Calculating the optimal cut-point of a predictor variable for a binary classification task is a crucial step in clinical practice, directly influencing diagnostic accuracy and patient management. This tutorial offers a practical, step-by-step guide for clinicians and researchers for optimal cut-point calculation and reporting, with a focus on transparency and reproducibility. We review key statistical concepts, including performance metrics and receiver operating characteristic (ROC) curve analysis, before detailing a recommended workflow encompassing exploratory data analysis, polarity identification, cut-point calculation, stability assessment, and post hoc verification. The approach is illustrated with real-world case studies using open clinical datasets, highlighting both the methodology and common challenges.
Background: Artificial intelligence (AI)-assisted imaging may support recognition of endometriosis, but perspectives on its use in adolescent care remain unclear. This study explored young people, parents, and health professionals’ views on its potential benefits, acceptable roles, perceived risks, and safeguards. Methods: Four semi-structured focus groups were conducted with adolescents aged 14–17 years (n = 3), young adults aged 18–25 years (n = 9), parents or carers (n = 5), and health professionals (n = 13). Data were analyzed using reflexive thematic analysis. Results: Four themes were developed: (1) diagnostic support, (2) human–AI collaboration, (3) automation bias and diagnostic harm, and (4) responsible AI use. Participants were cautiously hopeful that AI could improve image quality, identify subtle or missed findings, and reduce interpretive variation, but supported it only as an adjunct to skilled clinical care. They feared that apparently objective negative outputs could create false reassurance, reinforce diagnostic dismissal, and restrict further investigation. Trust depended on adolescent-relevant validation, representative data, transparent communication and consent, clear accountability, and clinicians retaining the expertise and authority to question or override AI outputs. Conclusion: AI-assisted imaging should strengthen rather than replace clinical judgment, and a negative AI-assisted imaging result should not prevent further assessment when symptoms persist.
Invasive fungal infections cause approximately 1.5 billion cases each year, with an estimated 1.6 million deaths. Classic diagnostic methods, such as culture, direct microscopy, and antigen testing, largely rely on professional experience and are often too slow. The field of fungal diagnostics is increasingly using deep learning (DL), a class of artificial intelligence based on multilayered neural architectures. Recent developments in DL applications to clinical mycology are examined in this review, including automated colony morphology analysis, MALDI-TOF mass spectrometry spectral classification, Raman spectroscopy, microscopy, molecular diagnostics, and antifungal resistance prediction. Aspergillus, Candida, Cryptococcus neoformans, Pneumocystis jirovecii, and dermatophytes are among the important pathogens evaluated for validated performance benchmarks. We also address the main barriers to clinical implementation, which include a lack of large annotated datasets, cross-site generalizability, interpretability requirements, and regulatory frameworks. We conclude that, depending on carefully selected open-access datasets and future clinical validation, DL has transformative potential across several mycological diagnostic platforms.
Analyzing physiological signals with artificial intelligence (AI) has become a strategic tool for clinical diagnosis, event prediction, risk stratification, monitoring, and functional analysis, among other tasks. However, the growing use of AI in healthcare raises a key concern: models can achieve high performance on specific tasks, but this does not necessarily translate into clinical utility, generalizability, or integration into real-world care settings. This work presents a structured review of recent and selected foundational research published between 2018 and 2026, aimed at characterizing AI applications to physiological signals. It considers the analyzed signals, AI families, implemented techniques, reported metrics, the degree of unimodal or multimodal integration, and gaps in validation, interpretability, and reproducibility. The final database included 83 evidence entries, organized as an evidence matrix with information on AI family, architecture, reported performance, and the physiological or functional signal used. This review identifies single-source approaches, multimodal predictive integration, and multimodal representation learning, highlighting their different methodological characteristics and potential clinical applications. The main contribution of this work is to show that the potential clinical contribution of AI applied to physiological signals depends not only on algorithmic performance but also on its integration with robust validation, traceability, interpretability, physiological coherence, and relevance to the clinical phenomenon.
(1) Background: Modeling signaling cascades is helpful in deciphering the structure and interactions within and between cascades, and testing our understanding of the cascades by using the models to make predictions and suggesting experiments to evaluate the efficacy of the models. Here, several interconnected toll-like receptor (TLR) signaling cascades were studied in combination to determine if there was any significant difference from studying them individually. (2) Methods: Our approach was to enhance a tool previously used to construct agent-based models of single-TLR signaling cascades with the ability to combine multiple individual models into a single, more complex model. The tool was used to combine three TLR signaling models, and predictions from the model were then compared against experimental results from all three cascades. The lab experiments focused on one of two different readouts: examining RNA expression or effector function. (3) Results: The results showed that if we used RNA as the readout, the model and experimental data were closely aligned. However, to fit the effector function data, we had to assign more importance to the role of the TLR7 cascade. (4) Conclusions: These results show that modeling can be used to decipher the structure and interactions within and between cascades, but the pattern of response is specific to the experimental readout. The data also show that modeling and experimenting with a combination of cascades is critical to (1) determining the relative importance of each cascade with respect to response to particular agonists and (2) understanding the interactions (both synergistic and antagonistic) between the cascades.
