection of developmental problems and guiding of their therapeutic interventions. MAIJU-based analyses rely fully on the classification of infant's posture and movement; it is hence essential to study ways to increase the accuracy of such classifications, aiming to increase the reliability and robustness of the automated analysis. Here, we investigated how self-supervised pre-training improves performance of the classifiers used for analyzing MAIJU recordings, and we studied whether performance of the classifier models is affected by context-selective quality-screening of pre-training data to exclude periods of little infant movement or with missing sensors. Our experiments show that i) pre-training the classifier with unlabeled data leads to a robust accuracy increase of subsequent classification models, and ii) selecting context-relevant pre-training data leads to substantial further improvements in the classifier performance.Clinical relevance- This study showcases that self-supervised learning can be used to increase the accuracy of out-of-hospital evaluation of infants' motor abilities via smart wearables.

Optimization of time series features to estimate brain age in children from electroencephalography.

Functional brain age measures in children, derived from the electroencephalogram (EEG), offer direct and objective measures in assessing neurodevelopmental status. Here we explored the effectiveness of 32 preselected 'handcrafted' EEG features in predicting brain age in children. These features were benchmarked against a large library of highly comparative multivariate time series features (>7000 features). Results showed that age predictors based on handcrafted EEG features consistently outperformed a generic set of time series features. These findings suggest that optimization of brain age estimation in children benefits from careful preselection of EEG features that are related to age and neurodevelopmental trajectory. This approach shows potential for clinical translation in the future.Clinical Relevance-Handcrafted EEG features provide an accurate functional neurodevelopmental biomarker that tracks brain function maturity in children.

Analysing heart rate variability in preterm infants: the effect of temporal adjustment of NN peaks and missing data.

The measurement of heart rate variability (HRV) in preterm infants provides important information on function to clinicians. Measuring the underlying electrocardiogram (ECG) in the neonatal intensive care unit is a challenge and there is a trade off between extracting accurate measurements of the HRV and the amount of ECG processed due to contamination. Knowledge on the effects of 1) quantization in the time domain and 2) missing data on the calculation of HRV features will inform clinical implementation. In this paper, we studied multiple 5 minute epochs from 148 ECG recordings on 56 extremely preterm infants. We found that temporal adjustment of NN peaks improves the estimate of the NN interval resulting in HRV features (m = 9) that are better correlated with age (median percentage increase in correlation of individual features: 0.2%, IQR: 0.0 to 5.6%; correlation with age predictor and age from 0.721 to 0.787). Improved (sub-sample) quantization of the NN intervals (via interpolation) reduced the overall value of HRV features (median percentage reduction in feature value: -1.3%, IQR: -18.8 to 0.0; m = 9), primarily through a reduction in the energy of high-frequency oscillations. HRV features were also robust to missing data, with measures such as mean NN, fractal dimension and the smoothed nonlinear energy operator (SNEO) less susceptible to missing data than features such as VLF, LF, and HF. Furthermore, age predictions derived from a combination of HRV measures were more robust to missing data than individual HRV measures.Clinical Relevance-Poor quantization in time when estimating the NN peak and the presence of missing data confound HRV measures, particularly spectral measures.

Networks of cortical activity show graded responses to perinatal asphyxia.

BACKGROUND: Perinatal asphyxia often leads to hypoxic-ischemic encephalopathy (HIE) with a high risk of neurodevelopmental consequences. While moderate and severe HIE link to high morbidity, less is known about brain effects of perinatal asphyxia with no or only mild HIE. Here, we test the hypothesis that cortical activity networks in the newborn infants show a dose-response to asphyxia. METHODS: We performed EEG recordings for infants with perinatal asphyxia/HIE of varying severity (n = 52) and controls (n = 53) and examined well-established computational metrics of cortical network activity. RESULTS: We found graded alterations in cortical activity networks according to severity of asphyxia/HIE. Furthermore, our findings correlated with early clinical recovery measured by the time to attain full oral feeding. CONCLUSION: We show that both local and large-scale correlated cortical activity are affected by increasing severity of HIE after perinatal asphyxia, suggesting that HIE and perinatal asphyxia are better represented as a continuum rather than the currently used discreet categories. These findings imply that automated computational measures of cortical function may be useful in characterizing the dose effects of adversity in the neonatal brain; such metrics hold promise for benchmarking clinical trials via patient stratification or as early outcome measures. IMPACT: Perinatal asphyxia causes every fourth neonatal death worldwide and provides a diagnostic and prognostic challenge for the clinician. We report that infants with perinatal asphyxia show specific graded responses in cortical networks according to severity of asphyxia and ensuing hypoxic-ischaemic encephalopathy. Early EEG recording and automated computational measures of brain function have potential to help in clinical evaluation of infants with perinatal asphyxia.

Neonatal cortical activity organizes into transient network states that are affected by vigilance states and brain injury.

Early neurodevelopment is critically dependent on the structure and dynamics of spontaneous neuronal activity; however, the natural organization of newborn cortical networks is poorly understood. Recent adult studies suggest that spontaneous cortical activity exhibits discrete network states with physiological correlates. Here, we studied newborn cortical activity during sleep using hidden Markov modeling to determine the presence of such discrete neonatal cortical states (NCS) in 107 newborn infants, with 47 of them presenting with a perinatal brain injury. Our results show that neonatal cortical activity organizes into four discrete NCSs that are present in both cardinal sleep states of a newborn infant, active and quiet sleep, respectively. These NCSs exhibit state-specific spectral and functional network characteristics. The sleep states exhibit different NCS dynamics, with quiet sleep presenting higher fronto-temporal activity and a stronger brain-wide neuronal coupling. Brain injury was associated with prolonged lifetimes of the transient NCSs, suggesting lowered dynamics, or flexibility, in the cortical networks. Taken together, the findings suggest that spontaneously occurring transient network states are already present at birth, with significant physiological and pathological correlates; this NCS analysis framework can be fully automatized, and it holds promise for offering an objective, global level measure of early brain function for benchmarking neurodevelopmental or clinical research.

Mechanisms underlying pathological cortical bursts during metabolic depletion.

Cortical activity depends upon a continuous supply of oxygen and other metabolic resources. Perinatal disruption of oxygen availability is a common clinical scenario in neonatal intensive care units, and a leading cause of lifelong disability. Pathological patterns of brain activity including burst suppression and seizures are a hallmark of the recovery period, yet the mechanisms by which these patterns arise remain poorly understood. Here, we use computational modeling of coupled metabolic-neuronal activity to explore the mechanisms by which oxygen depletion generates pathological brain activity. We find that restricting oxygen supply drives transitions from normal activity to several pathological activity patterns (isoelectric, burst suppression, and seizures), depending on the potassium supply. Trajectories through parameter space track key features of clinical electrophysiology recordings and reveal how infants with good recovery outcomes track toward normal parameter values, whereas the parameter values for infants with poor outcomes dwell around the pathological values. These findings open avenues for studying and monitoring the metabolically challenged infant brain, and deepen our understanding of the link between neuronal and metabolic activity.

Inter-site generalizability of EEG based age prediction algorithms in the preterm infant.

Objective. To overcome the effects of site differences in EEG-based brain age prediction in preterm infants.Approach. We used a 'bag of features' with a combination function estimated using support vector regression (SVR) and feature selection (filter then wrapper) to predict post-menstrual age (PMA). The SVR was trained on a dataset containing 138 EEG recordings from 37 preterm infants (site 1). A separate set of 36 EEG recordings from 36 preterm infants was used to validate the age predictor (site 2). The feature distributions were compared between sites and a restricted feature set was constructed using only features that were not significantly different between sites. The mean absolute error between predicted age and PMA was used to define the accuracy of prediction and successful validation was defined as no significant differences in error between site 1 (cross-validation) and site 2.Main results. The age predictor based on all features and trained on site 1 was not validated on site 2 (p< 0.001; MAE site 1 = 1.0 weeks,n= 59 versus MAE site 2 = 2.1 weeks,n= 36). The MAE was improved by training on a restricted features set (MAE site 1 = 1.0 weeks,n= 59 versus MAE site 2 = 1.1 weeks,n= 36), resulting in a validated age predictor when applied to site 2 (p= 0.68). The features selected from the restricted feature set when training on site 1 closely aligned with features selected when trained on a combination of data from site 1 and site 2.Significance. The ability of EEG classifiers, such as brain age prediction, to maintain accuracy on data collected at other sites may be challenged by unexpected, site-dependent differences in EEG signals. Permitting a small amount of data leakage between sites improves generalization, leading towards universal methods of EEG interpretation in preterm infants.

Charting infants' motor development at home using a wearable system: validation and comparison to physical growth charts.

BACKGROUND: Early neurodevelopmental care and research are in urgent need of practical methods for quantitative assessment of early motor development. Here, performance of a wearable system in early motor assessment was validated and compared to developmental tracking of physical growth charts. METHODS: Altogether 1358 h of spontaneous movement during 226 recording sessions in 116 infants (age 4-19 months) were analysed using a multisensor wearable system. A deep learning-based automatic pipeline quantified categories of infants' postures and movements at a time scale of seconds. Results from an archived cohort (dataset 1, N = 55 infants) recorded under partial supervision were compared to a validation cohort (dataset 2, N = 61) recorded at infants' homes by the parents. Aggregated recording-level measures including developmental age prediction (DAP) were used for comparison between cohorts. The motor growth was also compared with respective DAP estimates based on physical growth data (length, weight, and head circumference) obtained from a large cohort (N = 17,838 infants; age 4-18 months). FINDINGS: Age-specific distributions of posture and movement categories were highly similar between infant cohorts. The DAP scores correlated tightly with age, explaining 97-99% (94-99% CI 95) of the variance at the group average level, and 80-82% (72-88%) of the variance in the individual recordings. Both the average motor and the physical growth measures showed a very strong fit to their respective developmental models (R2 = 0.99). However, single measurements showed more modality-dependent variation that was lowest for motor (σ = 1.4 [1.3-1.5 CI 95] months), length (σ = 1.5 months), and combined physical (σ = 1.5 months) measurements, and it was clearly higher for the weight (σ = 1.9 months) and head circumference (σ = 1.9 months) measurements. Longitudinal tracking showed clear individual trajectories, and its accuracy was comparable between motor and physical measures with longer measurement intervals. INTERPRETATION: A quantified, transparent and explainable assessment of infants' motor performance is possible with a fully automated analysis pipeline, and the results replicate across independent cohorts from out-of-hospital recordings. A holistic assessment of motor development provides an accuracy that is comparable with the conventional physical growth measures. A quantitative measure of infants' motor development may directly support individual diagnostics and care, as well as facilitate clinical research as an outcome measure in early intervention trials. FUNDING: This work was supported by the Finnish Academy (314602, 335788, 335872, 332017, 343498), Finnish Pediatric Foundation (Lastentautiensäätiö), Aivosäätiö, Sigrid Jusélius Foundation, and HUS Children's Hospital/HUS diagnostic center research funds.

Comparison of End-to-End Neural Network Architectures and Data Augmentation Methods for Automatic Infant Motility Assessment Using Wearable Sensors.

Infant motility assessment using intelligent wearables is a promising new approach for assessment of infant neurophysiological development, and where efficient signal analysis plays a central role. This study investigates the use of different end-to-end neural network architectures for processing infant motility data from wearable sensors. We focus on the performance and computational burden of alternative sensor encoder and time series modeling modules and their combinations. In addition, we explore the benefits of data augmentation methods in ideal and nonideal recording conditions. The experiments are conducted using a dataset of multisensor movement recordings from 7-month-old infants, as captured by a recently proposed smart jumpsuit for infant motility assessment. Our results indicate that the choice of the encoder module has a major impact on classifier performance. For sensor encoders, the best performance was obtained with parallel two-dimensional convolutions for intrasensor channel fusion with shared weights for all sensors. The results also indicate that a relatively compact feature representation is obtainable for within-sensor feature extraction without a drastic loss to classifier performance. Comparison of time series models revealed that feedforward dilated convolutions with residual and skip connections outperformed all recurrent neural network (RNN)-based models in performance, training time, and training stability. The experiments also indicate that data augmentation improves model robustness in simulated packet loss or sensor dropout scenarios. In particular, signal- and sensor-dropout-based augmentation strategies provided considerable boosts to performance without negatively affecting the baseline performance. Overall, the results provide tangible suggestions on how to optimize end-to-end neural network training for multichannel movement sensor data.