Risk factors for cardiovascular disease in patients with metabolic-associated fatty liver disease: a machine learning approach

Anthropometric parameters, including height (meters) and weight (kilograms), as well as waist and hip circumference were measured (meters) by standard methods, and the body mass index (BMI) was calculated as weight/height² (kg/m²). Additionally, the waist-to-hip ratio (WHR) was obtained. The obesity was diagnosed when BMI ≥ 30 whereas the overweight was diagnosed when BMI ≥ 25 but < 30. The patients were considered to have T2DM based on a known history of this disease. Blood pressure was measured three times after 5 min of rest in a sitting position, at least 2 min apart, and the mean blood pressure of the three measurements was calculated. Arterial hypertension was defined as a systolic blood pressure ≥ 140 mmHg and/or a diastolic blood pressure ≥ 90 mmHg or previous treatment with antihypertensive medications. Hypercholesterolemia was recognized when a patient had this diagnosis present in the documented medical history and/or there was newly recognized plasma high density lipoprotein cholesterol (HDL-C) < 1.0 mmol/l for men and < 1.3 mmol/l for women and/or patient was on statin therapy. Hypertriglyceridemia was recognized when a patient had this diagnosis present in the documented medical history and/or there was newly recognized plasma triglyceride ≥ 1.7 mmol/l and/or patient was on fibrate therapy. The history of all concomitant diseases was obtained from the patient and confirmed on the basis of documented medical data. No self-reported diseases without medically confirmed diagnosis were recorded.

The patients were diagnosed with MAFLD [26] if there was an evidence of steatosis acquired by the hepatic ultrasonography and presence of one of the following criteria: T2DM, overweight or obesity defined as BMI greater than or equal to 25 kg/m² or at least two metabolic risk abnormalities, i.e., waist circumference ≥ 102 cm in men and ≥ 88 cm in women, blood pressure ≥ 130/85 mmHg or specific drug treatment, prediabetes, plasma triglycerides ≥ 1.7 mmol/l or specific drug treatment, plasma HDL-C < 1.0 mmol/l for men and < 1.3 mmol/l for women or specific drug treatment,  Homeostatic Model Assessment of Insulin Resistance (HOMA- IR) ≥ 2.5, serum C-reactive protein level > 2 mg/l.

The ultrasound examination of the carotid arteries was performed using a high-resolution Doppler ultrasound with double imaging and color coding of the flow (Color Doppler Duplex, CDD) exploiting the Esaote MyLab60 ultrasound equipment by the same certified neurologist. The examinations were done in the supine position, without any additional preparation, with a linear ultrasound head emitting an ultrasonic wave with a variable frequency of 4 MHz to 15 MHz. In the 2D presentation, the common carotid artery (CCA), the separation (bifurcation) of the common carotid artery into the internal carotid artery and the external and internal carotid artery (ICA), alongside the external carotid artery (ECA) were determined. Measurement of the intima-media complex (KIM) thickness and the assessment of atherosclerotic lesions in the carotid arteries were performed.

For these measurements, piezoelectric mechanical transducers in the cervical, femoral and radial areas were used (Complior, Alam Medical, France). The velocity of the carotid-femoral pulse wave (cfPWV) was used to assess the stiffness of the central artery and the velocity of the carotid-radial pulse wave (crPWV) was used to assess the stiffness of the peripheral arteries. With the Seca mod. 207 height meter, the right carotid-femoral and carotid-radial distances were measured. Blood pressure was measured in the supine position after at least 5 min of rest with the Microlife BP A1 sphygmomanometer immediately prior to the PWV assessment, and the mean of the three measurements on both arms was calculated and recorded. Derivative variables such as the central blood pressure, central pulse pressure and the gain index were analyzed, calculated by the integrated software on the basis of the carotid pulse waveform.

Hemoglobin A1c (HbA1c) was determined using a high-performance liquid chromatography (HPLC) method, and the results were expressed in the National Glycohemoglobin Standardization Program/Diabetes Control and Complications trial units [27]. Cholesterol and triglycerides were measured using the enzymatic methods, with the HDL-C measured after precipitation of the very low-density lipoprotein cholesterol (VLDL-C). The concentration of the low density lipoprotein cholesterol (LDL-C) was calculated using the Friedewald formula [28]. Serum creatinine was measured by means of the Jaffe’s method. The estimated glomerular filtration rate (eGFR) per 1.73 m² was calculated according to the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formula [29]. Blood cell morphology to obtain the platelet count (PLT) was determined using the fluorescent flow cytometry with the Sysmex XN-1000 (Sysmex) hematology analyzer [30]. Serum C reactive protein concentration were measured by a latex particle-enhanced turbidimetric immunoassay [31]. Alanine and aspartate aminotransferase activities in serum were assayed by the kinetic method according to the IFCC reference procedure. Analyzes of serum C reactive protein, alanine and aspartate aminotransferase were carried out on the Cobas 6000 analyzer, c 501 module (Roche). Fasting glucose was assessed using the enzymatic method with the Cobas 6000 hexokinase analyzer, c501 module (Roche). Insulin concentration was measured by electrochemiluminescence using the Cobas 6000 analyzer (module E601) [32].

To automatically identify the patients with a high risk of overt CVD, we investigated biochemical (8 parameters), demographical [2], clinical [17], carotid ultrasound [10], arterial stiffness-related [2] and steatosis stage in elastography [3] parameters (42 parameters in total)—the parameters are summarized in Table 1. To handle the missing data, we imputed the mean values for each parameter—the percentage of patients for whom the parameter was missing never exceeded 5% of all patients (mean: 0.68%). Afterwards, the parameters underwent univariate feature ranking. First, we examined whether each feature (predictor variable) is independent of a response variable (low- vs. high-risk patient) by using individual chi-square tests. Then, the parameters were ranked using the p-values of the chi-square test statistics—here, the importance of a feature is quantified as \((-\mathrm{log}(p))\), therefore a large score indicates that the corresponding predictor is important. The subsets of the most discriminative predictors were selected, and they were utilized to fitted the multiple logistic regression classifiers. Additionally, we fitted the classifiers over the principal components (PCs) extracted using PCA followed by the parallel analysis to determine the significant PCs [33]. For each model, we investigated the relationship between the model’s ability to correctly classify low- and high-risk patients using the ROC curve analysis, and we calculated the area under the ROC curve (AUC) as the summary metric to quantify the diagnostic ability of the corresponding classifier. To obtain the optimal cut-point value from each ROC curve, we exploited the (i) Index of Union (IoU), (ii) the closest to (0, 1) criteria (referred to as the Distance technique) and (iii) the Youden index methods [34]. For the selected cut-point values, we reported not only sensitivity and specificity of the corresponding classifier, but we also calculated its positive and negative predictive value (PPV and NPV, respectively), and the percentage of correctly classified low- and high-risk patients. The clinical utility of the developed ML models was investigated in the decision curve analysis. GraphPad Prism 9.4.1 was used for principal component, parallel analysis and other statistical processing, whereas MATLAB R2021b was exploited for feature selection (the fscchi2 function).

We are aware that there are important limitations of our study, which was a proof of concept for a wider program of work. This would include larger cohorts and further external validation, as well as prospective observations for all of them. Also, the patients were not asked whether they had lost their weight before participating in the study, which could explain the negative association of obesity with CVD. An important aspect of cardiovascular pharmacologic management is antiplatelet treatment. However, we were unable to gather reliable information related to this type of treatment in patients without CVD, hence we excluded this parameter from the analysis due to a large amount of missing data. Moreover we included patients with at least 50% of carotid artery stenosis qualifying them as those with overt CVD since 50% is a cut point of stroke related to large artery atherosclerosis [51‐53]. In this context, one should notice that the plaque score has been identified as one of the top 5 risk factors related to CVD so this information potentially could be biased. Finally, exploiting the recent deep learning advances which established the state of the art in various medical data analysis tasks could help us further improve the classification performance of the CVD patients, but it would also require focusing on the interpretability of such large-capacity learners, to make their deployment in clinical setting much easier [48, 49].