Antonio Gabriel Ravelo García

The rapid spread of COVID-19 has resulted in more than 635 million cases worldwide, highlighting the need for efficient, noninvasive, and cost-effective diagnostic methods. Traditional diagnostic methods such as RT-PCR are effective but often invasive and slow. This systematic review, following Preferred Reporting Items for Systematic reviews and Meta-Analyses PRISMA guidelines evaluates studies using machine learning techniques on physiological data from wearable devices for COVID-19 diagnosis. We included 20 peer-reviewed articles, conference papers, and preprints published between November 2020 and January 2024, focusing on studies using physiological parameters collected through commercial wearable devices and machine learning for COVID-19 diagnosis. An analysis of the variables used and feature extraction is presented, highlighting the time-domain heart rate, and statistically due to Heart Rate Variability HRV. The most frequently used Machine Learning ML algorithms (Random Forest, Support Vector Machine, Logistic Regression, and k-Nearest Neighbor) in this context are also shown, as well as their performance analyzed by variable, algorithm, features used, and study population. The 20 included studies reported Appropriate Use Criteria area under the curve AUC values ranging from 0.770 to 0.994, indicating high diagnostic accuracy. The meta-analysis showed a high level of heterogeneity ( I 2 = 84%) across studies. This review highlights the potential of combining commercially available wearable technology withmachine learning for early and accurate detection of COVID-19, suggesting directions for future research to improve noninvasive diagnostic methods and suggest improvements in the reporting of results.

Physiological signs are key indicators of cardiovascular health, which can be estimated using remote photoplethysmography. Their estimations in dark environments are particularly important, where infrared based methods were predominantly applied, since they are illumination resistant. However, the extracted signals have poor pulsatile strength with low signal-to-noise ratio, eventually resulting in spurious estimates. Conversely, RGB based methods exhibits stronger pulsatile strength, but hindered by poor illumination. To overcome these limitations, we propose 2E1D-Net, trained using a self-created database acquired in a dark environment with marginal illuminance ≤ 1 lux. It comprises dual encoders that take paired input images captured at different exposure levels, and project them to a latent. The decoder then, elevates the noise (darkness) component from the dark image, followed by multiscale feature fusion, to produce enhanced images. 2E1D-Net was trained using a linear combination of multiscale structured-similarity-index, L1 and L2 losses, respectively. Subsequently, RGB heart rate and oxygen saturation methods cascaded to trained 2E1D-Net, were tested on self-created and public databases. Experimental results proved the superiority of 2E1D-Net, over state-of-the-art, which ensured the extended ability of RGB methods for physiological measurements in dark, thereby proposing RGB as reliable and clinically relevant alternative to infrared methods without performance compromise.

Uncertainty analysis of classification or regression models is a key feature of probabilistic approaches to supervised learning, allowing the assessment of how trustworthy predictions are. Just as boosting algorithms aim at obtaining accurate ensembles of simple classifiers, using a process guided by the accuracy of each of these classifiers, the method proposed in this paper builds an ensemble guided by the uncertainty of each of its individual models. The proposed method, named ProBoost (probabilistic boosting), uses the epistemic uncertainty of each training sample to determine those about which each model is most uncertain; the importance of these samples is then increased for the next learner, producing a sequence that progressively focuses on samples found to have the highest uncertainty. In the end, the learned models are combined into an ensemble. Thus, the approach goes beyond standard boosting methods, which usually focus on deterministic error correction, by quantifying predictive uncertainty to guide sequential training through dataset manipulation. Three methods are proposed to update the importance of the samples according to the uncertainty estimates at each stage: undersampling, oversampling, and weighting. Furthermore, two approaches are studied regarding the final ensemble combination. The learners herein considered are standard convolutional neural networks, and the probabilistic models underlying the uncertainty estimation use either variational inference or Monte Carlo dropout. The experimental evaluation carried out on MNIST and CIFAR 10 benchmark datasets shows that ProBoost yields significant performance improvement, compared to not using ProBoost, and outperforms a wider single model with a similar number of parameters.

Traditional methods for water-level measurement usually employ permanent structures, such as a scale built into the water system, which is costly and laborious and can wash away with water. This research proposes a low-cost, automatic water-level estimator that can appraise the level without disturbing water flow or affecting the environment. The estimator was developed for urban areas of a volcanic island water channel, using machine learning to evaluate images captured by a low-cost remote monitoring system. For this purpose, images from over one year were collected. For better performance, captured images were processed by converting them to a proposed color space, named HLE, composed of hue, lightness, and edge. Multiple residual neural network architectures were examined. The best-performing model was ResNeXt, which achieved a mean absolute error of 1.14 cm using squeeze and excitation and data augmentation. An explainability analysis was carried out for transparency and a visual explanation. In addition, models were developed to predict water levels. Three models successfully forecasted the subsequent water levels for 10, 60, and 120 min, with mean absolute errors of 1.76 cm, 2.09 cm, and 2.34 cm, respectively. The models could follow slow and fast transitions, leading to a potential flooding risk-assessment mechanism.

SARS-CoV-2 infection has a wide range of clinical manifestations making its diagnosis difficult, which is an important problem to solve. We evaluated heart rate data extracted from the Stanford University database. The data set considers heart rate and step records of 118 patients, where 90 correspond to healthy individuals and 28 patients with COVID. Each daily record was divided into 5-minute segments, providing 288 data per patient. The date of symptom onset was considered as a reference point to extract subsets of data whose variability was considerable, such as 30 days before the date and 30 days after it. Each of the 60 segments of 288 data per patient was treated using Permutation Entropy, Approximate Entropy, Spectral Entropy and Singular Value Decomposition Entropy. The average of the data from each group was used to construct the circadian profiles which were analyzed using the Mann-Whitney-Wilcoxon test, determining the most relevant 5-minute segments, whose p-value was less than 0.05. In this way, the Spectral Entropy was discarded as it did not show any significantly different segment. The efficiency of the method was reflected in the performance of a logistic model for binary classification proposed in this work, which reflected an accuracy of 94.12% in the PE case, 88% in the ApEn case and 94% in the SVDE case. The proposed analysis turns out to be highly efficient when detecting significant segments that allow improving the classification tasks carried out by Machine Learning models, which provides a basis for the study of statistics such as entropy to delimit databases and improve the performance of classifier models.

This study presents a novel approach for kernel selection based on Kullback–Leibler divergence in variational autoencoders using features generated by the convolutional encoder. The proposed methodology focuses on identifying the most relevant subset of latent variables to reduce the model’s parameters. Each latent variable is sampled from the distribution associated with a single kernel of the last encoder’s convolutional layer, resulting in an individual distribution for each kernel. Relevant features are selected from the sampled latent variables to perform kernel selection, which filters out uninformative features and, consequently, unnecessary kernels. Both the proposed filter method and the sequential feature selection (standard wrapper method) were examined for feature selection. Particularly, the filter method evaluates the Kullback–Leibler divergence between all kernels’ distributions and hypothesizes that similar kernels can be discarded as they do not convey relevant information. This hypothesis was confirmed through the experiments performed on four standard datasets, where it was observed that the number of kernels can be reduced without meaningfully affecting the performance. This analysis was based on the accuracy of the model when the selected kernels fed a probabilistic classifier and the feature-based similarity index to appraise the quality of the reconstructed images when the variational autoencoder only uses the selected kernels. Therefore, the proposed methodology guides the reduction of the number of parameters of the model, making it suitable for developing applications for resource-constrained devices.

This study conducted a systematic review to determine the feasibility of automatic Cyclic Alternating Pattern (CAP) analysis. Specifically, this review followed the 2020 Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) guidelines to address the formulated research question: is automatic CAP analysis viable for clinical application? From the identified 1,280 articles, the review included 35 studies that proposed various methods for examining CAP, including the classification of A phase, their subtypes, or the CAP cycles. Three main trends were observed over time regarding A phase classification, starting with mathematical models or features classified with a tuned threshold, followed by using conventional machine learning models and, recently, deep learning models. Regarding the CAP cycle detection, it was observed that most studies employed a finite state machine to implement the CAP scoring rules, which depended on an initial A phase classifier, stressing the importance of developing suitable A phase detection models. The assessment of A-phase subtypes has proven challenging due to various approaches used in the state-of-the-art for their detection, ranging from multiclass models to creating a model for each subtype. The review provided a positive answer to the main research question, concluding that automatic CAP analysis can be reliably performed. The main recommended research agenda involves validating the proposed methodologies on larger datasets, including more subjects with sleep-related disorders, and providing the source code for independent confirmation.

Cyclic Alternating Pattern (CAP) is a sleep instability marker defined based on the amplitude and frequency of the electroencephalogram signal. Because of the time and intensive process of labeling the data, different machine learning and automatic approaches are proposed. However, due to the low accuracy of the traditional approach and the black box approach of the machine learning approach, the proposed systems remain untrusted by the physician. This study contributes to accurately estimating CAP in a Frequency-Time domain by A-phase and its subtypes prediction by transforming the monopolar deviated electroencephalogram signals into corresponding scalograms. Subsequently, various computer vision classifiers were tested for the A-phase using scalogram images. It was found that MobileNetV2 outperformed all other tested classifiers by achieving the average accuracy, sensitivity, and specificity values of 0.80, 0.75, and 0.81, respectively. The MobileNetV2 trained model was further fine-tuned for A-phase subtypes prediction. To further verify the visual ability of the trained models, Gradcam++ was employed to identify the targeted regions by the trained network. It was verified that the areas identified by the model match the regions focused on by the sleep experts for A-phase predictions, thereby proving its clinical viability and robustness. This motivates the development of novel deep learning based methods for CAP patterns predictions.

In this paper, we thoroughly analyze the detection of sleep apnea events in the context of Obstructive Sleep Apnea (OSA), which is considered a public health problem because of its high prevalence and serious health implications. We especially evaluate patients who do not always show desaturations during apneic episodes (non-desaturating patients). For this purpose, we use a database (HuGCDN2014-OXI) that includes desaturating and non-desaturating patients, and we use the widely used Physionet Apnea Dataset for a meaningful comparison with prior work. Our system combines features extracted from the Heart-Rate Variability (HRV) and SpO₂, and it explores their potential to characterize desaturating and non-desaturating events. The HRV-based features include spectral, cepstral, and nonlinear information (Detrended Fluctuation Analysis (DFA) and Recurrence Quantification Analysis (RQA)). SpO₂-based features include temporal (variance) and spectral information. The features feed a Linear Discriminant Analysis (LDA) classifier. The goal is to evaluate the effect of using these features either individually or in combination, especially in non-desaturating patients. The main results for the detection of apneic events are: (a) Physionet success rate of 96.19%, sensitivity of 95.74% and specificity of 95.25% (Area Under Curve (AUC): 0.99); (b) HuGCDN2014-OXI of 87.32%, 83.81% and 88.55% (AUC: 0.934), respectively. The best results for the global diagnosis of OSA patients (HuGCDN2014-OXI) are: success rate of 95.74%, sensitivity of 100%, and specificity of 89.47%. We conclude that combining both features is the most accurate option, especially when there are non-desaturating patterns among the recordings under study.

ProBoost, a new boosting algorithm for probabilistic classifiers, is proposed in this work. This algorithm uses the epistemic uncertainty of each training sample to determine the most challenging/uncertain ones; the relevance of these samples is then increased for the next weak learner, producing a sequence that progressively focuses on the samples found to have the highest uncertainty. In the end, the weak learners' outputs are combined into a weighted ensemble of classifiers. Three methods are proposed to manipulate the training set: undersampling, oversampling, and weighting the training samples according to the uncertainty estimated by the weak learners. Furthermore, two approaches are studied regarding the ensemble combination. The weak learner herein considered is a standard convolutional neural network, and the probabilistic models underlying the uncertainty estimation use either variational inference or Monte Carlo dropout. The experimental evaluation carried out on MNIST benchmark datasets shows that ProBoost yields a significant performance improvement. The results are further highlighted by assessing the relative achievable improvement, a metric proposed in this work, which shows that a model with only four weak learners leads to an improvement exceeding 12% in this metric (for either accuracy, sensitivity, or specificity), in comparison to the model learned without ProBoost.