Phenotypic Clusters of "Cardiometabolic Cycling" and High Variability Identified by Machine Learning Predict Progression of Kidney Diseases

Background: High variability and dynamic changes in haemoglobin A1c (HbA1c), blood pressure, body mass index (BMI), and cholesterol predicts cardiovascular disease. However the combined effect of high variability of all these indices on kidney disease progression is unknown. . Our aim was to use advanced machine learning methods to identify phenotypic clusters of people with type 1 diabetes (T1DM) based on long-term variability in metabolic and hemodynamic markers, and assess their prognostic value for kidney disease progression defined as eGFR fall >50% from baseline with final eGFR <30 ml/min with death as a competing risk. Methods: Electronic health data and records from from 2,735 people with T1DM (median age 35 years; 52% female; 11% Black, 81% White, 8% Other) all with baseline eGFR>45 ml/min were analysed. Longitudinal values for HbA1c, mean arterial pressure (MAP), BMI, and cholesterol were collected with their variability measured via Average Real Variability (ARV). ARV adjust for time differences between visits. Variability was computed over a 15-measurement maximum exposure window. Clustering on log-ARV values was performed using K-means, hierarchical, and Gaussian Mixture Models; K-means was selected based on silhouette score and Calinski-Harabasz index. Dimensionality reduction supported cluster structure. Results: Over observation period of upto 14 years the total number of renal events was 71 and the number of deaths 94. K-means identified four clusters (n=1096, 733, 577, 329 respectively). K-means cluster membership was added as a covariate to a multivariable Cox proportional hazards model predicting kidney disease progression. The baseline model (without clustering) yielded a concordance index of 0.821 (95% CI: 0.757–0.863), which improved to 0.864 (95% CI: 0.814–0.890) with cluster inclusion p<0.05. K-means clusters demonstrated significant hazard ratio separation even after adjusting for baseline HbA1c, MAP, eGFR, ACR, ethnicity, and cholesterol. Cluster 4 had the highest risk (HR: 2.45; 95% CI: 1.31–4.58; p=0.005), while Cluster 2 was protective (HR: 0.15; 95% CI: 0.04–0.66; p=0.01). Conclusion: Machine learning-derived clusters based on cardio-metabolic variability are strong predictors of kidney disease progression and may guide targeted risk stratification and intervention. Funding: Government Support – Non-U.S.