What is the clinical evidence from this study?

Study design: Observational. Population: Type 1 Diabetes (T1D) (n=603). Intervention: Machine learning risk model. Primary outcome: Progression to Stage 3 T1D within 2 years.

What does this research mean for the field?

A machine learning model utilizing noise-resistant, multidimensional clustering can predict two-year progression to Stage 3 Type 1 diabetes in individuals with multiple autoantibodies with high specificity and modest sensitivity. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

This research aims to develop a machine learning model to predict the progression to Stage 3 Type 1 Diabetes in individuals with multiple autoantibodies.

June 7, 2026

2223-P: Harnessing Machine Learning to Forecast Two-Year Progression to Stage 3 Type 1 Diabetes (T1D) in Individuals with Multiple Type 1 Diabetes Autoantibodies: A Noise-Resistant, Multidimensional Clustering Approach

Key Result

A machine learning model predicting 2-year progression to Stage 3 Type 1 diabetes in individuals with ≥2 autoantibodies achieved a mean specificity of 0.89, PPV of 0.71, and sensitivity of 0.50.

Key Points

This research aims to develop a machine learning model to predict the progression to Stage 3 Type 1 Diabetes in individuals with multiple autoantibodies.
Participants with ≥2 T1D-associated autoantibodies were analyzed, utilizing demographic, immunologic, metabolic, and genetic data.
Data was split into five subsets for feature selection and model training using cross-validation.
Synthetic samples were generated to address class imbalance, and a noise-resistant clustering algorithm was applied.
Among 603 participants, 205 (34.0%) progressed to Stage 3 T1D within 2 years, with key predictors being metabolic measures and risk scores.
The ML model achieved a mean specificity of 0.89 (±0.05), PPV of 0.71 (±0.07), and sensitivity of 0.50 (±0.13).
F1 score indicating balance between PPV and sensitivity was 0.58 (±0.09).

Study Design

Type

Observational (n=603)

Structured PICO

Can a machine learning model accurately predict 2-year progression to Stage 3 Type 1 Diabetes in individuals with multiple T1D autoantibodies?

Population

603 participants with ≥2 T1D-associated autoantibodies (median age 10 years, 45.8% female) evaluated for 2-year progression to Stage 3 Type 1 diabetes.

Exposure

Machine learning model using demographic, immunologic, metabolic, and genetic data to output 'Proximity Scores' (range: 0-200)

Outcome

Progression to Stage 3 T1D within 2 years

A machine learning model using multidimensional data demonstrated modest discriminative performance for predicting 2-year progression to Stage 3 Type 1 diabetes in high-risk individuals.

Abstract

Introduction and Objective: Early identification of individuals at high risk for progressing to Stage 3 (clinically diagnosed) Type 1 diabetes (T1D) is essential for designing prevention trials and recommending disease modifying therapies. Existing risk models often rely on limited data types and struggle to capture variability in complex data, limiting predictive accuracy. Using a novel algorithm designed to capture patterns in complex data, we developed a machine learning (ML) model to predict 2-year progression to Stage 3 T1D among TrialNet Pathway to Prevention (PTP) participants with ≥2 T1D-associated autoantibodies (AAb). Methods: Demographic, immunologic, metabolic, and genetic data for participants with ≥2 AAbs were divided into five stratified subsets for feature selection and model training via cross-validation. Synthetic samples were generated to address class imbalance. Using an algorithm that combines noise suppression with spatial clustering, we produced an ML model that outputs “Proximity Scores” (range: 0-200) for predicting whether Stage 3 T1D will occur within 2 years. Results: Of 603 participants with ≥2 AAbs (median age, 10 IQR 8 years; 45.8% female; 77.3% non-Hispanic White), 205 (34.0%) progressed to Stage 3 T1D within 2 years. Top predictors included metabolic measures (e.g., HbA1c) and T1D risk scores (e.g., Index60). Using a fixed decision threshold of 140 facilitated prioritization of specificity and positive predictive value (PPV) as performance metrics. Mean specificity was 0.89 (±0.05), PPV 0.71 (±0.07), sensitivity 0.50 (±0.13), and balance between PPV and sensitivity (F1 score) 0.58 (±0.09). Conclusion: Our results demonstrate consistent, modest discriminative performance for predicting progression to Stage 3 T1D within 2 years in PTP participants with ≥2 AAbs. These findings highlight the potential for ML-based risk models to aid early identification and guide therapeutic interventions. Disclosure E. Tallon: None. M. Shapiro: None. E. Paprocki: None. A.D. Cappeluti: Stock/Shareholder; Current; OTraces Inc, Noble Life Sciences Inc, Iterion Inc. T.A. Hogan: None. B. Lockee: None. A. Waghmode: None. T. Brusko: None. L. DiMeglio: Research Support; Ended; Dompé, Lilly. Stock/Shareholder; Ended; Lilly. Research Support; Current; MannKind Corporation. Research Support; Ended; Provention Bio, Inc. Research Support; Current; Sanofi. Research Support; Ended; Zealand Pharma A/S. Consultant; Current; Tandem Diabetes Care, Inc. Other - DSMB member; Current; Merck Current; Endsulin, Sernova, Corp. Board Member; Current; Diamyd Medical. Consultant; Current; Minutia, Quell Therapeutics, SAB Therapeutics, Vertex Pharmaceuticals Incorporated. Speaker's Bureau; Current; Sanofi. Stock/Shareholder; Current; Diamyd Medical. Other - Other; Current; Insulin for Life USA. P. Gottlieb: Consultant; Current; Eli Lilly and Company. Board Member; Current; IM Therapeutics. Other - CEO, CMO; Current; IM Therapeutics. Research Support; Current; Immune Tolerance Network, Nova Laboratories, National Institute of Diabetes and Digestive and Kidney Diseases. Advisory Panel; Current; Sanofi. Research Support; Current; Sanofi. Consultant; Current; SAB Biotherapeutics, Inc., Anaptys Bio, Cour, T1D Fund. Consultant; Ended; Imcyse, Viacyte, Abata. M. Clements: Employee; Current; Glooko, Inc. Research Support; Ended; Dexcom, Inc., Abbott Diabetes. W. Moore: None.

Bookmark