Perception Exploration on Robustness Syndromes With Pre-processing Entities Using Machine Learning Algorithm.

The majority of the current-generation individuals all around the world are dealing with a variety of health-related issues. The most common cause of health problems has been found as depression, which is caused by intellectual difficulties. However, most people are unable to recognize such occurrences in them, and no procedures for discriminating them from normal people have been created so far. Even some advanced technologies do not support distinct classes of individuals as language writing skills vary greatly across numerous places, making the central operations cumbersome. As a result, the primary goal of the proposed research is to create a unique model that can detect a variety of diseases in humans, thereby averting a high level of depression. A machine learning method known as the Convolutional Neural Network (CNN) model has been included into this evolutionary process for extracting numerous features in three distinct units. The CNN also detects early-stage problems since it accepts input in the form of writing and sketching, both of which are turned to images. Furthermore, with this sort of image emotion analysis, ordinary reactions may be easily differentiated, resulting in more accurate prediction results. The characteristics such as reference line, tilt, length, edge, constraint, alignment, separation, and sectors are analyzed to test the usefulness of CNN for recognizing abnormalities, and the extracted features provide an enhanced value of around 74%higher than the conventional models.

Introduction

Teenage mental illness is a major psychological condition that causes constant emotional distress as well as a loss of enthusiasm and energy in accomplishing any work. It has an impact on teenage thoughts, feelings, and behavior characteristics, resulting in emotional failure with functional and personal concerns. Teenage pupils who are experiencing depression have significant problems carrying out daily tasks and performing in various situations. According to a poll (1) more than 21% of teenagers in India reported suffering from serious depression in the previous year. Friendships and emotional bonds between families and friends are harmed by depression. It has a potential to slow down or halt the growth of the children. As a result of this, suicidal thoughts and attempts may result. Handwriting analysis, video and audio analysis, gesture and movement analysis, and psychological evaluation are among the many different approaches for detecting depression and other mental illnesses that might be created utilizing machine learning (ML) and deep learning methodologies (1). Depression is also one of the primary causes of the disability in teens. Often, a person who is going through this is hesitant to seek help from a counselor. In this case, technology can play a critical role and be the first line of defense. As people become more active on social media, their patterns can be exploited to identify physical and mental illnesses in simpler manner. Several ML algorithms for regression, such as random forest, linear regression, Bayesian network, Naive Bayes, support vector machine (SVM), k-nearest neighbor, and deep learning algorithms for feature extraction, such as Convolutional Neural Networks (CNN), have been used and successfully tested on various datasets containing video, audio, texts, and images fed to a classifier, applicable on affective and personality computing with feature analysis.

In the previous eras, there was an increase in the use of media information in an online method for recognizing, measuring, and tracking changes in the event of an illness. The universal design of web-based media presents an opulent case for improving the data available to psychological wellness therapists and analysts, allowing for an education expert and a well-organized expressive comfort field. Furthermore, entities are influenced by communicable cynical attitudes in informal associations, which stimulate dejection and other psychological diseases. Dysfunctional conduct is known to be a substantial risk factor for self-destruction; about 80% of those who attempt or die by self-destruction have been diagnosed with some form of psychological instability. Wretchedness is the most well-known psychological disorder, although it has remained unknown or untreated due to its recognition or denial. Early recognition of a significant sorrow's signs and therapy through optimal mediation can help to prevent the onset of a significant sorrow. Numerous studies have identified physical and psychological disorders as a result of the vast amounts of data available on the internet, including a few studies that focused specifically to depression.

Literature Survey

Anxiety, despair, stress, and other mental health concerns occur too frequent in youth. It is critical to detect such mental diseases earlier to avoid future effects. Currently, numerous methodologies based on AI-based approaches and algorithms for detecting depression and other mental diseases have been developed. The researchers created a ML model that analyzes social media posts and tweets to detect depression. The researchers used data gathered from social media profiles, activity, and tweets to train their model for this project. The test dataset findings reveal that the critical values for classification error metrics are determined by the sample-based features collected from social media analysis. The model's accuracy will increase as the number of features increases. This is a data-driven paradigm in which the performance and outcomes are determined by data and characteristics (2). The researchers aimed to develop a method for examining facial features and detecting depression scales using image processing techniques. They used characteristics and facial expressions to train the classifier for pleasant and bad emotions. A teen student's video is collected and sent into the system. Using Gabor filters, the model extracts the face and its features. The collected characteristics are then classified into different classes using an SVM technique with supervised learning classifier classification. The number and severity of depressive patients are computed using a neural network to analyze depressive emotions taken from a video sample. The class coordinator and parent are advised about the student's disturbed mental situation after the depression is calculated using the Beck's Depression Inventory (BDI-II) scale. Other characteristics used by the system include academic performance, extracurricular performance, and conduct opinion (3).

Deep learning algorithms have a wide range of applications in the treatment and analysis of mental health and other mental disorders to better identify biomarkers relevant to unique characteristics between them, and they can be better identified in a variety of situations, such as mental depression, where no models exist. Furthermore, the state of sadness is described as an event sequence, allowing users to compute psychological infection or severe ML circumstances. Traditional researches have stated that generated models are implemented with strategic tools that pre-process MRI data, such as MRI scanning tomography, for extracting all independent functionalities from deep learning models that are used as input data to validate the model. All of these aforementioned gadgets, on the other hand, collect data based on the prior knowledge of the reference set. As a result, certain datasets that might be used to predict mental disease or healthy cases of depression may be misplaced. Cooperation was incorporated in the analysis to improve the accuracy of the outcomes of this multi-modal investigation. Researchers worked with the National Institute of Neurological Disorders and Stroke, the Centers for Disease Control and Prevention (CDC), and the American Psychiatric Association (APA) to aggregate the data, in addition to the CAN–BIND study. This combined effort could hopefully result in easily available biological assays that can accurately and efficiently guide treatment selection for depression (4). Daily chores such as drawings or writings might be used to detect depression, anxiety, or any other mental problems. Emotional State Recognition from Handwriting and Drawing (EMOTHAW) is a public database that contains handwriting and drawing samples pertaining to an individual's emotions and mental state. This collection contains handwritten and drawn samples from 129 people who are facing mental and emotional problems. The participants were given the task of sketching a pentagon, a house, a circle, and a clock, as well as writing a cursive statement. The elements from the sample are used to analyze a person's personality and emotional aspects, resulting in a personality and character profile. Starting point, gradient, size, border, weight, spacing, strokes, and areas are some of the features analyzed from the handwriting sample that distinguishes it from other samples. The results of the sample analysis were uploaded to their database. To classify and evaluate the samples, the researchers employed a random forest approach. This ML approach is based on an ensemble-based decision tree learning, which includes a feature rank procedure. To study an individual's mental state, feature ranking is used to rank and find the key elements in the sample (5–18).

Artificial intelligence (AI) (i.e., ML) has been introduced into the therapeutic field as a means of improving exactness and accuracy while reducing the number of time-consuming tasks that require human intervention (19). There is emerging evidence that ML's innovation might potentially discover and considerably enhance therapy of complicated mental problems such as despair (20). The researchers devised a system capable of detecting doom with minimal human intervention: man-made reasoning mental assessment. This structure consists of a brief human-PC intelligent assessment that employs artificial consciousness, namely, profound learning, to predict if a member would be deterred by palatable execution. Because of its ease of use, this innovation could provide a useful tool for mental health professionals to spot signs of distress, allowing for quicker intervention. Furthermore, by providing a more targeted assessment, it may minimize the difficulty of identifying and decoding profoundly complicated the physiological and social biomarkers of misery (6). Based on the patient's condition and the doctor's viewpoint, a comprehensive literature review is conducted. Visits to various clinics are used to obtain information. With Doctor's consent, information is accessed and reviewed (21, 22). While conducting a depression analysis, several fields of doctors' opinions on depression are examined. Types and causes of depression are investigated to develop a solution for prediction (17, 23–27).

Depression is a psychiatric condition that obstructs a person's growth by causing behavioral, emotional, and mental alterations. Due to the lack of marking and measuring standards for the depression scale, the psychiatrist may be unable to recognize it. Electroencephalogram (EEG) signals can be used to identify detrimental alterations in the cerebrum's functioning. Thus, the suggested method can scale the downturn measure and recognize the degree of the illness using EEG data using AI methodologies. The highlights from the cerebrum's transitory region are isolated using six (FT7, FT8, T7, T8, TP7, and TP8) channels in this study. Delta, theta, alpha, beta, gamma1, and gamma2 band power, as well as their associated unevenness and matching lopsidedness, are the straight highlights used. Sample entropy (sample) and detrended fluctuation analysis (DFA) are two non-straight highlights used. The following classifiers were used: SVM bagging alongside three distinct piece capacities (polynomial, Gaussian, and sigmoidal). Relief-F is the highlight selection approach that was used. Using SVM (Gaussian Kernel function) and Relief-F as a highlight option, the most notable arrangement precision of 96.02 and 79.19% was achieved for the location and seriousness scaling of darkness. The investigation revealed that a decline has an impact on the worldly portion of the mind (temporo–parietal region). It is indicated that the presence of downturns in the system is having the capability to directly affect the equator reference band structures. It may also be deduced that, among the many components of SVM, the Gaussian bit is more capable than other bits. The combination of all combined unevenness and deviation revealed exceptional arrangement precision (exactness of 90.26% for depression discovery and exactness of 75.31% for seriousness scaling) (7) among the many highlights. Normally, determining gloom necessitates a thorough examination by a knowledgeable expert. Recently, much more thoughts have been devoted to programmed melancholy expectation to conduct a more thorough and fruitful examinations of unhappiness. In this research, a unique approach is proposed that uses a two-stream profound spatial–temporal architecture to estimate the level of downturn from video data. The Inception-ResNet-v2 network is used in our method to extract the spatial data. A volume neighborhood directional number is described, which is based on a powerful component description to capture facial movements. Then, the element map obtained from the Very-low-density lipoprotein (VLDN) is applied to a convolutional neural network to obtain more discriminative highlights. Furthermore, by combining the worldwide middle pooling strategy into the model, a multi-layer bidirectional long transient memory model is planned to obtain transitory data. On the transient parts of spatial and worldly highlights, the TMP method is used. Finally, a thorough study of two testing datasets, AVEC-2013 and AVEC-2014, demonstrates that the proposed strategy outperforms the current methodologies for predicting suffering levels (8). Table 1 indicates the Model Summary for feature extraction.

Table 1

Model Summary for feature extraction using Handwriting and Drawing sample.

Layer types	Output shape	Parametric values
2D Convolutional (Conv2D)-5	0, 222, 222, 32	898
2D Convolutional-6 (Conv2D)	0, 220, 220, 64	18,494
2D Maximum_pooling-4	0, 110, 110, 64	0
Dropout-5 (Dropout)	0, 110,110,64	0
2D Convolutional-7 (Conv2D)	0, 108,108,64	36,926
Dropout-6 (Dropout)	0,54,54, 64	0
2D Convolutional-8 (Conv2D)	0, 52,52,128	73,858
Dropout- 8	0,26, 26, 128	0
Flatten- 3	0, 86528	0
Dense- 4	0, 64	5,537,858
Dropout-7	0, 64	0
Dense-5	0, 1	65

Research Gap and Motivation

The major objective of the proposed work is to overcome the disadvantages that exists in the prediction model using handwritten statements where multiple formats are not present is any existing models (1–29). In addition, only occasional treatments are provided for predicting depression with low complexity algorithms. Even the researchers have examined the implementation model with handwritten statements where complexity of prediction is much higher. Moreover, in the presence of electrograph signals, it is much easier to predict different characteristics of depression but an electrical utensil is needed in this kind of detection. This in turn will create several back-end problems in the near future and the data gathering approach is much reduced in processing of signals. Therefore, to overcome the abovementioned gap, the proposed method on depression prediction using CNN has been implemented.

Proposed Methodology

The features derived from the output samples are used to analyze a person's personality and emotional components, resulting in a personality and character profile. Handwriting samples are examined for the properties such as reference line, tilt, length, edge, constraint, alignment, separation, and sectors that distinguish them from other samples. Handwriting samples of teens are collected and processed using a computerized deep learning utilizing the BDI-II scale. Samples are classified according to their rigorousness in the suggested model architecture to identify, scale, and classify depression using handwriting samples of teenagers. This study examines several depression scores and categories, such as anxiety or stress from 0 to 13, mild depression/anxiety from 14 to 19, moderate depression/anxiety from 20 to 28, and severe depression/anxiety from 29 to 63. In addition, for analyzing handwriting photographs, CNN is used with image recognition techniques to extract deep features from the sample. The CNN, or advanced neural network layers, is a type of deep learning technology that has been effectively used to solve real-world problems.

The network architecture is set up in such a way that extracting and analyzing features from images, audio signals, text, and video segments can be done quickly and efficiently. The proposed method on depression prediction is implemented using CNN for extracting all necessary features by processing the captured input images. In addition to input images, a handwritten sample is trained using hidden layer approach. This type of processing technique provides additional advantage as all dataset features are obtained using compatible mode of operation even if the images are provided at low resolutions. However, the images are converted to high resolution in the CNN model thus data is gathered after classification mechanism. All the steps are implemented under small time scale where errors are minimized at corresponding stages. Furthermore, during feature extraction, the accuracy of detection using CNN is maximized using rotation and scaling techniques.

Objectives

The proposed method on predicting different characteristics of depression aims to solve a multi-objective case study that can be described as follows:

To minimize the errors that are present in existing models using CNN and maximize the accuracy of prediction by implementing two different sample sets.

To implement BDI-II scaling for images during classification and extraction techniques where high-resolution images can be processed.

To integrate cloud storage for storing all monitored parametric values using reference line measurements.

Study Structure

The remainder of this study is laid out as follows. The proposed work's design technique is discussed in Section Design Methodology. The feature categorization technique is discussed in Section Optimization Algorithm: Classification and Regression Development with a step-by-step implementation. The results and overall consequences are discussed in Section Results and Discussions, and the importance of the projected model is described in Section Conclusions.

Design Methodology

A model has been constructed in this architecture to detect an individual's emotional state using handwriting and drawing examples. The model starts with handwriting and drawing samples as input, which it runs through a CNN-based classifier. A database is created for this system by collecting handwriting text and drawing samples from 250 people. A separate work was also provided, which required students to draw a pentagon, a house, a circle, and a clock, as well as write a statement in cursive. Individuals were given a tablet on which to draw and write for the assignment indicated in Figure 1. The samples are stored in a database once the process is completed. In addition, the database is separated into the following three categories: Training, testing, and validation, with an 80:10:10 ratio.

Figure 1

A sample task to draw a pentagon, a house, a circle and a clock and text in cursive.

Various characteristics of drawing and handwriting samples, such as reference line, tilt, length, edge, constraint, alignment, separation, and sectors, that distinguish one sample from another are examined. Wide separation between words/lines, increasing benchmark, unbalanced left and right edges, and less top edge were a few of the overall highlights of the database. The most commonly observed features for scale are as follows: Maximum weight, broadness at left-edge and top-edge, blended inclination, broadness dispersing in the middle of the lines, maximum slanting in letters such as F, G, H, K, and R, maximum slanting in letters such as F, G, H, K, and R, maximum slanting in letters such as F, G, H, K, and R, maximum slanting in letters such as F, G, H, K, and R. The features are accurate inclination, wide separation between lines/characters, and blended gauge on a scale ranging from 14 to 19, which are collectively defined as a mild depression/anxiety class. The highlights of the lopsided gauge, lopsided edge, long bending within alphabets such as Y, G, F, and so on, enormous dimension of alphabets, wideness in the left-edge, and broad dispersing in the middle of sentences for scale 20–28 are defined as a moderate depression/anxiety class. The top-most and leftmost edge wideness, blended/leftmost inclination, falling gauge, and asymmetrical dispersal in the midst of the sentences are the highlights for the nervousness–anxiety class. Maximum weight, rightmost inclination, blended standard, small size, rising benchmark, lopsided edge, huge bends with alphabets such as S, G, Y, F, and so on. are seen in the sadness–nervousness–stress class, while odd edges, linear gauge, linear/rightmost/leftmost inclination, and average sizing are seen in typical classes.

Handwriting samples were collected and used as a tool for validation, assessing intellectual deficiency, and assessing character attributes. In truth, collecting handwriting samples is non-intrusive, simple, and low-tech, and requires no administrative expertise. The following samples are saved in Switched Virtual Circuit (SVC) records, where SVC stands for record expansion. The SVC extension documents are typically presented in standard American Standard Code for Information Interchange (ASCII) formats that can be accessed using applications such as Microsoft Word, Google Docs, and other common trending applications. The following benchmarking parameters are applied to the data samples:

Location in x-hub.

Location in y-hub.

Time stamp.

Pen up = 0 or Pen down = 1.

Dimension of the pen as for the tablet.

Dimension of the pen as for the tablet.

Pen point pressure applied by the pen.

Additional data, such as speed highlights (quickening, speed), prompt direction point, brief dislodging, time highlights, and features highlights, can be produced using this arrangement of dynamic data.

Feature Extraction From Handwriting and Drawing Samples

The generated dataset is pre-processed by rotating, scaling, and translating it in order. The photos are resized and normalized to be compatible with the system. This dataset was used to train the convolutional neural network model. Following training, a new input image is given to the neural network to extract features and predict the BDI-II scale, as shown in Figure 2.

Figure 2

Network architecture for eight-layered CNN.

The component maps are used to lower the dimensionality of the convolutional layers using the created pooling layer. The framework executes the maximum pooling layer into the properly defined element map that are unpredicted upon the indicated size, and the largest component is chosen from the framework. There is a complete layer bound after the mixing of various layers, which bundles the highlights to address questionable classes utilizing the increased level features from the Convolutional Layer/Pooling Layer. Soft-max functions are engaged as the last layer, which fixes multiple yields from each unit to the features from beginning to final values. By assessing the vulnerability at each layer, the total cross-entropy job analyses the presentation of collecting ideals. The error that is corrected at the yield layer must be reduced.

The reverse engendering method, which is carried out by the inclination drive technique, attempts to reduce the error by progressing regressively over distinct levels by modifying multiple assessments of loads and tendencies. A cross-approval approach has been used in the development model for evaluating AI models by preparing a group of datasets and measuring using correlative separations on the dataset. The k-cover cross-endorsement condenses data into k bits of equal dimensions, with one segment adding separately for endorsement and the rest sections used for readiness in this model. The entire dataset is split into k wrinkles of similar size, and the model is built using k – 1 folds. The parametric test of the model on the kth overlay will begin immediately and continue in the foreseeable future. For each cover, the goof/setback, exactness is set aside. This cycle is repeated until all of the creases in the testing dataset have been applied. Accuracy stored in k-folds is used to resolve the typical exactness. By combining a few models, cluster knowledge representation facilitates the application of different representations in AI. A basic research control is executed a couple of times in a bunching calculation to outline for choosing from the hypotheses that are gathered. Cross-approval was used by placing persons with similar ensemble features on different planning circles and using different models for each participant in the approval process. Outfits are commonly used to obtain claimed categories that are obtainable from a collection of arrangements that produce low-precision results. As a result, the multiple layers that are used as a fundamental classifier in the proposed work will assemble the exactness after cross-approval an outfit methodology is functional where the top classifier loads out of the eight convolutional neural organization models that are used while making a gathering model as depicted in Figure 3.

Figure 3

Feature extraction for drawing and handwriting samples.

Optimization Algorithm: Classification and Regression Development

The process of predicting depression from textual data has been implemented with high-effect analysis using long short-term memory which is indicated as productive model (30). In this model, a handwritten content is placed and by using the frequency the deep learning model predicts the true and false positivity rates. However, using the frequency of recognition accuracy of content cannot be achieved to large extent. Therefore, to achieve high accuracy many challenges with ML techniques are deliberated with suitable application platforms (31). One best application for ML technique is depression prediction where all possible problems that affects the mental health of humans are discussed (31). However, a comparative analysis will not provide solutions to overcome the problems in health-related issues which indicates that a testing model must be placed during comparison state. To have a deep insight knowledge testing model comparisons are made (32) using long short-term memory in two different directions. In this bidirectional procedure, the neural networks are incorporated and the visualization approaches are built for extracting high-feature comparisons. After extracting the characteristics from the handwriting, drawing, and audio samples, the depression scale is predicted using several ML methods, and a comparison is made based on the results. The findings of this comparative research will provide the best performing algorithm for obtaining correct results. Random forest, logistic regression, Naive Bayes, k-nearest neighbor, and SVM methods are used to calculate the BDI-II scale using regression techniques. Comparative analysis is used to further discuss the algorithms.

A range of elements have been derived from distinct as well as the same modality in this design, as seen in Figure 4. Deep neural networks are used to extract information from handwriting and artwork samples. In addition, a fusion method was used to integrate the various retrieved feature sets. After the features are extracted, a fusion approach is used to combine feature vectors from various modalities in a sequential order. This sequence was repeated twice—once before and once after the dimensionality reduction. The fusion at the decision-making level necessitates the clustering of prediction or classification findings. Based on the extracted features, the model's prediction results can be dependent on a single or the multiple modalities. The fusion in regression is considered to be difficult to use due to the existing problem with multi-collinearity, which is used to characterize the observational features from the dataset used for assessing depression class and scale. As a result, in the multi-collinearity model for binary classification, logical operations and logical gates are used to apply the decision fusion technique. Even logics are utilized to detect the presence or absence of depression and categorize them. In addition, the participants are categorized using logic operators AND and OR to fuse the classification findings. A random forest algorithm is utilized as a classifier in this methodology, which is a common way for decision fusing classifier algorithm to train parameterized scores signifying distance. The random forest classifier's output is utilized as an input for the next level regressors. In the case of observation utilizing regression over choice fusion, the generalized linear models are used, and expectation maximization procedures are applied to the decision fusion level, maximizing the possibilities. The major intention for choosing CNN is that in proposed method an automatic detection is needed where all features must be extracted within a short period of time. Also, to avoid human interference as system must be chosen that learns the descriptive model by its own and high computational effects can only be observed if CNN is implemented. In addition, the location of implementation is much complex in other algorithms but in CNN the implementation process can be done without any reference location thus making the model much easier during training stages. Further, the reliability of operation is much higher with a few label sets from the operating ends; thus, CNN is preferred for predicting the depression state using written sets.

Figure 4

Feature fusion and depression scale prediction using handwriting and drawing samples

Design Algorithm for Handwriting and Drawing Feature Extraction

Divide the dataset in 80:10:10 for training, testing, and validation.

Using feature ranking, set the weights and features as key parameters.

On the input image, do scaling, dimensionality reduction, rotation, and translation.

Carry out the pooling and forward propagation operations.

Using the Softmax activation layer function, calculate probabilities for each class.

Determine the model's output error.

The inclinations of the error are determined in the back propagation step by taking the incomplete subordinate for all loads. For limiting the yield blunder, the boundary values are refreshed.

For each image in the training dataset, repeat steps V through IX.

Using the provided equations, calculate the model accuracy and model loss for the proposed model for the validation dataset.

Save the optimal weights for future computations and modeling.

Create an ensemble-based learning model using cross-validation modeling.

Using the provided equations, calculate the model accuracy and model loss for the proposed model for the testing dataset.

Model Training and Evaluation of CNN

The accuracy for prediction is calculated using Eq. (1) as follows:

where CP(i) and s(i) denote the total number of correct estimates and models, respectively.

The level of absolute number of right predictions to the all Outage tests is determined as accuracy. In the event that the model has a precision rate nearest to 1, it is considered as the model has more precise expectations and if the precision rate is nearest to 0, the model is viewed as less precise. The accuracy in terms of classification can be calculated as follows:

Here, tp(i), tn(i), fp(i), and fn(i) denote true positive, true negative, false positive, and false negative values, respectively.

Hence the values for sensitivity can be calculated using true positive and false negative values as follows:

The value for specificity is also measured using true negative and false positive values as follows:

Similarly, the positive predictive value and negative predictive value for the proposed prediction model can be calculated using Eq. (5) as follows:

To evaluate the percentage error in predicting the root mean square error values are designed using the depressed and non-depressed classes by the model and thus calculating the error difference with the actually predicted values. The root mean square Error (RMSE) and average estimation error (AEE) can be calculated as follows:

where, e and r denote expected and reference values, respectively.

where and z denote difference between class errors.

In analytical terms, the objective function can be defined using Eqs. (9) and (10), where the unconnected non-linear equations are established for both minimization and maximization problems.

In Eq. (9), both error values must be minimized by considering total sampling period. If errors are minimized, then the maximization problem can be framed in an automatic mode as follows:

There is no need for implementing manual procedures in Eq. (10) as the accuracy of prediction will reach maximum percentage by comparing the true and false values.

Results and Discussions

The experiment is carried out in real time on the Google Collaborator platform with an NVIDIA GeForce graphics card. The dataset consisted over 1,500 photos, which were separated into the following three groups in a ratio of 80:10:10: Train set, test set, and validation set. The samples are further classified into eight groups, with scores ranging from 0 to 13 for anxiety and stress, 14 to 19 for mild depression/anxiety, 20–28 for moderate depression/anxiety, and 29–63 for severe depression/anxiety. The training model also includes eight convolutional layers for feature extraction from handwriting and drawing samples as well as a batch activation layer and a batch normalization layer. With modifications in the training dataset, the layered architecture remains the same in each fold. The non-linearity functions in the model are achieved using the Rectified linear unit activation function. The batch normalization procedure is used to normalize the activation from the last layer. The mean value is set to 0 and the deviation to 1 after normalization. The Max pooling approach reduces the dimensionality of the pooling layer. The first, second, and third columns, respectively, show the layer type, input image shape, and number of parameters. While training the model, the output shape of the input image is 64 × 64 pixels with batch size of none, indicating the input image's dimensionality.

After completing cross-validation on the dataset, eight convolutional layers are blended using an ensemble-based learning process. The suggested ensemble-based model is evaluated on a testing dataset after it has been trained and validated. The convolutional neural filter extracts the features from the handwritten sample photos layer by layer, as shown in the architectural overview. Over a 64 × 64 image, there are a total of eight 3 × 3 filters applied. First, the curves and edges are retrieved by the first convolutional layer while the higher features are extracted later by the last convolutional layer. In addition, a BDI-II score column has been added to each of the files for model training. This indicates that features are distinct from one another. In addition, the impact of each feature on the score prediction target variable is investigated. The BDI-II scale is predicted for the tested participants using a random forest regressor with 40 estimators in comparison to the proposed model.

Furthermore, the following hypothesis is considered for analyzing the model's accuracy: A person with a depression scale value is depressed and otherwise not depressed. A binary classification column is added, with the depressed person being labeled as 1. The model's performance is validated by looking for commonalities in categorization and prediction using the participants' labels. Furthermore, a manual and device study are used to make simultaneous observations. If the question was marked “yes” in the recorded session for the domain and the model also gave a higher precision for the same class, it was considered a positive response. Any disagreement or dispute over classification is met with a negative response. Deep neural networks are used to extract information from handwriting and artwork samples. In addition, to merge these different retrieved feature sets, a fusion method is used. After the features are extracted, a fusion approach is used to combine feature vectors from various modalities in a sequential order. The findings of the comparative analysis will recommend the best performing algorithm for obtaining correct results. Following the extraction of features from handwriting and drawing samples, depression patients are accurately categorized using several ML algorithms, with comparative analysis based on the results.

The SVM algorithm has an accuracy of 82.54%, random forest algorithm has an accuracy of 88.97%, Naive Bayes algorithm has an accuracy of 78.21%, k-nearest neighbor algorithm has an accuracy of 71.36%, and logistic regression algorithm has an accuracy of 64.52%, according to Figure 5 and Table 2. With an accuracy of 88.23% and precision of 87.46%, the random forest method was judged to be the best performing algorithm.

Figure 5

Sample set vs. Measurement values (A) Total comparison, (B) accuracy, (C) precision, (D) recall, and (E) F1-Score.

Table 2

Comparative analysis of different ML algorithms used for predicting depression from samples.

Models	Accuracy	Precision	Recall	F1-Score
SVM	82.54	84.54	83.36	8.394
Random forest	88.97	87.46	88.77	8.811
Naive Bayes	78.21	76.96	75.12	7.6028
k-Nearest neighbor	71.36	73.49	72.48	7.2981
Logistic regression	64.52	66.16	68.83	6.744

As a result, utilizing handwriting and drawing examples to combine their properties to forecast depression scale is an effective strategy. Table 3 and Figure 6 show the BDI-II scale for depressive patients utilizing various ML algorithms, with a comparison analysis based on the results.

Table 3

Error values of different ML algorithms used for predicting depression from samples.

Models	RMSE	Mean average error	Mean error
SVM	8.8548	8.8942	9.6363
Random forest	6,541	8.4235	8.5896
Naive Bayes	10.2254	9.2152	10.2203
k-Nearest neighbor	4,141	8.4458	10.9299
Logistic regression	12.6656	11.5569	12.6112

Figure 6

Error values attained from sample dataset.

The RMSE for the SVM algorithm is 9.8548, random forest algorithm is 8.6541, Naive Bayes algorithm is 10.2254, k-nearest neighbor algorithm is 10.4141, and logistic regression algorithm is 12.6656, according to comparative analysis. With a mean average error of 8.4235 and a mean error of 8.5696, the random forest method was deemed to be the top performing algorithm.

A confusion matrix is used to visualize the performance of the prediction or classification model. A testing dataset for the known true values is used to generate the table. Figure 7 depicts the confusion matrix for the depressed and non-depressed classes.

Figure 7

Confusion matrix for depressed and non-depressed classes of the proposed model.

Convergence Characteristics

The depth of integration for an algorithm defines exact coupling strength in real time implementation cases therefore it is necessary to plot convergence of CNN using iteration periods. If a fast convergence rate is attained, then the high benefits can be assured and in turn the sample that is drawn for predicting depression can be identified quickly. In addition, the system time for executing the current state can also be defined using the same characteristic model and is simulated in Figure 8. From Figure 8, it can be perceived that number of epochs is varied from 10 to 100 and for each period time of convergence is plotted. In all iteration periods, the proposed method using CNN performs well as compared to existing system (32). It can be proved when the iteration reaches 50th period and during this case, the convergence is achieved at 5.77 s whereas the existing model ranges at 8.7 s and at 80th period convergence is attained. This proves that the proposed method is performed at much higher rate for more than 60% in terms of the faster parametric calculations.

Figure 8

Comparison of convergence characteristics.

Conclusions

The architecture for predicting the BDI-II scale and diagnosing depression using handwriting and drawing samples was devised and implemented in this study. The model starts with handwriting and drawing samples as input images, which it then runs through a CNN-based classifier. A cloud-based database was created for this system by collecting handwriting text and drawing samples from individuals. Then they were given a task of drawing a pentagon, a house, a circle, and a clock, as well as writing a sentence in cursive. Furthermore, cross-validation in the dataset is performed with many features and were studied from the following handwriting samples: Reference line, angle, distance, control, restraint, arrangement, separation, and sectors, which distinguishes a handwriting sample from other samples. The generated dataset is pre-processed by rotating, scaling, and translating it in order. The photos are resized and normalized to be compatible with the system. The same dataset is utilized to train the deep neural network model, which involves extracting the features from the input image using the neural network architecture and results in BDI-II scale prediction. A comparison of several ML techniques is also performed, and it is discovered that all parametric values are significantly higher than conventional models. In the future, the same approach might be used with high-resolution photos, allowing all flaws to be identified at an early stage of the transformation.

Data Availability Statement

The original contributions presented in the study are included in the article/supplementary files, further inquiries can be directed to the corresponding author/s.

Ethics Statement

Ethics review and approval/written informed consent was not required as per local legislation and institutional requirements.

Author Contributions

PK and HM: conceptualization and methodology. PK, HM, DS, and H-NL: software. SS: data curation. SS and HM: writing—original draft preparation. HA: visualization and investigation. DS and H-NL: supervision, validation, and writing—reviewing and editing. All authors contributed to the article and approved the submitted version.

Funding

This work was supported in part by the National Research Foundation of Korea (NRF) Grant funded by the Korean government (MSIP) (NRF-2021R1A2B5B03002118). This research was supported by the Ministry of Science and ICT (MSIT), Korea, under the ITRC (Information Technology Research Center) support program (IITP-2021-0-01835) supervised by the IITP (Institute of Information and Communications Technology—Planning and Evaluation).

Conflict of Interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher's Note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.