Leader: Licia Iacoviello, Giacomo Frati (NEUROMED); Other collaborator(s): Federica Agosta (UNISR); Nicola Baldini (UNIBO); RTDA PNRR TBA (UNICATT).
In this task, we aim to deploy AI algorithms to identify patients at risk of adverse aging-related health outcomes, using a high number of biomarkers from different omics layers. This will allow us to i) identify patients at higher risk of relapses of adverse health outcomes and ii) identify predictive biomarkers which may be used in common clinical practice to monitor their health status in the future.
Brief description of the activities and of the intermediate results: This Task is inextricably intertwined with the 3.5. Partner NEUROMED has computed a new marker of biological ageing (Phenotypic Age, or PhenoAge) within the Moli-sani cohort. PhenoAge is a measure previously developed to predict incident mortality risk in the North-American population, applying a Gompertz Proportional Hazard model predict incident mortality as a function of nine blood markers – albumin, creatinine, CRP, glucose, LY, MCV, RDW, WBC and alkaline phosphatase (ALP) – and of chronological age of subjects, followed by reconversion of the resulting mortality score in year units. We have tested the model after excluding the measure of ALP that was not available in our dataset. With data publicly available we have tested the algorithm with and without ALP, showing similar performances.
Meetings and updates with Cecchi’s group to work together on a systematic review on the relationship between digital health literacy and older people. After the preliminary screening procedures, 462 papers will be considered for the abstract/title screening, and we are in the process of assessing the full texts.
Main policy, industrial and scientific implications: This activity will generate new approach and algorithms od AI that will be used in other tasks to explore associations between complex exposures and outcomes.
Partner NEUROMED has investigated the blood-based index of biological aging acceleration, tagging different domains like renal, liver and heart functions (doi: 10.1007/s10654-021-00797-7), which was deployed within the Moli-sani cohort. In particular, it has been explored the influence of the different markers on the prediction of BA through SHAP (SHapley Additive exPlanations) values, computed using the ‘SHAPforxgboost’ library (v. 0.1.3). Then, we plots have been built to visualize the computed SHAP values using ‘shapviz’ library (v. 0.9.3), which were compared with scientific literature in the field.
The effect of min-max vs L2-normalization on the features of the algorithm have been evaluated, assessing and comparing their performance in terms of loss function (Root Mean Squared Error) and of accuracy metrics (Mean Absolute Error). BA and the relevant discrepancy with chronological age (hereafter called DeltaAge) has been computed for all the participants passing marker QC in the Moli-sani cohort (N=23,858).
Partner NEUROMED has further implemented the computation of the novel SHAP (SHapley Additive exPlanations) values, computed using the ‘SHAPforxgboost’ library (v. 0.1.3), and their visualization through the ‘shapviz’ library (v. 0.9.3)., on the novel BloodAge algorithm deployed in 23,858 participants from the Moli-sani cohort. This plot will be included in a novel manuscript assessing the joint influence of BloodAge and PhenoAge acceleration of cancer-related risks (in preparation).
Partner NEUROMED has further implemented the computation of the novel SHAP (SHapley Additive exPlanations) values, computed using the ‘SHAPforxgboost’ library (v. 0.1.3), and their visualization through the ‘shapviz’ library (v. 0.9.3)., on the novel BloodAge algorithm deployed in 23,858 participants from the Moli-sani cohort. This plot will be included in a novel manuscript assessing the joint influence of BloodAge and PhenoAge acceleration of cancer-related risks (in preparation).
Partner Neuromed, within a longitudinal subcohort of the Moli-sani study, we computed several DNA methylation aging clocks using epigenome wide methylation data of 865,918 CpGs, measured through the Illumina® EPIC array (v1). Specifically, 1,098 subjects underwent longitudinal methylation analyses using blood samples retrieved upon cohort recruitment (2005-2010, t0) and recall (2017-2020, t1). Quality Control (QC), normalization and batch effect removal of methylation (beta) signals were carried out in ChAMP v2.20.1 (see Quiccione et al., 2024 for further details; doi: 10.3390/ijms251910317). 668,413 probes and all the initial samples passed QC, which were used to compute different epigenetic aging clocks. Specifically, using the methylclock v1.12.0 package in R (https://github.com/isglobal-brge/methylclock) we computed two first generation DNA methylation (DNAm) aging clocks, namely the Hannum (Hannum et al 2013; 10.1016/j.molcel.2012.10.016) and the Horvath DNAm ages (Hannum et al 2013; 10.1016/j.molcel.2012.10.016), based on the application of elastic net regression models (as trained in other poppulations) on 71 and 353 CpG sites epigenome-wide. These clocks will be tested for association with several age-related incident outcomes in the Moli-sani cohort.