Development of a recommendation system and data analysis in personalized medicine: an approach towards healthy vascular ageing

Abstract

Purpose

Understanding early vascular ageing has become crucial for preventing adverse cardiovascular events. To this respect, recent AI-based risk clustering models offer early detection strategies focused on healthy populations, yet their complexity limits clinical use. This work introduces a novel recommendation system embedded in a web app to assess and mitigate early vascular ageing risk, leading patients towards improved cardiovascular health.

Methods

This system employs a methodology that calculates distances within multidimensional spaces and integrates cost functions to obtain personalized optimisation of recommendations. It also incorporates a classification system for determining the intensity levels of the clinical interventions.

Results

The recommendation system showed high efficiency in identifying and visualizing individuals at high risk of early vascular ageing among healthy patients. Additionally, the system corroborated its consistency and reliability in generating personalized recommendations among different levels of granularity, emphasizing its focus on moderate or low-intensity recommendations, which could improve patient adherence to the intervention.

Conclusion

This tool might significantly aid healthcare professionals in their daily analysis, improving the prevention and management of cardiovascular diseases.

Introduction

Early Vascular Ageing (EVA) is increasingly being recognized as a critical factor in assessing cardiovascular risk, representing a growing area of interest in preventive and clinical medicine [1]. When this condition, characterized by an accelerated deterioration of vascular health, manifests earlier than expected in the natural course of human ageing, it is significantly associated with an increased risk of adverse cardiovascular events, such as arrhythmia or stroke [2, 3]. The concept of EVA is based on a series of clinical and biochemical parameters that reflect vascular function and structure, including indicators such as arterial stiffness, pulse pressure, and markers of inflammation and metabolism [4]. Therefore, understanding EVA is crucial, not only for its ability to act as an early predictor of cardiovascular diseases but also for its potential to guide more effective preventive and therapeutic interventions. Furthermore, EVA may help elucidate the current gap between the predictions made by traditional cardiovascular risk models and the current prevalence of cardiovascular diseases, offering a more refined and comprehensive approach to cardiovascular health assessment and management [5].

The challenge in modelling EVA mostly comes from lacking a universal gold standard against which predictive models can be validated. Traditionally, studies in the literature have considered adverse cardiovascular events as the gold standard, focusing predominantly on populations already manifesting illness rather than on healthy individuals at risk [6]. However, this approach means a significant lack of preventive strategies, particularly for identifying and managing individuals who have not yet shown clinical symptoms but are on the trajectory towards cardiovascular diseases. Indeed, to achieve a sustainable, efficient, and cost-effective future in healthcare, it is necessary to transition from a disease-centric culture to one that emphasizes prevention and the promotion of health [7].

In this context, our research team has recently pioneered the development of an unsupervised predictive Artificial Intelligence (AI) model for EVA that offers new insights into cardiovascular risk in populations not yet affected by evident diseases [8]. Capable of predicting the risk of EVA in healthy individuals, our model shifts the focus from treatment to prevention. It is also worth noting this algorithm has been validated using external risk variables such as diabetes, cholesterol or smoking habits, among others, thus enhancing its reliability and applicability in a clinical context [9].

Nevertheless, the inherent complexity of AI algorithms poses a significant challenge, as their implementation and correct interpretation require a deep understanding of statistics and data analysis, thus limiting their accessibility to most healthcare professionals. This barrier highlights the imperative need for a medical application that could integrate our predictive EVA model and guide clinicians with patient-tailored interventions. Frequently, determining the most suitable type of intervention for each patient is facilitated through recommendation systems [7]. These systems are employed to deliver feedback and suggestions regarding health status and health behaviours, and encompassing areas such as lifestyle [10], nutrition [11], obesity [12], diabetes [13], or drug side effects [14], among others.

Therefore, the aim of this paper is to design and develop a recommendation system aimed at guiding clinical staff in their decision-making for the prevention of EVA. To the best of our knowledge, this is the first time this approach has been proposed in the scientific literature within this context. By introducing several clinical variables in the AI model, this work proposes a visual risk measurement system for evaluating a patient’s current health status, together with a cost function algorithm, to guide the patient from a state of EVA to a state of Health Vascular Ageing (HVA) in the fastest and safest way possible. This will be achieved through a mathematical approach that will adjust the input variables, supported by a recommendation system based on a collaborative filtering scheme [15]. The goal is for the system to offer personalized recommendations from a finite set of possible combinations, in accordance with their current health level. These recommendations should lead to specific medical interventions, related to their lifestyle, habits, or medication, depending on the intervention’s intensity in each case. Finally, the recommendations will be validated to confirm the system’s coherence and effectiveness.

The rest of the paper is organised as follows: “Cardiovascular risk prediction through machine learning techniques” section presents related studies in cardiovascular risk prediction using machine learning models. “Method” section describes the main steps towards the development and validation of the recommendation system. Next, “Results” section shows the results analysing the effectiveness of the recommendation system to improve the initial conditions of a set of patients. An experimental case study is presented in “Experimental evidence: a practical case” section, followed by a thorough discussion in “Discussion” section. Finally, the main conclusions are drawn in “Conclusion” section.

Cardiovascular risk prediction through machine learning techniques

EVA represents a relatively new and growing field of interest in cardiovascular medicine, highlighting the importance of understanding the biological processes that accelerate vascular deterioration beyond normal chronological ageing. Despite its significant implication in cardiovascular risk prediction and the early development of cardiovascular diseases, specific research on EVA and its accurate assessment remains limited. This is partly due to the complexity of the underlying mechanisms, the difficulty of directly measuring vascular changes without invasive techniques, and the need to integrate multiple risk factors and biological markers for effective risk assessment [16]. Furthermore, the criteria for defining EVA is still in the process of standardization, adding another level to this challenge for researchers and clinicians [16]. However, in the last years, with the advancement of information technologies and machine learning (ML), significant efforts have been made to address these issues, conducting studies that seek to better understand the causes and consequences of EVA and develop innovative methods for its detection and evaluation. Below, some key studies that have contributed to the assessment of vascular ageing risk are presented.

On the one hand, ML techniques for estimating risk parameters have shown promise for enhancing cardiovascular assessments. Thus, a significant contribution by Tavallali et al. [17] involved a ML method to estimate carotid-femoral pulse wave velocity (PWV) non-invasively, using routine clinical data. This method highlighted ML’s potential in improving cardiovascular risk prediction and early identification of EVA risk individuals. Also in the field of cardiovascular risk prediction, Ambale-Venkatesh et al. [18] utilised data from the Multi-Ethnic Study of Atherosclerosis (MESA) database to incorporate a wide range of variables, including biomarkers and clinical features, to enhance the prediction of cardiovascular events. They employed advanced ML techniques, such as neural networks and regression models, to analyze the data and identify patterns indicating a higher risk of cardiovascular events.

On the other hand, the majority of studies in the scientific literature that have addressed this topic focus on classifying cardiovascular risk [6]. This strategy aims to divide individuals into two distinct groups: those at high risk of developing EVA and those at low risk. Prominent research in this category includes the work of Alaa et al. [19], who used ML ensemble techniques to sift through extensive clinical and demographic datasets in search of patterns predictive of cardiovascular events. Similarly, Al’Aref et al. employed decision tree algorithms for coronary disease risk assessment using computed tomography scans and traditional risk factors [20], while Garcia-Carretero et al. utilized PWV and laboratory data to predict cardiovascular events, also through decision trees [21]. Noteworthy is the contribution by Jamthikar et al. [22] who explored how incorporating carotid ultrasound-based image phenotypes and the use of ML could enhance the assessment of vascular risk and stroke. Similarly, Kakadiaris et al. [23] used the MESA database to assess the risk of cardiovascular events using support vector machines. Interestingly, they clammed be able to outperform the clinical guidelines of the American College of Cardiology/American Heart Association by recommending fewer statin therapies while still identifying more cardiovascular events. Finally, Sorelli et al. [24] and Vallée et al. [25] investigated the potential of support vector machines and neural networks, respectively, in categorizing patients based on their cardiovascular risk, using pulse signal data and PWV.

Our study falls within this latter approach, where ML techniques are used for the assessment of the risk of EVA. However, the contribution of our research goes beyond mere risk prediction which is the common thread of all the previously cited works. What is proposed in this work extends over the utility of the risk model, moving towards the development of a novel methodology aimed at personalized medicine. This methodology is based on calculating distances to clusters characterized by optimal health patterns and, from there, generating specific and tailored recommendations for each individual. In this way, we make progress in the field of cardiovascular risk assessment, as well as propose a novel intervention strategy designed to enhance well-being and improve the patient’s health prognosis

Method

The methodology described in the following subsections is based on an EVA risk model previously developed using unsupervised learning techniques [8] and subsequently validated through external vascular risk variables [9] from the EVasCu study. This study adhered to the principles outlined in the Declaration of Helsinki and received prior approval from the Clinical Research Ethics Committee of the Cuenca Health Area (REG: 2022/PI2022). For the sake of understanding, the EVA model is briefly summarized in “Design and validation of the clustering model” section, and Fig. 1 shows a general scheme of the complete methodological approach from the conception of the EVA model to the calculation of recommendations for subjects potentially at vascular risk. The process starts with training a clustering model to ensure that healthy patients can be distinguished from non-healthy ones using the four variables proposed in the study. Next, the algorithm estimates the distance to the HVA cluster for each patient, which is next optimised to adjust the variables towards a healthy condition. To achieve this, the recommendation system issues recommendation intensity levels to adjust the variables towards a healthy status. The optimal recommendation levels are then determined, and finally, the overall system is validated.

Fig. 1.

General outline of the stages within the methodology

Design and validation of the clustering model

To build the AI model, an unsupervised K-means algorithm was employed to analyze a set of key clinical and biochemical parameters [8]. Specifically, the algorithm was trained using a set of 390 healthy subjects, where four variables were measured for each subject: Pulse Pressure (PP), Pulse Wave Velocity (PWV), Glycated Hemoglobin (HbA1c), and Advanced Glycation End Products (AGEs). The choice of these variables reflects a process grounded in prior evidence linking each variable independently to cardiovascular risk and mortality. For instance, PP indicates arterial stiffness and loss of vascular elasticity, while PWV is an established marker of arterial stiffness. AGEs are involved in the vascular ageing process and the pathogenesis of cardiovascular diseases. Conversely, HbA1c provides a measure of long-term glycemic control and has been linked to cardiovascular risk in both diabetic and non-diabetic populations. The four variables were shown to conform to a construct using a factorial confirmatory analysis, and each proved to be a risk index of cardiovascular disease in isolation [8]. After recollection, data was normalized using the z-score to ensure information comparability and consistency. Then, the K-means algorithm was utilized to stratify the data into distinct groups. Different strategies were used to determine the optimal number of conglomerates, resulting in the identification of two primary risk clusters labelled HVA and EVA conglomerates [8].

Given the multivariable approach of the resulting model, which includes the aforementioned four input variables, dimensionality reduction was performed using Principal Component Analysis (PCA) to enable 2-D visualization of patient dispersion across the two conglomerates, as depicted in Fig. 2. PCA simplifies the complexity of high-dimensional data spaces while simultaneously preserving as much of the original information as possible. This is achieved by identifying the directions (principal components) in which the data varies the most and projecting the original data onto a lower-dimensional space formed by these directions [26].

Fig. 2.

Graphical representation of patient dispersion using a 2-D Principal Component Transformation, across the two risk clusters: EVA and HVA

Finally, the model’s validation was conducted by comparing the identified clusters with external variables, which were categorised into sociodemographic, lifestyle, adiposity, lipid, vascular, and inflammatory factors [9]. This external validation proved crucial for verifying the robustness and clinical relevance of the model. Indeed, it was observed that the risk groups identified by the algorithm significantly correlated with aforementioned risk factors.

Estimation of patient’s distance to the healthy vascular age cluster

Once the risk stratification was computed, the health state of each patient was evaluated by computing the distance between each one of its variables and its corresponding HVA cluster centroid using the Euclidean Distance [27]. Mathematically, the distance d between a multidimensional sample point formed by the four input variables $v=(v_1,v_2,\dots ,v_n)$ and the HVA centroid $c=(c_1,c_2,\dots ,c_n)$ can be computed as:

$$\begin{array}{c}\hfill d(v,c)=\sqrt{\sum _{i=1}^n{(v_i-c_i)}^2}\end{array}$$ 1

A radial chart was considered to represent these multidimensional distances of the four key variables (HbA1c, PP, AGEs and PWV) to their respective HVA centroids more intuitively, forming a polygon when connected. In this respect, the area and shape of this polygon provide a visual representation of the patient’s health state at a glance. A larger area or extended polygonal shape indicates a greater deviation from the healthy state. This method allows for an immediate visual assessment of a patient’s health status in relation to the ideal HVA condition, with each axis offering specific insights into different health parameters. Figure 3 displays the graphical representation of two random patients’ clinical data using the radial chart method. Specifically, Fig. 3a represents a patient falling into the EVA group, while Fig. 3b depicts a patient categorized within the HVA group. It is worth noting that negative values on the charts indicate that the patient’s measurements are healthier than the centroid of the HVA cluster, such that these values are plotted towards the centre of the chart, thus reducing the polygon’s area.

Fig. 3.

Visual representation of the health status of two patients in EVA (a) and HVA (b), considering the multidimensional distance to the HVA centroid

Optimisation pathways for patient variable adjustment towards healthy vascular ageing

The aim of this optimisation consists of estimating the magnitude of changes required in each input variable so that a patient potentially at risk of EVA can understand how their initial variables should change to move to the healthy HVA cluster in an optimal way. To achieve this, a cost function f(v) was implemented to evaluate potential pathways from EVA to HVA. f(v) combines the Euclidean distance to the centroid of HVA with a penalty term $\lambda$ that accounts for the magnitude of changes in the clinical variables, mathematically defined as:

$$\begin{array}{c}\hfill f(v)=\sum _{i=1}^n{(v_i-c_i)}^2+\lambda \sum _{i=1}^n{(v_i-v{o}_i)}^2\end{array}$$ 2

where $v_i$ represents the vector of the patient’s variables, $c_i$ is the vector of the HVA centroid, $v{o}_i$ are the original values of the patient’s variables, and $\lambda$ is the factor that regulates the penalty for changes in the variables. In this respect, the optimisation process is tailored to each patient’s unique clinical situation by iteratively adjusting the value of $\lambda$. Thus, this approach begins with a high $\lambda$ ($\lambda =10$) value to ensure initial conservative changes, which is then gradually decreased (0.1 steps in each iteration) to allow for more significant adjustments as needed. The dynamic tuning of $\lambda$ modulates the penalty for changes in clinical variables, ensuring that the optimisation does not enforce uniform changes but rather adapts to each patient’s individual circumstances. In each iteration, new values of variables ($v_i$) are computed by using the Broyden, Fletcher, Goldfarb, and Shanno (BFGS) optimization algorithm. The BFGS method, which is a sophisticated quasi-Newton technique, leverages only the first derivatives or gradient of the cost function to efficiently find its minimum [28]. Interestingly, BFGS optimizes by approximating the inverse of the Hessian matrix, thereby guiding each iteration towards the optimal solution with remarkable precision. This iterative process continues until a transition from the EVA to the HVA cluster is observed, signifying a shift toward a healthier state as defined by the model. In the event that a patient is already in the HVA cluster, the optimisation process will exit right after the first iteration, where the value of lambda is large, causing minimal changes to the initial variables. Indeed, dynamically decreasing $\lambda$ in each iteration ensures the balance between moving closer to the optimal health state and avoiding excessive and potentially unsafe changes in clinical variables is maintained throughout the optimisation.

Design of the recommendation system

A quantitative approach was adopted to construct a patient-tailored recommendation system, computing the changes in distance from the initial to the optimized state derived from the cost function optimisation defined in “Optimisation pathways for patient variable adjustment towards healthy vascular ageing” section. The methodology entailed the computation of the 95th percentile for the change in distance of each clinical variable across the patient cohort. This percentile serves as a benchmark, representing a change threshold that encapsulates the significant yet realistic adjustments for most of the population while effectively excluding outliers. Thus, the percentage change for each variable $\mathrm{\Delta }\%(v)$ can be calculated as:

$$\begin{array}{c}\hfill \mathrm{\Delta }\%(v)=\left\{\begin{array}{cc}\left(\frac{D_{\text{initial}}(v)-D_{\text{optimized}}(v)}{P_{95}(v)}\right)\times 100,\hfill & \text{if}{D}_{\text{initial}}(v)>0\hfill \\ 0,\hfill & \text{if}{D}_{\text{initial}}(v)\le 0\hfill \end{array}\right)\end{array}$$ 3

where $D_{\text{initial}}(v)$ represents the initial distance of variable v from the HVA centroid for a patient, $D_{\text{optimized}}(v)$ indicates the optimized distance after the application of the cost function optimisation and $P_{95}(v)$ represents the 95th percentile of change for that variable across the patient cohort. Finally, based on $\mathrm{\Delta }\%(v)$, outcomes were stratified into different levels of recommendation granularity that corresponded to the intensity of the proposed interventions.

In those instances where the percentage change $\mathrm{\Delta }\%(v)$ for any variable was less or equal to zero, indicating that the initial distance for that variable was already healthier than the HVA centroid, the system did not issue any recommendation. This ensures that recommendations are reserved only for those instances where a patient’s health status can be realistically improved, thereby avoiding unnecessary or redundant interventions for patients who are already at an optimal health state in specific variables.

Determination of optimal recommendation levels

Various configurations were assessed to discern the optimal number of recommendation intensity levels for the system, decrementally adjusting the number of recommendation levels from 6 (more granularity) to 3 (less granularity). The configurations were systematically analyzed based on different $\mathrm{\Delta }\%(v)$, ranging from 0 to 100, as detailed in Table 1. The analysis also incorporated a specific no-recommendation level, which was particularly designated for instances where the initial assessment metric, denoted as $D_{\mathit{initial}}$, was less than zero. As can be observed, the greater the $\mathrm{\Delta }\%(v)$, the higher the associated intensity of the recommendation, and the greater the number of intensity levels, the more detailed or granular the recommendations.

Table 1.

$\mathrm{\Delta }\%(v)$ ranges and intensity of recommendation for different levels of granularity

$\mathrm{\Delta }\%(v)$	6 levels	5 levels	4 levels	3 levels
$\le 0$	No rec.	No rec.	No rec.	No rec.
[0, 20)	Mild	–	–	–
[20, 40)	Moderate	–	–	–
[40, 60)	Normal	–	–	–
[60, 80)	Intense	–	–	–
[80, 100]	Very intense	–	–	–
[0, 25)	–	Moderate	–	–
[25, 50)	–	Normal	–	Normal
[50, 75)	–	Intense	–	–
[75, 100]	–	Very intense	–	Intense
[0, 33)	–	–	Mild	–
[33, 66)	–	–	Normal	–
[66, 100]	–	–	Intense	–

It is worth noting that, although the recommendation system operates using a set of predefined recommendations, it is capable of identifying the most suitable combination of recommendations for each individual user. This allows that, even though the recommendations are fixed, the selection of which ones are presented to each user is highly personalized. For instance, for 5 levels of intensity, there would be a total of 625 possible unique combinations. Therefore, the system’s adaptability lies in its ability to combine these fixed recommendations in a way that aligns optimally with the specific health needs of the user, thus offering a personalized guide to each user.

Validation of the recommendation system

With the aim of validating the recommendation system, an extensive examination was conducted to discern the distribution and categorization of patients within the context of the generated unique recommendations. Thus, depending on $\mathrm{\Delta }\%(v)$, the number of unique recommendations provided for a cohort of 390 patients was computed. This analytical approach was based on the premise that a robust recommendation system should inherently exhibit the capability to offer congruent recommendations for patients bearing similar clinical variables, such that the system should maintain consistency across patient recommendations.

To corroborate this consistency, 2-dimensional scatter plot visualizations were used to elucidate the clustering patterns of patients according to the received recommendations. This visualization technique assisted in revealing the inherent grouping of patients, allowing for an intuitive understanding of the distribution of recommendations across the patient population. Given the multidimensional behaviour of this approach, a PCA analysis was performed on the data to allow the visualization, similar to the one performed in “Design and validation of the clustering model” section. This step transformed the complexity of the data into its two principal components, thus reducing dimensionality while preserving the intrinsic structure of the data. The PCA scatter plots then provided a concise visual representation of how patients were systematically grouped, reflecting how the recommendation system aligned patients with similar clinical profiles into coherent clusters.

Furthermore, three statistical indices were calculated to assist in the decision process regarding the optimal number of recommendation levels to provide: the Calinski-Harabasz index, the Davies-Bouldin index, and the Silhouette index. The Calinski-Harabasz index measures the variance ratio between clusters to the variance within clusters, serving as an indicator of cluster tightness and separation [29]. Higher values of this index suggest that the clusters are well-delineated and compact, implying a clear distinction between different patient groupings. On the other hand, the Davies-Bouldin index reflects the average similarity between each cluster and its most similar cluster [30]. Lower values are desirable in this context, as they indicate that the clusters are well-separated and distinct, with minimal overlap between them. Finally, the Silhouette index calculates how similar an object is to its own cluster compared to other clusters, offering insight into the cohesion and separation of the clusters [31]. A higher Silhouette score indicates that patients are well-matched to their own cluster and poorly matched to neighbouring clusters, thus reinforcing the appropriateness of the assigned clusters. In the context of this recommendation system, a positive Silhouette index would signify that patients with similar clinical profiles are grouped together consistently, affirming the system’s capability to provide coherent and targeted recommendations.

Results

In the following subsections, the results obtained after applying the recommendation system to a cohort of 390 healthy patients are presented, analyzing the system’s effectiveness in improving the initial conditions of some patients and evaluating the optimal intensity levels we should consider for the system.

Visual and statistical analysis

Figure 4 presents the visual representation of the optimisation method applied to two random patients, one from the EVA and the other from the HVA group. The radial charts depict the distance of each patient’s clinical variables from the HVA centroid before and after optimisation. Specifically, Fig. 4b shows the chart for the HVA patient with a green polygon, representing the original state of the patient’s variables. The green dotted lines represent the optimized state after the execution of the cost function, indicating, in this case, minor adjustments, as the patient’s variables were already close to the ideal state. Similarly, Fig. 4a shows the chart for the EVA patient coloured with a red polygon, indicating a greater initial distance from the HVA centroid. After optimisation, illustrated by the green dotted lines, the polygon’s area decreased significantly, indicating that the patient’s variables moved closer to the HVA profile. In this respect, while all variables have experienced some degree of optimisation, the most altered variables are AGEs and HbA1c, which were initially the most compromised ones. This indicates that the optimisation pathway has effectively suggested changes that could potentially improve the patient’s vascular health status, thus demonstrating the practical utility of the cost function in guiding clinical decisions towards achieving a healthier vascular ageing state for patients.

Fig. 4.

Visual representation of the health status of two patients in a EVA and b HVA, along with the distances optimized by the cost function, for health improvement

Additionally, Table 2 provides a detailed comparison between the original and optimized distances for the input variables. As the Shapiro–Wilks test corroborated the lack of normality of the data distribution, the table presents the median and range (minimum and maximum values) for both the original and optimized scenarios. Notably, a decrease in the median values is observed post-optimisation for all variables, indicating a shift towards more centralized values. Furthermore, the range of values for each variable narrows after optimisation, reflecting a reduction in variability and a tighter clustering of data points. The p values, derived from the Wilcoxon test for paired samples, were remarkable (p $\le$ 0.005), affirming the statistical significance of the differences observed before and after the optimisation of distances for all of the variables.

Table 2.

Median, range, and p value of the original and optimized distances for the input variables

Variable	Original state		Optimized state		p value
	Median	Range	Median	Range
PP	0.0552	(− 4.38, 5.77)	0.0460	(− 2.75, 2.35)	$3.53\times {10}^{-4}$
HbA1c	0.3823	(− 3.83, 4.69)	0.2932	(− 3.19, 2.13)	$4.60\times {10}^{-21}$
PWV	0.5435	(− 1.01, 4.54)	0.3913	(− 0.84, 1.62)	$2.86\times {10}^{-23}$
AGEs	0.4395	(− 1.52, 5.95)	0.2642	(− 1.27, 2.24)	$3.28\times {10}^{-25}$

Finally, Fig. 5 shows a comprehensive visual representation of the intensity levels of recommendations for the four variables. In this example, each variable is colour-coded distinctly across five recommendation levels (see “Validation analysis and determination of optimal recommendation levels” section), ranging from the least intense (light green) to the most intense (red). Visually, the distribution of recommendations across these variables appears remarkably balanced, indicating a detailed and equitable approach to the intensity of recommendations. In this respect, while each variable exhibits a unique distribution pattern, no single intensity level dominates for any variable. Hence, Chi-squared test revealed no significant differences in the distribution of these five recommendation levels among these variables, scoring a Chi-squared statistic and p value of 8.04 and 0.530, respectively. This lack of significant disparity among recommendation levels suggests that the system prevents from exhibiting a disproportionate bias towards any particular variable when issuing recommendations.

Fig. 5.

Patients with the same intensity of recommendation per variable considering five levels

Validation analysis and determination of optimal recommendation levels

Figure 6 presents different PCA scatter plots considering the four different approaches proposed in Table 1 (see “Determination of optimal recommendation levels” section). Each plot features the distribution of recommendations common to ten or more users (see Fig. 6) represented with different colours and symbols. As the number of recommendation levels decreases from six (Fig. 6a) to three (Fig. 6d), a notable increase in the number of clusters containing more than ten patients is observed, as depicted by the expanding legends. This phenomenon reflects the inherent consequence of reduced granularity: as fewer recommendation levels are available, patients naturally merge into broader categories, thus elevating the count of larger, more populated clusters. Conversely, when the recommendation levels are augmented, more finely divided clusters emerge, signifying a tailored approach to patient categorization.

Fig. 6.

Groups with ten (or more) patients sharing the same set of recommendations, using a 6 levels, b 5 levels, c 4 levels, and d 3 levels of granularity

It is important to note that, despite this increased detail in patient segmentation, the scatter plots reveal a visual consistency in the proximity of patients within the clusters. Thus, patients with comparable initial four variables are consistently grouped in close association with one another, regardless of the number of recommendation levels. This persistent pattern, visible across all levels of granularity, corroborates the recommendation system’s validity and robustness, as it confirms that the recommendation system is proficient in forming coherent clusters, highlighting that patients with similar characteristics receive similar recommendations.

Table 3 provides a structured overview of the clustering quality assessments conducted at various levels of recommendation granularity. It shows the number of unique sets of recommendations, the number of sets of recommendations with more than ten patients, and the values for three key metrics: the Calinski–Harabasz index, the Davies–Bouldin index, and the Silhouette score. The table showcases a trend where the Calinski–Harabasz index generally increases as the number of recommendation levels decreases, suggesting a trend toward more well-defined clusters. Starting with six levels of recommendations, where the Calinski–Harabasz score is 36.12, this index peaks at 48.47 when the recommendation levels are reduced to three, indicating the most distinct cluster separation at the lowest granularity level examined. Correspondingly, clusters with more than ten patients increase from 9 to 14 as the recommendation levels are changed from six to three, highlighting a tendency for patients to aggregate into larger, more generalized groupings as the number of unique recommendations decreases.

Table 3.

Analysis of patient dispersion with different granularity

Granularity	Unike sets of recom.	Sets with $\ge$ 10 patients	Calinski–Harabasz	Davies–Bouldin	Silhouette index
6	144	9	36.12	1.37	0.169
5	122	11	39.12	1.42	0.152
4	94	13	38.22	1.48	0.137
3	46	14	48.47	1.42	0.137

Higher values of Calinski–Harabasz index mean that the clusters are well-delineated and compact. Lower values of Davies–Bouldin indicate that the clusters are well-separated and distinct. A higher Silhouette index indicates that patients are well-matched to their own cluster and poorly matched to neighbouring clusters

Conversely, the Silhouette score, starting at 0.169 for six recommendation levels, shows a downward trend as the levels decrease, settling at 0.137 for both four and three levels of recommendations. This suggests that while clusters become more distinct, the tightness of the clusters may not proportionally increase with fewer recommendation levels. Finally, the Davies-Bouldin index fluctuates slightly but does not exhibit a consistent trend, varying between 1.37 and 1.48 across the different levels, which indicates variability in the average similarity between clusters. These changes reflect the complex balance between cluster separation, cohesion, and the granularity of the recommendation system.

Experimental evidence: a practical case

This section is conceived as a case study intended to illustrate how the implemented methodology is applied to individual samples, showing the application of the different phases and what recommendations the system proposes. For this practical case, three representative samples have been selected that cover a broad spectrum of situations related to EVA: a first patient with an advanced degree of EVA, another one with a milder manifestation of EVA, and a third case of an individual exhibiting HVA. This selection allows for a comprehensive comparison, demonstrating how the methodology adapts and responds to different degrees of risk and vascular conditions, thus underscoring its potential to significantly contribute to the personalized prevention and management of cardiovascular health. Table 4 displays the initial conditions of these three study patients chosen for this practical case, concerning their input variables.

Table 4.

Initial conditions of study patients

Patient type	PP (mmHg)	HbA1c (%)	PWV (m/s)	AGEs (arbitrary units)
Patient 1	57	5.53	7.0	2.4
Patient 2	43	5.17	7.7	2.2
Patient 3	39	5.04	6.6	1.9

After z-score normalization, the Euclidean distance for each variable is calculated relative to the healthy (HVA) centroid, using the coordinates computed from the previously trained clustering model. With this information, the patient’s belonging to the corresponding cluster is determined. Next, the optimized distances for each patient and variable are calculated following the methodological approach described in “Optimisation pathways for patient variable adjustment towards healthy vascular ageing” section. Thus, the optimization algorithm determines the shortest path required to transition from a situation of risk to a healthier condition. If the patient is already in a healthy situation, the $\lambda$ term will render the optimization practically negligible.

The Radial Fig. 7 shows the distances to the HVA cluster for each patient and variable, as well as the health status evaluation. Here, it can be observed that the algorithm has determined that both, patient 1 and patient 2, are in a state of vascular risk (red colour), whereas patient 3 is in a state of HVA (green colour). The figure also depicts the optimal pathway (green dotted lines) each patient should follow to transition from a risk to a health state in the most optimal manner. It is interesting to note that while for patients 1 and 2 the optimization algorithm proposes alternative health pathways to the initial state, the optimization for patient 3 is practically negligible.

Fig. 7.

Current vascular situation and its optimized version based on distances, for the patient a in an advanced EVA situation (patient 1), b in a moderate EVA situation (patient 2), and c in an HVA situation (patient 3)

Once the current and optimized vascular situations have been calculated for each patient and variable, the system calculates the percentage of change ($\mathrm{\Delta }\%$) according to Eq. (3), as detailed in “Design of the recommendation system” section. Table 5 displays these $\mathrm{\Delta }\%$ for each patient and variable under study. It can be observed that all $\mathrm{\Delta }\%$ for patient 1 are significant, especially for the PP, HbA1c, and AGEs variables, whereas for patient 2, only the PWV and AGEs variables show significant percentage changes. Finally, the changes proposed for patient 3 are minimal. Only the AGEs variable appears to have a slight proposed change.

Table 5.

Percentage of change $\mathrm{\Delta }\%$ estimated for each patient according to Eq. (3)

$\mathrm{\Delta }\%$	xPP (%)	HbA1c (%)	PWV (%)	AGEs (%)
Patient 1	90.30	75.73	47.69	82.35
Patient 2	0	15.98	50.17	54.91
Patient 3	0	2.39	0.30	12.13

Finally, these percentage changes are categorized into levels of intervention intensity, using the ranges proposed in Table 1 described in “Design of the recommendation system” section. Considering 6 levels of intervention, the recommendation results for these 3 patients are as shown in Table 6.

Table 6.

Intensity of recommendations categorised for each patient according to Table 1, and using 6 levels of intensities

Intensity	xPP	HbA1c	PWV	AGEs
Patient 1	Very intense	Intense	Normal	Very intense
Patient 2	No recommendation	Moderate	Normal	Normal
Patient 3	No recommendation	Mild	Mild	Mild

Discussion

The integration of AI in medical applications can lead to faster and more accurate diagnoses, optimize treatments, and improve patient outcomes by providing physicians with powerful, data-driven tools to inform and support their clinical decisions [32]. However, there is a reluctance among physicians to trust and adopt something they do not fully understand and consider a ’black box’ [33]. For this reason, authors in the literature have highlighted the need to develop user-friendly tools for physicians that allow them to focus on clinical decision-making rather than the technical details of AI [34]. In this work, the proposed AI system has been programmed into a computer application, as can be seen in Fig. 8, following some key design principles such as intuitiveness, ease of use, and interactivity [35].

Fig. 8.

Web app prototype for health status visualization and recommendation proposal

Our approach categorises the intensity of each intervention depending on the distance between the original and optimized health status, resulting in the recommendation system showing a clear predominance of moderate or medium interventions for the treatment of all input characteristics. This approach may be indicative of a carefully calibrated approach prioritizing incremental or moderated health improvements, which have been reported to be more effective and sustainable for long-term adherence compared to aggressive interventions [36, 37]. Indeed, a conservative strategy is employed by using a penalty term in the cost function, where the system deliberately opts for less drastic interventions and aims to minimize potential risks or disruptions in patients’ lives.

Furthermore, the granularity level of the recommendations was also evaluated, resulting in small differences between the clusters of patients receiving the same recommendation. In this case, the optimum approach is to propose a balance between cluster separation, cohesion, and the granularity of the recommendation system. Analysing the result, this threshold could be set at five recommendation levels, where the system articulates 122 unique recommendations, providing a substantial degree of personalisation without delving into excessive granularity that could potentially obfuscate clinical decision-making. Moreover, the relatively high Calinski-Harabasz index at this level indicates well-defined and separated clusters. Similarly, the Davies-Bouldin index stands competitively, indicating that the spread within clusters is not overly pronounced relative to their separation. Besides, the Silhouette index for five levels remains reasonably indicative of a coherent clustering structure, where patients are closer to others within their cluster than to those in neighbouring clusters. Finally, from a clinical point of view, this granularity allows for addressing diverse patient profiles without introducing overwhelming complexity.

It is also worth commenting on the rationale behind the use of z-score normalization, especially in the context of models that rely on data distances, such as the K-means clustering algorithm. In this respect, z-score normalization, by standardizing data based on their mean and standard deviation, mitigates the potential distortion caused by outliers, which is particularly important for distance-based models. Indeed, K-means model benefits from such normalization as it guarantees that each feature equally influences the algorithm’s clustering process, preventing the possible overemphasis of certain features due to outliers. This normalization is also essential for PCA, as it relies on the variance of the variables to determine the principal components. Without proper normalization, variables with higher variance could dominate the results, biasing the principal components towards these characteristics and potentially overlooking the importance of other variables with lower variance but equally significant [38]. Nevertheless, it is crucial to acknowledge that when contemplating the visual interpretation of patient clustering within the same recommendation clusters, the representation might not encapsulate the full complexity of the underlying relationships. This restriction arises from the inherent limitations of dimensionality reduction techniques such as PCA, which can sometimes obscure the subtle details present in the original multi-dimensional space [39].

Interestingly, most existing studies that utilize cardiovascular risk models adopt a supervised training approach, where the output variable is often the occurrence of a cardiovascular event [19] or the presence of indicators that lead to such events, like the occurrence of atheromas [22]. While these approaches are invaluable for identifying high-risk individuals within populations already recognized for their preexisting conditions or significant risk factors [6], systems based on unsupervised learning could offer a promising advantage, as they allow for the exploration and discovery of subgroups of individuals who may be on the path to accelerated vascular ageing without yet showing clear clinical signs [40].

Moreover, despite the existence of a variety of algorithms evaluating cardiovascular risk from different etiologies, few studies have translated this knowledge for general population use. Noteworthy is the web-based applications resulted from studies published by McClelland et al. on the calculation of arterial age [41] and coronary risk [42], thanks to the MESA study [43]. However, although these apps are publicly available, there are no intervention recommendations that supply the user with personalised health information. This gap highlights a crucial need for developing recommendation tools to offer actionable guidance that can empower individuals to proactively manage their cardiovascular health, as in Cai et al. [44]. This work proposes a collaborative filtering scheme where new recommendations are generated using the outcomes of similar users. To this respect, a variety of works utilising the same approach can be found, aimed at different health areas, such as physical activity recommendations [45], sleep routines recommendations [46], or the search for suitable hospitals and doctors [47], among others.

Nevertheless, only a few studies focus on recommendation systems for cardiovascular risk using ML techniques, and they predominantly rely on supervised learning approaches. This fundamental difference in methodology makes direct comparisons between these systems and our work challenging. For instance, Mustaqeem et al. [48] proposed an improved collaborative filtering technique based on clustering, applied to a supervised set of data for four different types of cardiovascular diseases. The generated recommendations were chosen from a list of previously labeled categories, based on the similarity of previous patients. To this respect, although our approach can also recommend similar interventions to patients with comparable starting situations, the level of customization for each patient is unique, and the recommendations they receive are calculated independently for that patient’s unique starting situation, not by similarity with previous cases. Thus, since the health situation of each patient is unique, the customization that receives will depend only on the granularity of the defined interventions, which can be tuned with different levels of importance, not on the similarity with previous patients.

Furthermore, Jabeen et al. [49] proposed a fog-based model to classify patients into one of eight types of cardiovascular diseases, using a community-based filtering algorithm to generate personalized lifestyle and dietary recommendations. Conversely, Mehmood et al. [50] developed a hybrid recommendation system that similarly provides personalized medical advice on diet and exercise after processing clinical tests and diagnoses. Both teams validated their models with labeled data collected from cardiologists and from the coronary care unit of a prominent hospital, respectively.

In contrast to these approaches, our work focuses on a healthy population that has not been previously diagnosed with cardiovascular diseases. This specificity limits our ability to directly assess the performance of the recommendations, yet underscores our emphasis on primary prevention over secondary intervention. Moreover, the systems developed by Jabeen et al. and Mehmood et al. are also based on a similar premise, relying on recommendations derived from patients previously labeled in a similar manner. However, unlike these approaches, our model aims to personalize interventions by considering the unique vascular situation of each individual, without depending solely on historical patient data. It is also worth noting that our work is not centered on patient recommendations but rather aims to guide clinical staff in evaluating and managing vascular risk, supporting clinicians in making more informed decisions. For instance, a clinician might hesitate to medicate a person with elevated glycated hemoglobin in a prediabetic state. However, if the patient were assessed to be at vascular risk according to our proposed model, the clinician could have the additional information necessary to prescribe medication and prevent possible cardiovascular events.

Conclusion

In recent years, recommendation systems have been widely used in various application areas, such as eCommerce, entertainment, or finances, among others. Thanks to advances in AI and data analysis, recommendation systems are also being used intensively in personalized medicine, providing tailored recommendations for treatments, medications, and other health services. In this work, a recommendation system based on collaborative filtering is proposed for the first time to assess the risk of suffering from EVA, where the recommendations provided to patients are based on an unsupervised AI model trained using the K-means algorithm.

The system features an optimisation approach with an associated cost function to determine the intensity of the recommendations, which allows a given patient to transition from a state of EVA to a more optimal health state. These methodological techniques have previously been used in personalized medicine to improve medication efficacy, safety, and cost reduction.

The effectiveness of the system was assessed using information from 390 healthy patients (not diagnosed with EVA), and the system issued recommendations, considering the four variables, to improve the overall condition. The optimization process results in a general improvement, modifying the variables proportionally to their degree of misalignment. This suggests that the optimization process has effectively proposed alterations that may enhance the vascular health status of the patient, thereby illustrating the practical usefulness of the cost function in directing clinical decisions toward attaining a healthier vascular ageing state for patients.

Future directions

Since the app developed for this work is a proof of concept of the integration of the AI technique (k-means) and the recommendation system, in the future, a user-centred app should be developed. Thus, it should cover issues such as research and active user participation from the early stages of development [51] or the evaluation and ongoing support to improve user adherence to the app [52]. Another future improvement is to translate the recommendation intensity levels for addressing interventions on each of the model’s variables into specific actions. To this end, a network meta-analysis study will be conducted, gathering scientific evidence on the most effective treatments that can reduce risk at different intensity levels. Finally, additional clustering and normalization methodologies should be analyzed, to assess the algorithm suitability to model this approach.

Author contributions

Conceptualization, IC-R and AS-L; methodology, AM-R and JCC; software, AM-R; validation, IO-L, AD-L and IC-R; formal analysis, AM-R and IC-R; investigation, AM-R; resources, IC-R and AS-L; data curation, IC-R and IO-L; writing—original draft preparation, AM-R and JCC; writing—review and editing, AM-R; visualization, AM-R, AS-L and IO-L; supervision, IC-R and AS-L. All authors have read and agreed to the published version of the manuscript.

Funding

This research has received financial support from several sources: the 42nd edition of the nursing award of the University Pontificia Comillas and Escuela de Enfermeria y Fisioterapia San Juan de Dios; Carvascare Research Group from the Universidad de Castilla-La Mancha (2023-GRIN-34459); PID2021-128525OB-I00 and TED2021-130935B-I00, funded by the Spanish Government in conjunction with the European Regional Development Fund (EU) jointly with SBPLY/21/180501/000186, provided by the Junta de Comunidades de Castilla-La Mancha, Spain and the European Regional Development Fund (EU).

Data availability

The database used in this study is available to all users within the Mendeley Data repository on demmand, with the following 10.17632/72xhjkvjk2.1.

Declarations

Conflict of interest

The authors declares no conflict of interest.

Ethical approval and consent to participate

The research protocol of this study was approved by the Clinical Research Ethics Committee of the Cuenca Health Area (REG: 2022/PI2022). Written informed consent to participate was obtained from all subjects included in the study.

Consent for publication

Written informed consent for publication was obtained from all subjects included in the study.