This study was approved by The Research and Ethics Committee of the Hospital Universiti Kebangsaan Malaysia with a registration number of UKM.PPI.800-1/2/21. The study population consisted of 113 nonsmoking subjects (60 men and 53 women) with age ranging from 19–80 years old (mean ± standard deviation, 37.1 ± 13.4 years); they participated in the data collection process at the UKM Medical Center (UKMMC), UKMMC’s primary care clinic, and university health centre, UKM. A total of 16.81% of the subjects had hypertension, 14.16% had hypercholesterolemia, 6.19% had diabetes, and the remaining 62.84% does not have any CVD risk factors based on the questionnaire provided for subjects.

Participants submitted their informed consent prior to the data collection process. Subjects arrived for the data collection in the morning after fasting overnight for the blood test. A random fasting blood test was required to gauge the individual’s UA level. A questionnaire, which consisted of the health status of the subjects, was provided. The weight, height, arm circumference, and distance between suprasternal notch (jugulum, jug) and pubic bone (symphysis, sy) were measured prior to the arteriographs’ protocol assessment. Arm circumference was measured and recorded for better selection of cuff size. The velocity of the PWV value is measured in m/s given the jug-sy measurement.

Arteriograph (TensioMed Ltd., Hungary) is a cuff-based oscillometry device that requires a cuff to be placed at the upper hand in measuring pulse wave. Every subject should be rested in a supine position for 10 mins before the arteriograph measurement is recorded. The subjects should avoid speaking and muscle movements, which may result in measurement failure. The measurement process was repeated three times for each arm, and the mean of these output was used for the next phase. Tab. 1 shows the explanation of hemodynamic output measured by the arteriograph (P₁: Amplitude of the early (forward) systolic wave, P₂: Amplitudes of the late (backward) systolic wave, R: Correlation coefficient). At the first cuff inflation, the systolic blood pressure (SBP) and diastolic blood pressure (DBP) were measured. The cuff was overinflated at 35–40 mmHg above the subject’s SBP for pulse wave configuration. The pulse wave produced was a combination of early and late systolic peaks. During systole, the blood volume ejected into the aorta resulted in pulse wave (early systolic peak). Then, this pulse wave was reflected at the bifurcation of the aorta when it propagates through periphery (late systolic peak). The early, late systolic, and diastolic waves were decomposed using arteriograph software, and the onset and peaks of the waves were determined [35].

The pulse oximeter (CMS50D+, Contec, China) with sampling frequency of 100 Hz was placed at the index finger of subjects for the PPG measurement from both hands for 10 mins each. It was placed at the opposite arm from the arteriograph-cuffed arm during the recording. Our preliminary result showed no PPG pulse from the pulse oximeter once the arteriograph is inflated. This finding may be due to the flow condition that stopped due to the occlusion in the brachial artery that prevented blood flow in the artery for better detection of late systolic wave to determine Aix and SPBao [30].

PPG signal preprocessing algorithm consists of baseline and high-frequency removal and signal quality indices (SQI). All PPG data were processed offline in MatLAB (The Mathworks Inc., USA). In this step, a fast Fourier transform technique was used as a band-pass filter with a cut-off frequency of 0.5–10.0 Hz. Based on the collected PPG signals, a frequency above 10 Hz was considered high-frequency noise and that below 0.5 Hz was attributed to baseline wander. Amplitude offset was found using the filtering process. To overcome this problem, auto-offsetting was used to bring back any y-axis below zero value to a positive value. It was performed by offsetting the signal by the difference between zero-amplitude and the largest negative value.

Then, the filtered PPG signal undergoes SQI in determining the reliable signals. This step is crucial because only high-quality signal is processed for the next phase. This type of signals is referred to as stable signal within a period of time, in which the three conditions proposed by Orphanidou et al. [36] are fulfilled, as follows: (1) the extrapolated 10 s PPG signal must be between 40 and 180 bpm, (2) the PPG pulse-peak gap must not exceed 3 s to avoid missing more than one beat, and (3) the ratio of maximum and minimum beat–beat interval within a sample must be less than 2.2. Fig. 3 shows that the poor quality PPG signal was described in red color, and the blue color indicates the good quality signal based on conditions previously described.

The good quality of PPG signals undergoes first and second differentiations, resulting in FDPPG and SDPPG, respectively. Fig. 4 shows the PPG signal and their derivative of first and second derivations. The FDPPG and SDPPG represent the velocity and acceleration of blood, respectively. Delineator [37] and bp_annotate [38] algorithms were applied to detect the fiducial points in the PPG signals. These points are pulse onset, systolic peak, dicrotic notch, and diastolic peak, as shown in Fig. 4a. Both algorithms were compared manually in detecting fiducial points toward our PPG data. Delineator algorithm shows accurate detection of onset and systolic peak, and bp_annotate provides better detection of notch and diastolic peak. The determination of pulse onset related to zero-crossing point prior to maximum inflection and systolic peak was defined as the zero-crossing point after inflection [37]. Maximum and minimum inflection points are represented in blue dotted line in Fig. 4c respectively. The dicrotic notch is the minimum of the subtraction of the signal and the straight line from systole to diastole [38]. For the diastolic peak, the minimum of the second derivative of the time series following the dicrotic notch relative to a window of radius systolic–systolic peak (median heartbeat interval calculated from the onset) is divided by 5 s.

The SDPPG consists of five waves and divided into systole region (a, b, c, and d waves) and diastole region (e wave). SDPPG waves were detected by adapting the algorithm proposed in [23] with some modification to improve ‘a’ and ‘e’ wave detection. The first and second derivatives of PPG signal are derived using the following equations: interval calculated from the onset divided by 5 s around itself. The first and second derivative of PPG signal are derived using the following Eqs. (1) and (2):

(1) FDPPG=ddt(PPG)=ddt[y(t+1)−y(t)]

(2) SDPPG=ddt(FDPPG)=ddt[y(t+1)+y(t−1)−2y(t)]

where y(t), y(t – 1), and y(t + 1) are the present sample, previous, and next sample, respectively. The maximum and minimum peaks from FDPPG were determined for references in detecting all SDPPG waves, as shown in Fig. 4b. For maximum peak identification, all positive peaks and the first peak obtained prior to systolic peak and after onset. The minimum peak is defined from all the negative peaks and peaks found after the maximum peak and before half of the current cycle duration. After maximum and minimum peak determination, five waves, namely, wave ‘a’, ‘b’, ‘c’, ‘d’, and ‘e’ from the SDPPG waveform were then extracted, as shown in Fig. 4c. The first positive peak prior to the maximum peak of FDPPG considered as the ‘a’ wave, which can be easily located using the maximum peak location in SDPPG. This approach was different from [23] that used threshold for the determination of the most significant peak. It can result in false peak detection due to the low amplitude of the maximum peak. The ‘b’ wave was defined from all negative peaks and the first negative peak that occurred after the zero-crossing of the ‘a’ wave. The ‘c’ wave was determined at the first positive peak prior to ‘d’ wave, and ‘d’ wave is the first negative peak prior to the ‘e’ wave; otherwise, ‘c’, ‘d’, and ‘e’ waves were merged into a single wave. Finally, ‘e’ wave lies within the minimum peak and local extreme peak in FDPPG given its maximum range, prior to the dicrotic notch detected in the PPG signal compared with [23]. Hence, finding the local extreme peak is a complicated process. Tab. 2 shows the features used in the proposed method (O: Onset, S: Systolic, N: Dicrotic notch, D: Diastolic, R: Peak ratio, A: Amplitude difference, T: Peak interval, J: Jerk, RCT: Relative crest time).

Twenty features can be extracted from PPG pulse consisting of two domain time (16 features) and amplitude (4 features) domain, as shown in Fig. 5a. The number ‘2’ indicates features from the next PPG pulse. For SDPPG, 17 features can be extracted on the basis of five waves from SDPPG after removing ‘c’ and ‘d’ waves because, for some signals, the ‘c’, ‘d’, and ‘e’ waves were merged into one wave (‘e’ wave) due to the high heart rate [38] bring to previous. Hence, features related to ‘c’ and ‘d’ waves were excluded because they result in non-accurate value of both features. Jerk is the accelerated volume change per unit time. It can be measured from the absolute height of two respective waves with the corresponding time interval, as summarized in Eq. (3). The relative crest time (RCT) feature is the ratio of extracted interval compared with the pulse width. Fig. 5b shows features that are available in the SDPPG pulse.

(3) Jerk=|pulse height1−pulse height2|time interval1,2

A total of 98 features were obtained from PPG, SDPPG, and arteriograph from the left and right hands. These features were divided as follows: 40 from PPG, 34 from SDPPG, and 24 from arteriograph’s output. A descriptive statistics of the features (mean ± SD) in ranking the features based on the p-value was applied using SPSS software. The normality of all features was assessed using the Shapiro–Wilk test, Q–Q plot, and equality of variance using Levine’s test. Shapiro–Wilk technique was preferable because of good power properties over a wide range of asymmetric distribution compared with the Kolmogorov–Smirnov for normality test [39]. Non-normally distributed features undergo Mann–Whitney U test to find the significant difference between two groups, hyperuricemia and normouricemia. Independent t-test was performed for normally distributed features. The features were then sorted by the least to the highest p-value with confidence interval of 95%. This step enhances the accuracy rate of the classifiers [40]. Significant features (p < 0.05) were used as the inputs for ML classification.

The final phase of the proposed method is to design an effective decision boundary for the classification of two different groups of normouricemia (control) and hyperuricemia (case). The respective features were divided randomly for training and test data with ratio of 70 and 30, respectively. UA level greater than 360 μmol/L for women and 420 μmol/L for male was under the case group, and UA level lower than those values was categorized as control group. ANN, LDA, kNN, and SVM were used to investigate which ML techniques produced better performance in classifying control and case group. The selection of these 4 types of ML is due to their wide application in the study of CVD prediction [41–46].

ANN is a very capable ML technique given that neural networks are data-driven self-adaptive methods [47]. ANN is composed of an input, hidden, and output layers. The input layer consists of neurons that represent different features. A three-layer feed-forward network was used with Levernberg–Marquadt training algorithm and log-sigmoid transfer function at the first and second hidden layers, respectively. The number of hidden neurons was set to 1–15 at each layer. The purelin transfer function was used at the output layer, and the outputs of control and case group were set to ‘0’ and ‘1’, respectively. The LDA technique allows transformation of features into a low-dimensional space that maximizes the ratio between-class to the within-class variance [48]. Three steps in transforming into a low-dimensional space are as follows: (1) calculate the separability between different classes, (2) calculate the distance between the mean and the samples of each class, and (3) construct the low-dimensional space that maximizes the calculation in step 1 and minimizes that in step 2. kNN identifies the similarities between the training inputs (control and case group). Each neighbor is set with a weight age to allow the nearer neighbors to contribute more than the distant ones to the average [49]. Euclidean distance of five neighbors was applied to determine the nearest neighbor of the test data corresponding to the control or case group. The SVM technique uses the hyperplane decision boundary to maximize the distance from the hyperplane to the nearest training data point [50]. The linear and nonlinear quadratic function kernels were used to find the linear and nonlinear relations, respectively, of the selected input to the corresponding groups.

Performance evaluation of four different types of ML used Se, Sp, and Acc. The details of performance evaluators in Eqs. (4)–(6) are as follows:

(4) Sensitivity(Se)=TPTP+FN

(5) Specificity(Sp)=TNTN+FP

(6) Accuracy(Acc)=TP+TNTP+FN+TN+FP

where TP = number of hyperuricemia accurately detected as case group, TN = number of normouricemia accurately identified as control group, FN = number of hyperuricemia inaccurately detected as control group, and FP = number of normouricemia inaccurately detected as case group.

Patient’s demographics are described in Tab. 3 (L: Left hand, R: Right hand). A total of 113 subjects were divided into control and case groups, which consist of 72 and 41, respectively. A total of 226 sets of PPG data were recorded with a duration of 10 min per data. Only PPG and SDPPG signals undergo signal analysis given that arteriograph produces parametric form of measurement, which is not signal-based. High-quality signals based on SQI process proceed to the next phase, which is feature extraction. A total of 40 features were extracted from the left and right hands of the PPG signal based on onset, systolic peak, dicrotic notch, and diastolic peak. For the SDPPG feature, given that ‘c’, ‘d’, and ‘e’ waves merge, any features that contain ‘c’ and ‘d’ waves are discarded to obtain accurate feature values. The total features after reduction of ‘c’ and ‘d’ waves is 17. The features are excluded for the next phase, which is feature ranking, because of the limited sample size. Fig. 6 shows an example of merged ‘c’ and ‘d’ waves.

Statistical analyses were performed using t-test (independent samples t-test) and Mann–Whitney U test for normal and non-normally distributed data. This type of t-test compared the mean between two groups, and the Mann–Whitney U test or Wilcoxon rank sum test is an alternative for nonparametric analysis, comparing median between two groups. Among the 98 features, 16 features with significant p-value < 0.05 are fed as the inputs for ML algorithm, as shown in Tab. 4. Arteriograph produced three significant features, as follows: LPWVao (0.036), RPWVao (0.025), and RRT (0.046). The four significant features from PPG include LO1_S1 (0.002), RO1_S1 (0.042), RO1_N1 (0.002), RO1_D1 (0.012). The highest number of significant features is 9 from SDPPG, namely, Lbe (0.019), RRae (0.042), RRbe (0.011), Rbe (0.040), Ra_b (0.011), Ra_e (0.003), Ra (0.037), RJae (0.030), and Rb_e (0.031).

The performance of the four classifiers was evaluated using RMSE, Se, Sp, and Acc. Different numbers of features provide different performance results. Tab. 5 shows the performance summary of the result using different types of ML techniques. The number of features from Tab. 4 is sorted from greatest to least significant p prior to feeding into the ML algorithm. All MLs use five-fold cross-validation and 16 features as the inputs. SVM compared four types of kernel functions, namely, linear, radial basis function, and third- and fourth-order polynomial. Linear kernel functions produced better performance than other kernel functions using six features. The LDA ML uses two different types of discriminants, which are linear and quadratic. Quadratic discriminants only require six features compared with linear type but obtain the same performance result. For kNN ML, the maximum number of neighbor is set to 100, and the best number of neighbor is 7–13 using six features that produce the same best performance for that neighbor. The least amount of neighbor was selected for the final comparison with other ML techniques. The ANN produced better performance than the three other ML techniques. Hidden layers one and two were set to 30. Random data division was used as the target. It was divided into three groups using random indexes (train, validation and test). Levenberg-Marquardt training function and the performance were expressed in terms of mean square error. For the training computation mode, MEX algorithm was used. The epoch number was set at 1000 and the algorithm has converged at a performance of 0.00. The threshold gradient and MU value were set at (1e−7) and (1e10), respectively. The maximum number of verification failures was set to 6 tests. The best performance of ANN was obtained when using 12 features with hidden layers one and two at 12 and 11, respectively. Although ANN used higher number of features than other ML techniques, the gap performance proved that ANN algorithm can predict better for hyperuricemia and normouricemia. Tab. 6 shows the confusion matrix of the classification result obtained by using ANN method.