2025 · International journal of molecular sciences · open access
Harnessing Machine Learning, a Subset of Artificial Intelligence, for Early Detection and Diagnosis of Type 1 Diabetes: A Systematic Review
Mittal et al.
The finding, in our words
This systematic review of ten studies with 49,172 participants found that machine-learning models integrating continuous glucose monitoring data with clinical, genetic and multi-omic biomarkers achieved AUC values up to 0.993 for early type 1 diabetes detection, with plasma CXCL10 and IL-1RA emerging as useful markers. The findings support decentralised diagnostics by validating wearable-derived digital biomarkers alongside plasma biomarkers, though data heterogeneity and limited generalisability remain barriers to clinical implementation.
A paraphrase to the Library’s standard, never the abstract. The source is one link away and is always the authority.
The High-Definition Oncology feasibility study in 30 women with metastatic cancer demonstrated high adherence to biospecimen collection (97.4% plasma, 80.7% stool) and wearable monitoring (70–95% of days captured for activity, heart rate, sleep, and oxygen saturation). This establishes a viable framework for decentralised, multimodal data collection in oncology to support individualised treatment models.
Garma et al., JCO precision oncology (paywalled) · source ↗
In adults with type 2 diabetes, exposure to natural daylight during office hours increased time in the normal glucose range and shifted whole-body substrate metabolism towards greater fat oxidation, suggesting that daylight could be a useful adjunct in managing glycaemic control.
Harmsen et al., Cell metabolism (paywalled) · source ↗
Continuous glucose monitoring of 55 participants showed considerable interindividual variation in postprandial glycaemic responses to carbohydrate meals, with rice producing the highest overall responses. Potato-spikers were more insulin resistant with lower beta cell function, grape-spikers were more insulin sensitive, and mitigators were less effective in insulin-resistant individuals. Multi-omics profiling revealed signatures linking glycaemic responses to triglycerides, metabolites and microbiome pathways.
The Human Phenotype Project enrolled 28,000 participants in a deep-phenotyping cohort that combines blood and microbiome sampling with continuous glucose and sleep monitoring. An AI model trained on dietary and glucose data predicted disease onset more accurately than existing methods, showing how integrating continuous digital signals with multi-omic data can support personalised risk assessment for decentralised diagnostics.
Reicher et al., Nature medicine (paywalled) · source ↗
The study demonstrates that frequent collection of 10 microlitres of capillary blood, combined with wearable sensor data, enables simultaneous measurement of thousands of metabolites, lipids, cytokines and proteins, revealing individualised metabolic and inflammatory responses to dietary interventions and molecular fluctuations linked to intra-day changes in heart rate, glucose, cortisol and physical activity, supporting dynamic health profiling outside the clinic.
Shen et al., Nature biomedical engineering · source ↗