IAES Inter national J our nal of Articial Intelligence (IJ-AI) V ol. 15, No. 4, August 2026, pp. 3888 3902 ISSN: 2252-8938, DOI: 10.11591/ijai.v15.i4.pp3888-3902 3888 Pr edicti v e model based on machine lear ning to identify sleep-r elated health pr oblems Laberiano Andrade-Ar enas 1 , Inoc Rubio P aucar 2 , Mar garita Giraldo Retuerto 1 , Cesar Y actay o-Arias 3 1 F aculty of Science and Engineering, Uni v ersidad de Ciencias y Humanidades, Lima, Per ´ u 2 F aculty of Engineering and Business, Uni v ersidad Pri v ada Norbert W iener , Lima, Per ´ u 3 Department of General Studies, Uni v ersidad Continental, Lima, Per ´ u Article Inf o Article history: Recei v ed Jul 19, 2025 Re vised Jul 13, 2026 Accepted Jul 21, 2026 K eyw ords: Extreme gradient boosting Kno wledge disco v ery in databases methodology Machine learning Predicti v e model Public health monitoring Sleep quality ABSTRA CT Sleep quality has become a gro wing public health issue w orldwide, mainly due to a lack of a w areness about its long-term consequences. Despite e xisting strate gies to address this problem, there remains a need for more ef fecti v e approaches. In this study , an early detection model for sleep disorders w as implemented using the e xtreme gradient boosting (XGBoost) algorithm, follo wing the kno wledge disco v ery in databases (KDD) methodology , which includes the phases of selection, preprocessing, transformation, data mining, and interpretation. A dataset e xtracted from the Kaggle platform in CSV format w as used, consisting of 374 records. W ith an o v erall accurac y of 91.5%, a recall of 100% for the insomnia class, and a precision of 100% for sleep apnea, the proposed model demonstrated e xceptional performance. It also recei v ed an area under the curv e (A UC) of 0.909 and an a v erage F1-score of 0.913. W ith a mean accurac y of 91%, a 95% condence interv al (0.89–0.94), and a p-v alue of 0.0012, cross-v alidation conrmed its rob ustness and sho wed a statistically signicant change from the baseline model. The error rates remained withi n clinically acceptable ranges, conrming its applicability as a diagnostic support tool. Ov erall, the results demonstrate the ef fecti v eness of the model in identifying patterns related to sleep disorders. This is an open access article under the CC BY -SA license . Corresponding A uthor: Cesar Y actayo-Arias Department of General Studies, Uni v ersidad Continental Lima, Per ´ u Email: c yactayo@continental.edu.pe 1. INTR ODUCTION The scientic community is currently becoming more interested in learning more about human sleep indicators. Numerous studies ha v e demonstrated that sleep quality signicantly af fects one’ s ph ysical and mental health. Lifestyle choices, eating habits, stress le v els, and particularly the frequenc y and intensity of personal ph ysical acti vity are among the primary determinants of this quality [1], [2]. In response to this issue, data science and machine learning (ML) ha v e emer ged as essential tools for addressing the la r ge- scale analysis of biometric and beha vioral data related to sleep. These technologies also enable the identication of comple x patterns and the e xtraction of k e y v ariables that f acilitate informed decision-making in the elds of public health and the pre v ention of sleep disorders. The a v ailability of v aluable data on sleep habits, the calculation of rele v ant ph ysiological indicators, and the assessment of lifestyle choices pro vide an unprecedented opportunity to in v estig ate risk f actors that af fect the o v erall health of the population. J ournal homepage: http://ijai.iaescor e .com Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 3889 Although there is e xtensi v e e vidence linking sleep to v arious aspects of health, there are stil l challenges in accurately predicting the potential ne g ati v e health consequences caused by poor sleep patterns. Moreo v er , man y studies rely on traditional statistical analyses, whose main limitation lies in their inability to capture nonlinear relationships between the studied v ariables [3], [4]. F or this reason, there is a gro wing interest in complementing these approaches with ML techniques, which enable the modeling of comple x patterns and the deli v ery of more precise and in-depth kno wledge. Sleep is a vital ph ysiological function for human beings, and its sustained depri v ation or disruption can signicantly compromise functional performance in daily acti vities and promote the onset of v arious non-communicable chronic diseases [5]. In this re g ard, scientic e vidence has sho wn that sleep disorders are closely associated with an increased risk of de v eloping h ypertension, t yp e 2 diabetes mellitus, obesity , as well as mental health disorders, including anxiety and depression. Despite thi s, man y indi v i duals prioritize their daily responsibilities without considering that chronic sleep restriction represents a signicant risk f actor for their o v erall well-being. This ne glect has contrib uted to a high pre v alence of chronic stress conditions and sleep-related pathologies [6]. In light of this, people must become a w are of the k e y aspects of achie ving proper sleep at appropriate times in order to maintain a health y life. In this conte xt, the application of ML techniques, framed within the kno wledge disco v ery in dat abases (KDD) met ho dol ogy , enables the e xploration and analysis of lar ge v olumes of data with the aim of identifying non-tri vial patterns and de v eloping highly accurate predicti v e models. This approach not only strengthens research in public health and beha vioral sciences b ut also pro vides adv anced analytical tools for pre v ention, early diagnosis, and the personalization of health interv entions [7], [8]. It is important to consider that ML is a k e y f actor in generating kno wledge that adds v alue in addressing a public health issue [9], [10]. The present research is justied by its capacity to generate applied kno wledge in the eld of pre v enti v e health through the design and implementation of an algorithmic model capable of predicting health risks based on the analysis of v ariables related to sleep quality and lifestyle. This contrib ution not only e xpands the s cientic e vidence base b ut also pro vides rele v ant inputs for decision-making by healthcare professionals, public polic y mak ers , and de v elopers of technologies aimed at promoting population well-being. The objecti v e of this research is to de v elop a predicti v e model using ML techniques specical ly the e xtreme gradient boosting (XGBoost) algorithm that enables the identication and pre diction of potential impacts on ph ysical and mental health based on the analysis of sleep patterns. 2. LITERA TURE REVIEW This section focuses on related w orks by v arious authors on the topic, thus pro viding a holi stic o v ervie w for the in v estig ation. Lik e wi se, the theoretical foundations s upp or t the identication and justication of the v ariables to be studied. This comprehensi v e approach ensures a solid frame w ork for the current study . 2.1. Related w orks The article e v aluates v arious ML models for predicting sleep disorders, considering their si gnicant impact on public health. The study utilizes a dataset of 400 indi vidual records classied by sleep disorder status (none, insomnia, and sleep apnea) and incorporates demographic, lifestyle, and health metrics. Dif ferent ML models, including logistic re gression, decision trees, ensemble methods (lik e random forest (RF) and gradient boosting), support v ector machines, and neural netw orks are assessed for their diagnostic accurac y . K e y ndings indicate that ensemble methods, particularly RF and e xtreme gradient boosting classier (XGBClassier), outperform other models, achie ving high accurac y and F1-scores (up to 0.93). The results highlight the importance of adv anced ensemble techniques in accurately diagnosing sleep disorders, suggesting their inte gration into clinical practices for impro v ed outcomes. The research methodology includes data collection and preprocessing, model de v elopment, trai ning, and e v aluation using metrics lik e accurac y , precision, recall, and F1-score. The study emphasizes the comple xity of sleep disorders and the potential of ML to enhance diagnostic tools in sleep medicine, adv ocating for future research into optimizing these models for real-w orld applications [11], [12]. Sleep problems pose an important challenge for human health. In this study , a cross -v alidated XGBoost–bidirectional long short-term memory (BiLSTM) model w as implemented using 75 v ariables and a dataset of 10,765 aging-related records. The proposed model reached 97% accurac y and demonstrated better performance than traditional long short-term memory (LSTM) models and other machine-learning approaches. The results demonstrate its ef fecti v eness in the diagnosis and treatment of sleep apnea by identifying risk Pr edictive model based on mac hine learning to identify sleep-r elated ... (Laberiano Andr ade-Ar enas) Evaluation Warning : The document was created with Spire.PDF for Python.
3890 ISSN: 2252-8938 f actors. In pre vious studies, the XGBoost algorithm w as also emplo yed, applying training strate gies focused on the most rele v ant features of the dataset. As in the pre vious study , k-fold cross-v alidation w as implemented to optimize model performance. The results were obt ained by retraining on subsets of the training set using test data, achie ving an accurac y of 90.6% in predicting the dif ferent stages of sleep [13], [14]. On the other hand, some authors in v estig ated obstructi v e sleep apnea (OSA) in the pediatric population, linking it to de v elopmental aspects such as ph ysical gro wth, cardio v ascular function, and cognition through ML techniques. Using a sample of 3,139 children, the XGBoost algorithm w as applied after randomly splitting the data into training and testing sets. In the testing set, the model achie v ed area under the curv e (A UC) of 0.95, 0.88, and 0.88 for the classication of mild, moderate, and se v ere OSA, respecti v ely , with accuracies of 90.45%, 85.67%, and 89.81%, outperforming re gression-based models. Similarly , complementary approaches ha v e been e xplored using techniques such as entrop y and linear feature analysis for the selection of rele v ant v ariables. In this conte xt, a model based on the XGBoost algorithm w as trained using the Sleep Heart Health Study 2 dataset. Through feature selection strate gies, the most signicant attrib utes for diagnosis were identied. When e v aluating the model with a v alidation subset composed of 422 participants not included in the trai n i ng phase, an accurac y of 68.83%, a precision of 71.61%, and a recall of 67.84% were achie v ed, demonstrating acceptable performance in scenarios in v olving pre viously unseen data, as reported by [15], [16]. In contrast to other studies [17], [18] conducted a longitudinal analysis on 239 patients with insomnia using the XGBoost algorithm, which achie v ed a multiclass A UC of 80% and a binary A UC of 83%, with adequate sensiti vity and specicity , f acilitating the identication of patients with lo w response to con v entional depression treatments. In line with pre vious studies on the relationship between sleep-w ak e c ycle disorders and major depressi v e disorders in adolescents, v arious ML models were e v aluated, with XGBoost standing out by achie ving an A UC of 85% in predicting the risk of depression. Mul tiple in v estig ations ha v e demonstrated that the XGBoost algorithm is highly ef fecti v e in anticipating the onset of depressi v e disorders associated with sleep-w ak e rh ythm disturbances, thus supporting the de v elopment of personalized treatments and more precise pre v enti v e strate gies. Lik e wise, conditions such as OSA, the comorbidi ty of insomnia and sleep apnea (COMISA), and insomnia a s an independent disorder represent signicant manifestations of sleep disorders. In order to predict the risk of these conditions, XGBoost w as implemented on clinical data collected from medical centers, selecting nine rele v ant v ariables through the Shaple y additi v e e xplanations (SHAP) e xplainability method. These features were used to construct the schedule, light, eat and drink, en vironment, ph ysiology , and stress (SLEEPS) model, which demonstrated outstanding predicti v e capacity , with A UC v alues e xceeding 89% across all three analyzed disorders. As a result, SLEEPS is positioned as an inno v ati v e diagnostic tool that enables early and accessible detection of sleep disorders without relying on in v asi v e methods such as polysomnograph y . In the present s tudy , three predicti v e models were de v eloped using the Python programming language: logistic re gression, XGBClassier , and support v ector classier (SVC). After e v aluating their performance, the XGBClassier model w as i dentied as the most ef fecti v e. The results sho wed that occupation is the main f actor associated with insomnia, while body mass inde x (BMI) cate gory signicantly inuences the occurrence of sleep apnea. Additionally , BMI, stress le v el, blood pressure, and sleep duration were identied as k e y v ariables af fecting sleep quality . These ndings suggest that interv entions aimed at controlling BMI and blood pressure could signicantly contrib ute to impro ving nighttime rest [19], [20]. On the other hand, a study focused on estimating the risk of insomnia in breast cancer patients used data obtained through surv e ys. In the predictions performed, logistic re gression models with L2 re gularization and the XGBoost algorithm achie v ed predicti v e accuracies of 71.5% and 70.6%, respecti v ely , with A UC v alues of 0.76 and 0.75. Additionally , population subgroups at high risk of insom n i a were found by using the RuleFit algorithm. The ndings demonstrated that among breast cancer survi v ors, cancer -related f atigue is a signicant predictor of the onset of sleeplessness. In contrast to the study that e xamined insomnia in patients with breast cancer , the current study used the XGBoost algorithm to nd risk f actors link ed to sleep disturbances. A total of 7,929 patients who satised the inclusion criteria were included in the sample; their mean age w as 49.2 years (standard de viation (SD) = 18.4), with 4,055 (51%) for w omen and 3,874 (49%) for men. The ethnic composition w as distrib uted as follo ws: 2,885 (36%) white indi viduals, 2,144 (27%) African Americans, 1,639 (21%) Hispanics, and 1,261 (16%) from other ethnic groups. Of the 684 v ariables analyzed, 64 were statistically signicant p < 0 . 0001 and were incorporated into the predicti v e model. The performance of the XGBoost model w as reected in an area under the recei v er operating characteristic curv e (A UR OC) of 0.87, with both sensiti vity and specicity equal to 0.77, demonstrating a rob ust capacit y to identify risk f actors related Int J Artif Intell, V ol. 15, No. 4, August 2026: 3888–3902 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 3891 to sleep disorders [21], [22]. In a dif ferent study , Haque et al. [23] created a predicti v e model to assess the connection between academic achie v ement and sleeplessness in kids and teenagers bet ween the ages of 4 and 17. F or this purpose, the “Y oung minds matter” database w as used, applying the Boruta algorithm in combination with the RF classier to select the most rele v ant features. Subsequently , the tree-based pipeline optimization tool-classier (TPO TClassier) w as used to identify the best-performing supervised models, e v aluating the RF , XGBoost, decision tree, and Gaussian nai v e Bayes algorithms. As a result, the RF model demonstrated the best performance, achie ving a precision of 99% and an accurac y of 95%, outperforming the other e v aluated models and sho wing greater ef fecti v eness in predicting insomnia in relation to academic performance. 2.2. Theor etical bases 2.2.1. Extr eme gradient boosting XGBoost is dened as a supervised ML algorithm based on decision trees. Its main goal is to perform classication and re gression tasks with a high le v el of accurac y . It is based on the idea of gradient boosting, a method that entails training models one after the other , with each ne w model xing the mistak es of the preceding ones. What mak es XGBoost particularly ef fecti v e is that it is optimized to deli v er high speed, performance, and computational ef cienc y . One of its most notable features is its ability to handle lar ge v olumes of data quickly , thanks to a highly optimized internal architecture that includes parallel processing, ef cient memory usage, and adv anced tree pruning techniques [24], [25]. In addition, XGBoost incorporates re gularization mechanisms (L1 and L2 penalties ), which help pre v ent o v ertting and enhance the model’ s generalization ability when applied to ne w data. Another signicant adv antage is its abili ty to automatically handle missing v alues, making it a rob ust and practical tool for real-w orld applications where incomplete data is common. The algorithm also allo ws ne-tuning of multiple h yperparameters, such as tree depth, learning rate, and the number of estimators, pro viding a high de gree of customization and f acilitating model optimization according to the specic problem [26], [27]. 2.2.2. Sleep quality Sleep quality is a multidimensional construct that refers to the ef cienc y , continuity , depth, and subjecti v e perception of nighttime rest. Medically , it is assessed by considering multiple objecti v e and subjecti v e parameters, including sleep latenc y , sleep ef cienc y , nocturnal a w ak enings, total sleep duration, and the presence or absence of sleep disorders such as insomnia, sleep apnea, or periodic limb mo v ements. From a clinical standpoint, health y sleep is typically dened as taking no longer than about 30 minutes to f all asleep and maintaining a sleep ef cienc y the percentage of tim e actually sleeping while in bed abo v e 85%, a total duration of 7 to 9 hours per night in adults, and a r educed number of prolonged nocturnal a w ak enings. Additionally , the indi vidual’ s subjecti v e perception upon w aking-such as feeling rested and alert is a rele v ant component in the clinical assessment of sleep quality [28], [29]. Sleep quality is closely related to multiple neuroph ysiological functions, including memory consolidation, emotional re gulation, cellular repair , and endocrine balance. Disruptions in sleep ha v e been link ed to an increased lik elihood of psychiatric conditions including major depressi v e disorder and anxiety as well as long-term medical problems such as h ypertension, type 2 diabetes, and cardio v ascular disease. F or this reason, e v aluating sleep quality is considered a k e y element both in clinical settings and in research on general health [30]. Blue light, with w a v elengths between 450 and 495 nanometers, is emitted both by the sun and by electronic de vi ces such as cell phones, computer s, and tele visions. While benecial during the day , i ts e xposure at night can ne g ati v ely af fect sleep. The circadian rh ythm, which re gulates the sleep-w ak e c ycle, is controlled by the suprachiasmatic nucleus (SCN) of the h ypothalamus and responds primarily to ambient light. Melatonin, a k e y hormone for inducing sleep, is secreted in darkness, and its production decreases when e xposed to light. When the retina detects blue light at night, specialized cells (intrinsically photosensiti v e retinal g anglion cells (ipRGCs) containing melanopsin) are acti v ated and send signals to the SCN indicating that light is still present. This inhibits melatonin secretion, delays sleep onset, and disrupts the circadian rh ythm. As a result, using blue light-emitting de vices before bedtime can lead to insomnia, f atigue, and a decrease in sleep quality [31], [32]. It is important to raise a w areness of this issue, as the majority of people use electronic de vices e xcessi v ely , often ignoring the health risks associated with blue light e xposure. Pr edictive model based on mac hine learning to identify sleep-r elated ... (Laberiano Andr ade-Ar enas) Evaluation Warning : The document was created with Spire.PDF for Python.
3892 ISSN: 2252-8938 3. METHOD 3.1. Denition of the kno wledge disco v ery in databases methodology The KDD methodology is an iterati v e and interacti v e process that inte grates e xpert kno wledge about a specic problem with a di v erse set of data analysis techniques, particularly those based on ML algorithms. This process consists of v e fundamental stages: selection, preprocessing, transformation, data mining, and interpretation/e v aluation of results. Figure 1 sho ws that each of these phases is articulated in a sequential and logical manner , allo wing input data to be cleaned, transformed, and subsequently analyzed in order to e xtract useful patterns or kno wledge. Finally , the results are interpreted and e v aluated according to the initial objecti v es [33]. One of the main adv antages of this approach is its e xibility , as it allo ws returning to pre vious stages of the process if the results are unsatisf actory or if it becomes necessary to adjust certain assumptions. This ensures greater accurac y and coherence in the e xtraction of kno wledge from lar ge v olumes of data. Ultimately , this methodology represents a rob ust and adaptable frame w ork for v arious projects and promotes continuous impro v ement thanks to its iterati v e nature. Figure 1. KDD methodology 3.1.1. Selection In this section, an e xhausti v e search w as conducted on the Kaggle platform, a widely used repository within the scientic and technological community due to its v ast v ariety of public datasets from dif ferent elds of kno wledge. During the search process, specic criteria were emplo yed to ensure the rele v ance and quality of the selected dataset. K e yw ords such as sleep health, sleep disorders, sleep quality , mental health, among others, were used within Kaggle’ s search engine to nd databases containing rele v ant v ariables related to sleep habits, associated disorders, and their possible connection to other dimensions of ph ysical and mental health [34], [35]. Lik e wise, priority w as gi v en t o selecting a dataset in CSV format due to its tab ular structure, compatibility with statistical analysis tools (such as Python), and ease of manipulation. This format allo ws for ef cient data processing, f acilitating cleaning, se gmentation, and descripti v e or inferential analysis according to the objecti v es of the study . 3.1.2. Pr epr ocessing In the preprocessing stage, descripti v e statistics re v eal that the 374 participants ha v e a mean age of 42.18 years (SD = 8.67; range = 27–59), with an a v era g e sleep duration of 7.13 hours (SD = 0.80; Q1 = 6.40; median = 7.20; Q3 = 7.80; minimum = 5.80; maximum = 8.50), and an a v erage sleep quality score of 7.31 (SD = 1.20; range = 4–9). The a v erage ph ysical acti vity le v el is 59.17 (SD = 20.83; Q1 = 45; median = 60; Q3 = 75; range = 30–90), and the mean stress le v el is 5.39 (SD = 1.77; range = 3–8). The a v erage heart rate is 70.17 bpm (SD = 4.14; Q1 = 68; median = 70; Q3 = 72; range = 65–86), and the a v erage daily steps are 6,816.84 (SD = 1,617.92; Q1 = 5,600; median = 7,000; Q3 = 8,000; range = 3,000–10,000), as sho wn in T able 1. Finally , all numerical v ariables were standardized using the StandardScaler() function, and this information is presented in T able 2. Int J Artif Intell, V ol. 15, No. 4, August 2026: 3888–3902 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 3893 T able 1. Descripti v e statistics of the dataset Statistics Age Sleep duration Quality of sleep Ph ysical acti vity le v el Stress le v el Heart rate Daily steps Count 374.00 374.00 374.00 374.00 374.00 374.00 374.00 Mean 42.18 7.13 7.31 59.17 5.39 70.17 6816.84 SD 8.67 0.80 1.20 20.83 1.77 4.14 1617.92 Min 27.00 5.80 4.00 30.00 3.00 65.00 3000.00 Q1 (25%) 35.25 6.40 6.00 45.00 4.00 68.00 5600.00 Q2 (Median) 43.00 7.20 7.00 60.00 5.00 70.00 7000.00 Q3 (75%) 50.00 7.80 8.00 75.00 7.00 72.00 8000.00 Max 59.00 8.50 9.00 90.00 8.00 86.00 10000.00 T able 2. T echnical steps for data preprocessing P assed T echnical operation 1 Con v ersion to num eric: pd.to numeric(df[’Age’], errors=’coerce’) 2 Imputat ion of N A: df[’Sleep Duration’].fillna(median, inplace=True) 3 Calcul ate IQR para Heart Rate : IQR = Q 3 Q 1 Outliers = { x < Q 1 1 . 5 × IQR } { x > Q 3 + 1 . 5 × IQR } 4 Standardi zation: StandardScaler().fit transform(df[column numbers]) 3.1.3. T ransf ormation During the data transformation phase, se v eral preprocessing techniques were applied to prepare the health and lifestyle dataset for analysis. Data con v ersion and standardization were performed e xtensi v ely , ensuring that the v ariables were transformed into a consistent format suitable for model training. Additional procedures, including missing v alue imputation and duplicate record remo v al, were required only to a limited e xtent, while outlier detection, v ariable transformation, and feature engineering were applied selecti v ely to impro v e data quality . Data visualization also played an important role in e xploring the dataset, together with the analysis of the v ariance e xplained by the principal components, as illustrated in Figure 2. Specically , Figure 2(a) summarizes the preprocessing w orko w , sho wing that standardization (100%) and data con v ersion (90%) were the most e xtensi v ely applied techniques, follo wed by data visualization (80%) and outlier detection using the interquartile range (IQR) method (70%). In contrast, missing v alue imputation (7.7%) w as only occasionally necessary , and no duplicate records (0%) were detected, highlighting the high quality of the dataset prior to preprocessing. Figure 2(b) illustrates the tw o-dimensional representation generated using principal component analysis (PCA). The rst tw o principal components e xplain 93.4% of the total v ariance (PC1 = 65.2% and PC2 = 28.2%), re v ealing a clear separation between the sleep apnea and insomnia classes with minimal o v erlap. This distrib ution suggests that the e xtracted features ef fecti v ely preserv e the underlying class structure, pro viding a solid foundation for the subsequent classication task. (a) (b) Figure 2. Data manipulation and transformation of (a) percentages of features in the data and (b) percentage of v ariance in components Pr edictive model based on mac hine learning to identify sleep-r elated ... (Laberiano Andr ade-Ar enas) Evaluation Warning : The document was created with Spire.PDF for Python.
3894 ISSN: 2252-8938 3.1.4. Data mining During the data mining procedure, the required XGBoost algorithm w as applied. The results in Figure 3 indicate that the XGBoost model demonstrates e xceptional performance according to the precision-recall and recei v er operating characteristic (R OC) curv es. The precision-recall curv e, with an a v erage precision of 0.95, indicates that the model maintains high precision e v en with a high recall, making it highly ef fecti v e at identifying positi v e classes, as in Figure 3(a). Simultaneously , the R OC curv e, with an A UC of 0.94, conrms e xcellent discriminati v e ability , meaning that the model is v ery capable of distinguishing between positi v e and ne g ati v e classes across a wide range of thresholds, as in Figure 3(b). T ogether , these results suggest a rob ust and reliable classier for the task at hand. Both plots demonstrate that the XGBoost model performs e xceptionally well for binary class ication tasks. The precision-recall curv e, with an a v erage pre cision of 0.95, sho ws that the model maintains high precision (fe w incorrect positi v e predictions) e v en while achie ving high recall (detecting most actual positi v e cases). Complementarily , the R OC curv e, with an A UC of 0.94, conrms the model’ s e xcellent capacity to distinguish between positi v e and ne g ati v e classes o v er a broad range of thresholds, signicantly outperforming a random classier [36], [37]. (a) (b) Figure 3. Classier performance metrics for XGBoost model: (a) precision-recall curv e and (b) R OC curv e 4. MA THEMA TICAL B ASIS OF THE EXTREME GRADIENT BOOSTING MODEL The follo wing section presents the mathematical foundation of the XGBoost algorithm, adapted to the binary classication problem aimed at identifying health-related conditions associated with sle ep. It describes the k e y concepts and equations that go v ern the model’ s lea rning process. The section also e xplains ho w the algorithm optimizes its objecti v e function to impro v e predicti v e performance. 4.1. General objecti v e function XGBoost b uilds an ensemble of additi v e decision trees by optimizing an objecti v e function that combines a loss function with a re gularization term to pre v ent o v ertting as (1) [38]. This approach allo ws the model to s equentially correct the errors of pre vious trees. It also balances model comple xity and predicti v e accurac y to achie v e rob ust performance on unseen data. L ( t ) = n X i =1 l ( y i , ˆ y ( t ) i ) + t X k =1 Ω( f k ) (1) Where l ( y i , ˆ y ( t ) i ) is loss function between the prediction and the actual label; and Ω( f k ) is penalty on tree comple xity f k . Int J Artif Intell, V ol. 15, No. 4, August 2026: 3888–3902 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 3895 4.2. Model r egularization Re gularization controls the comple xity of the model by penalizi ng the depth and the weight of the lea v es in each tree as in (2) [39]. This helps pre v ent o v ertting by discouraging o v erly comple x trees. It ensures that the model generalizes well to ne w , unseen data. Ω( f ) = γ T + 1 2 λ T X j =1 w 2 j (2) Where T : number of lea v es; w j is predicted v alue on the sheet j ; γ is penalty for each leaf; and λ is L2 re gularization parameter . 4.3. Logistic loss function F or binary classication, the l ogistic loss (log-loss) is used [40], which measures the error between the predicted probability and the actual class, as in (3). This loss function penalizes incorrect predictions more hea vily when the model is condent b ut wrong. It pro vides a smooth gradient that f acilitates optimization during the training of XGBoost. l ( y i , ˆ y i ) = [ y i log ( σ ( ˆ y i )) + (1 y i ) log (1 σ ( ˆ y i ))] (3) Where the sigmoid function is: σ ( ˆ y i ) = 1 1 + e ˆ y i 4.4. Optimization by second order expansion T o impro v e training ef cienc y , XGBoost uses a second-order (T aylor) e xpansion [41] around the current prediction, as in (4). This approximation allo ws the model to consider both the gradient and the Hessian when updating tree lea v es. It accelerat es con v er gence and enhances the accurac y of each iteration during training. L ( t ) n X i =1 g i f t ( x i ) + 1 2 h i f t ( x i ) 2 + Ω( f t ) (4) Where g i = l ( y i , ˆ y ( t 1) i ) ˆ y ( t 1) i is gradient and h i = 2 l ( y i , ˆ y ( t 1) i ) ˆ y ( t 1)2 i is Hessian. 4.5. Optimal weight of a sheet The optimal v alue that a leaf should predict is calculated as (5). This v alue is deri v ed by minimizing the re gularized objecti v e function for that specic leaf. It ensures that each leaf contrib utes optimally to reducing the o v erall prediction error of the model. w j = G j H j + λ (5) Where G j = P i I j g i is sum of gradients; H j = P i I j h i is sum of Hessians; and I j is set of samples o n the sheet j . 4.6. Pr ot fr om a di vision T o decide whether to perform a split at a node, the resulting information g ain is calculated (6). Thi s g ain measures ho w much the split reduces the o v erall loss of the model. It helps the algorithm choose the most informati v e splits, impro ving predicti v e accurac y while controlling model comple xity . Gain = 1 2 G 2 L H L + λ + G 2 R H R + λ ( G L + G R ) 2 H L + H R + λ γ (6) Where G L , H L is sum of gradients and Hessians to the left of the split; G R , H R is sum to the right of the split; and γ is penalty for creating an additional sheet. Pr edictive model based on mac hine learning to identify sleep-r elated ... (Laberiano Andr ade-Ar enas) Evaluation Warning : The document was created with Spire.PDF for Python.
3896 ISSN: 2252-8938 5. RESUL TS 5.1. Ev aluation of r esult Figure 4 compares three classication models (logistic re gression, RF , and XGBoost), sho wing that all three e xhibit e xcellent and comparable performance [42]. The R OC curv e demonstrates that all models are equally capable of distinguishing between classes (A UC = 0.95), as sho wn in Figure 4(a). The precision-recall curv e indicates that XGBoost is slightly superior in maintaining a balance between precision and recall (AP = 0.92), as sho wn in Figure 4(b), while the cumulati v e g ain chart conrms that all models are highly ef cient at quickly identifying the majority of positi v e cases, as sho wn in Figure 4(c). (a) (b) (c) Figure 4. Model performance comparison: (a) logistic re gression, (b) RF , and (c) XGBoost Furthermore, other results sho wn in Figure 5 include normalized confusion matrices for the three ML models e v aluated for the binary classication task, allo wing for a detailed comparison between the predicted and actual class labels. Figure 5(a) sho ws the performance of the logistic re gression model, which correctly classied 91% of the class 0 samples and 94% of the class 1 samples, with misclassication rates of 9% and 6%, respecti v ely . On the other hand, Figure 5(b) illustrates the results obtained with the RF classier , which achie v ed corr ect classication rates of 88% for class 0 and 94% for class 1, although it had a slightly higher f alse positi v e rate (12%) than the other models. Finally , Figure 5(c) presents the confusion matrix for the XGBoost model, whi ch of fered the best o v erall performance by correctly identifying 97% of class 0 instances Int J Artif Intell, V ol. 15, No. 4, August 2026: 3888–3902 Evaluation Warning : The document was created with Spire.PDF for Python.
Int J Artif Intell ISSN: 2252-8938 3897 and 94% of class 1 instances, while reducing the f alse positi v e rate to 3% and maintaining a lo w f alse ne g ati v e rate (6%). Ov erall, the results indicate that all three classiers achie v ed high predicti v e performance, with XGBoost demonstrating the most accurate and balanced classication. (a) (b) (c) Figure 5. Confusion matrices of classication models: (a) logistic re gression, (b) RF , and (c) XGBoost T able 3 reports t he o v erall error metrics. The model sho ws a lo w f alse positi v e rate (FPR) of 0.065 and a true positi v e rate (TPR) of 0.909, indicating high o v erall sensiti vity . The true ne g ati v e rate (TNR) is also high (0.935), reecting good specicity . The o v erall error rate is relati v ely lo w (0.085), which supports the model’ s general reliability . T able 3. Global error metrics Metric V alue FPR 0.065 FNR 0.182 TPR 0.909 TNR 0.935 Error rate 0.085 The class-wise metrics presented in T able 4 reect a balanced performance of the model in classi fying insomnia and sleep apnea. The insomnia class sho ws a precision of 0.862, a perfect recall of 1.000, and an F1-score of 0.926, indicating that all i nsomnia cases were correctly identied with no f alse ne g ati v es. Pr edictive model based on mac hine learning to identify sleep-r elated ... (Laberiano Andr ade-Ar enas) Evaluation Warning : The document was created with Spire.PDF for Python.