Dysphagia is a common geriatric syndrome that might induce serious complications and death. Standard diagnostics using the Videofluoroscopic Swallowing Study (VFSS) or Fiberoptic Evaluation of Swallowing (FEES) are expensive and expose patients to risks, while bedside screening is subjective and might lack reliability. An affordable and accessible instrumented screening is necessary. This study aimed to evaluate the classification performance of Transformer models and convolutional networks in identifying swallowing and non-swallowing tasks through depth video data. Different activation functions (ReLU, LeakyReLU, GELU, ELU, SiLU, and GLU) were then evaluated on the best-performing model. Sixty-five healthy participants (n = 65) were invited to perform swallowing (eating a cracker and drinking water) and non-swallowing tasks (a deep breath and pronouncing vowels: “/eɪ/”, “/iː/”, “/aɪ/”, “/oʊ/”, “/u:/”). Swallowing and non-swallowing were classified by Transformer models (TimeSFormer, Video Vision Transformer (ViViT)), and convolutional neural networks (SlowFast, X3D, and R(2+1)D), respectively. In general, convolutional neural networks outperformed the Transformer models. X3D was the best model with good-to-excellent performance (F1-score: 0.920; adjusted F1-score: 0.885) in classifying swallowing and non-swallowing conditions. Moreover, X3D with its default activation function (ReLU) produced the best results, although LeakyReLU performed better in deep breathing and pronouncing “/aɪ/” tasks. Future studies shall consider collecting more data for pretraining and developing a hyperparameter tuning strategy for activation functions and the high dimensionality video data for Transformer models.

Background: Dysphagia affects more than half of older adults with dementia and is associated with a 10-fold increase in mortality. The development of accessible, objective, and reliable screening tools is crucial for early detection and management. Objective: This systematic scoping review aimed to (1) examine the current state of the art in artificial intelligence (AI) and sensor-based technologies for dysphagia screening, (2) evaluate the performance of these AI-based screening tools, and (3) assess the methodological quality and rigor of studies on AI-based dysphagia screening tools. Methods: We conducted a systematic literature search across CINAHL, Embase, PubMed, and Web of Science from inception to July 4, 2024, following the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) framework. In total, 2 independent researchers conducted the search, screening, and data extraction. Eligibility criteria included original studies using sensor-based instruments with AI to identify individuals with dysphagia or unsafe swallow events. We excluded studies on pediatric, infant, or postextubation dysphagia, as well as those using non–sensor-based assessments or diagnostic tools. We used a modified Quality Assessment of Diagnostic Accuracy Studies–2 tool to assess methodological quality, adding a “model” domain for AI-specific evaluation. Data were synthesized narratively. Results: This review included 24 studies involving 2979 participants (1717 with dysphagia and 1262 controls). In total, 75% (18/24) of the studies focused solely on per-individual classification rather than per–swallow event classification. Acoustic (13/24, 54%) and vibratory (9/24, 38%) signals were the primary modality sources. In total, 25% (6/24) of the studies used multimodal approaches, whereas 75% (18/24) used a single modality. Support vector machine was the most common AI model (15/24, 62%), with deep learning approaches emerging in recent years (3/24, 12%). Performance varied widely—accuracy ranged from 71.2% to 99%, area under the receiver operating characteristic curve ranged from 0.77 to 0.977, and sensitivity ranged from 63.6% to 100%. Multimodal systems generally outperformed unimodal systems. The methodological quality assessment revealed a risk of bias, particularly in patient selection (unclear in 18/24, 75% of the studies), index test (unclear in 23/24, 96% of the studies), and modeling (high risk in 13/24, 54% of the studies). Notably, no studies conducted external validation or domain adaptation testing, raising concerns about real-world applicability. Conclusions: This review provides a comprehensive overview of technological advancements in AI and sensor-based dysphagia screening. While these developments show promise for continuous long-term tele-swallowing assessments, significant methodological limitations were identified. Future studies can explore how each modality can target specific anatomical regions and manifestations of dysphagia. This detailed understanding of how different modalities address various aspects of dysphagia can significantly benefit multimodal systems, enabling them to better handle the multifaceted nature of dysphagia conditions.

Assessing sleep posture, a critical component in sleep tests, is crucial for understanding an individual’s sleep quality and identifying potential sleep disorders. However, monitoring sleep posture has traditionally posed significant challenges due to factors such as low light conditions and obstructions like blankets. The use of radar technolsogy could be a potential solution. The objective of this study is to identify the optimal quantity and placement of radar sensors to achieve accurate sleep posture estimation. We invited 70 participants to assume nine different sleep postures under blankets of varying thicknesses. This was conducted in a setting equipped with a baseline of eight radars—three positioned at the headboard and five along the side. We proposed a novel technique for generating radar maps, Spatial Radio Echo Map (SREM), designed specifically for data fusion across multiple radars. Sleep posture estimation was conducted using a Multiview Convolutional Neural Network (MVCNN), which serves as the overarching framework for the comparative evaluation of various deep feature extractors, including ResNet-50, EfficientNet-50, DenseNet-121, PHResNet-50, Attention-50, and Swin Transformer. Among these, DenseNet-121 achieved the highest accuracy, scoring 0.534 and 0.804 for nine-class coarse- and four-class fine-grained classification, respectively. This led to further analysis on the optimal ensemble of radars. For the radars positioned at the head, a single left-located radar proved both essential and sufficient, achieving an accuracy of 0.809. When only one central head radar was used, omitting the central side radar and retaining only the three upper-body radars resulted in accuracies of 0.779 and 0.753, respectively. This study established the foundation for determining the optimal sensor configuration in this application, while also exploring the trade-offs between accuracy and the use of fewer sensors.

The objective of this review was to summarize the applications of sonoelastography in testicular tumor identification and inquire about their test performances. Two authors independently searched English journal articles and full conference papers from CINAHL, Embase, IEEE Xplore^®, PubMed, Scopus, and Web of Science from inception and organized them into a PIRO (patient, index test, reference test, outcome) framework. Eleven studies (n = 11) were eligible for data synthesis, nine of which (n = 9) utilized strain elastography and two (n = 2) employed shear-wave elastography. Meta-analyses were performed on the distinction between neoplasm (tumor) and non-neoplasm (non-tumor) from four study arms and between malignancy and benignity from seven study arms. The pooled sensitivity of classifying malignancy and benignity was 86.0% (95%CI, 79.7% to 90.6%). There was substantial heterogeneity in the classification of neoplasm and non-neoplasm and in the specificity of classifying malignancy and benignity, which could not be addressed by the subgroup analysis of sonoelastography techniques. Heterogeneity might be associated with the high risk of bias and applicability concern, including a wide spectrum of testicular pathologies and verification bias in the reference tests. Key technical obstacles in the index test were manual compression in strain elastography, qualitative observation of non-standardized color codes, and locating the Regions of Interest (ROI), in addition to decisions in feature extractions. Future research may focus on multiparametric sonoelastography using deep learning models and ensemble learning. A decision model on the benefits–risks of surgical exploration (reference test) could also be developed to direct the test-and-treat strategy for testicular tumors.

Natural disasters create complex environments where effective human detection is both critical and challenging, especially when individuals are partially occluded. While recent advancements in computer vision have improved detection capabilities, there remains a significant need for efficient solutions that can enhance search-and-rescue (SAR) operations in resource-constrained disaster scenarios. This study modified the original DINO (Detection Transformer with Improved Denoising Anchor Boxes) model and introduced the visibility-enhanced DINO (VE-DINO) model, designed for robust human detection in occlusion-heavy environments, with potential integration into SAR system. VE-DINO enhances detection accuracy by incorporating body part key point information and employing a specialized loss function. The model was trained and validated using the COCO2017 dataset, with additional external testing conducted on the Disaster Occlusion Detection Dataset (DODD), which we developed by meticulously compiling relevant images from existing public datasets to represent occlusion scenarios in disaster contexts. The VE-DINO achieved an average precision of 0.615 at IoU 0.50:0.90 on all bounding boxes, outperforming the original DINO model (0.491) in the testing set. The external testing of VE-DINO achieved an average precision of 0.500. An ablation study was conducted and demonstrated the robustness of the model subject when confronted with varying degrees of body occlusion. Furthermore, to illustrate the practicality, we conducted a case study demonstrating the usability of the model when integrated into an unmanned aerial vehicle (UAV)-based SAR system, showcasing its potential in real-world scenarios.

The accuracy of sleep posture assessment in standard polysomnography might be compromised by the unfamiliar sleep lab environment. In this work, we aim to develop a depth camera-based sleep posture monitoring and classification system for home or community usage and tailor a deep learning model that can account for blanket interference. Our model included a joint coordinate estimation network (JCE) and sleep posture classification network (SPC). SaccpaNet (Separable Atrous Convolution-based Cascade Pyramid Attention Network) was developed using a combination of pyramidal structure of residual separable atrous convolution unit to reduce computational cost and enlarge receptive field. The Saccpa attention unit served as the core of JCE and SPC, while different backbones for SPC were also evaluated. The model was cross-modally pretrained by RGB images from the COCO whole body dataset and then trained/tested using dept image data collected from 150 participants performing seven sleep postures across four blanket conditions. Besides, we applied a data augmentation technique that used intra-class mix-up to synthesize blanket conditions; and an overlaid flip-cut to synthesize partially covered blanket conditions for a robustness that we referred to as the Post-hoc Data Augmentation Robustness Test (PhD-ART). Our model achieved an average precision of estimated joint coordinate (in terms of PCK@0.1) of 0.652 and demonstrated adequate robustness. The overall classification accuracy of sleep postures (F1-score) was 0.885 and 0.940, for 7- and 6-class classification, respectively. Our system was resistant to the interference of blanket, with a spread difference of 2.5%.