Predicting stimulated C-peptide in type 1 diabetes using machine learning: a web-based tool from the T1D exchange...

Area of Science:

• Endocrinology
• Computational Biology
• Diabetes Research

Background:

• The mixed-meal tolerance test (MMTT) is the gold standard for assessing residual beta-cell function in type 1 diabetes (T1D) but is not practical for routine clinical use.
• Accurate assessment of beta-cell function is crucial for understanding T1D progression and guiding treatment strategies.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting MMTT-stimulated C-peptide categories using readily available clinical data.
• To provide a non-invasive and practical tool for estimating beta-cell function in individuals with T1D.

Main Methods:

• Utilized data from 319 T1D patients from the T1D Exchange Registry, splitting it into training (70%) and testing (30%) sets.
• Selected key clinical variables including age at diagnosis, diabetes duration, HbA1c, non-fasting glucose, and non-fasting C-peptide.
• Trained and validated four ML algorithms (Random Forest, XGBoost, LightGBM, Ordinal Logistic Regression) using 10-fold cross-validation.

Main Results:

• The Random Forest (RF) model achieved high performance, with an AUC of 0.94 in cross-validation and 0.97 in the test set.
• The RF model demonstrated excellent sensitivity (84% in CV, 88% in test) and specificity (92% in CV, 94% in test) for predicting C-peptide categories.
• A significant proportion (17.7%) of individuals with undetectable fasting C-peptide showed measurable levels post-MMTT, highlighting the model's ability to detect subtle function.

Conclusions:

• The developed ML model offers a practical and non-invasive method for estimating residual beta-cell function in T1D.
• This tool can potentially streamline clinical practice by reducing the need for the MMTT.
• The model is accessible online for broader application in T1D research and clinical settings.