We are moving our newsletter to Substack for a better experience!
In Week #249 of the Doctor Penguin newsletter, the following papers caught our attention:
1. Diabetes Subtyping. While diabetes and prediabetes are currently diagnosed based on elevated blood glucose levels alone, the underlying metabolic disruptions that cause these conditions can vary significantly between individuals.
Metwally et al. demonstrate that individuals with prediabetes and early type 2 diabetes can be classified according to their underlying metabolic physiology rather than just blood glucose levels. They developed machine learning models to predict metabolic subphenotypes based on the shape of glucose response curves measured during standardized oral glucose tolerance tests (OGTTs), where patients drink glucose solution and then have their blood glucose levels measured at 16 timepoints over 180 minutes. Using continuous glucose monitors (CGMs), they showed that these tests could be performed reliably at home, making the approach more accessible. The models achieved high prediction accuracy for key metabolic characteristics - with areas under the curve (AUCs) of 95% for muscle insulin resistance, 89% for β-cell deficiency and 88% for impaired incretin action when using OGTT data from the initial cohort of 32 individuals. Using CGM data from at-home OGTTs, the models predicted muscle insulin resistance and β-cell deficiency subphenotypes in 29 individuals with AUCs of 88% and 84% respectively. This approach represents a new method for early identification of specific metabolic dysfunctions that could enable more personalized prevention and treatment strategies for type 2 diabetes.
Read paper | Nature Biomedical Engineering
2. Biological Age. People of the same age often show marked differences in health status and physical aging. These variations reflect differences in biological aging—the accumulation of molecular and cellular damage that impairs physiological function. Metabolites, the end products of metabolism, offer a promising way to measure these differences by providing detailed snapshots of an individual's physiological state and potentially revealing patterns of accelerated or decelerated aging.
Mutz et al. benchmarked 17 different machine learning algorithms for predicting biological age based on 168 blood plasma metabolites from over 225,000 participants in the UK Biobank. Among all models tested, the difference between metabolite-predicted and chronological age from a Cubist rule-based regression model showed the strongest associations with health and aging markers. People with a higher predicted metabolic age compared to their chronological age (indicating accelerated aging) were more likely to be frail, had shorter telomeres, suffered more from chronic illnesses, rated their health worse, and had a 51% higher mortality risk. Notably, while accelerated metabolomic aging strongly correlated with poor health outcomes and higher mortality risk, decelerated aging did not equivalently translate to better health outcomes, suggesting these metabolomics-based risk scores should primarily be used to identify high-risk individuals. The authors emphasize that while perfect age prediction isn't the goal, it's the deviation between predicted and chronological age that meaningfully indicates biological aging patterns and health risks.
Read Paper | Science Advances
3. Brain Aging. How do aged cells in the brain affect neighboring cells and contribute to tissue decline?
Sun et al. generated an extensive spatial transcriptomics atlas of the aging mouse brain, profiling 4.2 million cells across 20 different ages throughout adult life and under two rejuvenating interventions: exercise and partial reprogramming. To assess cellular aging, they developed "spatial aging clocks" - machine learning models that predict biological age from spatially contextualized gene expression data. These tools enabled them to quantify how different interventions and disease states affect specific brain regions and cell types. Their analysis revealed two key findings about cell-cell interactions in aging: T cells, which become more prevalent in the aging brain, exert a pro-aging effect on neighboring cells, while neural stem cells demonstrate a rejuvenating effect on their neighbors. When comparing interventions, they found exercise to be more effective at rejuvenation than partial reprogramming, with particularly strong effects on brain vasculature cells. This comprehensive spatial and temporal map of brain aging provides new insights into how aging impacts different cell types and how cellular interactions influence aging states, suggesting potential therapeutic strategies for maintaining brain health during aging.
Read Paper | Nature
4. ECG Foundation Model. ECG interpretation remains a challenging task that requires extensive training (usually more than 12 years), is time-consuming and error-prone even for experienced cardiologists, and is especially difficult to deliver in remote and underserved regions due to the scarcity of specialists and limited resources.
Tian et al. developed KED (Knowledge-enhanced ECG Diagnosis), a foundation model for automated ECG diagnosis that demonstrated excellent zero-shot diagnostic performance. Unlike radiology reports that contain detailed descriptive information, ECG reports are typically concise and abstract, which limits the contextual information available for model training. To address this limitation, the model leverages large language models to enhance ECG reports with detailed medical terminology explanations and diagnostic background knowledge. Built on four key modules - an ECG signal encoder, a knowledge encoder, a label query network (LQN), and a classification head - KED can query and diagnose cardiac conditions through natural language, providing both conclusions and explanations. The model employs a novel augmented signal-text-label contrastive learning (AugCL) strategy that handles multi-label classification by creating independent contrastive spaces for each label. Despite being trained on single-center data, KED demonstrates robust performance across diverse regions, including various locations in China, the United States, and other regions. When compared to three experienced cardiologists on real clinical datasets, the model achieves comparable performance in zero-shot diagnosis of seven common clinical ECG types. It also shows good diagnostic ability for diseases not seen during training (e.g., SVT, AT, QAb, anterolateral myocardial ischemia).
Read Paper | Cell Reports Medicine
-- Emma Chen, Pranav Rajpurkar & Eric Topol