Unsupervised clustering of biochemical markers reveals health profiles associated with function and survival in active aging

Abstract

This study explores the relationships between biochemical phenotypes identified using machine learning, and key health outcomes, including body composition, physical function, and mortality risk. Data were collected from 536 physically active Spanish participants aged over 65 years (76.5% women) enrolled in the EXERNET cohort (2017-2018), with a 6-year mortality follow-up. Principal component analysis, and hierarchical and k-means clustering was used to identify distinct biochemical profiles. Associations between clusters and health outcomes were assessed using analysis of covariance and Cox proportional hazards models. Three distinct clusters emerged: 'Healthy', characterized by biochemical values within the normal range and used as the reference group; 'Metabolic', marked by dysregulated metabolic parameters; and 'Hepatic', which exhibited impaired liver function markers. Notably, all clusters showed subclinical levels of dysfunction. The 'Healthy Cluster' demonstrated the highest levels of organized physical activity (90%, p < 0.001), whereas the 'Metabolic Cluster' showed poorer body composition and reduced physical performance. Both the 'Metabolic' and 'Hepatic' clusters demonstrated a higher mortality risk, as confirmed through Cox regression analyses. Adjusted hazard ratios were significantly elevated when considering physical activity and adiposity, with values of 3.45 and 3.71 for the 'Metabolic Cluster', and 3.01 and 3.85 for the 'Hepatic Cluster' (p < 0.05). This study underscores the strong link between metabolic health, physical activity, body composition and 6-years mortality risk in older adults. Machine learning techniques for identifying phenotypic clusters offers a promising tool for early detection and targeted interventions to improve aging outcomes.

Keywords: Aging; Biochemical phenotypes; Body composition; Cluster analysis; Mortality risk; Physical function.

Fig. 1

Flowchart of participants included in the analysis.

Fig. 2

Principal Component Analysis (PCA) heatmap of the blood test profile. The plot represents the contribution of variables onto the principal components which explain X% (First component) and Y% (Second component) of the total variance, respectively. The colour of the points reflects the contribution of each variable to the PCA. K: potassium, Ca: calcium, ALT: alanine aminotransferase, GGT: gamma glutamyltransferase, LDH: lactate dehydrogenase, VLDL: very low-density lipoproteins, TyG index: triglycerides/glucose index, MCV: Mean Corpuscular Volume, MHC: Mean Corpuscular Hemoglobin, MCHC: Mean Corpuscular Hemoglobin Concentration, MPV: Mean Platelet Volume.

Fig. 3

Biochemical profiles of three phenotypes across four parameter categories. The figure shows the biochemical profiles of the ‘Healthy Cluster’, ‘Hepatic Cluster’, and ‘Metabolic Cluster’ across four categories: (a) metabolic, (b) renal, (c) immunological, and (d) hepatic parameters. The X-axis represents the parameters within each category, and the Y-axis displays the scaled values (− 1 to 1) to enable comparison across clusters. VLDL: very low-density lipoproteins, TyG index: triglycerides/glucose index, K: potassium, Ca: calcium, ALT: alanine aminotransferase, GGT: gamma glutamyl transferase, LDH: lactate dehydrogenase.

Fig. 4

Survival curves for each biochemical phenotype and gender, and Cox hazard ratios for influencing variables. The figure shows: (a) survival curves for the ‘Healthy’, ‘Metabolic’ and ‘Hepatic’ clusters, (b) by sex, and (c) Cox proportional hazard plot showing the log of the hazard ratio (HR) and 95% confidence intervals (CI) for variables influencing mortality, highlighting significant risk factors for mortality across groups. BMI: Body mass index.