A machine learning model identifies a functional connectome signature that predicts blood pressure levels: imaging insights from a large population of 35 882 patients

This editorial refers to ‘A functional connectome signature of blood pressure in >30 000 participants from the UK biobank’, by R. Jiang et al.,https://doi.org/10.1093/cvr/cvac116. Mounting evidence indicates that hypertension is associated with brain structural damage, including macroscopically evident white matter hyperintensities (WMHs) and microscale grey and white matter atrophy.1,2 These lesions have a detrimental role in the onset and development of cognitive impairment and dementia. Specifically, WMHs, also known as leukoaraiosis, are increased hyperintensity signals detected on T2-weighted magnetic resonance imaging (T2-MRI), fluid-attenuated inversion recovery (FLAIR), and proton density MRI sequences, which have been associated with cognitive decline, in particular of information processing speed, and may lead to executive dysfunction. The histopathology of WMHs is heterogeneous, with tissue damage ranging from slight disentanglement of the matrix to varying degrees of myelin and axonal loss. Intriguingly, WMHs are more strongly associated with previous than concurrent high blood pressure (BP).2 The above-mentioned alterations in brain structure are also associated with reductions in cognitive ability as well as with an increased risk of dementia. It is widely accepted that functional impairments are preceding the observation of macroscopic evidence of brain damage. In fact, subtle vascular brain injury develops insidiously during life, with discernible effects even in young adults.3 However, the exact neurobiological mechanisms underlying these complex processes are not fully understood. Hypertension has been associated with cognitive issues, and structural changes of some brain regions have been reported in hypertensive patients. The cerebral vasculature is a major target of damage induced by hypertension. Microvascular rarefaction, i.e. a reduction in vascular density including both capillaries and arterioles, may result from increased BP being transmitted to the microvascular bed, eventually contributing to the development of white matter lesions. Fibrinoid necrosis and lipohyalinosis (deposition of a glass-like material in the vascular wall) represent typical hypertensive microvascular lesions. Apart from clinically evident vascular events, midlife hypertension has been shown to be an important modifiable risk factor for cognitive decline, mild cognitive impairment, and vascular dementia late in life.1,3 The brain connectome refers to a comprehensive map of neural connections in the brain, including both structural and functional connections based on patterns of activity, typically derived from brain imaging data. The connectome is a complex network, with many interacting elements and multiple levels of organization, which is thought to underlie many aspects of brain function, including perception, memory, cognition, and behaviour. Machine learning has recently emerged as a powerful tool for processing and analysing large amounts of complex data. Machine learning techniques can be applied to analyse the brain connectome and extract patterns of connectivity that are relevant to different brain functions or neurological conditions. For example, graph theory-based approaches can be used to quantify the properties of the brain network, including the degree of connectivity between different regions or the efficiency of information transfer. Other machine learning techniques, such as clustering or classification algorithms, can be used to identify subnetworks or patterns of connectivity that are associated with specific cognitive or behavioural functions or that are altered in specific neurological conditions. Hence, machine learning provides a powerful tool for analysing the brain connectome and understanding the complex network of neural connections in the brain. Deep learning models can be trained to classify brain scans as belonging to healthy individuals or patients with different neurological disorders based on patterns of connectivity in the connectome. One common approach is to use graph theory, which involves representing the connectome as a network of nodes and edges and analysing its properties using mathematical tools. Machine learning algorithms can then be used to identify patterns in the network, such as hubs (nodes with many connections), modules (groups of nodes with strong connections to each other), and network motifs (repeating patterns of connectivity). Another approach is to use machine learning to predict cognitive or behavioural outcomes from the brain connectome. For example, machine learning algorithms have been used to predict individual differences in intelligence, memory performance, and personality traits based on the brain connectome.4 Machine learning can be harnessed to rapidly dissect individual differences in whole-brain functional connectivity profiles and to diagnose neurological disorders and/or predict clinical outcomes based on connectome data (Figure 1). A brain connectome signature identified via machine learning can be used to predict BP levels. Image credit: Prof. Gaetano Santulli, MD, PhD. In 2020, a neuroimaging study had established, through means of a large-scale analysis harnessing the UK Biobank, that a history of hypertension was associated with a diminished functional connectivity of the hippocampus, linked to a decrease in the prospective memory score,5 which had been previously associated with a history of hypertension.6 Specifically, increasing impairments of memory, which correlate with reduced functional connectivity of the hippocampus, were shown to be associated in a graded way with increased systolic BP.5 Fully in line with these works, in this issue of Cardiovascular Research, Jiang and colleagues elegantly implemented, within a 10-fold cross-validation framework, a data-driven, machine learning approach to accurately predict past, concurrent, and future BP levels separately based on whole-brain connectomes.7 Importantly, prediction correlations remained significant after adjusting for potential confounding factors including age, gender, smoking history, socioeconomic status, education attainment, framewise displacement, and imaging centres. The authors analysed spatial maps of 55 non-artificial independent components obtained from group independent component analysis. Each of the 55 group-level components was visualized in axial views and grouped into 9 typical large functional networks. Instead of simply focusing on current BP levels, they also examined the predictability of past and future BP levels from brain connectivity. With such a neat approach, the authors identified a brain signature capable of accurately predicting BP level from an individual’s unique connectivity profiles. Overall, predictions were robust enough to be generalizable across multiple contexts, corroborating the model’s robustness in capturing reliable brain–BP relationships rather than sample-specific idiosyncrasies, emphasizing the need for early and optimum control of BP. These observations are consistent with previous reports. For example, cerebrovascular reactivity, quantified as per cent change in blood oxygen level–dependent signal in activated voxels (mapped to a common brain template and log-transformed), has been found to be reduced in subjects with pre-hypertension or hypertension.8 Similarly, in another study led by Dr Lembo’s research group conducted in hypertensive patients, left ventricular hypertrophy (LVH) was shown to be associated with cerebral damage even in the absence of clinical symptoms;9 precisely, patients with concentric hypertrophy displayed more pronounced asymptomatic cerebrovascular damage compared with patients with eccentric hypertrophy. There are numerous challenges to using machine learning in connectomics, including issues with data quality and processing, variability in connectivity patterns across individuals and populations, and the need for more interpretable models. Nevertheless, machine learning has already led to many exciting advances in the biomedical field; for instance, a deep learning cancer risk model, Sybil (a name that has its origins in the oracles of Ancient Greece), was shown to predict lung cancers occurring 1–6 years after a screen by analysing a single low-dose chest computed tomography scan.10 In the near future, machine learning and artificial intelligence are likely to play an increasingly important role in understanding the brain connectome and its relationship to cognition and behaviour. The Santulli's Lab is supported in part by the National Institutes of Health (NIH): National Heart, Lung, and Blood Institute (NHLBI: R01-HL164772, R01-HL159062, R01-HL146691, T32-HL144456), National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK: R01-DK123259, R01-DK033823), National Center for Advancing Translational Sciences (NCATS: UL1-TR002556-06, UM1-TR004400) to G.S., by the Diabetes Action Research and Education Foundation (to G.S.), and by the Monique Weill-Caulier and Irma T. Hirschl Trusts (to G.S.). F.V. is supported in part by a postdoctoral fellowship of the American Heart Association (AHA: 22POST915561).