Mental health and substance use evolution in Swiss ED residents: a 6-month prospective longitudinal single-center study.

Analyses of the UK Biobank and EyePACS datasets show that deep-learning artificial intelligence (AI) can accurately predict several demographic and systemic features. For instance, it estimates age with a mean absolute error (MAE) of 3.26 years and determines sex with an area under the curve (AUC) of 0.97. Furthermore, the model achieved a mean absolute error of 11.23 mmHg in systolic blood pressure prediction and yielded AUC values of 0.71 and 0.70 for identifying smoking status and cardiac complications, respectively, by analyzing color fundus photographs using the Inception-v3 model []. Furthermore, refraction can be estimated by AI through the TensorFlow platform with an MAE of 0.56 diopters, as demonstrated using the UK Biobank and Age-Related Eye Disease Study datasets []. In addition, analysis of the Beijing Eye Study revealed that the Inception-ResNet-v2 architecture can predict axial length directly from fundus photographs, achieving an MAE of 0.56 mm []. Sex prediction by AI can reach an accuracy of up to 97% []. Nevertheless, determining the specific features responsible for this prediction remain elusive, even when attention heat maps or quantitative analyses are applied. A similar limitation is observed in board games such as chess, Go, and Shogi: deep-learning systems can consistently suggest optimal moves, yet the rationale behind these decisions is often unclear []. Because of this opacity, experts must analyze AI’s strategies, models frequently referred to as “black-box AI” []. To shed light on this black box, classical statistical approaches such as multivariable regression have been considered as explanatory alternatives to AI []. The ocular fundus, much like a fingerprint or facial features, is unique to each individual. Several characteristics show wide variation: the angle and path of retinal vessels [–], the location and morphology of the optic disc [, ], and the coloration of the peripapillary area [–]. In particular, the trajectories of the arcade vessels parallel those of the infra- and supra-temporal retinal nerve fiber layers, and both shift toward the fovea as axial length increases [–]. Our previous studies demonstrate that L2-regularized ridge regression based on fundus parameters could predict sex with accuracies of 77.9% in young adults in their 20s [], 63.2% in children with a mean age of 8.5 years [], and 80.4% in individuals older than 40 years []. Because a multivariable regression approach was employed, sex-related differences in ocular features were also identified: women exhibited more oval-shaped optic discs, retinal vessels positioned closer to the fovea, and a greener peripapillary fundus color compared with men [–]. We also revealed age-specific and axial length-specific fundus changes using this method [, ]. In a previous study, regression analysis yielded a predictive score (0–1) for each fundus image: lower values indicated a more masculine fundus, while higher values reflected a more feminine fundus. This approach treats sex as a continuous parameter based on fundus characteristics, rather than a binary classification. Thus, the probability score can be used to evaluate sex-related tendencies rather than categorical differences. The prediction score was designated as the fundus sex index (FSI), and its correlation with sex-specific axial length was assessed in the Kumejima population. Fundi with more masculine features were consistently linked to greater axial length across both sexes [].