Learning-based Classification Research Articles

From Stability to Variability: Classification of Healthy Individuals, Prediabetes, and Type 2 Diabetes using Glycemic Variability Indices from Continuous Glucose Monitoring Data.

This study aims to investigate the continuum of glucose control from normoglycemia to dysglycemia (HbA1c ≥ 5.7% / 39 mmol/mol) using metrics derived from Continuous Glucose Monitoring (CGM). Additionally, we aim to develop a machine learning-based classification model to classify dysglycemia based on observed patterns. Data from five distinct studies, each featuring at least two days of CGM, were pooled. Participants included individuals classified as healthy, with prediabetes, or with type 2 diabetes mellitus (T2DM). Various CGM indices were extracted and compared across groups. The dataset was split 70/30 for training and testing two classification models (XGBoost / Logistic Regression) to differentiate between prediabetes or dysglycemia and the healthy group. The analysis included 836 participants (healthy: n=282; prediabetes: n=133; T2DM: n=432). Across all CGM indices, a progressive shift was observed from the healthy group to those with diabetes (p<0.001). Statistically significant differences (p<0.01) were noted in mean glucose, Time Below Range, Time Above 140 mg/dl, Mmobility, Multiscale Complexity Index and Glycemic Risk Index when transitioning from health to prediabetes. The XGBoost models achieved the highest Receiver Operating Characteristic Area Under the Curve (ROC-AUC) values on the test dataset ranging from 0.91 [CI: 0.87-0.95] (prediabetes identification) to 0.97 [CI: 0.95-0.98] (Dysglycemia identification). Our findings demonstrate a gradual deterioration of glucose homeostasis and increased glycemic variability across the spectrum from normo- to dysglycemia, as evidenced by CGM metrics. The performance of CGM-based indices in classifying healthy individuals and those with prediabetes and diabetes is promising.

Simultaneous Utilization of Mood Disorder Questionnaire and Bipolar Spectrum Diagnostic Scale for Machine Learning-Based Classification of Patients With Bipolar Disorders and Depressive Disorders.

Bipolar and depressive disorders are distinct disorders with clearly different clinical courses, however, distinguishing between them often presents clinical challenges. This study investigates the utility of self-report questionnaires, the Mood Disorder Questionnaire (MDQ) and Bipolar Spectrum Diagnostic Scale (BSDS), with machine learning-based multivariate analysis, to classify patients with bipolar and depressive disorders. A total of 189 patients with bipolar disorders and depressive disorders were included in the study, and all participants completed both the MDQ and BSDS questionnaires. Machine-learning classifiers, including support vector machine (SVM) and linear discriminant analysis (LDA), were exploited for multivariate analysis. Classification performance was assessed through cross-validation. Both MDQ and BSDS demonstrated significant differences in each item and total scores between the two groups. Machine learning-based multivariate analysis, including SVM, achieved excellent discrimination levels with area under the ROC curve (AUC) values exceeding 0.8 for each questionnaire individually. In particular, the combination of MDQ and BSDS further improved classification performance, yielding an AUC of 0.8762. This study suggests the application of machine learning to MDQ and BSDS can assist in distinguishing between bipolar and depressive disorders. The potential of combining high-dimensional psychiatric data with machine learning-based multivariate analysis as an effective approach to psychiatric disorders.

Identifying sex from pharyngeal images using deep learning algorithm

The pharynx is one of the few areas in the body where blood vessels and immune tissues can readily be observed from outside the body non-invasively. Although prior studies have found that sex could be identified from retinal images using artificial intelligence, it remains unknown as to whether individuals’ sex could also be identified using pharyngeal images. Demographic information and pharyngeal images were collected from patients who visited 64 primary care clinics in Japan for influenza-like symptoms. We trained a deep learning-based classification model to predict reported sex, which incorporated a multiple instance convolutional neural network, on 20,319 images from 51 clinics. Validation was performed using 4869 images from the remaining 13 clinics not used for the training. The performance of the classification model was assessed using the area under the receiver operating characteristic curve. To interpret the model, we proposed a framework that combines a saliency map and organ segmentation map to quantitatively evaluate salient regions. The model achieved the area under the receiver operating characteristic curve of 0.883 (95% CI 0.866–0.900). In subgroup analyses, a substantial improvement in classification performance was observed for individuals aged 20 and older, indicating that sex-specific patterns between women and men may manifest as humans age (e.g., may manifest after puberty). The saliency map suggested the model primarily focused on the posterior pharyngeal wall and the uvula. Our study revealed the potential utility of pharyngeal images by accurately identifying individuals’ reported sex using deep learning algorithm.

A survey on machine learning system for intraductal papillary mucinous neoplasms detection

IPMN cysts, a pre-malignant risk to the pancreas, have the potential to develop into pancreatic cancer. Accurately identifying and evaluating the risk level is crucial for planning an efficient treatment strategy. However, this task is immensely challenging due to the varied and irregular shapes, textures, and sizes of IPMN cysts, as well as those of the pancreas itself. In this study, we introduce a new computer-aided diagnostic approach for classifying IPMN risk levels based on multi-contrast MRI scans. The proposed analysis framework comprises an efficient volumetric self-adapting segmentation strategy for delineating the pancreas, followed by a newly developed deep learning-based classification scheme incorporating a radiomics-based predictive approach. To evaluate the proposed decision-fusion model, we use multi-centre datasets and multi-contrast MRI scans, aiming to achieve superior performance compared to the current state of the art in this field. The ablation studies illustrate the importance of both radiomics and deep learning modules in achieving a new state-of-the-art (SOTA) performance compared to international guidelines and published studies (81.9% vs 61.3% in accuracy). These key findings carry significant implications for clinical decision-making, potentially revolutionizing the way IPMN risk levels are classified. Through a series of rigorous experiments on multi-centre datasets (involving more MRI scans from five centers), we attained unprecedented performance levels with moderate accuracy. The code will be made available upon publication.

A medical image classification method based on self-regularized adversarial learning.

Deep learning (DL) techniques have been extensively applied in medical image classification. The unique characteristics of medical imaging data present challenges, including small labeled datasets, severely imbalanced class distribution, and significant variations in imaging quality. Recently, generative adversarial network (GAN)-based classification methods have gained attention for their ability to enhance classification accuracy by incorporating realistic GAN-generated images as data augmentation. However, the performance of these GAN-based methods often relies on high-quality generated images, while large amounts of training data are required to train GAN models to achieve optimalperformance. In this study, we propose an adversarial learning-based classification framework to achieve better classification performance. Innovatively, GAN models are employed as supplementary regularization terms to support classification, aiming to address the challenges describedabove. The proposed classification framework, GAN-DL, consists of a feature extraction network (F-Net), a classifier, and two adversarial networks, specifically a reconstruction network (R-Net) and a discriminator network (D-Net). The F-Net extracts features from input images, and the classifier uses these features for classification tasks. R-Net and D-Net have been designed following the GAN architecture. R-Net employs the extracted feature to reconstruct the original images, while D-Net is tasked with the discrimination between the reconstructed image and the original images. An iterative adversarial learning strategy is designed to guide model training by incorporating multiple network-specific loss functions. These loss functions, serving as supplementary regularization, are automatically derived during the reconstruction process and require no additional dataannotation. To verify the model's effectiveness, we performed experiments on two datasets, including a COVID-19 dataset with 13958 chest x-ray images and an oropharyngeal squamous cell carcinoma (OPSCC) dataset with 3255 positron emission tomography images. Thirteen classic DL-based classification methods were implemented on the same datasets for comparison. Performance metrics included precision, sensitivity, specificity, and -score. In addition, we conducted ablation studies to assess the effects of various factors on model performance, including the network depth of F-Net, training image size, training dataset size, and loss function design. Our method achieved superior performance than all comparative methods. On the COVID-19 dataset, our method achieved , , , and in terms of precision, sensitivity, specificity, and -score, respectively. It achieved across all these metrics on the OPSCC dataset. The study to investigate the effects of two adversarial networks highlights the crucial role of D-Net in improving model performance. Ablation studies further provide an in-depth understanding of ourmethodology. Our adversarial-based classification framework leverages GAN-based adversarial networks and an iterative adversarial learning strategy to harness supplementary regularization during training. This design significantly enhances classification accuracy and mitigates overfitting issues in medical image datasets. Moreover, its modular design not only demonstrates flexibility but also indicates its potential applicability to various clinical contexts and medical imaging applications.

Harmonic enhancement to optimize EOG based ocular activity decoding: A hybrid approach with harmonic source separation and EEMD

Intelligent robotic systems for patients with motor impairments have gained significant interest over the past few years. Various sensor types and human-machine interface (HMI) methods have been developed; however, most research in this area has focused on eye-blink-based binary control with minimal electrode placements. This approach restricts the complexity of HMI systems and does not consider the potential of multiple-activity decoding via static ocular activities. These activities pose a decoding challenge due to non-oscillatory noise components, such as muscle tremors or fatigue. To address this issue, a hybrid preprocessing methodology is proposed that combines harmonic source separation and ensemble empirical mode decomposition in the time-frequency domain to remove percussive and non-oscillatory components of static ocular movements. High-frequency components are included in the harmonic enhancement process. Next, a machine learning model with dual input of time-frequency images and a vectorized feature set of consecutive time windows is employed, leading to a 3.8% increase in performance as compared to without harmonic enhancement in leave-one-session-out cross-validation (LOSO). Additionally, a high correlation is found between the harmonic ratios of the static activities in the Hilbert-Huang frequency spectrum and LOSO performances. This finding highlights the potential of leveraging the harmonic characteristics of the activities as a discriminating factor in machine learning-based classification of EOG-based ocular activities, thus providing a new aspect of activity enrichment with minimal performance loss for future HMI systems.

Machine learning allows robust classification of visceral fat in women with obesity using common laboratory metrics

The excessive accumulation and malfunctioning of visceral adipose tissue (VAT) is a major determinant of increased risk of obesity-related comorbidities. Thus, risk stratification of people living with obesity according to their amount of VAT is of clinical interest. Currently, the most common VAT measurement methods include mathematical formulae based on anthropometric dimensions, often biased by human measurement errors, bio-impedance, and image techniques such as X-ray absorptiometry (DXA) analysis, which requires specialized equipment. However, previous studies showed the possibility of classifying people living with obesity according to their VAT through blood chemical concentrations by applying machine learning techniques. In addition, most of the efforts were spent on men living with obesity while little was done for women. Therefore, this study aims to compare the performance of the multilinear regression model (MLR) in estimating VAT and six different supervised machine learning classifiers, including logistic regression (LR), support vector machine and decision tree-based models, to categorize 149 women living with obesity. For clustering, the study population was categorized into classes 0, 1, and 2 according to their VAT and the accuracy of each MLR and classification model was evaluated using DXA-data (DXAdata), blood chemical concentrations (BLDdata), and both DXAdata and BLDdata together (ALLdata). Estimation error and R2\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\hbox {R}^{2}$$\\end{document} were computed for MLR, while receiver operating characteristic (ROC) and precision-recall curves (PR) area under the curve (AUC) were used to assess the performance of every classification model. MLR models showed a poor ability to estimate VAT with mean absolute error ≥401.40\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge 401.40$$\\end{document} and R2≤0.62\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\hbox {R}^{2} \\le 0.62$$\\end{document} in all the datasets. The highest accuracy was found for LR with values of 0.57, 0.63, and 0.53 for ALLdata, DXAdata, and BLDdata, respectively. The ROC AUC showed a poor ability of both ALLdata and DXAdata to distinguish class 1 from classes 0 and 2 (AUC = 0.31, 0.71, and 0.85, respectively) as also confirmed by PR (AUC = 0.24, 0.57, and 0.73, respectively). However, improved performances were obtained when applying LR model to BLDdata (ROC AUC ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge $$\\end{document} 0.61 and PR AUC ≥\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\ge $$\\end{document} 0.42), especially for class 1. These results seem to suggest that, while a direct and reliable estimation of VAT was not possible in our cohort, blood sample-derived information can robustly classify women living with obesity by machine learning-based classifiers, a fact that could benefit the clinical practice, especially in those health centres where medical imaging devices are not available. Nonetheless, these promising findings should be further validated over a larger population.

Fairness in Classifying and Grouping Health Equity Information.

This paper explores the balance between fairness and performance in machine learning classification, predicting the likelihood of a patient receiving anti-microbial treatment using structured data in community nursing wound care electronic health records. The data includes two important predictors (gender and language) of the social determinants of health, which we used to evaluate the fairness of the classifiers. At the same time, the impact of various groupings of language codes on classifiers' performance and fairness is analyzed. Most common statistical learning-based classifiers are evaluated. The findings indicate that while K-Nearest Neighbors offers the best fairness metrics among different grouping settings, the performance of all classifiers is generally consistent across different language code groupings. Also, grouping more variables tends to improve the fairness metrics over all classifiers while maintaining their performance.

Non-Invasive Quality Control of Organoid Cultures Using Mesofluidic CSTR Bioreactors and High-Content Imaging.

Human brain organoids produce anatomically relevant cellular structures and recapitulate key aspects of in vivo brain function, which holds great potential to model neurological diseases and screen therapeutics. However, the long growth time of 3D systems complicates the culturing of brain organoids and results in heterogeneity across samples hampering their applications. We developed an integrated platform to enable robust and long-term culturing of 3D brain organoids. We designed a mesofluidic bioreactor device based on a reaction-diffusion scaling theory, which achieves robust media exchange for sufficient nutrient delivery in long-term culture. We integrated this device with longitudinal tracking and machine learning-based classification tools to enable non-invasive quality control of live organoids. This integrated platform allows for sample pre-selection for downstream molecular analysis. Transcriptome analyses of organoids revealed that our mesofluidic bioreactor promoted organoid development while reducing cell death. Our platform thus offers a generalizable tool to establish reproducible culture standards for 3D cellular systems for a variety of applications beyond brain organoids.

Mediodorsal thalamus and ventral pallidum contribute to subcortical regulation of the default mode network

Humans and other animals readily transition from externally to internally focused attention, and these transitions are accompanied by activation of the default mode network (DMN). The DMN was considered a cortical network, yet recent evidence suggests subcortical structures are also involved. We investigated the role of ventral pallidum (VP) and mediodorsal thalamus (MD) in DMN regulation in tree shrew, a close relative of primates. Electrophysiology and deep learning-based classification of behavioral states revealed gamma oscillations in VP and MD coordinated with gamma in anterior cingulate (AC) cortex during DMN states. Cross-frequency coupling between gamma and delta oscillations was higher during DMN than other behaviors, underscoring the engagement of MD, VP and AC. Our findings highlight the importance of VP and MD in DMN regulation, extend homologies in DMN regulation among mammals, and underline the importance of thalamus and basal forebrain to the regulation of DMN.

A Deep Learning-based Classification Model for Arabic News Tweets Using Bidirectional Long Short-Term Memory Networks

This research develops a classification model for Arabic news tweets using Bidirectional Long Short-Term Memory networks (BiLSTM). Tweets about Arabic news were gathered between August 2016 and August 2020 and divided into five categories. Custom Python scripts, Twitter API and the GetOldTweets3 Python library were used to collect the data. BiLSTM was used to train and test the model. The results indicated an average accuracy, precision, recall, and f1-score of 0.88, 0.92, 0.88, and 0.89, respectively. The results could have practical implications for Arabic machine learning and NLP tasks in research and practice.

A proposed secure framework for protecting cloud-based educational systems from hacking

Educational institutions and users involved in the whole learning process frequently have concerns about the storage and processing of sensitive data and essential apps in the cloud. Security and privacy issues have emerged as a major challenge, limiting cloud computing’s implementation in educational environments. Several users have yet to meet this security challenge, which is linked to the system’s multi-tenancy nature and the outsourcing of resources and data. This study proposes a secure framework for protecting cloud-based educational systems from hacking using a unique encryption technique, as well as a deep learning-based classification for cloud attack detection. Initially, we preprocess the data and extract features using a gray-level covariance matrix (GLCM). Next, we propose a classification based on multiple convolutional neural networks (M−CNN) to detect attacks in the cloud environment. Finally, we propose a modified digital signature algorithm (MDSA) for data encryption and decryption. The proposed technique achieved high security rates, with an accuracy of 97.7%, sensitivity of 96%, specificity of 94.3%, precision of 99.6%, and recall of 97%. Comparative evaluations showed that the proposed mechanism outperformed other encryption techniques. This novel model enhances the security of cloud-based educational systems and promotes users’ confidence in such platforms.

Deep learning-based classification and application test of multiple crop leaf diseases using transfer learning and the attention mechanism

Deep learning-based classification and application test of multiple crop leaf diseases using transfer learning and the attention mechanism

Contributions of brain regions to machine learning-based classifications of attention deficit hyperactivity disorder (ADHD) utilizing EEG signals

Objective The study presented focuses on the creation of a machine learning (ML) model that uses electrophysiological (EEG) data to identify kids with attention deficit hyperactivity disorder (ADHD) from healthy controls. The EEG signals are acquired during cognitive tasks to distinguish children with ADHD from their counterparts. Methodology The EEG data recorded in cognitive exercises was filtered using low pass Bessel filter and notch filters to remove artifacts, by the data set owners. To identify unique EEG patterns, we used many well-known classifiers, including Naïve Bayes (NB), Random Forest, Decision Tree (DT), K-Nearest Neighbors (KNN), Support Vector Machine (SVM), AdaBoost and Linear Discriminant Analysis (LDA), to identify distinct EEG patterns. Input features comprised EEG data from nineteen channels, individually and in combination. Findings Study indicates that EEG-based categorization can differentiate between individuals with ADHD and healthy individuals with accuracy of 84%. The RF classifier achieved a maximum accuracy of 0.84 when particular region combinations were used. Evaluation of classification performance utilizing hemisphere-specific EEG data yielded promising outcomes, particularly in the right hemisphere channels. Novelty The study goes beyond traditional methodologies by investigating the effect of regional data on categorization results. The contributions of various brain regions to these classifications are being extensively researched. Understanding the role of different brain regions in ADHD can lead to better diagnosis and treatment options for individuals with ADHD. The study of categorization ability, utilizing EEG data specific to each hemisphere, particularly channels in the right hemisphere region, provides further granularity to the findings.

In vitro metabolic studies and machine learning analysis of mass spectrometry data: A dual strategy for differentiating alpha-pyrrolidinohexiophenone (α-PHP) and alpha-pyrrolidinoisohexanophenone (α-PiHP) in urine analysis

Synthetic cathinones are some of the most prevalent new psychoactive substances (NPSs) globally, with alpha-pyrrolidinoisohexanophenone (α-PiHP) being particularly noted for its widespread use in the United States, Europe, and Taiwan. However, the analysis of isomeric NPSs such as α-PiHP and alpha-pyrrolidinohexiophenone (α-PHP) is challenging owing to similarities in their retention times and mass spectra. This study proposes a dual strategy based on in vitro metabolic experiments and machine learning-based classification modelling for differentiating α-PHP and α-PiHP in urine samples: (1) in vitro metabolic experiments using pooled human liver microsomes and liquid chromatography tandem quadrupole time-of-flight mass spectrometry (LC-QTOF-MS) were conducted to identify the key metabolites of α-PHP and α-PiHP from the high-resolution MS/MS spectra. After 5 h incubation, 71.4 % of α-PHP and 64.7 % of α-PiHP remained unmetabolised. Nine phase I metabolites were identified for each compound, including primary β-ketone reduction (M1) metabolites. Comparing the metabolites and retention times confirmed the efficacy of in vitro metabolic experiments for differentiating NPS isomers. Subsequently, analysis of seven real urine samples revealed the presence for various metabolites, including M1, that could be used as suitable detection markers at low concentrations. The aliphatic hydroxylation (M2) metabolite peak counts and metabolite retention times were used to determine α-PiHP use. (2) Classification models for the parent compounds and M1 metabolites were developed using principal component analysis for feature extraction and logistic regression for classification. The training and test sets were devised from the spectra of standard samples or supernatants from in vitro metabolism experiments with different incubation times. Both models had classification accuracies of 100 % and accurately identified α-PiHP and its M1 metabolite in seven real urine samples. The proposed methodology effectively distinguished between such isomers and confirmed their presence at low concentrations. Overall, this study introduces a novel concept that addresses the complexities in analysing isomeric NPSs and suggests a path towards enhancing the accuracy and reliability of NPS detection.

Deep Residual Learning-Based Classification with Identification of Incorrect Predictions and Quantification of Cellularity and Nuclear Morphological Features in Digital Pathological Images of Common Astrocytic Tumors.

Interobserver variations in the pathology of common astrocytic tumors impact diagnosis and subsequent treatment decisions. This study leveraged a residual neural network-50 (ResNet-50) in digital pathological images of diffuse astrocytoma, anaplastic astrocytoma, and glioblastoma to recognize characteristic pathological features and perform classification at the patch and case levels with identification of incorrect predictions. In addition, cellularity and nuclear morphological features, including axis ratio, circularity, entropy, area, irregularity, and perimeter, were quantified via a hybrid task cascade (HTC) framework and compared between different characteristic pathological features with importance weighting. A total of 95 cases, including 15 cases of diffuse astrocytoma, 11 cases of anaplastic astrocytoma, and 69 cases of glioblastoma, were collected in Taiwan Hualien Tzu Chi Hospital from January 2000 to December 2021. The results revealed that an optimized ResNet-50 model could recognize characteristic pathological features at the patch level and assist in diagnosis at the case level with accuracies of 0.916 and 0.846, respectively. Incorrect predictions were mainly due to indistinguishable morphologic overlap between anaplastic astrocytoma and glioblastoma tumor cell area, zones of scant vascular lumen with compact endothelial cells in the glioblastoma microvascular proliferation area mimicking the glioblastoma tumor cell area, and certain regions in diffuse astrocytoma with too low cellularity being misrecognized as the glioblastoma necrosis area. Significant differences were observed in cellularity and each nuclear morphological feature among different characteristic pathological features. Furthermore, using the extreme gradient boosting (XGBoost) algorithm, we found that entropy was the most important feature for classification, followed by cellularity, area, circularity, axis ratio, perimeter, and irregularity. Identifying incorrect predictions provided valuable feedback to machine learning design to further enhance accuracy and reduce errors in classification. Moreover, quantifying cellularity and nuclear morphological features with importance weighting provided the basis for developing an innovative scoring system to achieve objective classification and precision diagnosis among common astrocytic tumors.

Machine learning-based classification reveals distinct clusters of non-coding genomic allelic variations associated with Erm-mediated antibiotic resistance.

Antibiotic resistance (AR) poses a global health threat as the efficacy of available antibiotics has rapidly eroded due to the widespread transmission of AR genes. Using Erm-dependent MLS resistance as a model, this study highlights the significance of non-coding genomic allelic variations. Through a comprehensive analysis of upstream regulatory elements within the erm family, we elucidated the evolutionary emergence and development of AR mechanisms. Leveraging population-wide machine learning (ML)-based genomic analysis, we transformed substantial non-random allelic variations into discernible clusters of elements, enabling precise prediction of MLS phenotypes from non-coding regions. These findings offer deeper insight into AR evolution and demonstrate the potential of harnessing non-coding genomic allele data for accurately predicting AR phenotypes.

A Deep Learning Approach for Fear Recognition on the Edge Based on Two-Dimensional Feature Maps.

Applying affective computing techniques to recognize fear and combining them with portable signal monitors makes it possible to create real-time detection systems that could act as bodyguards when users are in danger. With this aim, this paper presents a fear recognition method based on physiological signals obtained from wearable devices. The procedure involves creating two-dimensional feature maps from the raw signals, using data augmentation and feature selection algorithms, followed by deep learning-based classification models, taking inspiration from those used in image processing. This proposal has been validated with two different datasets, achieving, in WEMAC, WESAD 3-classes, and WESAD 2-classes, F1-score results of 78.13%, 88.07%, and 99.60%, respectively, and 79.90%, 89.12%, and 99.60% in accuracy. Furthermore, the paper demonstrates the feasibility of implementing the proposed method on the Coral Edge TPU device, prepared to make inferences on the edge.

Detection and Classification of Obstructive Sleep Apnea Using Audio Spectrogram Analysis

Sleep disorders are steadily increasing in the population and can significantly affect daily life. Low-cost and noninvasive systems that can assist the diagnostic process will become increasingly widespread in the coming years. This work aims to investigate and compare the performance of machine learning-based classifiers for the identification of obstructive sleep apnea–hypopnea (OSAH) events, including apnea/non-apnea status classification, apnea–hypopnea index (AHI) prediction, and AHI severity classification. The dataset considered contains recordings from 192 patients. It is derived from a recently released dataset which contains, amongst others, audio signals recorded with an ambient microphone placed ∼1 m above the studied subjects and apnea/hypopnea accurate events annotations performed by specialized medical doctors. We employ mel spectrogram images extracted from the environmental audio signals as input of a machine-learning-based classifier for apnea/hypopnea events classification. The proposed approach involves a stacked model which utilizes a combination of a pretrained VGG-like audio classification (VGGish) network and a bidirectional long short-term memory (bi-LSTM) network. Performance analysis was conducted using a 5-fold cross-validation approach, leaving out patients used for training and validation of the models in the testing step. Comparative evaluations with recently presented methods from the literature demonstrate the advantages of the proposed approach. The proposed architecture can be considered a useful tool for supporting OSAHS diagnoses by means of low-cost devices such as smartphones.

A lightweight and privacy preserved federated learning ecosystem for analyzing verbal communication emotions in identical and non-identical databases

The lack of vocal emotional expression is a major deficit in social communication disorders. The current scenario of artificial intelligence focuses on collaborative training of deep learning models without losing data privacy. The primary objective of this paper is to propose a federated learning-based classification model to identify and analyze the emotional capabilities of individuals with vocal emotion deficits. The methodology has developed a collaborative and privacy-preserved approach using federated learning for training the deep learning models. The proposed methodology utilizes Mel-frequency Cepstral Coefficients (MFCC) to preprocess audio recordings. The four datasets (RAVDESS, CREMA, TESS, SAVEE) including emotion-based classified audio recordings were collected from open sources. The collected audio recordings are 3 s each and the total data set has 668376 audio files with happy - 175119 files, sad – 172611 files, angry – 176346 files, and normal - 144300 files. Further, the input audio was pre-processed to generate MFCC features. The study began with extracting features from multiple pre-trained DL models as its base model. Then, the performance of the federated learning (FL) model was tested on independent and identically distributed (IID) and non-IID data. Further, this paper presents a federated deep learning-based multimodal system for verbal communication emotions classification that uses audio datasets to meet data privacy requirements by DL on the FL ecosystem. As per the findings, the federated learning trained model provides nearly similar parametric results in comparison to base model training. For IID data, the model had 99.71 % validation accuracy, precision (99.73 %), recall (99.69 %), and validation loss (0.01). The FL architecture with non-IID data outperformed these measures with validation accuracy (99.97 %), precision (99.97 %), recall (99.97 %), and least loss (0). Hence the acquired results support the utilization of federated learning ecosystem-based trained models with identically and non-identically distributed audio features from emotion identification without losing parametric results. In conclusion, the proposed techniques could be applied to identify verbal emotional deficits in individuals and could support developing emerging technological interventions for their well-being.