AI and Machine Learning for Cognitive Decline: From Benchmarks to Clinical Practice

The application of artificial intelligence and machine learning to cognitive decline and neurodegenerative disorders has grown rapidly, with studies routinely reporting impressive accuracy, sensitivity, and specificity metrics. However, a critical gap exists between high-performing models on benchmark datasets and tools that are genuinely translatable to clinical practice. This Research Topic challenges the current paradigm of evaluating AI systems primarily through aggregate final metrics and calls for a fundamental shift toward understanding and reporting learning dynamics, that is, the processes by which models acquire, represent, and generalize knowledge in clinically relevant aging, cognitive decline, and dementia contexts. In particular, this Research Topic is interested in AI and machine learning studies that advance the understanding, assessment, prognosis, or clinical management of aging-related cognitive decline and neurodegenerative disease.In medicine, the stakes extend far beyond statistical performance. Lives, quality of life, and clinical decision-making depend on AI systems that are robust, interpretable, and grounded in meaningful physiological or pathological patterns. A model that achieves a high accuracy through algorithmic shortcuts, such as exploiting data leakage, dataset artifacts, demographic confounders, or spurious correlations, offers little value when applied to real-world patient populations. Worse, such models risk causing harm if deployed without rigorous evaluation of how and why they arrive at their predictions. This Research Topic therefore aims to foreground clinical applicability above statistical performance, with emphasis on translational relevance, interpretability grounded in neuroanatomy, cognitive domains, disease progression, or related mechanisms of brain aging and neurodegeneration, and rigorous assessment of robustness, generalizability, and failure modes.The scope of this Research Topic is confined to studies that provide clear relevance to cognitive decline in aging and dementia-related neurodegeneration and that go beyond final aggregate metrics by analyzing learning behavior, robustness, interpretability, clinical utility, or failure modes. We welcome articles addressing, but not limited to, the following themes:- Learning dynamics in deep learning models for Alzheimer’s disease, Parkinson’s disease, mild cognitive impairment, and other dementias- Detection and mitigation of shortcut learning in neuroimaging, cognitive assessment, and biomarker-based AI models- Curriculum learning, transfer learning, and continual learning strategies for cognitive decline applications- Interpretable and explainable AI approaches with demonstrated clinical relevance- Prospective validation studies and real-world deployment evaluations- Multi-modal learning dynamics across imaging, genetics, clinical data, and digital biomarkers- Fairness, bias detection, and demographic robustness in cognitive decline AI- Critiques and meta-analyses of existing performance-centric evaluation practicesTo maintain a high standard of scientific and clinical rigor, submissions are expected to employ cognitive assessment and/or model validation measures that extend beyond brief screening instruments (e.g., MMSE, MoCA), unless their exclusive use is very thoroughly justified with compelling evidence for their adequacy and relevance within the study context. Similarly, model validation should extend beyond aggregate performance metrics by demonstrating healthy learning dynamics (such as epoch-wise learning curves for individual data subsets) that provide evidence of generalizability and robustness. While external validation is welcomed, it is not strictly required; truly external datasets are not always available, and even when they are, they may not be genuinely independent from the training data. Transparent reporting of internal learning behavior is therefore considered essential for establishing model trustworthiness and clinical applicability.We encourage submissions to go beyond reporting aggregate performance metrics on benchmark datasets by also considering learning behavior, robustness, and clinical applicability. Studies with a primarily methodological focus should demonstrate clear relevance to aging neuroscience, neurodegeneration, or clinical cognitive decline in order to align well with the aims of this Research Topic.Topic Editor Dr. Dyrba receives lecturing honoraria from the European Fernhochschule Hamburg GmbH (Euro-FH), Hamburg, Germany. Topic Editor Dr. Toma declares no competing interests. Topic editors: Milan Toma, Martin Dyrba Source: https://www.frontiersin.org/research-topics/80605/ai-and-machine-learning-for-cognitive-decline-from-benchmarks-to-clinical-practice