Improved perfusion pattern score association with type 2 diabetes severity using machine learning pipeline: Pilot study

Abstract

Background: Type 2 diabetes mellitus (T2DM) is associated with alterations in the blood-brain barrier, neuronal damage, and arterial stiffness, thus affecting cerebral metabolism and perfusion. There is a need to implement machine-learning methodologies to identify a T2DM-related perfusion pattern and possible relationship between the pattern and cognitive performance/disease severity.

Purpose: To develop a machine-learning pipeline to investigate the method's discriminative value between T2DM patients and normal controls, the T2DM-related network pattern, and association of the pattern with cognitive performance/disease severity.

Study type: A cross-sectional study and prospective longitudinal study with a 2-year time interval.

Population: Seventy-three subjects (41 T2DM patients and 32 controls) aged 50-85 years old at baseline, and 42 subjects (19 T2DM and 23 controls) aged 53-88 years old at 2-year follow-up.

Field strength/sequence: 3T pseudocontinuous arterial spin-labeling MRI.

Assessment: Machine-learning-based pipeline (principal component analysis, feature selection, and logistic regression classifier) to generate the T2DM-related network pattern and the individual scores associated with the pattern.

Statistical tests: Linear regression analysis with gray matter volume and education years as covariates.

Results: The machine-learning-based method is superior to the widely used univariate group comparison method with increased test accuracy, test area under the curve, test positive predictive value, adjusted McFadden's R square of 4%, 12%, 7%, and 24%, respectively. The pattern-related individual scores are associated with diabetes severity variables, mobility, and cognitive performance at baseline (P < 0.05, |r| > 0.3). More important, the longitudinal change of individual pattern scores is associated with the longitudinal change of HbA1c (P = 0.0053, r = 0.64), and baseline cholesterol (P = 0.037, r = 0.51).

Data conclusion: The individual perfusion diabetes pattern score is a highly promising perfusion imaging biomarker for tracing the disease progression of individual T2DM patients. Further validation is needed from a larger study.

Level of evidence: 1 Technical Efficacy: Stage 1 J. Magn. Reson. Imaging 2019;49:834-844.

Keywords: machine learning; perfusion diabetes pattern score; type 2 diabetes mellitus.