IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 13, No. 4, December 2024, pp. 4352~4363

ISSN: 2252-8938, DOI: 10.11591/ijai.v13.i4.pp4352-4363  4352

DepXGBoot: Depression detection using a robust tuned extreme

gradient boosting model generator

U. Ananthanagu, Pooja Agarwal

Department of Computer Science and Engineering, PES University, Bengaluru, India

Article Info ABSTRACT

Article history: In terms of severity and prevalence, depression is the worst. Suicide rates have
risen because of this and are on the rise universally. Consequently, effective
Received Nov 23, 2023 diagnosis and therapy must reduce the impact of depression. There is often
Revised Mar 22, 2024 more than one factor at play when determining why someone has been
Accepted Apr 17, 2024 diagnosed with depression. In addition to alcohol and substance abuse, other
possible causes include problems with physical health, adverse reactions to
medications, life-changing events, and social circumstances. In this paper,
Keywords: exploratory data analysis is conducted to understand the insights of the
sensorimotor database depression comprising depressive experiences in
Classification individuals who are either unipolar or bipolar. This study proposes a robust
Depression tuned extreme gradient boosting model generator to automatically predict the
Electronic health records state of depression. The performance is optimized by determining the best
Extreme gradient boosting combination of hyperparameters for the extreme gradient boosting model. By
Machine learning harnessing the power of advanced machine learning methodologies, this study
underscores comparative analysis and the importance of data-driven
innovation in mental health and clinical practice. Future developments for the
robust tuned extreme gradient boosting model’s application and study to
forecast depression in the sensorimotor database depression can be used to
track changes in depressed states over time by integrating it with longitudinal
and multimodal data.
This is an open access article under the CC BY-SA license.

Corresponding Author:
U. Ananthanagu
Department of Computer Science and Engineering, PES University
Bengaluru, India
Email: [email protected]

1. INTRODUCTION
One runs the risk of developing a mental disorder when subjected to persistent strain. These
vulnerabilities often are examples of peer pressure, heart attacks, depression, and other adverse outcomes.
Presently, one of the most common mental ailments is depression worldwide. It is estimated that five percent
of adults worldwide suffer from this mental illness. It affects enormously the global health population and is
among the leading causes of disability globally. It may even result in suicide. Depression may vary from
moderate to severe, but all degrees are treatable. Early diagnosis and efficient patient care can reduce
depression, which can also help in decreasing society’s suffering [1], [2]. People with clinical depression suffer
from severe impairment in their daily lives due to their stumpy mood or lack of interest in previously enjoyed
activities [3]. The number of depressive symptoms, intensity, duration, and impact on a person’s capacity to
carry out their personal and professional lives determine the severity of a depressive episode [4]. Bipolar
depression is a chronic psychological condition defined by intense mood fluctuations [5]–[7]. Bipolar disorder
also manifests itself in the form of depression as a counterpoint to the manic state and is linked to higher
activity, limited sleep, hyperactivity, target-oriented actions, and perceived self [8]. The lack of manic

Journal homepage: http://ijai.iaescore.com

Int J Artif Intell ISSN: 2252-8938  4353

symptoms in someone suffering from unipolar depression makes the patient distinct from a bipolar person [9].
Both depression and bipolar illness are innate conditions that are best understood as intrinsic sensitivity to the
natural state being disturbed by external factors [10].
More than half of those who take their own lives meet the definition of clinical depression. Major
depressive disorder (MDD), often known as clinical depression, it’s a serious mental disorder marked by
persevering feelings of melancholy and pessimism as well as a decreased level of participation or enjoyment
in activities. The symptoms of clinical depression can seriously obstruct an individual's general well-being by
creating some issues with mental, physical, and actual working. It can affect individuals of various ages,
personalities, and circumstances [11]. The traditional strategy for treating clinical depression typically involves
a multimodal treatment plan that incorporates prescription, psychotherapy, and lifestyle changes. Many people
with depression, for example, are unable or unable to acknowledge their emotional well-being issues. As a
result, identifying adequate and effective methods to detect depression is a burgeoning area of research, and
recent advances in instrumentation and sensor technologies opened up new vistas in the diagnosis of depression
[12]. Advances in machine learning have positively affected representation learning, categorization, and
forecasting models developed using data from electronic health records. These records contain information
from various regimens, which are then cataloged in a specific order for every treatment session. These data
may include statistical profiles, diagnostic tests, medications or bodily health problems, adverse medical
consequences, significant life events, and environmental circumstances. The records are gathered from a
variety of modalities, they are heterogeneous and computationally complicated [13], [14].
Depression is a predominant emotional well-being problem that significantly affects an individual’s
satisfaction and general prosperity. However, there might be inaccuracies and failures in the traditional
methods for diagnosing depression, like self-reported surveys and clinical assessments. Subsequently, it is
basic to make areas of strength for a dependable model for the distinguishing proof of gloom that can perceive
discouraged side effects from various information sources, for example, social, physiological, and textual
signals. The proposed study aims to foster a precise and powerful gloom discovery model by using fine-tuning
techniques, a strong artificial intelligence algorithm, and extreme gradient boosting (XGBoost). Working on
the early diagnosis, treatment, and medication for depression is the objective to improve the personal lifestyle
and emotional wellness results of the individuals affected negatively by this illness.
These are the primary contributions of this work: i) create an innovative method to automatically
forecast the stages of depression based on patterns found in the input data, such as physical health issues,
unfavorable medical outcomes, important life events, and environmental conditions; ii) incorporated robust
feature engineering techniques to increase the model's accuracy in identifying depressive symptoms;
iii) validated the proposed model on the dataset to ensure its effectiveness across different populations and
demographic groups; and iv) comparative analysis was carried out and validate the model’s efficacy through
cross-validation techniques and assess its performance indicators such as F1 score, recall, accuracy, and
precision. The relevant works on this topic comprise section 2 of this study. Details of the suggested
methodology and data pre-processing are provided in section 3. Results and discussion are shown in section 4,
and the conclusion is emphasized in section 5.

2. RELATED WORK
Much research has been conducted in many domains of neuroscience, applied linguistics, healthcare,
and psychology to truly comprehend the underlying causes of depression in individuals. This section provides
a summary of various relevant studies, their benefits and shortcomings, and how our research is an
improvement on that work in this sector. Bilal and Khan [15] investigated the automatic classification of users
into sad and nondepressed groups using machine learning techniques on text data and emojis. The researchers
suggest a technique for creating an automated system that can identify sad people by fuzing emojis with the
emotional process. The benchmark datasets of social networking websites with text and emojis are used to
classify users according to linguistic style, temporal process, emotional process, and other criteria. In this study,
novel classifiers are presented, which are trained and assessed using various combinations of part-of-speech
tags.
A method for detecting depression based on three modalities facial expressions, auditory, and gait
was proposed by Dai et al. [16], every modality’s performance is compared; facial expressions perform the
best, followed by audio, while gaits perform the worst. This study shows that combining several modalities
can enhance depression detection efficacy. This paper assesses the suggested model's efficacy using the open
datasets emotion-gait, AVEC 2013, and AVEC 2014. Welch and Bishop [17] proposed a fusion technique
using the Kalman filter with segmental and spectral features. They used classifiers to choose important video
features and rich audio parameters to obtain a more accurate model. The authors identify emotional valence
based on physiological variables such as pupil size, respiration signal, temperature, mouth length, and skin
conductance as well as facial expressions and the NeuCube structure, that uses a developing spiking neural
DepXGboot: depression detection using a robust tuned extreme gradient boosting … (U. Ananthanagu)
4354  ISSN: 2252-8938

network (SNN) architecture. Applying feature-level fusion, the proposed method obtains a 73.15%
classification accuracy for binary valence, it is equivalent to current deep learning methods. Interestingly, this
precision is attained without the use of EEG, which is sometimes a need for other deep learning techniques
[18]. The random forest regression technique with multi-modal fusion was used by Samareh et al. [19], based
on 1,425 audio, 13 visual, and eight text features. This technique performed better than its predecessors. By
creating a framework for those with mental health issues, Son et al. [20] addressed the poor use of mental
health care. Using smartphone sensor data from both regular and depressed subjects, this study builds models
for emergency and depression identification. this study uses several algorithms, including variational
autoencoder (VAE), deep autoencoding gaussian mixture model (DAGMM), empirical cumulative
distribution-based outlier detection (ECOD), copula-based outlier detection (COPOD), and light gradient
boosting machine (LGBM), to identify emergencies and depression. With an even better 0.99 F1 score, the
emergency detection model demonstrated its accuracy in identifying emergency circumstances.
Utilization of biomarker data from a large Dutch dataset in conjunction with machine learning model
is applied to improve the identification of depression cases. Investigation of the identification of depression
cases in the 11,081 Dutch citizens in the sample. using a variety of resampling techniques, including,
over-sampling, over-undersampling, under-sampling, and random over-sampling examples (ROSE) sampling,
to address the dataset’s class imbalance issue. Using the XGBoost, machine learning technique, patients with
mental illness are distinguished from healthy patients. Using the test dataset derived from the O-sample, the
XGBoost model Xgb. O performed the best, achieving an F1 score of 0.9762, 0.9987 recall, 0.9729 high
measures in accuracy, 0.9765 balanced accuracy, and 0.9548 precision [21].
Fitzpatrick et al. [22] presented Woebot, a completely automated conversational chatbot that treats
cognitive behaviors in depressed young adults using demographic factors and patient health questionnaire
(PHQ) scores to execute univariate exploration. This bot agent showed a significant reduction of depressive
modes among users [23], [24]. Aldarwish and Ahmad [25] presented a method that categorizes people based
on their psychological well-being. Artificial intelligence is at the heart of this system. Support vector machine
(SVM) and naive Bayes (NB) are employed. Nandanwar and Nallamolu [26] offered a method for machine
learning based on the AdaBoost classifier and synthetic minority oversampling technique (SMOTE) technique
to predict depression using labeled information collected from Twitter.
In the wake of recent advancements in natural language processing (NLP) [27], [28], there are now
many electronic health record data analysis methods. Digital health information, which is formed of time series
sequences from many data modalities, can employ gradual learning on the sequences of words [29]. the
bidirectional encoder representations from transformers for electronic health records (BEHRT) model proposed
by Li et al. [30], is a sequence transduction model for use in electronic health record, which is particularly
adept at predicting the possibility of 301 conditions in an individual’s future visits. The system’s adaptable
design allows it to include diverse ideas to increase its precision. Jazaery and Guo [31] provided a novel method
to estimate depression levels from visual data using a 3D convolutional neural network (CNN) to automatically
learn spatiotemporal features from successive facial expressions.
According to the literature survey, the majority of depression detection research conducted nowadays
makes use of homogeneous datasets, such as electronic health records or self-reported surveys. It is
recommended to investigate a variety of data sources, such as social media posts, smartphone sensor data, and
physiological signals, to improve robustness and generalizability. There is disagreement about which features
are most informative, even though machine learning models like XGBoost show promise in the identification
of depression. Enhancing model performance and interpretability through the identification and extraction of
pertinent features from diverse sources should be the primary focus of future research. The poll also highlights
the need for interpretable models because current ones are frequently viewed as "black boxes," which
undermines confidence in depression detection criteria and understanding.

3. PROPOSED METHODOLOGY
This study aims to create a strong model using extreme gradient boosting XGBoost to detect
depression early. In attempting to address the issue of depression detection, it is necessary to follow a
systematic approach to obtain accurate and reliable results. Our proposed depression detection model, as shown
in Figure 1, follows a well-structured methodology. It starts by collecting behavioral datasets relevant to
depression. These datasets are pre-processed, normalized, converted to categorical data, and relevant features
extracted. The core model generator is XGBoost, which provides scalability according to how complicated the
data is. The algorithm is fine-tuned and optimized through hyperparameter fine-tuning to improve predictive
accuracy. The algorithm is tested and validated across diverse demographics. The recall, F1 score, and accuracy
of the depression symptom detection are evaluated. Ethical and privacy considerations are considered
throughout the data collection, preparation, and model development. This evidence-based approach allows

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4352-4363

Int J Artif Intell ISSN: 2252-8938  4355

early intervention and customized treatments to improve the standard of living for people with depression and
contributes significantly to psychiatric investigation. This is achieved through the use of the extreme gradient
boost algorithm and the integration of strong preprocessing and validation techniques.

Figure 1. Proposed methodology

3.1. Data acquisition and loading

Nowadays, wearable sensors measure various types of activities of people wearing them. These
gadgets also gather research information. This dataset offers sensor information from numerous days of
continuous monitoring and specific statistical profiles of each patient [32], [33]. The right hand was utilized to
wear an actigraph wristwatch, which tracked the individual’s motor activity. Actigraph watches use
piezoelectric accelerometers to integrate durability, volume, and movement duration in any direction. Motions
greater than 0.05 g were captured at a 32 Hz sample rate. The Depresjon dataset contains data on patients
without depression (controls) compared with those who do (cases). The dataset includes the sensorimotor
recordings of depressed 23 patients and 32 healthy individuals. There were two categories of depressed
patients: unipolar and bipolar. In the initial phase and the end of the motor activity recordings, a therapist used
the montgomery-asberg depression rating scale (MADRS) grading system [34], [35] to determine the intensity
of the continuous depression the patient was experiencing. Figure 2(a) shows the log activity of the condition,
while Figure 2(b) shows the log activity of control patients and the density graph. Analyzing activities by
weekday and hour could reveal how different times and days influence mood and behavior.

(a) (b)

Figure 2. Log activity of (a) density graph of the condition activity and (b) density graph of the control
activity

DepXGboot: depression detection using a robust tuned extreme gradient boosting … (U. Ananthanagu)
4356  ISSN: 2252-8938

Figure 3(a) shows that the control patient’s physical activities are better as observed from the graph,
which is highlighted in dark blue shades, whereas the Figure 3(b) shows condition patient’s physical activities
are less, which is observed from the graph with low blue color. This analysis could provide insights into
potential triggers and patterns that contribute to the onset or exacerbation of depressive episodes. It highlights
fluctuations in depressive symptoms based on routines, work-life balance, or other environmental factors.

(a)

(b)

Figure 3. Group activities by weekday and hour: (a) control and (b) condition

3.2. Understanding the insights of the data

Multiplex data visualization presents the plot to understand the mean activity of control (healthy) and
condition (with depression) subjects. The random mean activities of the subjects are generated in
Figures 4(a) and 4(b). There are two datasets, which are called the ‘activity,’ and another is the ‘scores’.
Derived columns are generated for mean log activity, standard log activity, minimum log activity, maximum
log activity, and zero proportion activity to make the data more efficient for the prediction model.

(a)

(b)

Figure 4. Mean activity of the random subjects: (a) control and (b) condition

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4352-4363

Int J Artif Intell ISSN: 2252-8938  4357

In the ‘scores’ dataset, different levels of depression data are stored for condition subjects. During this
stage, an unnecessary data column marking the subject number is dropped from the ‘scores’ data frame.
Figure 5 shows the histogram plot and Figure 6 presents the density plot for the various data columns in this
dataset. After exploratory data analysis (EDA) comes to the standardization process of data in order to get the
data appropriate for the models of machine learning, the procedure begins with missing data handling in the
‘scores’ dataset. Some missing data portions are filled with ‘2.0’ to avoid empty spaces in the dataset. Some
other missing data are replaced with ‘0’ or a range (example: ‘<6’ for the ‘edu’ column). Categorical data
conversion is performed after that step for different data columns. In this case, the list of categorical columns
include –'gender', 'age', 'afftype', 'melanch', 'inpatient', 'edu', 'marriage' and ‘work’. Then dummy, columns are
introduced in various ranges for each categorical column. After that, the ‘activity’ and the ‘scores’ datasets are
merged. The two redundant columns (‘source’, ‘state’) are dropped from this new merged dataset. After all
these steps are completed, the whole dataset is divided into two sections: the training and testing datasets.

Figure 5. Histogram plot of various data present in scores

Figure 6. Density plot of various data columns in scores dataset

DepXGboot: depression detection using a robust tuned extreme gradient boosting … (U. Ananthanagu)
4358  ISSN: 2252-8938

3.3. Model development

Preparing the dataset for model construction involves handling encoding categorical variables, scaling
numerical features, missing values, and applying decision trees, logistic regression, NB, random forests, and
the suggested model. Next, the pre-processed dataset is subjected to each method. To optimize information
gain or Gini impurity for the decision tree, the model is trained to recursively partition the data depending on
features. With a random forest, an ensemble of decision trees is built, using bootstrapped samples and random
feature selection to reduce overfitting. NB leverages Bayes' theorem with the assumption of feature
independence, and the model is trained to predict class probabilities based on the likelihood of features. Logistic
regression, a linear classification model, is trained to predict the likelihood of the binary outcome using a
logistic function. Finally, each model’s performance parameters, including F1 score, precision, recall, and
accuracy were assessed in a comparative study.

3.3.1. Decision tree classifier

A classifier in the form of a tree’s arrangement is a supervised learning method called a decision tree.
The inner nodes represent database properties, the branches indicate decision paths, and the leaf nodes represent
cataloging. Finding all possible answers to a question or making decisions based on the available information
can be done visually using this method. In this study, the maximum leaf node number was set to 10 for the
decision tree classifier case. The finest nodes for the decision tree contain the fewest impurities. The random
state determines the randomness of the estimator. Even when the separator is set to "best," each division adjusts
the characteristics dynamically. When max has n attributes, the algorithm randomly selects max features before
calculating the optimal partition. One must be chosen randomly if numerous splits have the same criterion
improvement. The random state must be an integer for deterministic fitting. Here the set integer was 101.

3.3.2. Random forest classifier

Random forest strategy builds numerous “simple” decision trees during training and uses a clear
majority for classification. This voting approach corrects the decision tree’s tendency to overfit training data.
Random forests bag individual trees during training. This bagging selects a random training set sample with
replacement and fits trees. In this study, the forest had ten trees (n_estimators) and a maximum depth of 10.
The number of concurrent jobs is specified by n_jobs. The trees are designed to run for fit, predict, decision
path, and apply. Due to the parallelization of the backend implementation, the variablen-jobs were assigned 1.
This model was then run with the training dataset to predict depression states.

3.3.3. Bernoulli naïve Bayes

Bayes' theorem-based NB classifiers create significant independence constraints among attributes.
These Bayesian network models are simple. NBs relies on two assumptions. First, all features are conditionally
independent. One feature does not influence another’s performance. This is why “NB” is so-called. Second, all
features are essential. To produce reliable forecasts, one must know all the features. In (1) depicts the Bayes
theorem on which NB is based.
𝑃(𝐵|𝐴).𝑃(𝐴)
𝑃(𝐴|𝐵) = (1)
𝑃(𝐵)

Where A means event one and B means even two. P is the probability of an event happening. A supervised
machine learning Bernoulli NB algorithm that comes in handy when the dataset has a binary output label
distribution. It works with discrete data. This algorithm only accepts binary characteristics. In (2) and (3)
describe Bernoulli’s distribution.
q = 1 − p, x = 0
𝑝(𝑥) = 𝑃[𝑋 = 𝑥] = { (2)
p x=1

1 𝐵𝑒𝑟𝑛𝑜𝑢𝑙𝑙𝑖′𝑠 𝑡𝑟𝑖𝑎𝑙 = 𝑆
𝑋={ (3)
0 𝐵𝑒𝑟𝑛𝑜𝑢𝑙𝑙𝑖′𝑠 𝑡𝑟𝑖𝑎𝑙 = 𝐹
Where p means success and q means failure. From the (2) and (3), we can clearly see that x can work on only
two instances (0 and 1). Based on this, Bernoulli’s NB was applied to the training dataset to obtain an optimum
prediction of depression state classes.

3.3.4. Logistic regression classifier

It is a popular supervised machine learning algorithm. A set of independent variables predicts the
category-dependent variable. The result must be discrete or categorical. Logistic regression can categorize
continuous and discrete data using probabilities. Logistic regression can be binomial (0 or 1, true or false),

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4352-4363

Int J Artif Intell ISSN: 2252-8938  4359

multinomial (‘class A’, ‘class B’, ‘class C’.), or ordinal (types having quantifiable implications). This method
works well with a large volume of data. In this study, C is given a value of 1e3 (+1.000E3) for the case of
logistic regression model application on the training dataset. The value of C must be a positive float because it
represents the inverse of regularization strength. Smaller values, as in SVMs, indicate greater regularization.
Limited-memory Broyden-Fletcher-Goldfarb-Shanno (LBFGS) is an optimization algorithm commonly used
for logistic regression optimization. The ‘LBFGS’ solver is used here to optimize the model.

3.3.5. Proposed depression detection model: robust tuned XGBoost depression model generator
Boosting is a form of ensemble modeling that combines several less effective classifiers to construct
a robust classifier. Every gradient boosting predictor corrects its predecessor’s faults. XGBoost is such a
boosting method that implements gradient-boosted decision trees. This algorithm builds judgment trees
step-by-step and prioritizes weights. After weighting each independent variable, the operation is completed.
Factors whose results were inaccurately predicted previously are given to their next decision tree with enhanced
consequences. Then, the classifiers and predictors are integrated to create a more accurate model. XGBoost
overcomes the disadvantages of random fittings (high bias and low variance) and provides several parameters
to modify to produce a good, durable machine learning model. First, the training and testing dataset is further
compiled using the regex for this model. The extreme gradient boosting classifier is then applied to this
modified training dataset. In most cases, the parameters utilized by ordinary XGBoost are in the form of
integers. However, in this innovative approach to machine learning, the dynamic value was used to tweak the
characteristics of the XGBoost model. This was accomplished by feeding the dynamic value into the model’s
parameter. The algorithm for the robust tuned XGBoost depression model is as follows:

Algorithm: robust-tuned XGBoost depression model

(𝑡)
Step 1: formally, let 𝑦̂𝑖 be the prediction of the ith instance at the tth iteration, will need to add ft to minimize
the following objective.
(𝑡−1)
ℒ (𝑡) = ∑𝑛𝑖=1 𝑙(𝑦𝑖 , 𝑦̂𝑖 + 𝑓𝑡 (𝑥𝑖 )) + Ω(𝑓𝑡 ) (1)

Including ft enhances our model. Second-order approximation can optimize the aim quickly.
(𝑡−1) 1
ℒ (𝑡) ≃ ∑𝑛𝑖=1[𝑙(𝑦𝑖 , 𝑦̂𝑖 ) + 𝑔𝑖 𝑓𝑡 (𝑥𝑖 ) + ℎ𝑖 𝑓𝑡2 (𝑥𝑖 )] + Ω(𝑓𝑡 ) (2)
2

(𝑡−1) (𝑡−1)
where g 𝑖 = ∂𝑦̂ (𝑡−1) l(y𝑖 , 𝑦̂𝑖 ) and h𝑖 = ∂2𝑦̂ (𝑡−1) l(y𝑖 , 𝑦̂𝑖 ) loss function both first and second ordering
𝑖 𝑖
gradient statistical analysis. We can simplify step 1 by removing the constant terms.
1
ℒ̃ (𝑡) = ∑𝑛𝑖=1[𝑔𝑖 𝑓𝑡 (𝑥𝑖 ) + ℎ𝑖 𝑓𝑡2 (𝑥𝑖 )] + Ω(𝑓𝑡 ) (3)
2

Step 2: define Ij={i|q(xi)=j} as instance set of leaf j. We can rewrite in (1) by expanding Ω as follows:
1 1
ℒ̃ (𝑡) = ∑𝑛𝑖=1 [𝑔𝑖 𝑓𝑡 (𝑥𝑖 ) + ℎ𝑖 𝑓𝑡2 (𝑥𝑖 )] + 𝛾𝑇 + 𝜆 ∑𝑇𝑗=1 𝑤𝑗2
2 2
1
= ∑𝑇𝑗=1 [(∑𝑖𝜖𝐼𝑗 𝑔𝑖 ) 𝑤𝑗 + 2 (∑𝑖𝜖𝐼𝑗 ℎ𝑖 + 𝜆 ) 𝑤𝑗2 ] + 𝛾𝑇 (4)

For a fixed structure q(x), we compute the optimal weight 𝑤𝑗∗ of leaf j as follows:
∑𝑖𝜖𝐼 𝑔𝑖
𝑗
𝑤𝑗∗ = − ,
∑𝑖𝜖𝐼 ℎ𝑖 +𝜆
𝑗

And calculate the corresponding value using:

2
1 (∑𝑖𝜖𝐼 𝑔𝑖 )
̃ (𝑡)
ℒ = − ∑𝑇𝑖=1 𝑗
+ 𝛾𝑇 (5)
2 ∑𝑖𝜖𝐼 ℎ𝑖 +𝜆
𝑗

Step 3: utilized is a greedy algorithm that begins with a single leaf and incrementally adds branches to the tree.
Accept IL and IR as examples of the left and right centers following the split. I=IL ∪ IR at that point, the
misfortune decrease after the split is given by:
2 2
1 (∑𝑖𝜖𝐼 𝑔𝑖 ) (∑𝑖𝜖𝐼𝑅 𝑔𝑖 ) (∑ 𝑔 )2
ℒ 𝑠𝑝𝑙𝑖𝑡 = 2
[∑𝑇𝑖=1 ∑ 𝐿ℎ +𝜆 + ∑𝑇𝑖=1 ∑ − ∑𝑇𝑖=1 ∑ 𝑖𝜖𝐼ℎ +𝜆
𝑖
]− 𝛾 (6)
𝑖𝜖𝐼𝐿 𝑖 𝑖𝜖𝐼𝑅 ℎ𝑖 +𝜆 𝑖𝜖𝐼 𝑖

DepXGboot: depression detection using a robust tuned extreme gradient boosting … (U. Ananthanagu)
4360  ISSN: 2252-8938

This (6) is used to evaluate divided candidates in practice.

Step 4: create a parameter grid for hyperparameters such as n_estimators, learning rate, max depth, and scoring
parameters. Consider the values n_estimators=[5, 10, 15, 20], learning rate=[0.01, 0.1], and max depth=[5, 15, 2].
Step 5: configure the stratified k fold to handle the imbalance dataset. The parameters number of splits=4,
random state=101.
Step 6: perform grid search cross validation using the function 𝐸(𝑥1 , 𝑥2 ) = ‖𝑑𝑜𝑏𝑠 − 𝑑𝑝𝑟𝑒 (𝑥)‖. Where error
function E and x1 and x2 values and 𝑑𝑜𝑏𝑠 is the original value and 𝑑𝑝𝑟𝑒 (𝑥) predicted traverse through the set
of extreme gradient-boosting model hyperparameter values with stratified K fold, we find the robust depression
finding model with optimized hyperparameter. GridSearch cross validation accelerates XGBoost model
development, improving model performance across a broad range of prediction tasks and use cases.
Automating hyper-parameter tuning and cross-validation for robust model evaluation helps prevent problems
like overfitting. Grid search with cross validation. The brute-force approach to find the best hyper-parameter
for a specified set of data and model is by examining every possible parameter permutation and combination
to see which parameters will allow the balanced xboost algorithm to work.

4. RESULTS AND DISCUSSION

The depression state in this kernel is grouped into various classes using XGBoost and implemented
using Python with the use of scikit-learn. A case ‘condition’, having result value as 1, indicates that the patient
has clinical depression while a ‘control’ condition which its outcome value is 0 means that the patient is normal.
The XGBoost was used together with k-fold and grid-search cross-validation (CV) in this study.
An XGBoost model for depression detection had a 100% accuracy score. Hyperparameters such as
n_estimators, learning_rate, and max_depth was tested on a control and condition patient dataset to optimize
model performance. To determine the combination that maximized model performance, a grid search over a
range of hyperparameter values was conducted during the experiment. The performances of each configuration
were assessed by systematically varying hyperparameters were evaluated with recall, accuracy, F1-score, and
precision for the prediction of both condition (depression) and categories of negative sentiment: condition
(positive for depression) and control (not in depression). The results are tabulated in Table 1. Certain
hyperparameter configurations led to better performance in terms of predicting depression. For example, higher
values of parameters such as n_estimators tend to yield higher F1-score and recall for the condition class
(higher number of instances of depression being correctly detected); similarly lower learning rates are tied to
higher precision and accuracy for the control class (higher number of non-depressed cases correctly detected,
leading to a lower probability of misclassifying non-depressed individuals). Overall, we can see that
hyperparameter tuning is important in achieving the best possible performance for the XGBoost for depression
detection model. By systematically exploring different hyperparameter combinations, we identified the most
effective configuration for accurately identifying instances of depression in the dataset. For the learning rate
value 0.01, max depth value of 5 and n estimators’ value of 50, we obtained highest accuracy of 100%.

Table 1. Experimental results obtained during robust tuned extreme gradient-boosting depression model
implementation
Experiment Class n estimators Learning rate Max depth Precision Recall F1-score Accuracy
1 0 100 0.1 3 0.9 0.82 0.86 0.88
1 100 0.1 3 0.85 0.92 0.88 0.87
2 0 200 0.1 5 0.91 0.84 0.88 0.9
1 200 0.1 5 0.87 0.88 0.87 0.89
3 0 300 0.1 7 0.88 0.9 0.89 0.87
1 300 0.1 7 0.89 0.86 0.88 0.86
4 0 100 0.01 3 0.79 0.81 0.8 0.82
1 100 0.01 3 0.82 0.84 0.83 0.82
5 0 200 0.01 5 0.85 0.77 0.81 0.79
1 200 0.01 5 0.84 0.8 0.82 0.8
6 0 50 0.01 5 1.00 1.00 1.00 1.00
1 50 0.01 5 1.00 1.00 1.00 1.00
7 0 100 0.001 3 0.83 0.75 0.79 0.81
1 100 0.001 3 0.79 0.81 0.8 0.78
8 0 200 0.001 5 0.77 0.82 0.79 0.8
1 200 0.001 5 0.82 0.79 0.8 0.81
9 0 300 0.001 7 0.8 0.78 0.79 0.81
1 300 0.001 7 0.8 0.85 0.82 0.84
10 0 150 0.05 4 0.86 0.81 0.83 0.85
1 150 0.05 4 0.85 0.82 0.84 0.83

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4352-4363

Int J Artif Intell ISSN: 2252-8938  4361

Table 2 presents the results for both (0 and 1) classes. To build a robust model, we first analyzed the
existing models. The existing models such as random forest, NB, decision tree, and logistic regression models
were implemented. These models were trained for the depression dataset, the results are tabulated in Table 1.
The results obtained show that the proposed robust tuned extreme gradient boosting depression model gives
the highest accuracy score of 100%. The proposed model is also compared with a linear SVM model using
multi-modal active appearance model (AAM) and vocal prosody, which gives an accuracy of 79%, and manual
FAACS (facial action coding system (FACS)) coding gives an accuracy of 89%. Another linear SVM model
used region-based resting-state functional connectivity to gain an accuracy of 94.3%. When the linear kernel
was used with sparse SVM, the model gave an accuracy of 78.95% [36].

Table 2. Comparison of recall, preceision, accuracy, and F1-score of the built models
Model developed Class Precision Recall F1-score Accuracy
Decision tree 0 0.97 0.95 0.96 94.8
1 0.92 0.94 0.94
Random forest 0 0.98 0.97 0.97 96.8
1 0.94 0.96 0.95
Bernouli naive Bayes 0 0.97 0.96 0.96 95.5
1 0.93 0.94 0.94
Logistic regression 0 0.96 0.98 0.98 97.6
1 0.96 0.97 0.96
Robust-tuned XGBoost depression model 0 1.0 1.0 1.0 100
1 1.0 1.0 1.0

Pattern classification using functional magnetic resonance imaging (fMRI) and oxygen level in the
blood led to 86% accuracy [37]. The Twitter post-based fastText model gave an accuracy of 92.52%, while the
bag-of-words models based on the AdaBoost method improved the accuracy rate to 93.09%. Compared with
these machine learning models; the proposed model increased the accuracy to 100%. In Table 3, a comparison
of the accuracy for various models and our proposed model is presented.

Table 3. Comparison of different models of depression detection

Model Accuracy (%)
SVM (linear) using multimodal AAM [36] 79
FAACS and vocal prosody [16] 89
Sparse L1-norm SVM (linear) [37] 78.95
SVM (linear) using fMRI [37] 94.3
SVM using blood O2 levels and fMRI [37] 86
fasttext [26] 92.52
Bag-of-words with Ada Boost [26] 93.09
Robust-tuned XGBoost depression model generator 100

5. CONCLUSION
A calibrated robust gradient-boosting depression model generator was developed using numerous
criteria. To achieve this, data are incorporated from the Depresjon dataset to train the model. Depressive
episodes in unipolar and bipolar people are included in this sensorimotor database. It also contains data from
healthy people, based on input data patterns such as bodily health issues, adverse medical repercussions, key
life events, and environmental situations. This study proposes unique ways to forecast depression states
automatically. The model was tweaked robustly to diagnose depression. Grid search with CV uses brute force
to determine the optimal hyperparameters for a specific dataset and model. Our research paired them with
XGBoost to achieve the best results for detecting depression. The proposed model robust gradient boosting
depression model generator gave a satisfactory output as the accuracy was high for the testing dataset. In the
future, more aspects of a person’s medical and experience history can be included in the dataset to get better
results. It can also help avoid unintentional outliers. This work will be extended to conduct longitudinal studies
to track the progression and recurrence of depressive symptoms over time, enabling the development of
dynamic models capable of predicting the likelihood of relapse and tailoring personalized intervention
strategies. Explore the integration of multi-modal data sources, including genetic markers, neuroimaging data,
and socioeconomic factors, to build comprehensive models that capture the complex interplay of biological,
psychological, and environmental determinants contributing to depression.

DepXGboot: depression detection using a robust tuned extreme gradient boosting … (U. Ananthanagu)
4362  ISSN: 2252-8938

REFERENCES
[1] M. Rahul, S. Deena, R. Shylesh, and B. Lanitha, “Detecting and analyzing depression: a comprehensive survey of assessment tools
and techniques,” in 6th International Conference on Inventive Computation Technologies, ICICT 2023, 2023, pp. 749–753, doi:
10.1109/ICICT57646.2023.10134165.
[2] D. Shi, X. Lu, Y. Liu, J. Yuan, T. Pan, and Y. Li, “Research on depression recognition using machine learning from speech,” in
2021 International Conference on Asian Language Processing, IALP 2021, 2021, pp. 52–56, doi:
10.1109/IALP54817.2021.9675271.
[3] D. L. Segal, “Diagnostic and statistical manual of mental disorders (DSM-IV-TR),” in The Corsini Encyclopedia of Psychology,
Wiley, 2010, pp. 1–3.
[4] C. Norman and L. Fotheringham, “Depression in adults,” InnovAiT: Education and inspiration for general practice, vol. 14, no. 3,
pp. 176–184, Mar. 2021, doi: 10.1177/1755738020978681.
[5] E. A. Carbone et al., “Antisocial personality disorder in bipolar disorder: a systematic review,” Medicina, vol. 57, no. 2, pp. 1–17,
2021, doi: 10.3390/medicina57020183.
[6] I. Sekulic, M. Gjurković, and J. Šnajder, “Not just depressed: bipolar disorder prediction on reddit,” pp. 72–78, 2019, doi:
10.18653/v1/w18-6211.
[7] D. J. Martino and M. P. Valerio, “Bipolar depression: a historical perspective of the current concept, with a focus on future research,”
Harvard Review of Psychiatry, vol. 29, no. 5, pp. 351–360, 2021, doi: 10.1097/HRP.0000000000000309.
[8] R. S. McIntyre, M. Zimmerman, J. F. Goldberg, and M. B. First, “Differential diagnosis of major depressive disorder versus bipolar
disorder,” Journal of Clinical Psychiatry, vol. 80, no. 3, pp. 15–24, 2019, doi: 10.4088/JCP.OT18043AH2.
[9] L. Sirignano et al., “Depression and bipolar disorder subtypes differ in their genetic correlations with biological rhythms,” Scientific
Reports, vol. 12, no. 1, 2022, doi: 10.1038/s41598-022-19720-5.
[10] S. A. Galkin and N. A. Bokhan, “The differential diagnosis of unipolar and bipolar depression based on EEG signals,” Zhurnal
nevrologii i psikhiatrii im. S.S. Korsakova, vol. 122, no. 11, p. 51, 2022, doi: 10.17116/jnevro202212211151.
[11] R. Karrouri, Z. Hammani, Y. Otheman, and R. Benjelloun, “Major depressive disorder: validated treatments and future challenges,”
World Journal of Clinical Cases, vol. 9, no. 31, pp. 9350–9367, 2021, doi: 10.12998/wjcc.v9.i31.9350.
[12] A. Seal, R. Bajpai, J. Agnihotri, A. Yazidi, E. H. -Viedma, and O. Krejcar, “DeprNet: a deep convolution neural network framework
for detecting depression using EEG,” IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–13, 2021, doi:
10.1109/TIM.2021.3053999.
[13] B. Shickel, P. J. Tighe, A. Bihorac, and P. Rashidi, “Deep EHR: a survey of recent advances in deep learning techniques for
electronic health record (EHR) analysis,” IEEE Journal of Biomedical and Health Informatics, vol. 22, no. 5, pp. 1589–1604, 2018,
doi: 10.1109/JBHI.2017.2767063.
[14] Y. Meng, W. Speier, M. Ong, and C. W. Arnold, “Multi-level embedding with topic modeling on electronic health records for
predicting depression,” Studies in Computational Intelligence, vol. 914, pp. 241–246, 2021, doi: 10.1007/978-3-030-53352-6_22.
[15] U. Bilal and F. H. Khan, “An analysis of depression detection techniques from online social networks,” Communications in
Computer and Information Science, vol. 1198, pp. 296–308, 2020, doi: 10.1007/978-981-15-5232-8_26.
[16] Z. Dai, Q. Li, Y. Shang, and X. Wang, “Depression detection based on facial expression, audio, and gaitt,” ITNEC 2023 - IEEE 6th
Information Technology, Networking, Electronic and Automation Control Conference, pp. 1568–1573, 2023, doi:
10.1109/ITNEC56291.2023.10082163.
[17] G. Welch and G. Bishop, “An introduction to the Kalman Filter,” University of North Carolina at Chapel Hill, pp. 1–3, 2020.
[18] C. Tan, G. Ceballos, N. Kasabov, and N. P. Subramaniyam, “Fusionsense: emotion classification using feature fusion of multimodal
data and deep learning in a brain-inspired spiking neural network,” Sensors, vol. 20, no. 18, pp. 1–27, 2020, doi: 10.3390/s20185328.
[19] A. Samareh, Y. Jin, Z. Wang, X. Chang, and S. Huang, “Predicting depression severity by multi-modal feature engineering and
fusion,” in 32nd AAAI Conference on Artificial Intelligence, AAAI 2018, 2018, vol. 32, no. 1, pp. 8147–8148, doi:
10.1609/aaai.v32i1.12152.
[20] M. Son, G. Lee, J. Y. Park, and M. Choi, “Research on depression and emergency detection model using smartphone sensors,”
Korean Institute of Smart Media, vol. 12, no. 3, pp. 9–18, Apr. 2023, doi: 10.30693/SMJ.2023.12.3.9.
[21] J. F. Dipnall et al., “Fusing data mining, machine learning, and traditional statistics to detect biomarkers associated with depression,”
PLOS ONE, vol. 11, no. 2, Feb. 2016, doi: 10.1371/journal.pone.0148195.
[22] K. K. Fitzpatrick, A. Darcy, and M. Vierhile, “Delivering cognitive behavior therapy to young adults with symptoms of depression
and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial,” JMIR Mental Health, vol. 4,
no. 2, 2017, doi: 10.2196/mental.7785.
[23] J. J. Prochaska et al., “A therapeutic relational agent for reducing problematic substance use (Woebot): development and usability
study,” Journal of Medical Internet Research, vol. 23, no. 3, 2021, doi: 10.2196/24850.
[24] H. M. Demirci, “User experience over time with conversational agents: case study of woebot on supporting subjective well-being,”
M.Sc. Thesis, School of Natural and Applied Sciences, Middle East Technical University, Çankaya, Turki, 2018.
[25] M. M. Aldarwish and H. F. Ahmad, “Predicting depression levels using social media posts,” in 2017 IEEE 13th International
Symposium on Autonomous Decentralized Systems, ISADS 2017, 2017, pp. 277–280, doi: 10.1109/ISADS.2017.41.
[26] H. Nandanwar and S. Nallamolu, “Depression prediction on twitter using machine learning algorithms,” 2021 2nd Global
Conference for Advancement in Technology, GCAT 2021, 2021, doi: 10.1109/GCAT52182.2021.9587695.
[27] Y. Juhn and H. Liu, “Artificial intelligence approaches using natural language processing to advance EHR-based clinical research,”
Journal of Allergy and Clinical Immunology, vol. 145, no. 2, pp. 463–469, 2020, doi: 10.1016/j.jaci.2019.12.897.
[28] Y. Luo et al., “Natural language processing for EHR-based pharmacovigilance: a structured review,” Drug Safety, vol. 40, no. 11,
pp. 1075–1089, 2017, doi: 10.1007/s40264-017-0558-6.
[29] T. T. V. Vleck et al., “Augmented intelligence with natural language processing applied to electronic health records for identifying
patients with non-alcoholic fatty liver disease at risk for disease progression,” International Journal of Medical Informatics, vol.
129, pp. 334–341, Sep. 2019, doi: 10.1016/j.ijmedinf.2019.06.028.
[30] Y. Li et al., “BEHRT: transformer for electronic health records,” Scientific Reports, vol. 10, no. 1, pp. 1–12, 2020, doi:
10.1038/s41598-020-62922-y.
[31] M. Al Jazaery and G. Guo, “Video-based depression level analysis by encoding deep spatiotemporal features,” IEEE Transactions
on Affective Computing, vol. 12, no. 1, pp. 262–268, 2021, doi: 10.1109/TAFFC.2018.2870884.
[32] I. Peis et al., “Actigraphic recording of motor activity in depressed inpatients: a novel computational approach to prediction of
clinical course and hospital discharge,” Scientific Reports, vol. 10, no. 1, 2020, doi: 10.1038/s41598-020-74425-x.

Int J Artif Intell, Vol. 13, No. 4, December 2024: 4352-4363

Int J Artif Intell ISSN: 2252-8938  4363

[33] Y. Tazawa et al., “Actigraphy for evaluation of mood disorders: a systematic review and meta-analysis,” Journal of Affective
Disorders, vol. 253, pp. 257–269, 2019, doi: 10.1016/j.jad.2019.04.087.
[34] E. G. -Ceja et al., “Depresjon: a motor activity database of depression episodes in unipolar and bipolar patients,” in Proceedings of
the 9th ACM Multimedia Systems Conference, MMSys 2018, 2018, pp. 472–477, doi: 10.1145/3204949.3208125.
[35] L. Z. -Calzada et al., “Feature extraction in motor activity signal: towards a depression episodes detection in unipolar and bipolar
patients,” Diagnostics, vol. 9, no. 1, Jan. 2019, doi: 10.3390/diagnostics9010008.
[36] J. H. Majed, S. A. Nasser, A. Alkhayyat, and I. A. Hashim, “Artificial intelligent algorithms based depression detection system,” in
IICETA 2022 - 5th International Conference on Engineering Technology and its Applications, 2022, pp. 408–413, doi:
10.1109/IICETA54559.2022.9888485.
[37] X. Wang, Y. Ren, and W. Zhang, “Depression disorder classification of fMRI data using sparse low-rank functional brain network
and graph-based features,” Computational and Mathematical Methods in Medicine, vol. 2017, 2017, doi: 10.1155/2017/3609821.

BIOGRAPHIES OF AUTHORS

U. Ananthanagu currently pursuing a Ph.D. at PES University, India. He holds a

master’s degree in computer science and engineering from Jawaharlal Nehru Technological
University, a bachelor’s degree in Computer Science and Engineering from Visvesvaraya
Technological University, and a Diploma in Computer Science and engineering from the
Technical Board of Education. With 14 years of experience in academics, his primary areas of
research interest include data science, machine learning, NLP, social media analysis, and
transfer learning. He has published over 20 papers in international journals and conferences. He
is a life member of the Computer Society of India (CSI) and the Indian Society for Technical
Education (ISTE). He can be contacted at email: [email protected] or
[email protected].

Pooja Agarwal completed her bachelor’s degree in 1998 from CCS University,
Meerut, and her master’s in computer science in 2001 from Banasthali Vidyapith, Rajasthan,
and was awarded the Ph.D. degree in Computer Science primarily focusing in the area of
machine learning and deep learning from Visvesvaraya Technological University in 2018. Since
July 2008, she has been associated with PES institutions at PESIT Bangalore South Campus
(now PES University-EC Campus). She has worked with over a hundred students on various
projects at the undergraduate and master levels in the application areas of data science, machine
learning, soft computing, natural language processing, data analytics, and deep architectures.
She has around 35 publications in various reputed international conferences and journals. She
is a member of IEEE and member IAENG. She can be contacted at email:
[email protected].

DepXGboot: depression detection using a robust tuned extreme gradient boosting … (U. Ananthanagu)