JMIR Publications

ABSTRACT

Background:

Background:

Foundation models can map continuous glucose monitoring (CGM) data into learned glucose representations, offering new potential for cardiovascular disease (CVD) risk prediction.

Objective:

This study investigated the glucose patterns captured by a foundation model and evaluated whether CGM data can improve the prediction of elevated N-terminal pro-BNP (NT-proBNP), a heart failure biomarker

Methods:

Method: Two datasets, including either people with type 1 diabetes (n=205) or type 2 diabetes (n=159), were analyzed. Representations derived from CGMformer were evaluated using uniform manifold approximation and projection (UMAP) and pairwise cosine distance analysis. Logistic regression with bootstrap optimism correction predicted elevated NT-proBNP (>125pg/mL) using clinical risk factors, the SCORE2-Diabetes score, CGM-derived metrics, and learned representations.

Results:

Results:

Cosine distance and UMAP analyses demonstrated the foundation model captured individual-specific glycemic patterns (HbA1c and time-in-range(TIR)). UMAP dimensions for CANCAN dataset showed strong correlations with TIR (ρ=-0.71 to 0.81, p<0.001) and HbA1c (ρ=0.48 to -0.65, p<0.001), whereas correlations with NT-proBNP were weak (ρ=-0.04, p=0.063 and ρ=0.18, p<0.001). Similar and statistical significant (p<0.001) but weaker correlations with TIR and HbA1c were observed in T1DX. Prediction models with clinical predictors or SCORE2-Diabetes score achieved optimism-corrected area under the receiver operating characteristic curve (AUROC) of 0.79 [95%CI:0.76-0.87] and 0.74 [0.65-0.81] respectively. Adding the CGM-derived metrics time-in-range and mean amplitude of glycemic excursions achieved AUROCs of 0.81 [0.77-0.88] and 0.77 [0.70-0.85], replacing them with CGMformer based representations resulted in AUROCs to 0.79 [0.75-0.87] and 0.72 [0.65-0.81], and increased overfitting. The UMAP dimensions from the CANCAN data were strongly correlated with TIR (UMAP dimension 1: ρ=-0.71, p<0.001; UMAP dimension 2: ρ=0.81, p<0.001), and moderately correlated with HbA1c (UMAP dimension 1 (UMAP 1)ρ=0.48, p<0.001; UMAP dimension 2(UMAP 2): ρ=-0.65, p<0.001), while only weak correlations were observed with NT-proBNP (UMAP 1: ρ=-0.04, p=0.063; UMAP 2: ρ=0.18, p<0.001). Weaker correlations were observed for the T1DX dataset for both TIR (UMAP 1: ρ=-0.39, p<0.001; UMAP 2: ρ=-0.65,p<0.001) and HbA1c (UMAP 1: ρ=0.30, p<0.001; UMAP 2: ρ=0.50,p<0.001).

Conclusions:

Conclusions:

Learned representations capture both individual-level glucose patterns and meaningful structures when mapped against HbA1c and TIR. No patterns were observed upon inspection for CVD-related markers and CGM data did not improve prediction of NT-proBNP beyond established clinical variables. Larger prospective studies are warranted to examine whether individual glucose patterns may improve CVD prediction in diabetes