Photos Shared on Facebook in the Context of Safe Sleep Recommendations: Content Analysis of Images.

Background: Sudden unexpected infant death (SUID) remains a leading cause of infant mortality; therefore, understanding parental practices of infant sleep at home is essential. Since social media analyses yield invaluable patient perspectives, understanding sleep practices in the context of safe sleep recommendations via a Facebook mothers' group is instrumental for policy makers, health care providers, and researchers. Objective: This study aimed to identify photos shared by mothers discussing SUID and safe sleep online and assess their consistency with infant sleep guidelines per the American Academy of Pediatrics (AAP). We hypothesized the photos would not be consistent with guidelines based on prior research and increasing rates of accidental suffocation and strangulation in bed. Methods: Data were extracted from a Facebook mothers' group in May 2019. After trialing various search terms, searching for the term "SIDS" on the selected Facebook group resulted in the most relevant discussions on SUID and safe sleep. The resulting data, including 20 posts and 912 comments among 512 mothers, were extracted and underwent qualitative descriptive content analysis. In completing the extraction and subsequent analysis, 24 shared personal photos were identified among the discussions. Of the photos, 14 pertained to the infant sleep environment. Photos of the infant sleep environment were then assessed for consistency with safe sleep guidelines per the AAP standards by 2 separate reviewers. Results: Of the shared photos relating to the infant sleep environment, 86% (12/14) were not consistent with AAP safe sleep guidelines. Specific inconsistencies included prone sleeping, foreign objects in the sleeping environment, and use of infant sleeping devices. Use of infant monitoring devices was also identified. Conclusions: This study is unique because the photos originated from the home setting, were in the context of SUID and safe sleep, and were obtained without researcher interference. Despite study limitations, the commonality of prone sleeping, foreign objects, and the use of both infant sleep and monitoring devices (ie, overall inconsistency regarding AAP safe sleep guidelines) sets the stage for future investigation regarding parental barriers to practicing safe infant sleep and has implications for policy makers, clinicians, and researchers.

Tailored Prompting to Improve Adherence to Image-Based Dietary Assessment: Mixed Methods Study.

Background: Accurately assessing an individual's diet is vital in the management of personal nutrition and in the study of the effect of diet on health. Despite its importance, the tools available for dietary assessment remain either too imprecise, expensive, or burdensome for clinical or research use. Image-based methods offer a potential new tool to improve the reliability and accessibility of dietary assessment. Though promising, image-based methods are sensitive to adherence, as images cannot be captured from meals that have already been consumed. Adherence to image-based methods may be improved with appropriately timed prompting via text message. Objective: This study aimed to quantitatively examine the effect of prompt timing on adherence to an image-based dietary record and qualitatively explore the participant experience of dietary assessment in order to inform the design of a novel image-based dietary assessment tool. Methods: This study used a randomized crossover design to examine the intraindividual effect of 3 prompt settings on the number of images captured in an image-based dietary record. The prompt settings were control, where no prompts were sent; standard, where prompts were sent at 7:15 AM, 11:15 AM, and 5:15 PM for every participant; and tailored, where prompt timing was tailored to habitual meal times for each participant. Participants completed a text-based dietary record at baseline to determine the timing of tailored prompts. Participants were randomized to 1 of 6 study sequences, each with a unique order of the 3 prompt settings, with each 3-day image-based dietary record separated by a washout period of at least 7 days. The qualitative component comprised semistructured interviews and questionnaires exploring the experience of dietary assessment. Results: A total of 37 people were recruited, and 30 participants (11 male, 19 female; mean age 30, SD 10.8 years), completed all image-based dietary records. The image rate increased by 0.83 images per day in the standard setting compared to control (P=.23) and increased by 1.78 images per day in the tailored setting compared to control (P≤.001). We found that 13/21 (62%) of participants preferred to use the image-based dietary record versus the text-based dietary record but reported method-specific challenges with each method, particularly the inability to record via an image after a meal had been consumed. Conclusions: Tailored prompting improves adherence to image-based dietary assessment. Future image-based dietary assessment tools should use tailored prompting and offer both image-based and written input options to improve record completeness.

Prediction of cardiovascular risk factors from retinal fundus photographs: Validation of a deep learning algorithm in a prospective non-interventional study in Kenya.

AIM: Hypertension and diabetes mellitus (DM) are major causes of morbidity and mortality, with growing burdens in low-income countries where they are underdiagnosed and undertreated. Advances in machine learning may provide opportunities to enhance diagnostics in settings with limited medical infrastructure. MATERIALS AND METHODS: A non-interventional study was conducted to develop and validate a machine learning algorithm to estimate cardiovascular clinical and laboratory parameters. At two sites in Kenya, digital retinal fundus photographs were collected alongside blood pressure (BP), laboratory measures and medical history. The performance of machine learning models, originally trained using data from the UK Biobank, were evaluated for their ability to estimate BP, glycated haemoglobin, estimated glomerular filtration rate and diagnoses from fundus images. RESULTS: In total, 301 participants were enrolled. Compared with the UK Biobank population used for algorithm development, participants from Kenya were younger and would probably report Black/African ethnicity, with a higher body mass index and prevalence of DM and hypertension. The mean absolute error was comparable or slightly greater for systolic BP, diastolic BP, glycated haemoglobin and estimated glomerular filtration rate. The model trained to identify DM had an area under the receiver operating curve of 0.762 (0.818 in the UK Biobank) and the hypertension model had an area under the receiver operating curve of 0.765 (0.738 in the UK Biobank). CONCLUSIONS: In a Kenyan population, machine learning models estimated cardiovascular parameters with comparable or slightly lower accuracy than in the population where they were trained, suggesting model recalibration may be appropriate. This study represents an incremental step toward leveraging machine learning to make early cardiovascular screening more accessible, particularly in resource-limited settings.

Better-than-chance prediction of cooperative behaviour from first and second impressions.

Could cooperation among strangers be facilitated by adaptations that use sparse information to accurately predict cooperative behaviour? We hypothesise that predictions are influenced by beliefs, descriptions, appearance and behavioural history available for first and second impressions. We also hypothesise that predictions improve when more information is available. We conducted a two-part study. First, we recorded thin-slice videos of university students just before their choices in a repeated Prisoner's Dilemma with matched partners. Second, a worldwide sample of raters evaluated each player using videos, photos, only gender labels or neither images nor labels. Raters guessed players' first-round Prisoner's Dilemma choices and then their second-round choices after reviewing first-round behavioural histories. Our design allows us to investigate incremental effects of gender, appearance and behavioural history gleaned during first and second impressions. Predictions become more accurate and better-than-chance when gender, appearance or behavioural history is added. However, these effects are not incrementally cumulative. Predictions from treatments showing player appearance were no more accurate than those from treatments revealing gender labels and predictions from videos were no more accurate than those from photos. These results demonstrate how people accurately predict cooperation under sparse information conditions, helping explain why conditional cooperation is common among strangers.

Oral Cannabidiol for Seborrheic Dermatitis in Patients With Parkinson Disease: Randomized Clinical Trial.

BACKGROUND: Seborrheic dermatitis (SD) affects 18.6%-59% of persons with Parkinson disease (PD), and recent studies provide evidence that oral cannabidiol (CBD) therapy could reduce sebum production in addition to improving motor and psychiatric symptoms in PD. Therefore, oral CBD could be useful for improving symptoms of both commonly co-occurring conditions. OBJECTIVE: This study investigates whether oral CBD therapy is associated with a decrease in SD severity in PD. METHODS: Facial photographs were collected as a component of a randomized (1:1 CBD vs placebo), parallel, double-blind, placebo-controlled trial assessing the efficacy of a short-term 2.5 mg per kg per day oral sesame solution CBD-rich cannabis extract (formulated to 100 mg/mL CBD and 3.3 mg/mL THC) for reducing motor symptoms in PD. Participants took 1.25 mg per kg per day each morning for 4 ±1 days and then twice daily for 10 ±4 days. Reviewers analyzed the photographs independently and provided a severity ranking based on the Seborrheic Dermatitis Area and Severity Index (SEDASI) scale. Baseline demographic and disease characteristics, as well as posttreatment SEDASI averages and the presence of SD, were analyzed with 2-tailed t tests and Pearson χ2 tests. SEDASI was analyzed with longitudinal regression, and SD was analyzed with generalized estimating equations. RESULTS: A total of 27 participants received a placebo and 26 received CBD for 16 days. SD severity was low in both groups at baseline, and there was no treatment effect. The risk ratio for patients receiving CBD, post versus pre, was 0.69 (95% CI 0.41-1.18; P=.15), compared to 1.20 (95% CI 0.88-1.65; P=.26) for the patients receiving the placebo. The within-group pre-post change was not statistically significant for either group, but they differed from each other (P=.07) because there was an estimated improvement for the CBD group and an estimated worsening for the placebo group. CONCLUSIONS: This study does not provide solid evidence that oral CBD therapy reduces the presence of SD among patients with PD. While this study was sufficiently powered to detect the primary outcome (efficacy of CBD on PD motor symptoms), it was underpowered for the secondary outcomes of detecting changes in the presence and severity of SD. Multiple mechanisms exist through which CBD can exert beneficial effects on SD pathogenesis. Larger studies, including participants with increased disease severity and longer treatment periods, may better elucidate treatment effects and are needed to determine CBD's true efficacy for affecting SD severity. TRIAL REGISTRATION: ClinicalTrials.gov NCT03582137; https://clinicaltrials.gov/ct2/show/NCT03582137.

Smile aesthetics in Pakistani population: dentist preferences and perceptions of anterior teeth proportion and harmony.

BACKGROUND: This study aimed to evaluate dentist perceptions of attractive smiles in the Pakistani population, considering different dental proportions. METHODS: Maxillary casts and digital images were used to create symmetrical representations of anterior teeth. dentists' preferences for good and bad teeth proportions, width/height ratios, and various dental proportions (golden, recurring esthetic dental (RED), golden percentage, Preston, and local/observed) were assessed using one sample and paired t-test. The Chi-square test was used to determine the gender disparities and factors affecting smile attractiveness. A p-value of ≤ 0.05 was taken as significant. RESULTS: The RED proportion emerged as the preferred choice for normal-sized teeth, with specialists and general dentists favoring it over the golden proportion. For tall teeth, the golden proportion was predominantly preferred. The golden percentage received limited preference for aesthetic smile construction. CONCLUSIONS: The smiles created using the principles of RED proportion were opted as the most attractive by local dentists. Factors such as tooth arrangement, color, and midline were highlighted as essential considerations in aesthetic smile construction.

Using AI Text-to-Image Generation to Create Novel Illustrations for Medical Education: Current Limitations as Illustrated by Hypothyroidism and Horner Syndrome.

Our research letter investigates the potential, as well as the current limitations, of widely available text-to-image tools in generating images for medical education. We focused on illustrations of important physical signs in the face (for which confidentiality issues in conventional patient photograph use may be a particular concern) that medics should know about, and we used facial images of hypothyroidism and Horner syndrome as examples.

Intelligent diagnosis of retinal vein occlusion based on color fundus photographs.

AIM: To develop an artificial intelligence (AI) diagnosis model based on deep learning (DL) algorithm to diagnose different types of retinal vein occlusion (RVO) by recognizing color fundus photographs (CFPs). METHODS: Totally 914 CFPs of healthy people and patients with RVO were collected as experimental data sets, and used to train, verify and test the diagnostic model of RVO. All the images were divided into four categories [normal, central retinal vein occlusion (CRVO), branch retinal vein occlusion (BRVO), and macular retinal vein occlusion (MRVO)] by three fundus disease experts. Swin Transformer was used to build the RVO diagnosis model, and different types of RVO diagnosis experiments were conducted. The model's performance was compared to that of the experts. RESULTS: The accuracy of the model in the diagnosis of normal, CRVO, BRVO, and MRVO reached 1.000, 0.978, 0.957, and 0.978; the specificity reached 1.000, 0.986, 0.982, and 0.976; the sensitivity reached 1.000, 0.955, 0.917, and 1.000; the F1-Sore reached 1.000, 0.955 0.943, and 0.887 respectively. In addition, the area under curve of normal, CRVO, BRVO, and MRVO diagnosed by the diagnostic model were 1.000, 0.900, 0.959 and 0.970, respectively. The diagnostic results were highly consistent with those of fundus disease experts, and the diagnostic performance was superior. CONCLUSION: The diagnostic model developed in this study can well diagnose different types of RVO, effectively relieve the work pressure of clinicians, and provide help for the follow-up clinical diagnosis and treatment of RVO patients.

Performance of deep learning for detection of chronic kidney disease from retinal fundus photographs: A systematic review and meta-analysis.

OBJECTIVE: Deep learning has been used to detect chronic kidney disease (CKD) from retinal fundus photographs. We aim to evaluate the performance of deep learning for CKD detection. METHODS: The original studies in CKD patients detected by deep learning from retinal fundus photographs were eligible for inclusion. PubMed, Embase, the Cochrane Library, and Web of Science were searched up to October 31, 2022. The Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool was used to assess the risk of bias. RESULTS: Four studies enrolled 114,860 subjects were included. The pooled sensitivity and specificity were 87.8% (95% confidence interval (CI): 61.6% to 98.3%), and 62.4% (95% CI: 44.9% to 78.7%). The area under the curve (AUC) was 0.864 (95%CI: 0.769, 0.986). CONCLUSION: Deep learning based on retinal fundus photographs has the ability to detect CKD, but it currently has a lot of room for improvement. It is still a long way from clinical application.

"Cornflake nevoid" lesions, mimicking choroidal lesions of neurofibromatosis 1, in the absence of neurofibromatosis 1: An imaging-based hypothesis.

Whitish hypermelanocytic flake-like lesions on scanning laser ophthalmoscopy (SLO) multicolor posterior pole imaging (PPI) can correspond to several conditions, including simple nevi or shallow choroidal melanomas, paraneoplastic fundus lesions like bilateral diffuse uveal melanocytic proliferation, or choroidal melanocytic lesions found in neurofibromatosis type 1 (NF1). We report three cases with unilateral flake-shaped choroidal lesions on SLO multicolor PPI, similar to choroidal NF1 lesions, monitored their evolution and analyzed their potential nature using multimodal imaging including SLO multicolor and classical PPI, infrared autofluorescence (IRAF), spectral-domain-optical coherence tomography (SD-OCT), enhanced-depth imaging-OCT (EDI-OCT), OCT-angiography as well as fluorescein angiography, and indocyanine green angiography (ICGA). Two oncologic patients and one healthy patient presented unilateral whitish cornflake-shaped lesions on SLO multicolor and IRAF PPI, faintly or not visible on fundus photography, hypofluorescent on the intermediate-phase ICGA, but isofluorescent on the late-phase ICGA corresponding to hyperreflective areas in the choroid immediately under the retinal pigment epithelium on SD-OCT. The lesions were nonevolutive. Multimodal imaging determined that these "nevoid" lesions were melanocytic but could not be assimilated to classical nevi, having a looser structure that allowed some indocyanine green impregnation explaining the isofluorescence on the late-phase ICGA. The lesions were similar to those described in NF1 cases and were unrelated to the oncologic status.

Vision transformers: The next frontier for deep learning-based ophthalmic image analysis.

Deep learning is the state-of-the-art machine learning technique for ophthalmic image analysis, and convolutional neural networks (CNNs) are the most commonly utilized approach. Recently, vision transformers (ViTs) have emerged as a promising approach, one that is even more powerful than CNNs. In this focused review, we summarized studies that applied ViT-based models to analyze color fundus photographs and optical coherence tomography images. Overall, ViT-based models showed robust performances in the grading of diabetic retinopathy and glaucoma detection. While some studies demonstrated that ViTs were superior to CNNs in certain contexts of use, it is unclear how widespread ViTs will be adopted for ophthalmic image analysis, since ViTs typically require even more training data as compared to CNNs. The studies included were identified from the PubMed and Google Scholar databases using keywords relevant to this review. Only original investigations through March 2023 were included.

Automatic detection and differential diagnosis of age-related macular degeneration from color fundus photographs using deep learning with hierarchical vision transformer.

Age-related macular degeneration (AMD) is a leading cause of vision loss in the elderly, highlighting the need for early and accurate detection. In this study, we proposed DeepDrAMD, a hierarchical vision transformer-based deep learning model that integrates data augmentation techniques and SwinTransformer, to detect AMD and distinguish between different subtypes using color fundus photographs (CFPs). The DeepDrAMD was trained on the in-house WMUEH training set and achieved high performance in AMD detection with an AUC of 98.76% in the WMUEH testing set and 96.47% in the independent external Ichallenge-AMD cohort. Furthermore, the DeepDrAMD effectively classified dryAMD and wetAMD, achieving AUCs of 93.46% and 91.55%, respectively, in the WMUEH cohort and another independent external ODIR cohort. Notably, DeepDrAMD excelled at distinguishing between wetAMD subtypes, achieving an AUC of 99.36% in the WMUEH cohort. Comparative analysis revealed that the DeepDrAMD outperformed conventional deep-learning models and expert-level diagnosis. The cost-benefit analysis demonstrated that the DeepDrAMD offers substantial cost savings and efficiency improvements compared to manual reading approaches. Overall, the DeepDrAMD represents a significant advancement in AMD detection and differential diagnosis using CFPs, and has the potential to assist healthcare professionals in informed decision-making, early intervention, and treatment optimization.

A pipeline for the rapid collection of color data from photographs.

Premise: There are relatively few studies of flower color at landscape scales that can address the relative importance of competing mechanisms (e.g., biotic: pollinators; abiotic: ultraviolet radiation, drought stress) at landscape scales. Methods: We developed an R shiny pipeline to sample color from images that were automatically downloaded using query results from a search using iNaturalist or the Global Biodiversity Information Facility (GBIF). The pipeline was used to sample ca. 4800 North American wallflower (Erysimum, Brassicaceae) images from iNaturalist. We tested whether flower color was distributed non-randomly across the landscape and whether spatial patterns were correlated with climate. We also used images including ColorCheckers to compare analyses of raw images to color-calibrated images. Results: Flower color was strongly non-randomly distributed spatially, but did not correlate strongly with climate, with most of the variation explained instead by spatial autocorrelation. However, finer-scale patterns including local correlations between elevation and color were observed. Analyses using color-calibrated and raw images revealed similar results. Discussion: This pipeline provides users the ability to rapidly capture color data from iNaturalist images and can be a useful tool in detecting spatial or temporal changes in color using citizen science data.

Detection of AI-Created Images Using Pixel-Wise Feature Extraction and Convolutional Neural Networks.

Generative AI has gained enormous interest nowadays due to new applications like ChatGPT, DALL E, Stable Diffusion, and Deepfake. In particular, DALL E, Stable Diffusion, and others (Adobe Firefly, ImagineArt, etc.) can create images from a text prompt and are even able to create photorealistic images. Due to this fact, intense research has been performed to create new image forensics applications able to distinguish between real captured images and videos and artificial ones. Detecting forgeries made with Deepfake is one of the most researched issues. This paper is about another kind of forgery detection. The purpose of this research is to detect photorealistic AI-created images versus real photos coming from a physical camera. Id est, making a binary decision over an image, asking whether it is artificially or naturally created. Artificial images do not need to try to represent any real object, person, or place. For this purpose, techniques that perform a pixel-level feature extraction are used. The first one is Photo Response Non-Uniformity (PRNU). PRNU is a special noise due to imperfections on the camera sensor that is used for source camera identification. The underlying idea is that AI images will have a different PRNU pattern. The second one is error level analysis (ELA). This is another type of feature extraction traditionally used for detecting image editing. ELA is being used nowadays by photographers for the manual detection of AI-created images. Both kinds of features are used to train convolutional neural networks to differentiate between AI images and real photographs. Good results are obtained, achieving accuracy rates of over 95%. Both extraction methods are carefully assessed by computing precision/recall and F1-score measurements.

Reversing the eyes and reverse perspectives: Pseudoscopic amplification of reverspectives.

Stereoscopic photographs of works in reverse perspective do not reveal their three-dimensional structure whereas pseudoscopic photographs enhance the apparent depth effects.

The missing: Exploring the use of photographs in "working through" the natal body with transgender youth.

This paper focuses on how for some young people who identify as transgender, the anticipation, and/or the actual process, of transitioning represents a movement away from something in themselves that feels wrong, painful, or traumatic and that has not yet been consciously recognised as such. This becomes a 'missing' part of the self's experience, locked into the body. I suggest that the process of identifying and restitution of 'the missing' part requires working through the natal body in its metaphorical and literal senses, in the service of expanding autonomous choice about how to find a hospitable home in the body. Building on Money-Kyrle's three 'facts of life', I propose a fourth one, namely the inescapable fact of our embodied nature, to underscore that our personal history always includes our embodied history, hence the importance of working through what the natal body unconsciously represents. I describe the use of photographs during psychoanalytic psychotherapy with young people who have commenced social transitioning, to work through visual representations of the natal body in the service of facilitating the working through, in its psychoanalytic sense, of the natal body's unconscious narrative. I suggest that deploying this visual mode may be especially helpful in engaging young people on the autistic spectrum who nowadays comprise a significant minority of transgender young people.

Field Study on the Prevalence of Ovine Footrot, Contagious Ovine Digital Dermatitis, and Their Associated Bacterial Species in Swedish Sheep Flocks.

Ovine footrot and contagious ovine digital dermatitis (CODD) cause lameness in sheep, affecting welfare and economics. Previous Swedish studies focused on individual slaughter lambs, leaving flock-wide prevalence less explored. This study examined the prevalence of footrot and CODD in Swedish sheep flocks, focusing on adult sheep. From 99 flocks, 297 swabs were analysed using real-time PCR for Dichelobacter nodosus, Fusobacterium necrophorum, and Treponema spp. Sampled feet were photographed and assessed using scoring systems for footrot and CODD. Results indicated footrot prevalences (footrot score ≥ 2) of 0.7% and 2.0% at the individual and flock levels, respectively, whereas there were no signs of CODD. The individual footrot prevalence was lower than that from a 2009 study but aligned with a 2020 study, both conducted on slaughter lambs. Dichelobacter nodosus, F. necrophorum, and Treponema spp. were found in 5.7%, 1.3%, and 65.0% of sheep, and in 9.1%, 3.0%, and 82.8% of flocks, respectively. Compared to the 2020 study, there was a notable decrease in F. necrophorum and Treponema spp., while D. nodosus was consistent. In conclusion, the findings show a low prevalence of footrot, CODD, D. nodosus, and F. necrophorum in Swedish sheep flocks. Continuous surveillance and owner education are important to maintain this favourable status.

The Correlation of a Novel Photographic Parameter for Facial Profile Assessment in Subjects With Different Sagittal Malocclusions: A Prospective Study.

Aim The soft tissue paradigm shift is the current trend in orthodontic diagnosis and treatment planning. This study's aim was to assess the correlation of newly derived photographic Frankfort horizontal plane-subnasale to soft tissue pogonion (FSA) angle with other established soft tissue cephalometric angles, such as the Z angle and the Holdaway (H) angle, for estimating facial profile convexity in subjects with all classes of sagittal malocclusions. Materials and methods This prospective study included a sample of 60 Dravidian population subjects consisting of 30 males and 30 females with different skeletal sagittal malocclusions (Class I, Class II, and Class III) based on the radiographic criteria (ANB angle). The Z and Holdaway angles on lateral cephalograms were compared with the FSA angles in cephalograms and digital profile photographs. Statistical analysis was done using the Statistical Package for Social Sciences (SPSS) software version 23.0 (IBM SPSS Statistics, Armonk, NY). Pearson's correlation was done to assess the correlation between soft tissue FSA angle on digital photographs and cephalometric angle (Z angle and Holdaway angle). Results The overall Pearson's correlation was significant (p < 0.05) between the Z and FSA angles in Class I, II, and III malocclusions, which had a high positive correlation. There was a significant positive correlation (p < 0.05) between the Holdaway and FSA angles in subjects with Class I and Class II malocclusions. A moderate positive correlation was noted between the Holdaway and FSA angles in Class III. Conclusion Photographic FSA angle can be used to evaluate the facial profile of subjects with different sagittal malocclusions. This angle has a good correlation with other cephalometric profile measures, such as the Z and Holdaway angles used to assess facial profile convexity.

Deep learning technique to detect craniofacial anatomical abnormalities concentrated on middle and anterior of face in patients with sleep apnea.

OBJECTIVES: The aim of this study is to propose a deep learning-based model using craniofacial photographs for automatic obstructive sleep apnea (OSA) detection and to perform design explainability tests to investigate important craniofacial regions as well as the reliability of the method. METHODS: Five hundred and thirty participants with suspected OSA are subjected to polysomnography. Front and profile craniofacial photographs are captured and randomly segregated into training, validation, and test sets for model development and evaluation. Photographic occlusion tests and visual observations are performed to determine regions at risk of OSA. The number of positive regions in each participant is identified and their associations with OSA is assessed. RESULTS: The model using craniofacial photographs alone yields an accuracy of 0.884 and an area under the receiver operating characteristic curve of 0.881 (95% confidence interval, 0.839-0.922). Using the cutoff point with the maximum sum of sensitivity and specificity, the model exhibits a sensitivity of 0.905 and a specificity of 0.941. The bilateral eyes, nose, mouth and chin, pre-auricular area, and ears contribute the most to disease detection. When photographs that increase the weights of these regions are used, the performance of the model improved. Additionally, different severities of OSA become more prevalent as the number of positive craniofacial regions increases. CONCLUSIONS: The results suggest that the deep learning-based model can extract meaningful features that are primarily concentrated in the middle and anterior regions of the face.