Leader: Licia Iacoviello (NEUROMED); Other collaborator(s): Rosa Campopiano (Neuromed) Carla Letizia Busceti (Neuromed)
Molisani biobank will be tested for telomeric ncRNA accumulation during aging and disease
Brief description of the activities and of the intermediate results: finalization of the DNN algorithm used for Heart Age estimation, based on dichotomous ECG tracing alteration features (Minnesota coding system). Performance evaluation in terms of loss (root mean squared error, RMSE) and accuracy (mean absolute error, MAE) with respect to the label variable (chronological age). The best performing algorithm shows RMSE = 84.86 and MAE = 7.27 years, Pearson correlation coefficient r = 0.62 and R2 = 0.39 in the test set. A comparative analysis of the literature in the field of heart age estimation based on similar algorithms revealed how the algorithm thus produced performs comparably – if not better – than neural networks trained on cohorts 1.5x to 65x times larger ( doi: 10.1038/s41467-021-25351-7; 10.1161/CIRCEP.119.007284; Calculation of the discrepancy between biological and chronological age as an index of biological aging in the test set.
Update: validation of the Heart Age measure and the relative difference between HeartAge and chronological age (hereinafter DeltaHeartAge), through comparison between men and women, subjects with prevalent cardiovascular disease and unaffected subjects, as well as through Cox proportional hazards models for the prediction of the incident risk of death from all causes and from specific causes (cardiovascular, cancer, other cause), hospitalizations for specific causes (cardiovascular, cancer), as well as for incident fatal/non-fatal cases of type 2 diabetes and various types of cancer (breast, prostate, lung and colorectal). To this end, incremental Cox models were constructed, adjusted for i) age and sex (Model 1), ii) Model 1 + prevalent diseases (cardiovascular, diabetes, dyslipidemia, hypertension, cancer and BMI) (Model 2), iii) Model 2 + education level (Model 3), iv) Model 4 + lifestyle factors (adherence to the Mediterranean diet, physical activity, daily caloric and alcohol intake). These analyses revealed significant risk associations of increasing DeltaHeartAge with all-cause mortality, coronary heart disease and stroke, hospitalizations for atrial fibrillation and with incident risk of occurrence of fatal/non-fatal type 2 diabetes.
For the same purpose, it was decided to invest in estimating epigenetic clocks, which aim to assess the degree of aging based on methylation patterns in the genome and are the most tested and successful estimators of biological aging in the past 10 years. Specifically, after extracting DNA from buffy coat samples stored at the Neuromed biobanking Center and bisulfiting the DNA, methylation levels at approximately 935,000 CpG sites in the epigenome began to be tested using the Illumina MethylationEPIC V2.0 Kit array, for about 1,400 participants from the study out of the ~2,500 with data available from both baseline recruitment (2005-2010) and the first active cohort follow-up (2017-2020). These subjects will allow the analysis of how epigenetic aging correlates with biological aging measured through other blood-based clocks (BioAge and PhenoAge) and with organ-specific estimators (e.g., HeartAge). Additionally, it will be possible to test whether these clocks independently predict fractions of risk for mortality and age-related chronic diseases. To better understand the mechanisms and pathways of the biological aging process at the brain level, we have initiated studies on induced stem cells obtained from a cohort of subjects over the age of 65, consisting of healthy individuals and patients with neurodegenerative diseases. These cells are differentiated into different neuronal populations in order to study the neuronal connection network under physiological and pathological conditions and their ability to respond to external stimuli. These studies will be conducted using advanced technological platforms based on MEA technology and confocal fluorescence microscopy.
We have made progress in deploying a novel algorithm for the estimation of HeartAge. In particular, we improved training by implementing hyperparamter tuning of DNN models through Bayesian approaches using ‘rBayesianOptimization’ package (v. 1.2.1) in R. This approach is expected to be more efficient and to outperform other optimization strategies (grid and random searches). Moreover, we have investigated the inclusion of novel potential features for the prediction of HeartAge, like Heart Rate variability.
Two lines of iPSCs derived from healthy individuals and two derived from PD patients were selected. Differentiation of iPSCs into dopaminergic neurons was performed as described (PMID 28858290), and the expression of specific markers were analyzed through Immunofluorescence and Confocal Imaging, at various stages of differentiation. Specifically, we analyzed: on day 0, pluripotency markers NANOG, OCT4, and SOX2, before initiating neuronal differentiation; on day 16, LMX1a/b and FOXA2, markers of dopaminergic precursor neurons of the substantia nigra; on day 45, at the end of differentiation, the expression of dopaminergic neuron marker TH and pan-neural marker MAP2. Additionally, we have set up plating conditions on MEA chips to evaluate the neuronal activity of PD cells and control cells at different stages of differentiation.
We made progress in the deployment of HeartAge, by further refining the hyperparameter tuning through Bayesian approaches and reaching a better accuracy (MAE=7.0 years), comparable to the one reached in much larger cohorts. This measure confirmed significant associations with incident cardiovascular risk (coronary heart disease, myocardial infarction, both ischemic and haemorrhagic stroke). Moreover, an accelerated heart aging was observed in subjects with prevalent cardiometabolic conditions (including diabetes, hypertension and all the above mentioned cardiovascular disorders). Further analyses are ongoing to deepen associations with incident risks other than cardiovascular ones and to clarify the influence of each feature on the resulting measure through ShAP values.
For "in vitro" experiments, four lines of iPSCs derived midbrain precursors at 16 days of differentiation, two from healthy individuals and two derived from PD patients, carrying pathogenic mutations in novel PD genes (SLC6A3, HMOX2, KIF21B), were plated on single-well CorePlate™ 1W 27/42 (3Brain). This is a high-density microelectrode array (HD-MEA) integrating 4,096 recording electrodes on a small and compact area of 2.7 x 2.7 mm2 (50.000 cells for each array) that allow for simultaneous electrophysiological recording. Cells were maintained in culture until the sixtieth day of differentiation with the addition of specific factors for differentiation into mdDA neurons. Electrophysiology recording was performed on the live-cell imaging platform BioCAM X (3Brain) at 40, 47, 55 and 60 days of differentiation. Preliminary data show a faster differentiation of PD cells compared to control cells with earlier formation of foci and a higher number of cells positive for the TH marker at the end of differentiation (66% PD vs 44% CNT). Analysis of spike signals demonstrates a higher neuronal activity of control cells compared to PD cells. The analysis of data at different times is still ongoing.
After carefully tuning hyperparameters via Bayesian optimization, we developed a multi-layer perceptron DNN with two hidden layers to estimate an individual’s age based on sex, heart rate, systolic and diastolic blood pressure, and standard dichotomus encodings (Minnesota) from standard 12-lead resting ECGs. The model was trained on a randomly selected training set from the Moli-sani cohort (baseline recruitment: 2005–2010; follow-up until 31 December 2022; N = 24,325; age ≥35 years; 51.9% women) and evaluated on the remaining 20% of the sample. The difference between the predicted biological age and the chronological age was calculated and referred to as the Heart Aging (HA) clock.
Notably, after excluding individuals with missing ECG data or prevalent cerebro-cardiovascular disease at baseline, Cox Proportional Hazards models—incrementally adjusted for age, sex, education, lifestyle factors, quality of life, adiposity, and prevalent health conditions—revealed a significant association between HA and the incidence of both fatal and non-fatal cardiovascular events during follow-up, including cardiovascular disease (HR: 1.03), coronary heart disease (HR: 1.04), stroke (HR: 1.03), and diabetes (HR: 1.03). No significant associations were found with cancer, major depression, Parkinson’s disease, or Alzheimer’s disease. The HA clock was also significantly associated with increased risk of all-cause mortality (HR: 1.03), and mortality due to cardiovascular disease (HR: 1.03) and coronary heart disease (HR: 1.04), while associations with mortality from COVID-19 or cancer approached but did not reach statistical significance. Furthermore, when evaluated alongside other aging clocks such as PhenoAge and BloodAge, the HA clock remained significantly associated with several outcomes, suggesting that these biomarkers may capture distinct aspects of the aging process.
These preliminary results highlight the prognostic potential of HA - alone or in combination with other aging clocks- as a non-invasive, cost effective, largely available, ECG-based biomarker for the risk of incident morbidity and mortality even beyond cerebro-cardio-vascular disease.
These interesting findings are going to be presented at the 18th European Public Health Conference 2025 and at the Age-IT 2nd General Meeting.
In the reference period, once the experimental protocol was set up, we expanded the number of experiments in order to obtain at least 3 biological replicates for each sample on which to perform the statistical analysis. In particular, the functional analyses were performed through simultaneous electrophysiological recording of 4 lines of dopaminergic neurons derived from iPSCs from 2 patients and 2 healthy controls. Electrophysiology recording was performed on the live-cell imaging platform BioCAM X (3Brain) at 40, 47, 55 and 60 days of differentiation. Patient cells showed an earlier differentiation, but were smaller, with a less stable network, and with a significantly reduced number (p<0.01) of electrophysiologically active neurons. At day 60 of differentiation, cells were treated with 100 microM glutamate to test for glutamate neurotoxicity. TH-positive cells from PD patients were more sensitive to glutamate-induced toxicity. The analysis will be extended to other patient lines and controls.
Gialluisi A.: “Machine learning approaches for the estimation of biological aging”, EUPHA International Symposium «Society, Nutrition and Chronic Disease Prevention: Time for a Paradigm Shift», June 5-6, 2025, Pozzilli, Italy
Pepe et al. «Potential of an ECG-based heart aging index for risk stratification in the general population», accepted as poster at the 18th European Public Health Conference 2025, Helsinki, Finland.
Pepe et al. «Electrocardiography-based heart aging: a novel potential index for risk stratification in the general population», Accepted as oral presentation at the Age-IT 2nd General Meeting, October 29-31 2025, Naples.