Machine learning analysis of continuous glucose monitoring identifies a novel dysglycemic phenotype found in most people with cystic fibrosis.

BACKGROUND

Cystic fibrosis related diabetes (CFRD) is a common complication in people with cystic fibrosis (PwCF), yet traditional diagnostic tools such as fasting glucose, HbA1c, and the oral glucose tolerance test (OGTT) often fail to detect early dysglycemia. Continuous glucose monitoring (CGM) generates high resolution glucose data, but analytic methods for extracting meaningful phenotypes remain limited.

METHODS

CGM data from 82 PwCF aged 6 to 78 years were compared with 166 healthy controls (HC). Thirty-two glycemic features were extracted from 24-hour CGM segments. Uniform Manifold Approximation and Projection (UMAP) was trained using HC and CFRD data, and Silhouette scores quantified the alignment of each daily profile with these clusters. Group differences were evaluated using linear mixed effects models.

RESULTS

UMAP showed complete separation between HC and CFRD. Mean Silhouette score values were +0.35 (95% CI: 0.31 to 0.38) for HC and -0.77 (95% CI: -0.83 to -0.71) for CFRD. PwCF classified as normal glucose tolerance (NGT) or impaired glucose tolerance (IGT) had negative Silhouette score values, -0.58 (95% CI: -0.67 to -0.49) and -0.56 (95% CI: -0.63 to -0.49), with no difference between groups.

CONCLUSIONS

Machine learning analysis of CGM data revealed a pervasive dysglycemic phenotype in PwCF. NGT and IGT individuals showed similar glycemic profiles shifted toward the CFRD phenotype, indicating that OGTT categories underestimate early metabolic dysfunction. CGM based digital phenotyping offers a more sensitive and continuous assessment of dysglycemia and may improve early detection and risk stratification in PwCF.