JMIR Publications

ABSTRACT

Background:

Distinguishing type 1 diabetes (T1D) from type 2 diabetes (T2D) guides treatment, but adult presentations overlap. Continuous glucose monitoring (CGM) captures glucose fluctuations and day-to-day consistency, but CGM studies differ in recruitment, treatment, and sensor manufacturer, complicating classifier evaluation.

Objective:

This study aimed to test whether 7-day CGM features distinguish source-labeled T1D, T2D, and no diabetes in adults when entire datasets are held out, compare interpretable and neural models with simple baselines, and measure how sensor substitution changes predictions.

Methods:

We trained models on 4 adult datasets (AI-READI, ShanghaiT2DM, T1DiabetesGranada, and T1DEXI) and tested them on 4 others (Glucotypes, ShanghaiT1DM, AZT1D, and CGMacros), with every T1D/T2D-by-Dexcom/Abbott combination in both partitions. A cap of 100 participants per dataset and diagnosis yielded 492 training and 107 test participants. Each 7-day window provided 47 CGM features; predictions were averaged within participant. We compared mean-glucose and clinical-summary (mean, standard deviation, coefficient of variation, and glucose zones) logistic regression with elastic net, boosted trees, and 4 neural networks. The primary end point was participant-level macro-averaged area under the receiver operating characteristic curve (AUROC) with bootstrap 95% CIs. Additional analyses covered pooled, within-Shanghai, cross-cohort, and manufacturer-restricted evaluation. In 29 CGMacros participants (15 without diabetes, 14 with T2D) wearing Abbott FreeStyle Libre Pro and Dexcom G6 Pro sensors simultaneously, we substituted Dexcom for Libre input.

Results:

The 599 adults comprised 237 with T1D, 210 with T2D, and 152 without diabetes. Elastic net achieved a macro-AUROC of 0.943 (95% CI 0.914-0.969) and a T1D-versus-T2D AUROC of 0.935 (95% CI 0.869-0.980), exceeding mean-glucose and clinical-summary logistic regression (paired gains over the latter 0.046, 95% CI 0.015-0.079, and 0.105, 95% CI 0.017-0.212). Pooled mean glucose was similar (T1D-minus-T2D difference 5.28 mg/dL, 95% CI -1.24 to 11.40), whereas the mean of daily differences (MODD), the leading elastic-net feature, was higher in T1D (21.11 mg/dL, 95% CI 18.31-23.83). T1D sensitivity was 78.4% overall, 92% in AZT1D, and 50% in ShanghaiT1DM; T2D precision was 45.7%. No dataset contained both T1D and T2D; when the T1D-versus-T2D contrast came from the shared Shanghai registry and sensor, elastic-net AUROC was 0.82-0.86, and clinical-summary logistic regression led within Shanghai (0.907 vs 0.859) and in Abbott-only evaluation (0.976 vs 0.887). Substituting Dexcom for Libre input reduced primary elastic-net log loss by 0.389 (95% CI 0.140-0.673) but increased Abbott-only log loss by 0.567 (95% CI 0.257-0.923).

Conclusions:

In primary external testing, an interpretable 47-feature elastic-net model ranked diabetes categories better than mean-glucose and clinical-summary baselines, with day-to-day glucose variability as the leading signal. T1D errors concentrated in 1 adult-onset cohort, and sensor substitution changed predictions depending on the training setting, so classifiers pooling manufacturers can be checked with paired-sensor recordings as well as external cohorts.