Indonesian J our nal of Electrical Engineering and Computer Science V ol. 38, No. 2, May 2025, pp. 1231 1244 ISSN: 2502-4752, DOI: 10.11591/ijeecs.v38.i2.pp1231-1244 1231 T extual and numerical data fusion f or depr ession detection: a machine lear ning framew ork Mohammad T ar ek Aziz 1 , T anjim Mahmud 2 , Md F aisal Bin Abdul Aziz 3 , Md Ab u Bakar Siddick 4 , Md. Maskat Sharif 1 , Mohammad Shahadat Hossain 5,6 , Karl Andersson 6 1 Department of Computer Science and Engineering, Chittagong Uni v ersity of Engineering and T echnology , Chittagong, Bangladesh 2 Department of Computer Science and Engineering, Rang amati Science and T echnology Uni v ersity , Rang amati, Bangladesh 3 Department of Computer Science and Engineering, Comilla Uni v ersity , Comilla, Bangladesh 4 Department of Computer Science and T echnology , Beijing Institute of T echnology , Haidian, China 5 Department of Computer Science and Engineering, Uni v ersity of Chittagong, Chittagong, Bangladesh 6 Cybersecurity Laboratory , Lule ˚ a Uni v ersity of T echnology , Sk ellefte ˚ a, Sweden Article Inf o Article history: Recei v ed Mar 28, 2024 Re vised Oct 13, 2024 Accepted Oct 30, 2024 K eyw ords: Chit-squared test CO VID-19 Depression Machine learning algorithms Random forest ABSTRA CT Depression, a wi despread mood disorder , signicantly af f ects global mental health. T o mitig ate the risk of recurrence, early detection is crucial. This study e xplores socioeconomic f actors contrib uting to depression and proposes a no v el machine learning (ML)-based frame w ork for its detection. W e de v elop a tailored questionnaire to collect te xtual and numerical data, follo wed by rigorous feature selection using methods lik e bac kw ard remo v al and Pearson’ s chi-squared test. A v ariety of ML algorithms, including random forest (RF), support v ector ma- chine (SVM), and logistic re gression (LR), are emplo yed to create a predicti v e classier . The RF model achie v es the highest accurac y of 96.85%, highlighting its e f fecti v eness in identifying depression risk f actors. This research adv ances depression detection by inte grating socioeconomic analysis with ML, of fering a rob ust tool for enhancing predicti v e accurac y and enabling proacti v e mental health interv entions. This is an open access article under the CC BY -SA license . Corresponding A uthors: T anjim Mahmud Rang amati Science and T echnology Uni v ersity , Rang amati 4500, Bangladesh Email: tanjim cse@yahoo.com 1. INTR ODUCTION The rapid adv ancement of technology and human skills has profoundly inuenced both ph ysical and mental health [1]. Despite the gro wing a w areness of ph ysical well-being, mental health often remains under - v alued, leading to a rise in mental health issues that e v olv e into disorders, including depression. Depression, a pre v alent and debilitating mental illness, af fects indi viduals of all ages, impairing their emotions, thoughts, and beha viors [2]. In the 2019–2020 academic year alone, a staggering 20.78% of adults e xperienced mental health issues, with o v er 50 million Americans af fected [3]. The global CO VID-19 pandemic has further e xacerbated mental health challenges, with anxiety and hopelessness increasing by 25% w orldwide [4]. Ev en post-reco v ery from the virus, indi viduals may still f ace lingering ef fects contrib uting to depression [5]. The pandemic’ s multif aceted impact, spanning mental, ph ysical, and economic realms, has manifes ted in heightened le v els of anxiety , insomnia, subs tance ab use, and other adv erse outcomes. Socioeconomic f actors such as unemplo yment, f amilial disconnection, and social isolation ha v e compounded feelings of despair and disillusionment [6]. Alarmingly , approximately 280 million indi viduals w orldwide suf fer from depression, with a signicant portion recei ving inadequate treatment, particularly in lo w- and middl e-income countries. J ournal homepage: http://ijeecs.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
1232 ISSN: 2502-4752 The global suicide rate, a tragic consequence of untreated depression, underscores the ur genc y of addressing mental health issues. In the realm of depression research, se v eral studies ha v e tackled v arious aspects of mental health detection using machine learning (ML) techniques. Ferdo wsy et al. [7] focused on predicting obesity risk, emplo ying nine ML algorithms on a dataset comprising 1,100 instances and 28 features. Logistic re gression (LR) emer ged with the highest accurac y of 97.09%, while gradient boosting (GB) classier yielded the lo west accurac y at 64.08%. Feature se lection tech- niques such as correlation and principal component analysis (PCA) were utilized to identify criti cal features. Arif et al. [8] proposed a ML technique for detecting drug addiction in the population of Bangladesh. Nine ML classiers were applied to a dataset containing 510 instances and 23 attrib utes, with LR achie ving the highest accurac y of 97.91%. Correlation and PCA were also emplo yed for feature selection to enhance model performance. Khatun et al. [9] in v estig ated the detection of betel nut addiction using ML te chniques. Among six classiers applied to a dataset comprising 1001 samples and 19 features, random forest (RF) achie v ed the highest accurac y of 99.00%, while Nai v e Bayes (NB) obtained the lo west accurac y at 91.04%. Feature selec- tion methods, including PCA and chi-square, wer e emplo yed for impro v ed a ccurac y . Mia et al. [10] introduced a ML approach to determine the re gistration status of students in pri v ate uni v ersities in Bangladesh. Se v en ML algorithms were applied to a dataset with v e features, achie ving an accurac y of 85.76% with support v ector machine (SVM) and 79.65% with RF . Shahriar et al. [11] aimed to predict vulnerability to drug addiction using ML algorithms . Three classiers were applied to a dataset comprising 498 samples and 60 attrib utes, with RF achie ving the highest accurac y of 94.00%. Lee and Kim [12] focused on predicti ng problematic smartphone use using ML tech- niques. Among three classiers applied to a dataset with 29,712 instances and 27 features, RF achie v ed the highest accurac y of 82.59%, while decision tree (DT) had the lo west accurac y at 74.56%. K e ya et al. [13] emplo yed ML approaches to analyze the performance of g arment w omen’ s w orking status in Bangladesh. Fi v e ML algorithms were applied to a dataset with 512 instances and 13 features, with LR achie ving the highest accurac y of 69%. Go vindasamy and P alanichamy [14] utilized ML techniques on T witter data to detect depres- sion. T w o classiers were applied to datasets with 1,000 and 3,000 instances respecti v ely , with both models achie ving an accurac y of around 92.34% and 97.31%. Sadeque et al. [15] e xplored depression detection in so- cial media by focusing on indi viduals e xhibiting signs of depression within online communities. Amanat et al. [16] emplo yed deep learning algorithms to predict depression, achie ving an accurac y of 99% using long short- term memory (LSTM) and recurrent neural netw ork (RNN) algorithms on te xtual data. Ag ainst this backdrop, our research endea v ors to contrib ute to the early detection of depression us- ing ML algorithms [17], [18] and feature selection techniques [19]. By analyzing a dataset comprising 6,186 instances and 28 features related to depression. W e aim to identify indi viduals at risk of depression and distin- guish them from non-depressed indi viduals. The primary objecti v es of our study include detecting depression using mix ed data (both numerical and t e xt ual), g athering comprehensi v e data through tailored questionnaires, and applying feature selection techniques to enhance model accurac y . The main contrib utions of this study are belo w: i) De v elopment of a combined dataset incorporating both te xtual and numerical data, augmented through o v ersampling and chi-square technique to increase dataset rob ustness. ii) Utilization of the de v eloped dataset to train ML classiers for accurate prediction of depression, achie v ed by con v erting dataset into binary v alues for labeling. iii) Proposal of a RF model, demonstrating s uperior performance with training accurac y of 97.37% and testing accurac y of 96.85%, surpassing other classiers. i v) Application of nine ML classiers to ef fecti v ely detect depression from mix ed data, achie ving accurac y rates e xceeding 78% in most cases. In this study , the proposed system architecture in section 2, the result analysis in section 3, and nally the conclusion and future plan in section 4 are discussed sequentially . 2. METHOD The w orking procedure of the proposed system architecture is illustrated in Figure 1, and an e xplana- tion of the mechanism is described in belo w . Indonesian J Elec Eng & Comp Sci, V ol. 38, No. 2, May 2025: 1231–1244 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 1233 Figure 1. Proposed system architecture of depression detection 2.1. Dataset Data collection w as conducted through online sources including websites, web portals, and soci al media platforms such as F acebook, Instagram, and T witter , utilizing the data-collecting tool F acepager . Addi- tionally , f ace-to-f ace intervie ws were conducted with indi viduals af liated with medical institutions, hospitals, and clinics. This approach in v olv ed interacting with indi viduals who were diagnosed with depression or e x- hibited symptoms of mental disorder , al beit being comparati v ely challenging and time-consuming. Despite the challenges, a total of 520 records were obtained through these tw o methods [20]. T o comprehend the re gional f actors contrib uting to depression, a questionnaire w as de v eloped by re vie wing literature, rele v ant websites, and consulting re gional psychiatrists. Initially , a draft quest ionnaire w as created, which w as later rened to in- clude a total of 28 items. These questions are structured as multiple choice questions (MCQs) with tw o or more possible solutions. A dataset comprising 520 cases w as collected based on the responses to these questions, which serv e as input data for the proposed system. Each question in the questionnaire presents respondents with specic answer options, such as binary choices lik e “yes” or “no”. The data collection process is depicted in Figure 2, encompassing v arious steps outlined belo w . D a t a   C o l l e ct i o n   Me t h o d s On l i n e   S u r v e y Qu e s t i on n a i r e   An a l y s i s R a w   D a t a se t Da ta   P r e p r o ce ssi n g Da ta   A u g me n ta tio n F e a t u r e   A n a l ysi u si n g   C h i - sq u a r e d   T e st E l i m i n a t e d   L e ss  N e ce ssa r Fe a tu re s Figure 2. W orking process of data collection and feature analysis with chi-squared test 2.1.1. Questionnair e analysis W e adminis tered a set of 28 questions to users to g ather data, with the intention of e xtracting i nforma- tion for future analysis based on their responses [21]. These questions, detailed in T able 1, encompass v arious aspects related to the indi vidual’ s circumstances, including inquiries about age, f amily crises, and mental pres- sure [22]. The solutions to these questions typically in v olv e more than tw o options, such as “yes” or “no”, with each response contrib uting to the assessment of the indi vidual’ s lik elihood of e xperiencing depression. F or T e xtual and numerical data fusion for depr ession detection: ... (Mohammad T ar ek Aziz) Evaluation Warning : The document was created with Spire.PDF for Python.
1234 ISSN: 2502-4752 instance, questions pertaining to age, f amily dynamics, and mental well-being are particularly rele v ant indica- tors. F ollo wing the collection of responses, the model can then determine whether the indi vidual e xhibits signs of depression or not. F or a comprehensi v e list of questions, please refer to T able 1, while T able 2 presents an analysis of the corresponding solutions. T able 1. Questions to collect data from person intervie w or website making a dataset S.L User’ s information to nd out depression 1 Y our Age range? 2 Y our gender? 3 Y our occupation? 4 What kind of relationship is between you and your f amily? 5 Can you share your personal matter with your f amily members easily? 6 Do you ha v e a f amily crisis? 7 Is your f amily supporti v e? 8 Do you ha v e f aced an y V iolence in your f amily? 9 Ho w did you spend your lockdo wn time during CO VID 19? 10 During CO VID 19 lockdo wn, did you feel lonely? 11 If you are a student, are you satised with your academic result? 12 If you are a student, Online class during CO VID 19 af fect your mental health? 13 Did you retreat from your study because of CO VID 19? 14 If you are an under graduate, ha v e you f allen in Session clutter because of CO VID 19? 15 Are you afraid to apply abroad for your higher study because of CO VID 19? 16 What kind of relationship is between you and your friends/ colleagues? 17 What is your relationship status? 18 Are you happ y with your partner? 19 Are you tensed about an ything that you cannot for get an ymore? 20 Are you w orried about the uncertainty of getting a job because of CO VID 19? 21 Are you happ y with your current situation? 22 Did you attack yourself for some reason during CO VID 19? 23 Did you loss your closer an yone due to CO VID 19 and you cannot for get this situation till no w? 24 Which things are al w ays pressurized on you? 25 Are you addicted to an y drugs? 26 Ha v e you e v er e xperie nced b ullying from friends or through social media? 27 Do you spend most of your time on Social media? 28 Do you ha v e s leeping (insomnia) problem? T able 2. T op 10 features of collected ra w dataset with top v e v alue Features Data 1 Data 2 Data 3 Data 4 Data 5 Age 18-25 18-25 25-40 18-25 18-25 Gender Female Female Male Male Male Occupation St udent Student Service holder Student Student Relationships Strong Strong Normal Strong Strong Data sharing No Y es Y es Y es Y es F amily-crisis No No No No No Mental pressure Y es No Y es No No Supporti v e-f amily Y es Y es Y es Y es Y es F amily-violence No No No No Y es V iolence pressure No Y es No No Y es 2.1.2. Raw dataset After nishing the data collection, we got a total of 520 records with 28 features. The features are: Age’, ‘gender’, ‘occupation’, ‘relation-with-f amily’, ‘shared-personal-matter -with-f amily’, ‘f amily-crisis’, ‘mental-pressure-due-to-crisis’, ‘supporti v e-f amily’, ‘f aced-V iolence’, ‘mental-pressure-due-to-violence’, ‘spend-lockdo wn’, ‘feel-lonely-during-CO VID19’, ‘Satised-academic-result’, ‘mental-pressure-due-to- online-class’, ‘retreat-from-study-due-to-CO VID19’, ‘mental-pressure-due-to-retreat’, ‘f allen-into-session- clutter’, ‘mental-pressur e-due-to-session-clutter’, ‘afraid-apply-abroad’, ‘relati on-with-friends’, ‘relationship- status’, ‘happ y-with-partner’, ‘tensed-something’, ‘w orried-uncertainty-getting-job-CO VID-19’, ‘mental- pressure-due-to-getting-job’, ‘happ y-with-current-situation’, ‘attack-yourself-in-CO VID-19-situation’, ‘lost- Indonesian J Elec Eng & Comp Sci, V ol. 38, No. 2, May 2025: 1231–1244 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 1235 closer -during-CO VID-19’, ‘pressurized-on-you’, ‘addicted’, ‘b ullied-social media’, ‘mental-pressure-due-to- b ullied’, ‘spend-much-time-social-media’, ‘insomnia-problem’. Only “age” contains the numerical v alue, and the other features contain the te xtual v alue. In most cases, te xtual-based features ha v e tw o v alues: yes or no. The top 10 features are sho wn in T able 2 with their v alues. The other 18 features also ha v e almost the same v alue, lik e data sharing and violent pressure. That means the y contain tw o Boolean v alues: yes and no. So, we ignored adding these features here due to length. 2.2. Data labeling Data labeling, also kno wn as data annotation, plays a pi v otal role in ML tasks [23]. Labeled data is ess ential to pre v ent issues such as o v ertting or undertting, which may occur when using unlabel ed data. W ithout proper labeling, classiers may misclassify or e xhibit reduced accurac y , thereby disrupting the desired outcome [24]. In our proposed method, we opted for binary transformation as mos t features possess tw o distinct v al- ues, such as “yes” or “no”. T able 3 displays the binary-transformed v alues for a total of 28 features, sho wcasing the top 5 v alues. While presenting the top 10 features in this table, we omitted the remaining features as the y e xhibit similar characteristics. T able 3. Binary transformed v alue for some features with top v e v alues Features Data 1 Data 2 Data 3 Data 4 Data 5 Relationships 1 1 0 1 1 Data sharing 0 1 1 1 1 F amily-crisis 0 0 0 0 0 Mental pressure 1 0 1 0 0 Supporti v e-f amily 1 1 1 1 1 F amily-violence 0 0 0 0 1 V iolence pressure 0 0 0 0 1 2.3. Data pr epr ocessing The questionnaire serv es the purpose of g athering essential insights into the primary causes of de p r es- sion pre v alent in Bangladesh. Data coll ection w as conducted through tw o methods: online search and personal intervie ws. Ultimately , a total of 520 instances were amassed. The subsequent step in our methodology in v olv es data preprocessing, which encompasses four fundamental tasks [25], [26]: 1. Null v alues were lled electronically to ensure completeness of the dataset. 2. Missing v alues labeled as “nan” were manually addressed to ensure compatibility with subsequent oper - ations. 3. Some feature v alues were updated to binary “yes” or “no” format to align with the classication method to be used in the model. 4. Redundant v alues were eliminated from the dataset to streamline the analysis process. 2.4. Data augmentation Gi v en the imbalanced nature of our dataset, we emplo yed o v ersampling techniques [27], [28] to rec- tify the imbalance. A commonly used method in the literature for generating additi on a l samples is the synthetic minority o v er -sampling technique (SMO TE). Through o v ersampling, our dataset e xpanded by approximately 6,186 samples, resulting in a well-balanced dataset suitable for training our ML model. Specically , T able 4 sho wcases a selection of se v en randomly chosen features along with their corres po ndi ng matrix format, fol- lo wing o v ersampling. Other features were disre g arded due to their similarity . T able 4. Ov er sampled matrix format of the features Feature Matrix format Gender (610, 26) shared-personal Matter -with-f amily (702, 26) Supporti v e-f amily (1100, 26) Mental Pressure-due-to-violence (1592, 26) Addicted (5846, 26) Insomnia-problem (6258, 26) Attack-yourself-in-CO VID19 situation (2948, 26) T e xtual and numerical data fusion for depr ession detection: ... (Mohammad T ar ek Aziz) Evaluation Warning : The document was created with Spire.PDF for Python.
1236 ISSN: 2502-4752 2.5. Chi-squar ed test analysis and featur e selection W ithout assuming an y particular distrib ution, the Pearson’ s chi-squared test of independence is a sta- tistical technique used to e v aluate the relationship between tw o cate gorical v ariables based on their frequencies [29]. This test enables us to determine whether there is a signicant correlation between the predictor v ariables and the tar get v ariable in our dataset, where all of the v ariables are cate gorical. This test’ s null h ypothesis presupposes that there is no relationship between the v ariables. The chi-squared test, sho wn in Figure 3, can be used to ascertain whether the tar get v ariable and the feature v ariables are dependent on one another . This is accomplished by creating a contingenc y table, which arranges the v ariable frequencies in an or g anized manner and is also referred to as a cross-tab ulation or tw o- w ay table. Based on the v ariations between observ ed and e xpected frequencies in the contingenc y table, the test computes a chi-squared statistic [30]. W e reject the null h ypothesis and conclude that there is a signicant association between the v ariables if the computed chi-squared statistic is greater than a critical v alue determined by the de grees of freedom and selected signicance le v el. In contrast, if the computed statistic is less than the critical v alue, we are unable to reject the null h ypothesis, indicating that there is no meaningful correlation. The analytical form of the chi-squared test statistic for a contingenc y table with r ro ws and c columns is gi v en by: χ 2 = r X i =1 c X j =1 ( O ij E ij ) 2 E ij where: O ij represents the observ ed frequenc y in the i th ro w and j th column of the contingenc y table. E ij represents the e xpected frequenc y in the i th ro w and j th column, calculated under the assumption of independence between the tw o cate gorical v ariables. The chi-squared test statistic follo ws a chi-squared distri b ut ion with de grees of freedom equal to ( r 1)( c 1) under the null h ypothesis of independence between the v ariables. By comparing the calcu- lated chi-squared statistic with the critical v alue from the chi-squared distrib ution, one can determine whether to reject or f ail to reject the null h ypothesis. Using the chi-square test, based on the relationship between each independent feature and tar get v ariable, we selected 25 v ariables. F or one independent v ariable, the null h y- pothesis w as f alse. So we remo v ed this feature. Among the original 26 features in our data collection, the 25 special v ariables used for creation ha v e been e v aluated as ha ving the highest impact. Figure 3. Feature analysis with chi-test for tar get and other v ariables 2.6. Classiers In this study , we applied a total of 9 ML classiers to detect depression from both te xtual and numerical datasets. The classiers are used to nd out the depression with optimum solutions. The used classiers are described belo w . 2.6.1. Random f or est RF is a popular ensemble learning algorithm that constructs multiple DTs during training and com- bines their predictions to reduce o v ertting and impro v e accurac y for both re gression and classication tasks [31], [32]. Figure 4 sho ws w orking process of the proposed RF classiers. Indonesian J Elec Eng & Comp Sci, V ol. 38, No. 2, May 2025: 1231–1244 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 1237 Figure 4. W orking process of the proposed RF classiers 2.6.2. Multilay er per ceptr on The multilayer perceptron (MLP) is a basic articial neural netw ork (ANN) with a layered archite cture consisting of an input layer , one or more hidden layers, and an output layer , enabling it to model comple x non- linear relationships between inputs and outputs. 2.6.3. Decision tr ee The DT algorithm is a popular and straightforw ard ML method that constructs a tree-lik e s tructure for re gression and classication tasks, of fering ease of use, interpretability , and the ability to handle both cate gorical and numerical data. 2.6.4. Logistic r egr ession LR is a statistical technique used for binary classication problems, modeling the probability of a binary outcome based on predictor v ariables and con v erting ra w predictions into probabilities using the logistic function [33]. 2.6.5. Gradient boosting GB is an ensemble learning technique that b uilds a strong predicti v e model by iterati v ely adding weak learners, typically DTs, each impro ving on the residuals of the pre vious models to enhance o v erall accurac y . 2.6.6. Gaussian Nai v e Bay es Gaussian NB is a probabili stic classication algorithm based on Bayes’ theorem, assuming feat ures are independent and normall y distrib uted, making it ef fecti v e for tasks lik e te xt classication and spam ltering, especially with continuous data. 2.6.7. Support v ector machine SVM is a po werful supervised learning algorithm that determines the optimal h yperplane to sepa- rate classes in the feature space, useful for both linear and non-linear data, particularl y in high-dimensional spaces [34]. 2.6.8. K-near est neighbors K-nearest neighbors (KNN) i s a non-parametric ML algorithm that assigns a data point to the m ajority class among its nearest neighbors, commonly used for tasks with non-linear decision boundaries [35]. T e xtual and numerical data fusion for depr ession detection: ... (Mohammad T ar ek Aziz) Evaluation Warning : The document was created with Spire.PDF for Python.
1238 ISSN: 2502-4752 2.6.9. AdaBoost AdaBoost (AD A-B), or adapti v e boosting, is an ensemble technique that combines se v eral weak learn- ers by adjusting their weights based on misclassied instances, enhancing classication performance through iterati v e impro v ement. 3. RESUL TS AND DISCUSSION 3.1. Data splitting The dataset underwent a di vision into training and testing subsets, allocated at an 80% to 20% ratio, respecti v ely . Both subsets utilized binary e xpressed v alues for training and testing purposes. V arious ML techniques were emplo yed to construct a predicti v e binary cla ssier using the training set. These techniques included GB, AD A-B classier , KNN, g aussian NB, SVM, DT , RF , and LR. The selected tar get v ariable for the classier w as “F amily crisis, chosen due to the signicant role of nancial issues in male depression. The model’ s output w as compared ag ainst the tar get v ariable of the test set to e v aluate the accurac y and ef fecti v eness of the classier . 3.2. Used classier parameters W e re vie w the man y parameters utilized to b uild the classiers and run e xperiments in this part. The parameters v ary based on the classiers. See the parameter details in T able 5 that are used in the study . T able 5. Classier parameter details Classiers P arameter LR Dual=F alse, solv er=‘lbfgs’, penalty=‘l2’ DT threshold=‘gini’, maximum depth=25, random state=0 splitter = “optimal” RF The smallest samples are “4 for departs 9 for splits in half, 114 for estimations, and 22 for random states, while the maximum depth is 25 and the maximum features are “log2. SVM random state=0, k ernel=‘rbf KNN weights=‘uniform’ Gaussian NB v arsmoothing=1e-9 MLP Random state = 0, a cti v ation = logistic, solv er = lbfgs GB classiers Loss = “de viance, learning rate = 0.1, and n estimators = 100 AD A-B classier n estimators= 50 3.3. Ev aluation metric In this study , we used classiers to predict sad indi vidual s. W e used the Scikit-learn function accurac y score() [36], [37], which uses the test dataset’ s classiers’ projected results, to determine the accurac y . TP (true positi v e): it indicates the amount of accurately predicted dataset by the classier . FP (f alse positi v e): the amount of dataset had the classier incorrectly predicted as being in good health when the y are depressed. FN (f alse ne g ati v e): the classier mislabeled a number of outcomes as ha ving addictions when, in f act, the y were health y in the sample. TN (true ne g ati v e): the number of outcomes in the dataset is correctly classied as depressed by the classier . Accur acy = T P + T N T P + T N + F P + F N (1) P r ecision = T P T P + F P (2) R ecal l = T P T P + F N (3) F 1 = 2 P r ecision R ecal l P r ecision + R ecal l = 2 T P 2 T P + F P + F N (4) In addition to guring out these numbers, we also measured the A UC v alue and produced the R OC curv e (recei v er operator characteristic). In a R OC diagram, the X- and Y -ax es for the TP rate (recall) and Indonesian J Elec Eng & Comp Sci, V ol. 38, No. 2, May 2025: 1231–1244 Evaluation Warning : The document was created with Spire.PDF for Python.
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752 1239 FP rate (1-TNR) are displayed, respecti v ely . Comparing the potenc y of dif ferent prediction models is done using the tw o-dimensional area under the R OC curv e, or A UC. A higher A UC v alue typically indicates bet- ter prediction performance. The classier that performs t he best is chosen after taking i nto account all of these v ariables. 3.4. Confusion matrix analysis If the e xplanation of error analysis were used, it w ould be simpler to comprehend ho w well the chosen classier performed. A confusion matrix mak es error analysis easier to understand. The v alue of it, the best classier after training with 28 features is displayed in Figure 5. Out of the 1,238 total predictions, 848 re- spondents were correctly recognized as being in good health, while 17 health y respondents were misclassied as being in bad health. In the instance of depression predictions, 41 depressed respondents were mistak enly classed as health y , while the remaining 346 respondents were accurately recognized. Compared to depression responses, the classier mak es fe wer mistak es in the health y class. Figure 5. Confusion matrix of RF classier 3.5. Accuracy scor es Since the feature analysis is performed using the chi-squar ed test, and we selected 28 features based on it, we performed nine algorithms sho wn in T able 6 and analyzed their accurac y score. W e got the highest accurac y in DT classiers at 99.49% for training and 96.44% for testing accurac y . 3.6. R OC and A UC v alue analysis A classier , named RF , has the best accurac y among trained classiers, according to an analysis of the data in T able 6. W e need to look at the R OC signal for e v ery classier to test performance in terms of sensiti vity and specicity in addition to classication accurac y . Comparing classiers w as done using A UC v alues and R OC curv es, which pro vide reliable descriptions of discriminating skills. Accurac y ratings may not fully e xplain the situation, thus A UC v alues are necessary to comprehend ho w the classiers performed in both the health y and addicted classes in our study . By comparing accurac y scores and A UC v alues, we may arri v e at a con vincing decision on which model is best. The results presented in T able 6 and Figure 6 sho w that the RF classier has the highest correlation when we look at the A UC result for the learners who ha v e the highest accurac y scores, with a success rate of 96.85% and a mean A UC of 0.99. Figure 7 sho ws R OC curv es for SVM, KNN, and AD A-B classiers. T able 6. T raining and testing accurac y of the used classiers Classier T raining accurac y % T esting accurac y % LR 85.79 87.80 DT 99.49 96.44 RF 97.37 96.85 SVM 95.98 95.64 KNN 97.33 96.20 Gaussian NB 78.48 79.89 MLP 99.49 96.61 GB classiers 92.92 93.38 AD A-B classier 86.50 88.61 T e xtual and numerical data fusion for depr ession detection: ... (Mohammad T ar ek Aziz) Evaluation Warning : The document was created with Spire.PDF for Python.
1240 ISSN: 2502-4752 Figure 6. R OC curv es for MLP , RF , DT , L T , GB, and g aussian NB classiers Figure 7. R OC curv es for SVM, KNN, and AD A-B classiers 3.7. Pr ecision, r ecall, and F1-scor e analysis of the used classiers The classier with the best accurac y and A UC v alue among all of them is a RF classier trained with 28 feature v ariables. Consider T able 7, which contains e xplanations of each of the classiers mentioned abo v e along with additional metrics lik e recall and precision [38], [39]. T able 7. Precision, recall, and F1-score v alue for all classier Classier Class Precision Recall F1-score LR Health y 0.90 0.92 0.91 Depressed 0.82 0.78 0.80 DT Health y 0.97 0.97 0.97 Depressed 0.94 0.94 0.94 RF Health y 0.97 0.99 0.98 Depressed 0.97 0.92 0.95 SVM Health y 0.96 0.98 0.97 Depressed 0.96 0.90 0.93 KNN He alth y 0.97 0.98 0.97 Depressed 0.94 0.93 0.94 Gaussian NB Health y 0.91 0.79 0.84 Depressed 0.63 0.82 0.72 MLP Health y 0.97 0.98 0.98 Depressed 0.95 0.94 0.94 GB Health y 0.94 0.97 0.95 Depressed 0.93 0.85 0.89 AD A-B Health y 0.90 0.94 0.92 Depressed 0.85 0.76 0.80 Indonesian J Elec Eng & Comp Sci, V ol. 38, No. 2, May 2025: 1231–1244 Evaluation Warning : The document was created with Spire.PDF for Python.