Predicting Analyte Concentrations from Electrochemical Aptasensor Signals Using LSTM Recurrent Networks

Nanomaterial-based aptasensors are useful devices capable of detecting small biological species. Determining suitable signal processing methods can improve the identification and quantification of target analytes detected by the biosensor and consequently improve the biosensor’s performance. In this work, we propose a data augmentation method to overcome the insufficient amount of available original data and long short-term memory (LSTM) to automatically predict the analyte concentration from part of a signal registered by three electrochemical aptasensors, with differences in bioreceptors, analytes, and the signals’ lengths for specific concentrations. To find the optimal network, we altered the following variables: the LSTM layer structure (unidirectional LSTM (LSTM) and bidirectional LSTM (BLSTM)), optimizers (Adam, RMSPROP, SGDM), number of hidden units, and amount of augmented data. Then, the evaluation of the networks revealed that the highest original data accuracy increased from 50% to 92% by exploiting the data augmentation method. In addition, the SGDM optimizer showed a lower performance prediction than that of the ADAM and RMSPROP algorithms, and the number of hidden units was ineffective in improving the networks’ performances. Moreover, the BLSTM nets showed more accurate predictions than those of the ULSTM nets on lengthier signals. These results demonstrate that this method can automatically detect the analyte concentration from the sensor signals.


Introduction
Aptamer-based biosensors have been widely used in various fields, such as environmental monitoring [1], food quality and safety [2], and medical diagnostics and therapy [3], due to the physical and chemical features of aptamers for detection and small binding substances [4]. Moreover, the advancement in nanostructured materials has attracted much attention in recent years due to their potential applications and unique properties, including high reactivity, high functionalization, large surface-area-to-volume ratio, and small size [5]. Thus, advanced nanostructured materials have been utilized to improve the sensing capacities of aptasensors [6], lower the limits of detection of analytes [7], and amplify the sensors' signals [8]. Nanomaterial-based aptasensors have been used as effective instruments for recognizing small analytes in clinical health diagnostics [6,7], medical therapy [9], and disease biomarker detection [10]. In addition, using and improving these analytical devices for identifying and quantifying a target analyte is beneficial due to their having higher two categories: original data (OD) network data and control group data. In the next step, the OD network was used to make prediction models, and the control group was used to assess the prediction models. Then, a data augmentation method was developed to increase the data size. Then, both original and augmented data were used to make and train the LSTM models, and the test set and control group were responsible for assessing the performance of the prediction models.

Materials and Methods
This section describes the methods applied in this study, including those for data collection, data preprocessing, data augmentation, LSTM architecture and optimization, and evaluation of the prediction models.

Dataset Descriptions
The datasets used in this study were (1) 35-mer adenosine, (2) 31-mer oestradiol, and (3) 35-mer oestradiol. The datasets contained several time-series signals representing the drain current of three different aptasensors. Table 1 describes three key features of the sensors used for data collection to provide a quick and brief comparison of the datasets. The distinguishing features of these datasets were two main components of their sensors: their target analytes and their bioreceptors. However, their transducers, another main component used for these aptasensors, were carbon nanotube (CNT) field-effect transistors (FETs). Explaining all of details of the functionalization of these sensors is beyond the scope of this paper. However, detailed information on the 35-mer adenosine sensor, including transistor fabrication and aptamer functionalization, can be found in [24].  1 The bioreceptor of this dataset was created by Erica S. Cassie and was a modification of that in the mentioned reference. This sequence took the common part of the best three oestradiol aptamers, and then some extra mers were added on either side.
As the sensing protocols for the drain current measurements might provide a clear insight into the registered signals, the following explains the method of measuring the signals. The sensing protocols for measuring the 31-mer and 35-mer oestradiol sensors' responses were similar, but they were different from those of the adenosine sensors. Table 2 summarizes and compares the sensing protocols of the adenosine and oestradiol datasets.
The sensing responses for adenosine and oestradiol were measured in time intervals of 1 and 1.081 s with a standard deviation of 5 × 10 −3 , respectively, with similar gate and drain voltages, i.e., V G = 0 V and V D = 100 mV. The buffer selected for the adenosine sensor was 2 mM Tris-HCI, and that for the oestradiol sensors was 0.05 × PBS (phosphate-buffered saline) with 5% ethanol (EtOH).
Regarding the adenosine sensor, the initial load for each measurement was 110 µM of 2 mM Tris-HCI in a polydimethylsiloxane (PDMS) well, which lasted for 1000 s. Then, the adenosine solution was added to the PDMS well every 500 s in successively greater concentrations, considering the adenosine concentration in the PDMS well before each addition. The process of adding the adenosine solution increased the adenosine concentration, which varied from 1 pM to 10 µM in the PDMS well. Table 2. Comparison of the sensing protocols of the adenosine and oestradiol aptasensors. Note that the sensing protocols for both the 31-mer and 35-mer oestradiol aptasensors were completely the same. Thus, their relevant information is merged into one column.

Characteristics
Adenosine Aptasensor Oestradil Aptasensors Regarding the oestradiol sensors, the initial load for each measurement was 100 µL of 0.05 × PBS 5% EtOH in the well, which lasted for 300 s. Then, in the next 300 s, 20 µL of 0.05 × PBS 5% EtOH was added, while the oestradiol concentration did not increase. Then, the oestradiol solution was added to the well every 300 s in successively greater concentrations, considering the oestradiol concentration in the well before each addition. In addition, each time the oestradiol concentration was increased, a solution of 20 µL of 0.05 × PBS 5% EtOH was added to the well. The process of adding the oestradiol solution increased the oestradiol concentration, which varied from 1 nM to 10 µM in the well.

Contextual Outlier Detection
A contextual outlier, also known as a conditional anomaly, is defined as a data instance whose pattern does not conform to that of other well-defined data instances with similar contextual information [27,28]. Regarding the experiments related to this study, factors that could cause contextual outliers were background noises in the lab, the use of broken transistors, issues in immobilizing the aptamers on carbon nanotube surfaces, issues in fabricating the sensing interface, and so on.
The patterns of signals affected by these factors deviate from the patterns of well-defined and normal signals. The purpose of removing outliers is to eliminate non-informative signals or segments. As there were a few signals in the datasets, removing the outliers was performed with prior knowledge of the biosensors' behaviors and through data visualization.
In this paper, the signals were preprocessed with data normalization before being fed into the DL models. It needs to be mentioned that data normalization was applied to the entire signal. Data normalization or feature scaling puts all of the signals in a dataset on the same scale and prevents a feature from dominating and controlling the others. The data normalization applied in this paper was a Z-score scaling that used the mean (µ) and standard deviation (σ) of a signal.
Suppose that X = [x 1 , x 2 , . . . , x i , . . . , x n ] is an entire signal, where the n is the number of data points within the given signal or the length of the signal. Then, Equation (1) shows the new signalX created by Z-score scaling.

Segmentation and Labeling
After rescaling the signal, it was split into different segments. Each segment was a part of the signal for which the concentration of analyte remained constant from its beginning to its end. Then, each segment was labeled with its corresponding analyte concentration. This means that the labels for the three datasets were: No Analyte, 1 nM, 10 nM, 100 nM, 1 µM, and 10 µM. As shown in Table 3, these six labels and concentrations were considered as the six different classes, in the same order.

Data Split
The data fed into the DL model needed to be split into three subsets, namely, the training, validation, and test set, with the proportions of 60%, 20%, and 20%, respectively. The training set was used to extract meaningful information and find the optimal parameters, the validation set was used to tune the parameters, and the test set was used to assess the model's performance [13].
In this paper, the original data were split into two sets-network and test sets-with proportions of approximately 70% and 30%, respectively. These sets were named the original data (OD) network and OD test set, respectively. The former set was used to make DL models based on original data and for data augmentation. The latter, the OD test set, assessed the DL models and acted as a control group. The reasons for considering this OD test set were to prove the functionality of the data augmentation method in making the prediction models and to avoid biased results. In order to complete the information related to the data split, it must be mentioned that the augmented data and OD networks were randomly shuffled and separated again into the training set (60%), validation set (20%), and network test set (20%).

Data Augmentation
In machine learning, small amounts of training data might cause overfitting and might not be enough for training models [29]. The need for data augmentation is more critical for real-world data, since acquiring large enough real-world datasets has not always been possible due to cost or time limitations. Generating synthetic data, which is also known as data augmentation, is a solution for overcoming the problem of insufficient data samples [30] or compensating for datasets with imbalanced classes [31]. Data augmentation helps to increase the generalization capability of an ML prediction model and improve the model's performance by increasing the variability of the training and validation data samples [32,33].
In this paper, we utilized a data augmentation method to increase the size of the available datasets. Suppose theŜ i andŜ j are two preprocessed segments from an identical dataset with similar analyte concentrations. Then, S aug is an augmented segment generated with the following Equation (2): where w ∈ (0, 1) and is a normally distributed random number generated by the randn function in MATLAB R2021b.

Background of LSTM
This subsection explains long short-term memory (LSTM) and its architecture. Then, the unidirectional and bidirectional LSTM structures, as well as their similarities and differences, are discussed.
The advantage of using an LSTM network over a recurrent neural network (RNN) is that LSTM can capture the temporal dependency of input sequences during the training process [21,34]. An LSTM network is an RNN that prevents the long-term dependency problem by utilizing gates and calculating a hidden state with an enhanced function [17]. The building blocks of LSTM networks are LSTM cells, which means that an LSTM layer consists of recurrently connected cells [16,17,34]. An LSTM cell, or an LSTM hidden unit, consists of four parts: the forget gate, input gate, output gate, and a cell candidate. Figure 1 presents the structure of a cell. This cell decides to ignore or remember something in its memory by using a gating mechanism. The role of the three gates is to selectively control and transfer needed information into and out of the cell [34]. This figure can also be considered an LSTM layer consisting of only one memory cell or hidden unit, where X t and h t , respectively, are the input and output of the LSTM layer. Consider X t as a sequence input into the memory block at time step t; the forget gate selects which data to erase and which data to remember. As shown in Equation (3), these decisions are made by the sigmoid layer: The input gate is responsible for controlling the level at which the cell state is updated by using another sigmoid layer. As shown in Equation (4), the sigmoid layer of the input gate decides which data need to be updated. In the next step, as shown in Equation (5), the cell candidate (c t ) is responsible for adding information to the cell state by using the tanh layer. Now, the cell state is ready to be updated with the combination of the forget and input gates and new candidate values ofc t . Equation (6) describes the mathematical formula for calculating the cell state: The output gate, which is shown in Equation (7), utilizes a sigmoid layer to decide which part of the cell state contributes to the output. Now, the hidden state or output of the memory cell is ready to be calculated. The output gate and the cell state are contributors to the hidden state. Equation (8) presents its mathematical formula: Note that W ix , W f x , W ox , and W cx refer to the input weight matrices for the input gate, forget gate, output gate, and cell value, respectively, and W ih , W f h , W oh , and W ch are the recurrent weights for the gates and the cell value in the same order. Their corresponding bias vectors are b i , b f , b o , and b c .
Moreover, it can be seen that the cell state and gate activation functions (AFs) are, respectively, tanh (Equation (9)) and sigmoid (Equation (10)); these map the nonlinearity and make decisions: An LSTM layer in a deep neural network consists of a set of LSTM cells. LSTM layers can be categorized into unidirectional LSTM (ULSTM) and bidirectional LSTM (BLSTM) layers. Figure 2 represents a ULSTM structure. It can be said that a ULSTM structure is an RNN that uses LSTM cells instead. The unfolded figure of the ULSTM shows that the output of each cell is the input for the next cell in the same layer. It should be mentioned that an LSTM block refers to several LSTM cells or hidden units.   Figure 4 presents the flow of information in an LSTM layer during different time steps. In this figure, N is the length of the sequential input for the LSTM layer, L is the number of hidden units in the LSTM layer, and T is the length of the training set. Note that in this figure, h t can be considered as just the forward movement ( h t ) in the ULSTM layer or as a concatenation of both forward ( h t ) and backward ( h t ) movements in the BLSTM layer.

LSTM Architecture
In this work, two LSTM networks were employed to classify the analyte concentrations. The objective was to classify the input data into six different concentration classes: 0 M, 1 nM, 10 nM, 100 nM, 1µM , and 10 µM. The target outputs of each class were labeled in a binary vector format, where the desired class was labeled with "1" and the others were labeled with "0". Recall that the input data were the corresponding concentration segments of the signals, as well as the original and/or augmented segments. Figure 5 visualizes the architectures of both networks. The networks comprised five successive layers: a sequential input layer, an LSTM layer, a fully connected layer, a softmax layer, and a classification layer. The only difference between the networks was in their LSTM layers. The LSTM layer in the first network was a unidirectional LSTM layer, while this was a bidirectional LSTM layer in the second network. It should be taken into consideration that the fully connected (FC) layers were affected by the previous LSTM layers and the number of output classes in the classification layer. Figure 5. A schematic of the network architecture. Note that the only difference between the two networks was in their LSTM layer, which used either a unidirectional or a bidirectional LSTM layer. Table 4 describes and compares the layers and the properties of the unidirectional and bidirectional LSTM networks depicted in Figure 5. Recall that the size of the input layer entering the networks was equal to the length of the segments and was considered as one sequence, and the output size of the networks (m) was identical to the number of classes in the data. It should be mentioned that all of the input weights, the recurrent weight, and the bias matrices were concatenated together to form the input weights (W

LSTM Optimization
We adopted the following variables to find the optimal prediction model: the number of hidden units in the LSTM layers, the optimizers (Adam, RMSPROP, and SGDM) [35], and the number of segments fed into the networks. Initially, the original datasets were used to train the networks with the total size of a segment. The altered variables were: the number of hidden units in the LSTM layers-starting from 50 and increasing to 500 in increments of 50-and the use of the three optimizers mentioned. In the next step, we used both the original and augmented signals to train the networks. The sums of the original and augmented segments per class were 50 and 100.
It should be noted that all deep learning algorithms were implemented with the MATLAB R2021b Deep Learning Toolbox.

Evaluation Metrics
After training the prediction models, the classification performance of the neural networks needed to be assessed with relevant metrics. In this paper, the LSTM networks' performances were assessed using two standard metrics: overall accuracy (ACC) and macro F1-score (MF1) [21,36]. In this work, we created a confusion matrix using the predictions from the test data, and then the overall accuracy and the macro F1-score were calculated with the confusion matrix.
The overall accuracy was calculated from the sum of the diagonal numbers of the confusion matrix divided by the total number of elements. In other words, the overall accuracy (Equation (11)) was the proportion of correctly classified elements among all elements in the test set.

ACC =
Number of true testing outputs Total number of elements in the test set .
Before mentioning the method for calculating the macro F1-score, its building blocks-recall, precision, and F1-score-need to be defined. The recall, which is the true positive rate (TPR) and is presented in Equation (12), is the number of correctly classified positive elements among all positive elements. Precision, which is the positive predicted value (PPV) and is presented in Equation (13), is the number of correctly classified positive elements among the elements classified as positives by the model. Then, the harmonic mean of the recall and precision is called the F1-score (Equation (14)). The macro F1-score (Equation (15)) is the mean of the class-wise F1-score of each concentration.
where m is the number of classes in a given dataset. Figure 6 presents the typical raw signals of the available datasets. Vertical dashed lines separate the different analyte concentrations (ACs). Note that the initial ACs for the 35-mer adenosine experiments were not necessarily identical. Figure 6A,B represent the drain current in two experiments from the adenosine dataset in which the initial AC for the former experiment was 1 nM and that for the latter one was 1 µM. However, the initial ACs for all of the experiments of the 31-mer and 35-mer oestradiol datasets were completely similar; these are depicted in Figure 6C,D, respectively. In addition, we need to clarify the notions of an entire signal and a segment, as these notions will be repeatedly used in the rest of this study. An entire signal refers to all of the data points registered from the beginning to the end of an experimental measurement. For example, in Figure 6A, the entire signal comprises the data points at t ∈ [1, 3500]. On the other hand, a segment refers to a part of a signal that represents the sensor response regarding a specific analyte concentration. For example, Figure 6B contains three segments: the No Analyte segment for t ∈ [1, 1000], the 1 µM segment for t ∈ [1001, 1500], and the 10 µM segment for t ∈ [1501, 2000].  Figure 7A shows the normal pattern of the 35-mer oestradiol signals that were registered with a well-fabricated device. It can be seen that these signals had similar trends when the analyte concentration increased. On the other hand, Figure 7B shows that the signals that did not conform to normal signals. The red and blue lines were registered with non-sensing and broken transistor devices. All of the signals that were registered by the non-sensing and broken transistor devices were completely removed from the datasets.

Contextual Outlier Detection
In addition, Figure 7C shows two instances of signals that showed temporal abnormalities although the sensors were normal and other parts of the signals showed reasonable sensing. In this figure, the drain currents registered between 300-750 s were considered the contextual outliers. Thus, there was partial removal of the signals that were related to the abnormal segments. Consequently, the other segments of the signals remained in their relevant datasets for utilization in the deep learning models. It should be noted that removing the segments that showed temporal abnormalities was performed after data normalization, since the mean and standard deviation used for data normalization depended on the information of an entire signal. Table 5 shows the number of available normal segments in each dataset that were considered normal and suitable for utilization in the DL networks. However, as there were insufficient segments of adenosine with a concentration of 100 pM, these segments were excluded from the adenosine dataset. 3.3. Data Preprocessing Figure 8 presents and compares the instances of preprocessed signals according to the Z-score scaling. Recall that this preprocessing method was applied to the entire signal. Figure 8A shows a raw signal from the 35-mer adenosine dataset, and Figure 8B shows the preprocessed signal according to Equation   Table 6 presents the numbers of segments used in the OD network and the OD test sets in each dataset after the original data split. The network sets were used for feeding into the networks, and the OD test sets, which consisted of original data, were used to assess the networks. The data split in the oestradiol datasets was done by considering three and four entire signals for the 31-mer and 35-mer as their test sets, respectively. However, the data split in the 35-mer adenosine dataset was not as straightforward as that for the oestradiol datasets. The OD test set was selected segment-wise.

Evaluation
It was mentioned that we used the overall accuracy (ACC) and macro F1-score (MF1) as the performance metrics. Moreover, all of the networks were assessed with two datasets: the network test set and the OD test dataset. This means that there were two tables of evaluation results for each dataset after altering the effective parameters for optimization. Thus, these tables were organized according to the network and OD test datasets. Tables 8 and 9 present the performance metrics of the test and control data, respectively, from the networks trained with the 35-mer adenosine dataset. Tables 10 and 11 present the performance metrics for the network and control sets from the networks trained with the 31-mer oestradiol dataset, respectively. Tables 12 and 13 present the evaluation metrics for the test data when trained with the 35-mer oestradiol signals. For a better understanding of the tables, the boldface numbers indicate the highest performance of a network based on the optimizers and the LSTM layer structure for 50 and 100 augmented segments. Moreover, we ignored the best classifiers when the numerical metrics for almost all hidden units were similar. For example, in Table 9, the best classifiers were not selected for accuracy when using 50 segments in the ULSTM and BLSTM structures and the SGDM optimizer. The reason for this was that the use of more than four hidden units resulted in the same accuracy. Figure 10 shows the overall accuracy of the OD test set from the network trained with the original segments from the 35-mer adenosine dataset. The accuracies of the prediction models trained with the original data varied approximately from 30% to 55% for the three datasets.  Figure 11 depicts the effect of increasing the number of augmented segments on the prediction models. This figure shows the accuracy of the test data from the networks trained with the ADAM optimizer and the 35-mer adenosine dataset. In general, it can be seen that the performance of the prediction models significantly improved by utilizing the data augmentation method. However, the prediction models were slightly improved with the increase in the amount of augmented data.  In more detail, regarding the 35-mer adenosine dataset, as shown in Tables 8 and 9, the MF1 for the original data changed approximately from 50% to 72%. Then, the ACC and MF1 reached 85% and 94%, respectively, by utilizing the augmented segments for the control data, as shown in Table 9. For the 31-mer oestradiol dataset, Tables 10 and 11 show that the accuracy reached 90% and the MF1 increased approximately from 50-83% to 54-93%. For the 35-mer oestradiol dataset, Tables 12 and 13 depict that the accuracy range of 25-50% reached 70% by using the augmented data, and the MF1 increased from the range of 30-63% to 76% after data augmentation.

Discussion
Nanomaterial-based aptasensors are useful biosensors that are capable of detecting small chemicals and species. A vital goal in the advancement of biosensors is the identification and measurement of low levels of target analytes. Deep learning methods are attractive tools for the advancement of biosensor technology and the analysis biosensing data. RNNs exploit temporal information in time-series inputs to make prediction models for classification and regression problems.
In our work, we successfully employed LSTM networks to automatically predict analyte concentrations from parts of drain current signals registered by three different electrochemical aptasensors. The differences in these sensors were their bioreceptors, their analytes, and the lengths of the signals for specific concentrations. Among RNN-based models, LSTM networks, which contain a gating schema for data analysis, are suitable models due to their ability to learn temporal dependencies. Thus, we utilized ULSTM and BLSTM networks with different optimizers and hidden units to identify the optimal classification models for various concentrations.
Moreover, similarly to most real-world problems, the available signals registered by these sensors were insufficient for training the networks. To overcome this limitation, we proposed a data augmentation method in order to increase the size of the available datasets and improve the prediction model's generalization ability and overall performance. The augmentation method improved the model's performance.
By comparing the evaluation results obtained with the original and augmented data, it can be seen that the applied data augmentation method significantly improved the classification performance. In addition, the results showed that the number of hidden units might not be very effective in enhancing LSTM models. Considering the optimizers, the LSTM networks that used the SGDM optimization algorithm showed a lower prediction performance than that obtained with the ADAM and RMSPROP algorithms [35].
According to the results showing the networks' performance on the test data, the BLSTM networks used for the three datasets presented more accurate classification than that of the ULSTM networks [37]. However, this result was not identical to that obtained with the OD test data on the 31-mer and 35-mer oestradiol datasets, i.e., the prediction performance of the BLSTM models on both the test and control segments of the 35-mer adenosine dataset was higher than that of the ULSTM structures, but not for the oestradiol datasets. The reason for this failure to improve the performance of the BLSTM over that of the ULSTM for the oestradiol datasets might be the shorter length of the oestradiol segments compared to that of the adenosine segments.
Regarding the datasets, we observed that the least accurate models belonged to the 35-mer oestradiol dataset, and the most accurate networks were trained with the 35-mer adenosine dataset. This low accuracy might have occurred for one of two reasons: the aptamer length used to detect the analyte or the criteria chosen for the detection of contextual outliers. It was shown that successful detection of the analyte depended on the aptamer length, and the analyte-aptamer binding must occur within the Debye length to affect a CNT FET's drain current [38]. Thus, the 35-mer oestradiol signals might not have shown obvious evidence of sensing. In the case that this low-accuracy model might be the result of the detection of contextual outliers, it can be assessed in a future study.
In future work, we can evaluate the effects of the segment length and automatic detection of contextual outliers on the prediction models. In order to address the unequal size of segments in future studies, data augmentation methods based on artificial neural networks might be a solution. In fact, by defining a regression model and employing deep learning networks [39], we can extend the length of available time-series data. Consequently, this technique can produce new datasets containing segments of similar lengths. Regarding the detection of contextual outliers, neural networks such as autoencoder-based [40] or LSTM-based models [41] can be utilized to assess and compare the effects of automatic anomaly detection on the prediction models.
The insufficiency of the data exposed a limitation in the preprocessing method applied in this work. Typical data preprocessing methods apply identical changes to the training and test sets. However, we used Z-score scaling for each signal based on its mean and standard deviation. It was impossible to estimate the mean and standard deviation with the available statistical methods, such as with simulations and hypothesis tests.

Conclusions
In this work, we exploited a data augmentation method and LSTM networks to analyze the drain current signals of three similar CNT FET aptasensors in the 35-mer adenosine dataset and 31-mer and 35-mer oestradiol datasets. The drain current signals reflected the sensing responses of the aptasensors, while the concentrations of target analytes successively increased from 1 nM to 10 µM. The ultimate purpose of analyzing the signals was to automatically detect and classify the analyte concentration according to the corresponding sensing response.
The scaling-based data augmentation method was proven to be highly effective for improving and increasing the generalization ability of LSTM-based classification models. In addition, the results suggest that the applied data augmentation method might be more effective and useful in capturing the features from lengthier signals, e.g., 35-mer adenosine signals.
Moreover, the LSTM networks were successful in sensing response classification and in predicting the analyte concentration. In addition, the results suggest that using BLSTM networks does not necessarily result in making more accurate prediction models when using augmented data. Higher scores for evaluation metrics, namely, accuracy and MF1, of the BLSTM over the ULSTM might have resulted due to overfitting. Thus, using a control group when evaluating prediction models seems to be vital in order to obtain robust results. In addition, changing the number of hidden units was not effective with respect to the performance of the prediction models, and the SGDM algorithm was not a suitable optimisation algorithm for them.

Data Availability Statement:
The data presented in this study might be available on request from the corresponding author. There are restrictions on data availability due to their necessity for our future work.