Prediabetes risk classification algorithm via carotid bodies and K-means clustering technique

Participants and data collection

The participants were selected according to the methodology described in the CBmeter study protocol, approved on 10 January 2019 by the Ethics Committee of the Leiria Hospital Centre. Volunteers with prediabetes or T2DM and volunteers without prediabetes were recruited in a multicentric, non-randomized controlled interventional study conducted in Portugal during 18 months, resulting in the inclusion of eight volunteers with prediabetes and 25 healthy controls. Data from eight volunteers with prediabetes and 25 volunteers without prediabetes (control volunteers) were collected according to the protocol previously mentioned. Briefly, the volunteers were assessed for the variables of interest in baseline conditions during 10 min and afterwards, submitted to 100% oxygen inhalation during 10 s, and given a standardized meal after the twentieth minute (21). HR and RR data collection was performed by CBmeter via CBView software (Brito et al., 2018), a real-time physiological signal acquisition and processing system at an acquisition rate of 500 Hz. Interstitial glucose (iGlu) was collected every minute by means of a flash glucose monitoring system (FreeStyle Libre, Abbott, Lake County, IL, USA). The study was conducted in accordance with the Declaration of Helsinki, been previously approved by the Ethics Committee of the Leiria Hospital Centre (Protocol number PI.NC.EC.2018.01). Written informed consent was obtained via a specific form from all the participants and/or their legal guardians. The full protocol of the research can be seen at Lages et al. (2021).

Data dimensionality

As indicated in the previous subsection, the data comprises two groups of voluntaries: patients with prediabetes and patients control. The algorithms created use time in the clustering processes. Thus, the data used comprises time series ranging from 1 to 80 min, with samples taken every minute for the HR, RR and RR×HR variables. It is therefore important to note that the dataset is not two-dimensional, i.e., $patients\times variables$, but three-dimensional, that is, $patients\times variables\times time$. In formal mathematical notation: $\mathbf{X}\subset {\mathbb{R}}^{N\times V\times T}$ where:

• N represents the number of patients,

• V represents the number of variables,

• T represents time.

Therefore, the dataset is described by:

$$\mathbf{X}=\left({\mathbf{X}}_{i,v,t}\right)\mathrm{f}\mathrm{o}\mathrm{r}i\in \{1,2,\dots ,N\},v\in \{1,2,\dots ,V\},t\in \{1,2,\dots ,T\}.$$

In this work, N can be 8 (group of people with prediabetes) or 25 (control group), V is 3 (for the HR, RR and RR×HR variables) and T is 80 (time ranges from minute 1 to minute 80). This characteristic of multidimensionality and its computational operationalization can be clearly seen in Algorithm 1, presented in the Results section.

The K-means algorithm

To perform the clustering, the K-means algorithm from the Python library sklearn was used. This algorithm is based on unsupervised learning, which learns the relationships between data identifying when a group begins and ends another through a mathematical metric (Hartigan & Wong, 1979; Bock, 2007). In its traditional version, the K-means algorithm has the number K of clusters as input. From that, there is a random initialization of centroids. Then, for each point in the database, a calculation of the distance from the point to the centroids is performed, then this point is associated with the nearest centroid. After this, the algorithm calculates the average of all the points associated to each centroid and establishes a new centroid. The algorithm repeats the previous steps until there are no more updates of elements in groups.

For the analysis of HR and RR, the K-means method stands out for its simplicity and computational efficiency. K-means is particularly suitable when the number of clusters can be adequately defined, as will be shown later in the article, providing a clear segmentation of the different physiological states based on the patterns observed in the biosignals. Unlike complex methods such as Gaussian mixture models (GMM) and spectral clustering (Srivastava, Sarkar & Hanasusanto, 2023), which require intensive parameter adjustments and can be sensitive to the initial choice of parameters, K-means allows for straightforward implementation and intuitive interpretation of clustering results. Compared to DBSCAN (Bushra & Yi, 2021), known for its robustness in dealing with variable densities and outliers, K-means is more appropriate when a clear segmentation of the data into defined clusters is desired. While approaches based on deep learning, such as Autoencoders (Guo et al., 2017), offer the capacity to capture non-linear complexities, they are more computationally complex and less straightforward in terms of interpretability. Thus, K-means remains a viable choice for biosignal analysis, standing out for its practical applicability and ease of use.

Number of clusters

The determination of the number of clusters to be used in the classification algorithm is based on analysis of the results of the section Analysis, arriving at a number of six clusters for the data with the characteristics provided. Moreover, corroborates with this number, analysis performed via algorithm within-cluster sum-of-squares (WCSS), which measures the variation within each cluster. The WCSS technique establishes that the most appropriate number of clusters for a given sample is obtained when there is less variation in the decreasing curve. An example can be verified in Fig. 1.

Data augmentation technique

As the samples are unbalanced (eight volunteers with prediabetes and 25 without), data augmentation techniques are needed for more reliable validation. In this work, the adaptive synthetic sampling approach for unbalanced learning (ADASYN) (He et al., 2008) was used, which is a variant of the synthetic minority over-sampling technique (SMOTE) (Chawla et al., 2002). This method was implemented using MATLAB® R2020a (The MathWorks, Inc., Natick, MA, USA).

The synthetic minority oversampling technique (SMOTE) is a method used to address class imbalance in machine learning datasets by generating synthetic instances of the minority class. Rather than simply duplicating existing minority examples, SMOTE creates new, plausible samples by interpolating between existing data points. It selects a minority instance, identifies its nearest neighbors from the same class, and generates synthetic samples along the line segments connecting the instance to its neighbors. This approach increases the diversity of the minority class and helps improve model performance, particularly when the imbalance between classes leads to poor learning of minority patterns.

On the other hand, adaptive synthetic sampling (ADASYN) extends SMOTE by introducing an adaptive component that focuses on the most challenging examples in the minority class. Instead of creating samples uniformly, ADASYN generates more synthetic examples in regions where the minority class is harder to distinguish from the majority class. This targeted approach enhances the model’s ability to learn from difficult areas, refining the decision boundary and reducing bias in favor of the majority class.

Clustering process and analysis

In this first part of the results, using the K-means clustering technique, clusters analyses were performed to identify patterns and determine strategies for the construction of the classification algorithm, as well as for the determination of the most adequate number of clusters. The variables (features) under analysis are time, HR, RR and RR×HR. These were considered within the groups of people with prediabetes and control patients.

The clusters were generated using Algorithm 1, which receives a matrix $P_{g,var}$ with the data of a given variable for all volunteers of a given group. Since there were two groups (individuals with prediabetes and control) and three variables (HR, RR and RR×HR), it was assigned the following values: 1 for the group of people with prediabetes and 2 for the control group; 1 for the HR variable, 2 for RR and 3 for RR×HR. For example, for $P_{1,1}$ one has the matrix of the group of patients with prediabetes with HR data, where the first column of the matrix is the time in minutes (from 1 to 80) and the following columns are the HR of each patient with prediabetes (eight patients). So, in this example, one has a matrix with 80 lines and nine columns, that is, $P_{1,1}\in {\mathbb{R}}^{80\times 9}$ (see Table 1). In the case of the variable RR×HR from the group of patients with prediabetes, the matrix will be $P_{1,3}\in {\mathbb{R}}^{80\times 17}$ (see Table 2). It is important to highlight that the k-means clustering algorithm runs using HR, RR and RR×HR during the entire time interval, i.e., from minute 1 to minute 80.

Algorithm 1 returns the labels of the clusters between the time and the variable of each volunteer in the matrix $CPO{X}_{g,var}$ already in the 3 min after 100% oxygen administration, that is, times 11, 12 and 13 min. For this work, the term “variability” was defined as the index in percentage of the variation between the clusters of the volunteers of each group and each variable. The formula for calculating variability is given by Eq. (1). A lower index of variability means that in each group, for a given variable, a higher number of a given pattern of equal clusters was observed. In the HR analysis, this account is given in detail in order to clarify the concept of variability presented in this work.

(1) $$Variability=\frac{C_d}{N_s}\times 100,$$where $C_d$ represents the total number of clusters of different colors for the same variable and the same types of volunteers, and $N_s$ denotes the total number of volunteers in the same group (people with prediabetes or control) for the same variable.

Dejours test and time analysis

CB chemosensitivity was evaluated by two breaths of 100%O2, the double-breath Dejours test, as with this test is possible to obtain a change in the oxygen drive almost free of secondary factors since they are secondary in time (Dejours, 1962). To assess peripheral chemosensitivity respiratory was measured using the CBmeter, since it was already demonstrated that increased CB chemosensitivity observed in patients with prediabetes is seen in the RR, but not in tidal volume (Cunha-Guimaraes et al., 2020).

Thus, RR was assessed with the CBmeter while the subjects breathed room air (21% O2; normoxia), followed by two breaths of 100% O2 (hyperoxia) delivered at a 10 L/min flow, and then normoxia again. Hyperoxia applied during a few seconds resulted in a decrease in ventilation that reflects CB chemosensitivity exclusively, without interference of the central nervous system chemoceptors (Honoring, 2003). The maneuver was repeated three times in each patient to assess reproducibility of the test. If 100% O2 is breathed during prolonged exposures (several minutes), ventilation does not change or could even have increased.

The magnitude of the ventilatory depression caused by hyperoxia, often used as an index of carotid body sensitivity although without definition of a gold standard protocol. Herein, it was assessed the ventilatory responses to a brief hyperoxic test of 10 s to avoid interference of the central nervous system chemoreceptors. RR was monitored before and during the hyperoxic test with the CBmeter. Figure 3 shows that the mean of the RR of volunteers with prediabetes, in general are lower than those of control volunteers. Furthermore, by the Independent-Samples Mann-Whitney U Test (p = 0.007), it was concluded that the first 3 min after hyperoxia (11, 12, and 13) are linked to a greater ventilation decrease in response to the hyperoxic test (see red line in Fig. 3).

Therefore, this analysis verifies significant RR differences at times 11, 12 and 13. In the following analyses, via the K-means algorithm, the term “variability” is designated to name this pattern.

HR analysis

Basically, it is each person’s data that is clustered by the Algorithm 1. For example, the HR data and time of a volunteer with prediabetes is submitted to the K-means algorithm with a K value of 6. It should be noted that the method consists of clustering Time and HR; Time and RR; and Time and RR×HR. This procedure can be checked in detail in “# OBTAINS THE CLUSTERS FOR K=6” of the Python code “code_clusters_CBmeter.txt” provided as Supplemental Material. The algorithm returns these data separated and labeled into six groups. Figure 4 is a result of Algorithm 1 for volunteers with prediabetes in the HR data.

Looking only at the first three columns of Fig. 4, HR1 is the column in which the HR data is recorded over the 80-min Time (1st column) of subject 1. As the number of clusters chosen for K-means was 6, the CL1 column shows how the HR1 data was clustered over time by K-means. Figure 4 shows that the Time 11 and 12 min and HR1 data of 58 were placed in group 2 by K-means. The pair of minutes 13 and HR1 of 62 was placed in group 5. This process is carried out for each minute (between 1 and 80), where for each pair ( $Tim{e}_i,HR{1}_i$) K-means indicates to which group it belongs. During K-means processing, the pairs ( $Tim{e}_i,HR{1}_i$), for $i=1,2,\dots ,80$, can change groups until they reach the point where no more updates are possible (for example, when the centroids no change).

This method of using K-means allows to find a pattern in minutes 11, 12 and 13, which are the 3 min after inhaling oxygen, and this is the return from Algorithm 1 for the variables (HR, RR, RR×HR) of all the subjects. Also note that the HR4 data for minutes 11, 12 and 13 remain in the same group (group 5). On the other hand, for all the other volunteers, there is a variation between groups, between minutes 11, 12 and 13, which is the case for volunteer HR1. As explained before, note that for this volunteer the HR data fell into groups 2 and 5. This procedure enabled a clear observation that, in minutes 11, 12, and 13, the HR data exhibit patterns that distinguish volunteers with prediabetes from volunteers without prediabetes.

Remark 1. It is important to note that the method is not real-time; it is necessary to have all the data (a complete collection of HR and RR over the 80 min) before applying Algorithm 1 and obtaining the clusters.

In the HR analysis, grater variability is observed between minutes 11 and 13 (post-oxygenation) for the volunteers. From Fig. 4, only the fourth volunteer (all blue) of the cluster of peoples with prediabetes did not present any group variation, and seven volunteers had a variation of groups (clusters of different colors, that is, $C_d$) between minutes 11, 12 and 13. Applying Eq. (1), one has:

$$Variability=\frac{C_d}{N_s}\times 100=\frac{7}{8}\times 100=87.5\mathrm{\%}.$$

Figure 5 shows the clusters (CL) for the 25 control volunteers during minutes 11, 12 and 13. One can notice less variability for this group. That is, of the 25 volunteers, 12 had an inter-group variation which corresponds to $48\mathrm{\%}$ of the total, that is:

$$Variability=\frac{C_d}{N_s}\times 100=\frac{12}{25}\times 100=48\mathrm{\%}.$$

To determine the ideal number of clusters, the analysis was carried out with various numbers of clusters (K) as shown in Fig. 6, and the WCSS algorithm was applied, whose graph is shown in Fig. 1. Figure 6A shows the variability percentages for other cluster numbers. The graph has a trend line obtained by means of polynomial regression of degree 3, where the horizontal axis represents the number of clusters and the vertical axis represents the variability in percentage.

Remark 2. Figures 4 and 5 show the 3 min after oxygen inhalation (11, 12 and 13), where HR_ and CL_ represent the HR and the cluster indicated by the algorithm respectively for each volunteer. Colors were added for better visualization and to distinguish the types of groups found. It should be noted that each series of clusters covers a time series between minute 1 and minute 80 for each volunteer. For example, CL1 represents the group labels returned by K-means for volunteer 1’s HR data, indicating that minutes 11 and 12 of HR1 belong to group 2 (blue) and minute 13 belongs to group 5 (orange). The same analysis is carried out for RR and RR×HR.

Clusters names

As observed in the previous sections, the clusters follow certain number/color patterns. Thus, one can name these clusters according to their pattern as per Fig. 7. One can create a function to convert the data in the matrix CPOX into a matrix with the cluster names, attributing values to them. This is the idea of the Convert of Algorithms 2 and 3.

In Fig. 8, for the K parameter of the K-means algorithm equal to 6, the variabilities for the HR, RR and RR×HR variables of the volunteers with prediabetes and without prediabetes are compared. Note in the Fig. 8 (generated from the data in Fig. 6), and according to the variability Formula (1), that the clusters of different colors (DD, DUD, DU and T) represent the value of variability for the different variables. In these graphs, one can clearly see a different quantification of clusters between the control and the volunteers with prediabetes. This quantitative relationship enables the creation of a score in order to rank the volunteers with prediabetes, their clusters and the glucose data collected. In the Score and maximum risk subsection, the clusters are quantified and differentiated according to their different types. From the establishment of this score, one can obtain a risk classification for other patients through their clusters.

Training process: score matrix and maximum risk number

In this second part of the results, one presents the training process of the PRCA, which is systematized by Algorithm 2. After the entire clustering process performed by Algorithm 1 using K-means, Algorithm 2 takes as input the clusters of the volunteers with prediabetes and their glucose values. As a result, the score matrix and the maximum risk number are generated.

Firstly, the association matrix $\mathrm{\Delta }$ (Table 3) must be obtained, where the samples used for training, the weights derived from their glucose values, and the clusters provided by Algorithm 1 are related. The weights are obtained by quantifying the severity level (SL), which is calculated based on the glucose levels of volunteers with prediabetes. Figure 9 depicts a graph of the glucose level of people with prediabetes and control volunteers.

The graph in Fig. 9 was used to establish a parameter to measure the SL of each volunteer with prediabetes, where the red vertical lines are the time markers that represent the disconnection of the means of the volunteers with prediabetes and the control volunteers. Then the SL of each volunteer, calculated by Eq. (2), is obtained by the mean glucose outside the interval $(t_g,T_g)$ (border markers of separation of glucose data from the means of volunteers with prediabetes–see Fig. 9), where $g_{i,j}$ is the glucose level (mg/dl) of volunteer $i$ at time $j$.

(2) $$S{L}_i=\frac{\sum _{j=1}^{t_g}{g}_{i,j}}{t_g}+\frac{\sum _{j=T_g}^{80}{g}_{i,j}}{T_g-t_g},\mathrm{f}\mathrm{o}\mathrm{r}i=1,\dots ,8,\mathrm{a}\mathrm{n}\mathrm{d}j=1,\dots 80.$$

From the SL, according to Eq. (3), a normalized weight (W) between 1 and 10 can be established to be used to obtain the score.

(3) $$W_i=\frac{S{L}_i}{100}.$$

For example, the weight (W) of the patient 20180101001 is calculated as follows:

1. As shown in Fig. 9, take $t_g=34$ and $T_g=55$, and consider for this patient $i=1$;

2. Apply Eq. (2) using the glucose data collected from this patient, then:

$$S{L}_1=\frac{\sum _{j=1}^{34}{g}_{i,j}}{34}+\frac{\sum _{j=55}^{80}{g}_{i,j}}{55-34}=93+125.15=218.15;$$

3. Apply Eq. (3):

$$W_1=\frac{S{L}_1}{100}=\frac{218.15}{100}=\mathrm{2.1815.}$$

The volunteers with prediabetes were organized according to their (W) and associated with their clusters within each variable. For this, the association matrix $\mathrm{\Delta }$ was defined, where the rows represent each volunteer with prediabetes. The first column represents the associated weight and the other columns the clusters of the variables HR, RR and RR×HR. Table 3 shows this representation.

Finally, to construct the score matrix ( $\mathrm{\Psi }$) one must take the cluster names as rows and the variables as columns, in order to obtain a matrix $\mathrm{\Psi }\in {\mathbb{R}}^{n\times m}$ where $n$ is the number of cluster types (names) and $m$ is the number of variables. Each term of the matrix ( ${\mathrm{\Psi }}_{i,j}$) is calculated from the matrix $\mathrm{\Delta }$ (Table 3). The calculation is performed by adding the weights of the clusters of the same name that appear within a given variable. For example, cluster DU appears four times in variable HR in Table 3, being its weights 2.1543, 2.1815, 2.476 and 2.8356 respectively. Then the sum of these weights provides the term ${\mathrm{\Psi }}_{1,1}=9.64$. Table 4 gives the values obtained for all the terms of the $\mathrm{\Psi }$ matrix.

Given the Table 4, it is observed that the maximum score that is possible to obtain for the whole $\mathrm{\Psi }$ matrix is $\chi =28.41$, considering the maximum value of each of the variables. In general, this result can be obtained by the expression 4:

(4) $$\chi =\sum _{j=1}^m\underset{i}{max}\{{\mathrm{\Psi }}_{i,j}\}.$$

Algorithm 2 establishes all the procedures for constructing the $\mathrm{\Delta }$, $\mathrm{\Psi }$ matrices and obtaining the number $\chi$.

Prediabetes risk classification algorithm

The third and final part of the PRCA is characterized by Algorithm 3, whose function is to indicate the level of risk in any given subject. To do this, Algorithm 3 uses the clustering process of Algorithm 1 and the training information provided by Algorithm 2 (score matrix and maximum risk number). Algorithm 1 is used to obtain clusters of HR, RR and RR×HR data for the individuals under analysis, who were submitted to the CBmeter protocol (which includes inhaling $O_2$ and eating a specific meal). From this, Algorithm 3 calculates the individual’s score (from the Score Matrix, adding up the value of each cluster, i.e., HR, RR and RR×HR of the subject under analysis) and then indicates the risk by checking which of the ranges in Table 5 their score falls into. Note that the process in Algorithm 3 is not invasive, as the algorithm is already trained.

This table is built according to the maximum risk number ( $\chi$), which is divided into three parts, where the first part indicates the score range for low risk; the second part, for medium risk; and the third part, for high risk. Next, Algorithm 3 is presented, which establishes the procedures for classification.

Two examples are given. These examples illustrate how the risk classification algorithm for prediabetes works. The objective is to verify what the risk level is for a given patient to enter the group of people with prediabetes from the HR, RR, RR×HR clusters generated by the K-means algorithm. Furthermore, this example compares two volunteers from the control group and correlates the level of risk found with the level of severity via glucose (SL) calculated from Eq. (2) of the respective volunteer.

Control volunteer 20180102001

The HR, RR and RR×HR clusters generated for this volunteer are shown in Fig. 10A. In the same figure, at the bottom, the score is associated according to Table 4 for each variable (HR, RR and RR×HR). Summing the score for each variable, the value 9.005 is obtained, which fits a low risk level for this volunteer to enter the group of people with prediabetes. For this volunteer, the SL is equal to 180.25.

Performance and validation

The algorithm’s performance can be seen from the graph in Fig. 11. This graph shows the trend in the risk of control volunteers (blue dashed line with circular markers) after being subjected to the algorithm. It can be seen that as blood glucose increases (SL), the risk also increases, but to a more moderate extent compared to volunteers with prediabetes (orange continuous line with square markers). The trend lines were obtained by simple linear regression from the markers that were classified by the PRCA Algorithm.

At this stage of the CBmeter investigation, the validation of the PRCA can be approached by treating it as an artificial intelligence classification task. In this case, volunteers labeled as High Risk are considered to have prediabetes, while those labeled as Low Risk are considered to be without prediabetes. For simplification, volunteers labeled as Medium Risk are regarded as successful classifications.

The validation was carried out using the four-fold cross-validation technique (see more in Wong & Yeh (2019)). This means the volunteers with prediabetes were randomly divided into four groups, each containing two samples, resulting in a total of eight samples. Four rounds of testing were performed, with each round using one group for testing (two samples) and the remaining three groups (six samples) for training. Table 6 illustrates this process. The proportion for testing and training is similar to that used by Ahmed et al. (2022) of 70:30, i.e., 70% of the data for training and 30% of the data for validation.

The training process involves using the volunteers with prediabetes to determine the maximum risk number ( $\chi$) in each round. Since the PRCA algorithm employs unsupervised learning via k-means clustering, it groups data based on distance measures, forming distinct clusters. Although the algorithm is trained only with data from volunteers with prediabetes, it can also classify synthetic data (generated from the volunteers with prediabetes) and control group data. This is possible because PRCA identifies patterns within the data and separates groups based on their proximity or distance from the training data. Generally, the farther a patient’s clusters are from the training data clusters, the lower their risk of having prediabetes.

In the validation approach presented in this study, control group data and synthetic data generated using the ADASYN technique were added. ADASYN increased the number of data points for volunteers with prediabetes to match the number of control group volunteers. Since there are eight volunteers with prediabetes and 25 volunteers without prediabetes, the algorithm generated 17 synthetic samples from the volunteers with prediabetes.

As a result, there are 17 artificial samples from volunteers with prediabetes, plus 2 original data samples, totaling 19 samples per round. For the volunteers without prediabetes, 25 samples are available. Of these, 19 are randomly selected for each round, balancing the dataset for validation. The final validation result is the average of the performance metrics across the four rounds.

The metrics used for evaluation are accuracy, precision, recall, and F1-Score, which are described in more detail below:

• Accuracy shows the overall performance of the model, indicating, among all diagnostics, how much the model indicated correctly. Accuracy is a good overall indication of how the model performed. However, there may be situations in which it is misleading, so it is necessary to check the other metrics as well.

• Precision checks how many of the positive classifications the model has made are correct. Precision can be used in a situation where false positives are considered more harmful than false negatives.

• Recall/Revocation/Sensitivity calculates among all the positive class situations as expected value, how many are correct. The recall can be used in a situation where the false negative is considered more harmful than the false positive.

• F1-Score is the harmonic mean between precision and recall. The F1-Score is simply a way of looking at only one metric instead of two (precision and recall) in a given situation. It is a harmonic mean between the two, meaning that when you have a low F1-Score, it is an indication that either precision or recall is low.

The formula for each metric is obtained from the confusion matrix structure shown in Fig. 12.

Figure 13 shows the confusion matrix for the four rounds of cross-validation. Subsequently, according to the formulas in Fig 12, A, P, R and F1-Score ( $F1$) are calculated for each round, with the final result being the simple arithmetic mean of each metric. Thus, using the four-fold cross-validation technique, an accuracy of 86%, precision of 95%, recall of 78%, and F1-Score of 85% are obtained.

Comparison with other methods

Considering the PRCA as a classifier, it can be compared with other classification algorithms. Using MATLAB® R2020a and the same dataset, Pinheiro, Guarino & Fonseca-Pinto (2024) developed a classification algorithm via support vector machine (SVM) with linear, polynomial, and RBF kernels. The same data balancing method (ADASYN) was used. In that work, the results were obtained as shown in Table 7. Upon examining the table, it is observed that the PRCA has better accuracy.

Although a reasonable comparison for PRCA is presented in Table 7, for a better overview of what is available in the literature, Table 8 presents some other studies aimed at supporting diabetes diagnosis. It is important to note that the Table 8 should not be used to directly compare performance with PRCA, as the studies listed were conducted under substantially different conditions, using different datasets, methodological approaches and parameters from those applied in the development of PRCA. The table includes information on the method, sample size, number of resources, accuracy and whether the method is invasive or non-invasive. The table shows that all algorithms use a greater number of features and the vast majority of them are invasive. In addition, they all have a larger number of samples, which is a limitation of this study.

Notably, in addition to being a non-invasive method, PRCA uses fewer features, which indicates a lower computational cost. Moreover, with the acquisition of additional samples in the continuation of this investigation in the future, the algorithm’s performance could improve, as a larger dataset provides more examples for the model to learn from (Elfatimi, Eryiğit & Elfatimi, 2024).