Non-overlapping Windows Research Articles

0270 Obstructive Sleep Apnea detection using ECG morphology and Machine Learning

Abstract Introduction Obstructive Sleep Apnea (OSA) is characterized by reduction in airflow. Hypopnea is a smaller reduction in airflow compared to apnea but also accompanied by drop in oxygen saturation leading to sympathetic activation. Frequent OSA events are correlated with incidence of cardiovascular diseases. Measuring changes in airflow requires overnight polysomnography that is expensive. Electrocardiogram (ECG) is available through wearable devices at home and therefore, detecting OSA events using ECG can make OSA diagnostics more accessible. This paper studies the use of ECG morphology and an ensemble machine learning algorithm to detect OSA events. The algorithm designed is light weight to ensure that it can be implemented on an embedded processor. Methods The data from the Apnea, Bariatric surgery, and CPAP (ABC) study provided by the National Sleep Research Resource is used for the analysis. Twenty-six subjects diagnosed with severe OSA are considered with single night polysomnography before treatment. Since OSA events have a minimum 10s duration, 10s non-overlapping windows with OSA labels from a technician scoring is used. Features are extracted from the PQRST complex and majority voting across multiple algorithms is used to select the top 7 features (SDNN, RMSSD, P-P interval, P-duration, P-R interval, T-duration and T-P interval) with highest explanatory power. We train a random forest classifier using leave-one-out methodology where 25 subjects are used for training and the 26th subject is used for testing (all permutations are used). We use sensitivity, precision and F1 score for evaluation. Results The algorithm detects the precise occurrence (onset and offset) as well as the total number of events during the night. The precision, sensitivity and F1 scores are 70%, 96% and 80% respectively. The prediction leans towards higher occurrence of OSA events as the subjects suffer from severe OSA. We uncover 32% of false positives are events that are accompanied by significant SPO2 desaturation pointing to the inaccuracy of manual scoring. Assuming these false positives are correct predictions the precision increases to 79%. Conclusion The classification of OSA events using ECG features and machine learning demonstrates the feasibility of using wearables to measure OSA in a home setting. Support (if any)

Child Survival and Annual Crop Yield Reductions in Rural Burkina Faso: Critical Windows of Vulnerability Around Early-Life Development.

Populations that are reliant on subsistence farming are particularly vulnerable to climatic effects on crop yields. However, empirical evidence on the role of the timing of exposure to crop yield deficits in early-life development is limited. We examined the relationship between child survival and annual crop yield reductions at different stages of early-life development in a subsistence farming population in Burkina Faso. Using shared frailty Cox proportional hazards models adjusting for confounders, we analyzed 57,288 children under 5 years of age followed by the Nouna Health and Demographic Surveillance System (1994-2016) in relation to provincial food-crop yield levels experienced in 5 nonoverlapping time windows (12 months before conception, gestation, birth-age 5.9 months, ages 6.0 months-1.9 years, and ages 2.0-4.9 years) and their aggregates (birth-1.9 years, first 1,000 days from conception, and birth-4.9 years). Of the nonoverlapping windows, point estimates were largest for child survival related to food-crop yields for the time window of 6.0 months-1.9 years: The adjusted mortality hazard ratio was 1.10 (95% confidence interval: 1.03, 1.19) for a 90th-to-10th percentile yield reduction. These findings suggest that child survival in this setting is particularly vulnerable to cereal-crop yield reductions during the period of nonexclusive breastfeeding.

Route efficiency implications of time windows and vehicle capacities in first- and last-mile logistics

In this paper, we analyze the route efficiency effects that emerge from combining realistically constrained first-mile pickup and last-mile delivery operations into joint vehicle routes. Specifically, we examine (i) the individual effect of discrete, non-overlapping time window constraints; (ii) the individual effect of vehicle capacity constraints; and (iii) the joint impact of both constraint types on local route efficiency gains due to the integration of pickup and delivery operations. Extending the existent literature on continuum approximation of route distances, we propose closed-form adjustment factors which accurately capture these non-trivial route efficiency effects. To derive the adjustment factors, we conduct extensive numerical experiments and apply a novel hybrid data analysis approach which combines exploratory data analysis with symbolic regression. Our analysis suggests that these real-world constraints affect the optimal stop sequence and diminish or even eliminate the expected efficiency gains from integrating first- and last-mile operations. The proposed adjustment factors are particularly relevant for the optimal strategic design and operational planning of modern, industrial-scale distribution networks in light of growing and increasingly fragmented e-commerce volumes and a high cost pressure on the logistics industry. They help researchers and practitioners to efficiently quantify the expected benefits from integrating pickup and delivery operations, and to assess under which circumstances such an integration is (not) desirable.

Time of Day Preferences and Daily Temporal Consistency for Predicting the Sustained Use of a Commercial Meditation App: Longitudinal Observational Study

The intensive data typically collected by mobile health (mHealth) apps allows factors associated with persistent use to be investigated, which is an important objective given users' well-known struggles with sustaining healthy behavior. Data from a commercial meditation app (n=14,879; 899,071 total app uses) were analyzed to assess the validity of commonly given habit formation advice to meditate at the same time every day, preferably in the morning. First, the change in probability of meditating in 4 nonoverlapping time windows (morning, midday, evening, and late night) on a given day over the first 180 days after creating a meditation app account was calculated via generalized additive mixed models. Second, users' time of day preferences were calculated as the percentage of all meditation sessions that occurred within each of the 4 time windows. Additionally, the temporal consistency of daily meditation behavior was calculated as the entropy of the timing of app usage sessions. Linear regression was used to examine the effect of time of day preference and temporal consistency on two outcomes: (1) short-term engagement, defined as the number of meditation sessions completed within the sixth and seventh month of a user's account, and (2) long-term use, defined as the days until a user's last observed meditation session. Large reductions in the probability of meditation at any time of day were seen over the first 180 days after creating an account, but this effect was smallest for morning meditation sessions (63.4% reduction vs reductions ranging from 67.8% to 74.5% for other times). A greater proportion of meditation in the morning was also significantly associated with better short-term engagement (regression coefficient B=2.76, P<.001) and long-term use (B=50.6, P<.001). The opposite was true for late-night meditation sessions (short-term: B=-2.06, P<.001; long-term: B=-51.7, P=.001). Significant relationships were not found for midday sessions (any outcome) or for evening sessions when examining long-term use. Additionally, temporal consistency in the performance of morning meditation sessions was associated with better short-term engagement (B=-1.64, P<.001) but worse long-term use (B=55.8, P<.001). Similar-sized temporal consistency effects were found for all other time windows. Meditating in the morning was associated with higher rates of maintaining a meditation practice with the app. This is consistent with findings from other studies that have hypothesized that the strength of existing morning routines and circadian rhythms may make the morning an ideal time to build new habits. In the long term, less temporal consistency in meditation sessions was associated with more persistent app use, suggesting there are benefits from maintaining flexibility in behavior performance. These findings improve our understanding of how to promote enduring healthy lifestyles and can inform the design of mHealth strategies for maintaining behavior changes.

An interpretable transformer network for the retinal disease classification using optical coherence tomography

Retinal illnesses such as age-related macular degeneration and diabetic macular edema will lead to irreversible blindness. With optical coherence tomography (OCT), doctors are able to see cross-sections of the retinal layers and provide patients with a diagnosis. Manual reading of OCT images is time-consuming, labor-intensive and even error-prone. Computer-aided diagnosis algorithms improve efficiency by automatically analyzing and diagnosing retinal OCT images. However, the accuracy and interpretability of these algorithms can be further improved through effective feature extraction, loss optimization and visualization analysis. In this paper, we propose an interpretable Swin-Poly Transformer network for performing automatically retinal OCT image classification. By shifting the window partition, the Swin-Poly Transformer constructs connections between neighboring non-overlapping windows in the previous layer and thus has the flexibility to model multi-scale features. Besides, the Swin-Poly Transformer modifies the importance of polynomial bases to refine cross entropy for better retinal OCT image classification. In addition, the proposed method also provides confidence score maps, assisting medical practitioners to understand the models’ decision-making process. Experiments in OCT2017 and OCT-C8 reveal that the proposed method outperforms both the convolutional neural network approach and ViT, with an accuracy of 99.80% and an AUC of 99.99%.

Cryptocurrencies and Long-Range Trends

In this study we investigate possible long-range trends in the cryptocurrency market. We employed the Hurst exponent in a sample covering the period from 1 January 2016 to 26 March 2021. We calculated the Hurst exponent in three non-overlapping consecutive windows and in the whole sample. Using these windows, we assessed the dynamic evolution in the structure and long-range trend behavior of the cryptocurrency market and evaluated possible changes in their behavior towards an efficient market. The innovation of this research is that we employ the Hurst exponent to identify the long-range properties, a tool that is seldomly used in analysis of this market. Furthermore, the use of both the R/S and the DFA analysis and the use of non-overlapping windows enhance our research’s novelty. Finally, we estimated the Hurst exponent for a wide sample of cryptocurrencies that covered more than 80% of the entire market for the last six years. The empirical results reveal that the returns follow a random walk making it difficult to accurately forecast them.

CUR Transformer: A Convolutional Unbiased Regional Transformer for Image Denoising

Image denoising is a fundamental problem in computer vision and multimedia computation. Non-local filters are effective for image denoising. But existing deep learning methods that use non-local computation structures are mostly designed for high-level tasks, and global self-attention is usually adopted. For the task of image denoising, they have high computational complexity and have a lot of redundant computation of uncorrelated pixels. To solve this problem and combine the marvelous advantages of non-local filter and deep learning, we propose a Convolutional Unbiased Regional (CUR) transformer. Based on the prior that, for each pixel, its similar pixels are usually spatially close, our insights are that (1) we partition the image into non-overlapped windows and perform regional self-attention to reduce the search range of each pixel, and (2) we encourage pixels across different windows to communicate with each other. Based on our insights, the CUR transformer is cascaded by a series of convolutional regional self-attention (CRSA) blocks with U-style short connections. In each CRSA block, we use convolutional layers to extract the query, key, and value features, namely Q , K , and V , of the input feature. Then, we partition the Q , K , and V features into local non-overlapped windows and perform regional self-attention within each window to obtain the output feature of this CRSA block. Among different CRSA blocks, we perform the unbiased window partition by changing the partition positions of the windows. Experimental results show that the CUR transformer outperforms the state-of-the-art methods significantly on four low-level vision tasks, including real and synthetic image denoising, JPEG compression artifact reduction, and low-light image enhancement.

Radiomic phenotyping of the lung parenchyma in a lung cancer screening cohort

High-throughput extraction of radiomic features from low-dose CT scans can characterize the heterogeneity of the lung parenchyma and potentially aid in identifying subpopulations that may have higher risk of lung diseases, such as COPD, and lung cancer due to inflammation or obstruction of the airways. We aim to determine the feasibility of a lung radiomics phenotyping approach in a lung cancer screening cohort, while quantifying the effect of different CT reconstruction algorithms on phenotype robustness. We identified low-dose CT scans (n = 308) acquired with Siemens Healthineers scanners from patients who completed low-dose CT within our lung cancer screening program between 2015 and 2018 and had two different sets of image reconstructions kernel available (i.e., medium (I30f.), sharp (I50f.)) for the same acquisition. Following segmentation of the lung field, a total of 26 radiomic features were extracted from the entire 3D lung-field using a previously validated fully-automated lattice-based software pipeline, adapted for low-dose CT scans. The lattice in-house software was used to extract features including gray-level histogram, co-occurrence, and run-length descriptors. The lattice approach uses non-overlapping windows for traversing along pixels of images and calculates different features. Each feature was averaged for each scan within a range of lattice window sizes (W) of 4, 8 and 20 mm. The extracted imaging features from both datasets were harmonized to correct for differences in image acquisition parameters. Subsequently, unsupervised hierarchical clustering was applied on the extracted features to identify distinct phenotypic patterns of the lung parenchyma, where consensus clustering was used to identify the optimal number of clusters (K = 2). Differences between phenotypes for demographic and clinical covariates including sex, age, BMI, pack-years of smoking, Lung-RADS and cancer diagnosis were assessed for each phenotype cluster, and then compared across clusters for the two different CT reconstruction algorithms using the cluster entanglement metric, where a lower entanglement coefficient corresponds to good cluster alignment. Furthermore, an independent set of low-dose CT scans (n = 88) from patients with available pulmonary function data on lung obstruction were analyzed using the identified optimal clusters to assess associations to lung obstruction and validate the lung phenotyping paradigm. Heatmaps generated by radiomic features identified two distinct lung parenchymal phenotype patterns across different feature extraction window sizes, for both reconstruction algorithms (P < 0.05 with K = 2). Associations of radiomic-based clusters with clinical covariates showed significant differences for BMI and pack-years of smoking (P < 0.05) for both reconstruction kernels. Radiomic phenotype patterns were more similar across the two reconstructed kernels, when smaller window sizes (W = 4 and 8 mm) were used for radiomic feature extraction, as deemed by their entanglement coefficient. Validation of clustering approaches using cluster mapping for the independent sample with lung obstruction also showed two statistically significant phenotypes (P < 0.05) with significant difference for BMI and smoking pack-years. Radiomic analysis can be used to characterize lung parenchymal phenotypes from low-dose CT scans, which appear reproducible for different reconstruction kernels. Further work should seek to evaluate the effect of additional CT acquisition parameters and validate these phenotypes in characterizing lung cancer screening populations, to potentially better stratify disease patterns and cancer risk.

Pro-Ictal State in Human Temporal Lobe Epilepsy

BackgroundStudies of continuous electroencephalography (EEG) suggest that seizures in individuals with focal-onset epilepsies preferentially occur during periods of heightened risk, typified by pathologic brain activities, termed pro-ictal states; however, the presence of (pathologic) pro-ictal states among a plethora of otherwise physiologic (e.g., sleep–wake cycle) states has not been established.MethodsWe studied a prospective, consecutive series of 15 patients with temporal lobe epilepsy who underwent limbic thalamic recordings in addition to routine (cortical) intracranial EEG for seizure localization. For each participant, pro-ictal (45 minutes before seizure onset) and interictal (4 hours removed from all seizures) EEG segments were divided into 10-minute, nonoverlapping windows, which were randomly distributed into training and validation cohorts in a 1:1 ratio. A deep neural classifier was applied to distinguish pro-ictal from interictal brain activities in a patient-specific fashion.ResultsWe analyzed 1800 patient-hours of continuous thalamocortical EEG. Distinct pro-ictal states were detected in each participant. The median area under the receiver-operating characteristic curve of the classifier was 0.92 (interquartile range, 0.90–0.96). Pro-ictal states were distinguished at least 45 minutes before seizure onset in 13 of 15 participants; in 2 of 15 participants, they were distinguished up to 35 minutes prior.ConclusionsOn the basis of thalamocortical EEG, pro-ictal states — pathologic brain activities during periods of heightened seizure risk — could be identified in patients with temporal lobe epilepsy and were detected, in our small sample, more than one half hour before seizure onset.

AccNet24: A deep learning framework for classifying 24-hour activity behaviours from wrist-worn accelerometer data under free-living environments

ObjectiveAlthough machine learning techniques have been repeatedly used for activity prediction from wearable devices, accurate classification of 24-hour activity behaviour categories from accelerometry data remains a challenge. We developed and validated a deep learning-based framework for classifying 24-hour activity behaviours from wrist-worn accelerometers. MethodsUsing an openly available dataset with free-living wrist-based raw accelerometry data from 151 participants (aged 18–91 years), we developed a deep learning framework named AccNet24 to classify 24-hour activity behaviours. First, the acceleration signal (x, y, and z-axes) was segmented into 30-second nonoverlapping windows, and signal-to-image conversion was performed for each segment. Deep features were automatically extracted from the signal images using transfer learning and transformed into a lower-dimensional feature space. These transformed features were then employed to classify the activity behaviours as sleep, sedentary behaviour, and light-intensity (LPA) and moderate-to-vigorous physical activity (MVPA) using a bidirectional long short-term memory (BiLSTM) recurrent neural network. AccNet24 was trained and validated with data from 101 and 25 randomly selected participants and tested with the remaining unseen 25 participants. We also extracted 112 hand-crafted time and frequency domain features from 30-second windows and used them as inputs to five commonly used machine learning classifiers, including random forest, support vector machines, artificial neural networks, decision tree, and naïve Bayes to classify the 24-hour activity behaviour categories. ResultsUsing the same training, validation, and test data and window size, the classification accuracy of AccNet24 outperformed the accuracy of the other five machine learning classification algorithms by 16%–30% on unseen data. ConclusionAccNet24, relying on signal-to-image conversion, deep feature extraction, and BiLSTM achieved consistently high accuracy (>95 %) in classifying the 24-hour activity behaviour categories as sleep, sedentary, LPA, and MVPA. The next generation accelerometry analytics may rely on deep learning techniques for activity prediction.

Inertial Sensor-Based Sport Activity Advisory System Using Machine Learning Algorithms.

The aim of this study was to develop a physical activity advisory system supporting the correct implementation of sport exercises using inertial sensors and machine learning algorithms. Specifically, three mobile sensors (tags), six stationary anchors and a system-controlling server (gateway) were employed for 15 scenarios of the series of subsequent activities, namely squats, pull-ups and dips. The proposed solution consists of two modules: an activity recognition module (ARM) and a repetition-counting module (RCM). The former is responsible for extracting the series of subsequent activities (so-called scenario), and the latter determines the number of repetitions of a given activity in a single series. Data used in this study contained 488 three defined sport activity occurrences. Data processing was conducted to enhance performance, including an overlapping and non-overlapping window, raw and normalized data, a convolutional neural network (CNN) with an additional post-processing block (PPB) and repetition counting. The developed system achieved satisfactory accuracy: CNN + PPB: non-overlapping window and raw data, 0.88; non-overlapping window and normalized data, 0.78; overlapping window and raw data, 0.92; overlapping window and normalized data, 0.87. For repetition counting, the achieved accuracies were 0.93 and 0.97 within an error of ±1 and ±2 repetitions, respectively. The archived results indicate that the proposed system could be a helpful tool to support the correct implementation of sport exercises and could be successfully implemented in further work in the form of web application detecting the user's sport activity.

Adaptive Window Based 3-D Feature Selection for Multispectral Image Classification Using Firefly Algorithm

Feature extraction is the most critical step in classification of multispectral image. The classification accuracy is mainly influenced by the feature sets that are selected to classify the image. In the past, handcrafted feature sets are used which are not adaptive for different image domains. To overcome this, an evolutionary learning method is developed to automatically learn the spatial-spectral features for classification. A modified Firefly Algorithm (FA) which achieves maximum classification accuracy with reduced size of feature set is proposed to gain the interest of feature selection for this purpose. For extracting the most efficient features from the data set, we have used 3-D discrete wavelet transform which decompose the multispectral image in all three dimensions. For selecting spatial and spectral features we have studied three different approaches namely overlapping window (OW-3DFS), non-overlapping window (NW-3DFS) adaptive window cube (AW-3DFS) and Pixel based technique. Fivefold Multiclass Support Vector Machine (MSVM) is used for classification purpose. Experiments conducted on Madurai LISS IV multispectral image exploited that the adaptive window approach is used to increase the classification accuracy.

VDTR: Video Deblurring With Transformer

Video deblurring is still an unsolved problem due to the challenging spatio-temporal modeling process. While existing convolutional neural network (CNN)-based methods show a limited capacity of effective spatial and temporal modeling for video deblurring. This paper presents VDTR, an effective Transformer-based model that makes the first attempt to adapt pure Transformer for video deblurring. VDTR exploits the superior long-range and relation modeling capabilities of Transformer for both spatial and temporal modeling. However, it is challenging to design an appropriate Transformer-based model for video deblurring due to the complicated non-uniform blurs, misalignment across multiple frames and the high computational costs for high-resolution spatial modeling. To address these problems, VDTR advocates performing attention within non-overlapping windows and exploiting the hierarchical structure for long-range dependencies modeling. For frame-level spatial modeling, we propose an encoder-decoder Transformer that utilizes multi-scale features for deblurring. For multi-frame temporal modeling, we adapt Transformer to fuse multiple spatial features efficiently. Compared with CNN-based methods, the proposed method achieves highly competitive results on both synthetic and real-world video deblurring benchmarks, including DVD, GOPRO, REDS and BSD. We hope such a Transformer-based architecture can serve as a powerful alternative baseline for video deblurring and other video restoration tasks. The source code will be available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/ljzycmd/VDTR</uri> .

A Temporal Dependency Learning CNN With Attention Mechanism for MI-EEG Decoding.

Deep learning methods have been widely explored in motor imagery (MI)-based brain computer interface (BCI) systems to decode electroencephalography (EEG) signals. However, most studies fail to fully explore temporal dependencies among MI-related patterns generated in different stages during MI tasks, resulting in limited MI-EEG decoding performance. Apart from feature extraction, learning temporal dependencies is equally important to develop a subject-specific MI-based BCI because every subject has their own way of performing MI tasks. In this paper, a novel temporal dependency learning convolutional neural network (CNN) with attention mechanism is proposed to address MI-EEG decoding. The network first learns spatial and spectral information from multi-view EEG data via the spatial convolution block. Then, a series of non-overlapped time windows is employed to segment the output data, and the discriminative feature is further extracted from each time window to capture MI-related patterns generated in different stages. Furthermore, to explore temporal dependencies among discriminative features in different time windows, we design a temporal attention module that assigns different weights to features in various time windows and fuses them into more discriminative features. The experimental results on the BCI Competition IV-2a (BCIC-IV-2a) and OpenBMI datasets show that our proposed network outperforms the state-of-the-art algorithms and achieves the average accuracy of 79.48%, improved by 2.30% on the BCIC-IV-2a dataset. We demonstrate that learning temporal dependencies effectively improves MI-EEG decoding performance. The code is available at https://github.com/Ma-Xinzhi/LightConvNet.

Extended Detrended Fluctuation Analysis of Coarse-Grained Time Series

A coarse-graining procedure, which involves averaging time series in non-overlapping windows followed by processing of the obtained multiple data sets, is the initial step in the multiscale entropy computation method. In this paper, we discuss how this procedure can be applied with other methods of time series analysis. Based on extended detrended fluctuation analysis (EDFA), we compare signal processing results for data sets with and without coarse-graining. Using the simulated data provided by the interacting nephrons model, we show how this procedure increases, up to 48%, the distinctions between local scaling exponents quantifying synchronous and asynchronous chaotic oscillations. Based on the experimental data of electrocorticograms (ECoG) of mice, an improvement in differences in local scaling exponents up to 41% and Student's t-values up to 34% was revealed.

TC-Net: A Transformer Capsule Network for EEG-based emotion recognition

Deep learning has recently achieved remarkable success in emotion recognition based on Electroencephalogram (EEG), in which convolutional neural networks (CNNs) are the mostly used models. However, due to the local feature learning mechanism, CNNs have difficulty in capturing the global contextual information involving temporal domain, frequency domain, intra-channel and inter-channel. In this paper, we propose a Transformer Capsule Network (TC-Net), which mainly contains an EEG Transformer module to extract EEG features and an Emotion Capsule module to refine the features and classify the emotion states. In the EEG Transformer module, EEG signals are partitioned into non-overlapping windows. A Transformer block is adopted to capture global features among different windows, and we propose a novel patch merging strategy named EEG-PatchMerging (EEG-PM) to better extract local features. In the Emotion Capsule module, each channel of the EEG feature maps is encoded into a capsule to better characterize the spatial relationships among multiple features. Experimental results on two popular datasets (i.e., DEAP and DREAMER) demonstrate that the proposed method achieves the state-of-the-art performance in the subject-dependent scenario. Specifically, on DEAP (DREAMER), our TC-Net achieves the average accuracies of 98.76% (98.59%), 98.81% (98.61%) and 98.82% (98.67%) at valence, arousal and dominance dimensions, respectively. Moreover, the proposed TC-Net also shows high effectiveness in multi-state emotion recognition tasks using the popular VA and VAD models. The main limitation of the proposed model is that it tends to obtain relatively low performance in the cross-subject recognition task, which is worthy of further study in the future.

527. Antimicrobial stewardship for empirical treatment of bloodstream infection using machine learning clinical decision support

Abstract Background Undertreatment of sepsis is associated with increased mortality, but only 12 to 18% of people require broad-spectrum Gram negative or positive coverage. Unnecessary use of broad-spectrum treatments represents a modifiable driver for antimicrobial resistance (AMR). Here, we test the hypothesis that a recommendation algorithm built on ML models trained on electronic health record data can safely reduce the spectrum of activity for empiric antibiotic treatments relative to clinicians for patients with community-onset sepsis. Methods We included 29,508 people with a community-onset Adult Sepsis Event in the Mass General Brigham health system. We split the data into a training and test set by time, June 2016 to June 2020 and July 2020 to February 2021, respectively. We extracted 1,200 predictors and windowed time-varying features into non-overlapping windows. We used logistic regression, random forest and XGBoost to predict pathogen, and then predict AMR to treatments relevant to the most likely pathogen. We used dynamic thresholding to convert probabilities into AMR phenotypes and then chose the antibiotic of the narrowest spectrum among those susceptible. Outcomes were the percentage of decisions that were inappropriately narrow versus broad spectrum. We compared model performance to that of clinicians. Figure 1:Analytic pipeline Schematic of machine learning architecture, thresholding algorithm and calculation of outcomes. Results The highest performing models for pathogen prediction were XGBoost, with area under the receiver operator curve (AUROC) statistics ranging from 0.60 for Staphylococcus spp. to 0.68 for Proteus spp. Random forests performed best for AMR prediction, with AUROCs ranging from 0.59 for ampicillin-sulbactam to 0.74 for penicillin. The model utilized broad-spectrum therapy in 92.4% (92.2% - 92.7%) of patients compared to 89.3% (88.9% - 89.6%) for clinicians, a relative increase of 3.5%, but chose inappropriately narrow treatments in 7.3% (7.1% - 7.6%) of patients, compared to 9.4% (9.1% - 9.6%) for providers, a 29% reduction. Figure 2:Model performance for pathogen prediction AUROCs for prediction of pathogen groupings for patients with community-onset sepsis. All models are XGBoost. Figure 3:Model performance for AMR prediction AUROCs for prediction of antimicrobial resistance based on most likely predicted pathogens for patients with community-onset sepsis. All models are random forests. Figure 4:Post-hoc outcome analysis Comparison of rates of inappropriately narrow antibiotic treatment (IAT) and use of broad spectrum therapy for the machine learning decision algorithm versus clinicians for patients with community onset sepsis. Conclusion An ML decision support tool for antibiotic prescription in community-onset sepsis did not result in decreased broad spectrum use but was able to reduce inappropriately narrow treatment. Work is underway to refine feature sets, improve the thresholding algorithm and better account for time-varying confounders. Disclosures Sanjat Kanjilal, MD, MPH, GlaxoSmithKline: Advisor/Consultant|Roche Diagnostics: Honoraria David Sontag, PhD, Adobe: Grant/Research Support|ASAPP: Advisor/Consultant|Cureai Health: Stocks/Bonds|Facebook: Grant/Research Support|Google: Grant/Research Support|IBM: Grant/Research Support|SAP: Grant/Research Support|Takeda: Grant/Research Support.

Features derived from blood pressure and intracranial pressure predict elevated intracranial pressure events in critically ill children

Clinicians frequently observe hemodynamic changes preceding elevated intracranial pressure events. We employed a machine learning approach to identify novel and differentially expressed features associated with elevated intracranial pressure events in children with severe brain injuries. Statistical features from physiologic data streams were derived from non-overlapping 30-min analysis windows prior to 21 elevated intracranial pressure events; 200 records without elevated intracranial pressure events were used as controls. Ten Monte Carlo simulations with training/testing splits provided performance benchmarks for 4 machine learning approaches. XGBoost yielded the best performing predictive models. Shapley Additive Explanations analyses demonstrated that a majority of the top 20 contributing features consistently derived from blood pressure data streams up to 240 min prior to elevated intracranial events. The best performing prediction model was using the 30–60 min analysis window; for this model, the area under the receiver operating characteristic window using XGBoost was 0.82 (95% CI 0.81–0.83); the area under the precision-recall curve was 0.24 (95% CI 0.23–0.25), above the expected baseline of 0.1. We conclude that physiomarkers discernable by machine learning are concentrated within blood pressure and intracranial pressure data up to 4 h prior to elevated intracranial pressure events.

Low-power multilevel resistive switching in β-Ga2O3 based RRAM devices

In this study, multilevel switching at low-power in Ti/TiN/Ga2O3/Ti/Pt resistive random-access memory (RRAM) devices has been systematically studied. The fabricated RRAM device exhibits an excellent non-overlapping window between set and reset voltages of ∼1.1 V with a maximum R off/R on ratio of ∼103. Moreover, to the best of our knowledge, the multi-bit storage capability of these RRAM devices with a reasonably high R off/R on ratio is experimentally demonstrated, for the first time, for lower compliance currents at 10 μA, 20 μA and 50 μA. The multi-bit resistive switching behavior of the Ga2O3 RRAM device at a low compliance current paves the way for low-power and high-density data storage applications.

Age-sensitive high density surface electromyogram indices for detecting muscle fatigue using core shape modelling

The purpose of this preliminary study was to examine age-sensitive High Density surface Electromyogram (HD-sEMG) features by Core Shape Modelling (CSM) method. Fatiguing low force isometric contractions of the biceps brachii was performed by eight young (age, 24.40 ± 2.42 years) and five elderly (72.90 ± 2.21 years) males, while HD-sEMG recorded signals from the biceps brachii. The task was performed at 20 % maximal voluntary contraction (MVC). From the recorded HD-sEMG signals, three Probability Density Function (PDF) shape distances (SD) measures the departure from Gaussianity, i.e. Left (LSD), Right (RSD), and Central (CSD), were derived by the CSM method from non-overlapping five-second windows until task failure. A linear regression analysis was then used to quantify the change of these shape parameters throughout the contraction. The resultant slopes revealed that the elderly group showed a decreasing trend in PDF shape parameters as the contraction approached task failure. In contrast, the young showed an increasing trend. Statistical differences between the two groups were found for LSD (p = 0.006) and RSD (p = 0.001). No such age-sensitivity was detected using conventional sEMG fatigue features. These results suggest that the proposed CSM method can be used to obtain fatigue-related features from HD-sEMG that are age-sensitive and possibly related to different motor unit recruitment and synchronization schemes.