**Table 4.** *Array-based technologies in biomarker studies in T1D.*


*AUC: area under the curve, ZnT8A: Zn transporter 8 autoantibody, PTPRN22: protein tyrosine phosphatase receptor 22, MLH1: MutL protein homolog 1, PPIL2: Peptidylprolyl Isomerase Like 2, QRFPR: Pyroglutamylated RFamide Peptide Receptor.*

### **Table 5.**

*Novel T1D AAb panel identified by NAPPA array: AUC and sensitivity compared to ZnT8A [8].*

Advances in omic technology have allowed researchers to discover several potential biomarkers. Validation of these biomarkers requires thousands of samples, irrespective of the technique, and thus biomarker development is often hindered by limited availability of biological samples. Recently, T1D repositories such as TEDDY and TrialNet have addressed this problem by providing a large pool of available sample data for omic analysis. However, high computational power is required to analyze numerous biomarkers in these large data sets and identify the optimal markers for a panel.

AI machine learning has been used for a broad range of applications in cancer treatment, including diagnosis and classification of cancer as well as prediction of progression and treatment outcomes [62]. This suggests that AI may help to solve the issue of early prediction of T1D as well. The main goal of AI in T1D biomarker discovery is to combine information on the small differences in serum biomarkers between T1D and healthy patients and utilize this information to predict which patients are at high risk of developing T1D in the future.

Recent studies have utilized various machine-learning techniques to analyze large amounts of omics data (**Table 6**). Repeated Optimization for Feature Interpretation (ROFI) is one such technique that uses a repeated selection algorithm 500 times to generate important matrices for each feature [65, 67]. These matrices define the importance of a feature as the percentage of times it was selected to be included out of the 500 times the algorithm was run [65, 66]. After the feature importance values have been established, the final model is produced [65]. These studies illustrate the potential utility of machine learning in analyzing large amounts of data from diverse fields of omics data (genomics, proteomics, metabolomics, and lipidomics), clinical risk factors, and environmental factors for early prediction and prevention of T1D. Machine learning also has the potential to improve the prediction of T1D complications such as diabetic nephropathy, retinopathy, and peripheral neuropathy [11, 57]. The goal is that machine learning-derived risk scores can be used to identify T1D patients who would benefit the most from targeted preventative therapies before the development of these complications.
