Comparative performance of twelve machine learning models in predicting COVID-19 mortality risk in children: a population-based retrospective cohort study in Brazil

Study design and dataset description

This retrospective cohort study used data from the Surveillance Information System for Influenza (SIVEP-Gripe) to examine COVID-19 hospitalizations among individuals under 18 years of age in Brazil. Established in 2009 by the Ministry of Health, the SIVEP-Gripe is a nationwide database that captures data on severe acute respiratory infections. Since the COVID-19 pandemic, it has been the primary source of hospitalization data. Mandatory reporting from both public and private hospitals ensures comprehensive coverage. The database contains the demographic and clinical information of all hospitalized patients. We analyzed data from epidemiological week 08, 2020, to week 08, 2023, encompassing individuals aged under 18 years with confirmed SARS-CoV-2 infection via RT-qPCR or antigen testing upon hospital admission.

Data preparation

Over the designated period, 56,330 patient records with verified RT-PCR test outcomes for SARS-CoV-2 infection were documented. After completing the required procedures for preprocessing data for the machine learning algorithms, 24,097 records were chosen for the training, validation, and testing stages of the models. The subset of data from the SIVEP-Gripe dataset, which includes information about children and adolescents, is hereafter referred to as the SIVEP-Kids dataset.

In the SIVEP-Kids dataset, there are 37 primary features in four main categories: patient demographics (four features), clinical features (12 features), personal disease/comorbidity history (14 features), virus strain information (one feature), vaccine information (two features), a feature indicating the number of different comorbidities a patient has, a feature indicating whether a patient has comorbidities or not, a feature categorizing the number of comorbidities a patient has, a feature indicating the time of the outcome, and an output variable (0: survived and 1: deceased) for COVID-19 patients. The primary features of the SIVEP-Kids dataset are presented in Table 1.

Regarding the primary features presented in the SIVEP-Kids dataset, the ethnicity feature had five categories: Asian, Black, Brown, Indigenous, and White. Similarly, the region was divided into five regions: Central West, North, Northeast, South, and Southeast. The virus strain feature identified four types of strains in the dataset: ancestral, delta, gamma, and omicron. For features 6 through 32, all are of the nominal type and have values of “Yes” or “No,” indicating the presence or absence of a specific disease or clinical condition in the patient. The total comorbidity feature records the total number of comorbidities per patient in the SIVEP-Kids dataset. Feature 34 (number of vaccine doses) had valid values ranging from zero to three doses. Feature 38 is the target variable of this study, with three types of outcomes: discharge, death, and in-hospital, with the latter referring to cases in which the patient is still in the hospital in an ongoing clinical situation. In the present study, we considered only two types of outcomes in the target variable: death and discharge. This decision aimed to enhance the accuracy of machine learning algorithms, as multi-class problems (those with more than two classes in the target variable) are challenging and tend to reduce the accuracy of ML models because of the large number of decision boundaries to navigate, often failing to accurately separate instances across more than two classes (Bengio, Weston & Grangier, 2010; Del Moral, Nowaczyk & Pasham, 2022). Detailed information on the clinical, demographic, and epidemiological covariates recorded in the SIVEP-Gripe is described elsewhere (Oliveira et al., 2021, 2023b, 2024, 2022).

Data pre-processing

Data preprocessing is a critical step in addressing the influence of irrelevant, redundant, and unreliable data, ultimately improving data quality and resolving inconsistencies (Garcia, Luengo & Herrera, 2105). In this study, data preprocessing was conducted prior to training the machine learning models. Initially, the patient records with missing data were removed from the dataset. For example, records of sex, ethnicity, and reduced oxygen saturation were excluded if any missing values were detected. Missing values for the target variable were treated as the absence of the outcome of interest (death). Additionally, we utilized categorical encoding to transform nominal data into numerical representations. By applying one-hot encoding, we ensured that our analysis was guided by intrinsic relationships within the data rather than by the constraints of non-numerical representations (Xiang et al., 2021).

After applying the criteria for excluding data in the pre-processing step, we obtained a final sample consisting of 24,097 records. The dataset comprised 22,586 and 1,511 cases in the discharge and death classes, respectively. An imbalanced input distribution can lead to a bias in the results towards the dominant class, potentially skewing model performance and reducing generalizability. To address the problem posed by an imbalanced dataset, we employed the Synthetic Minority Over-sampling Technique (SMOTE) method, as outlined in https://imbalanced-learn.org/stable/. The SMOTE algorithm, which is widely utilized for synthetic oversampling, generates artificial samples for the minority class by randomly selecting instances from the minority class and their k-nearest neighbors. In this approach, a random data instance along with its k-nearest neighbors is chosen. Subsequently, the second data instance was selected from this set of k-nearest neighbors (Dorn et al., 2021). The synthesis of a new sample occurred along the line connecting these two instances as a convex combination. This process was iterated until a balance was achieved between minority and majority classes. The SMOTE method mitigates the risk of overfitting, distinguishing it from the random oversampling technique, and it is recognized for its potential to produce better results (Erol et al., 2022; Wang et al., 2021; Wongvorachan, He & Bulut, 2023).

Feature selection

Chi-square tests were used to discern statistically significant differences between the outcomes of discharged and deceased patients. Feature importance scores derived from XGBoost and random forests (as detailed in Fig. S1) were utilized to identify the essential variables for forecasting COVID-19 mortality. This methodology aims to increase the interpretability and steadfastness of mortality prediction models.

Feature selection techniques exhibited elevated scores for robust predictors such as overall comorbidities, diminished oxygen saturation, and age. Nevertheless, some disparities were evident in the importance scores between XGBoost and random forest for specific parameters. XGBoost showed considerable importance in reducing oxygen saturation and overall comorbidities, whereas random forest allocated minimal importance. A statistically significant difference (P < 0.01) in oxygen saturation and total comorbidities was observed between patients who survived and those who died. Chi-square tests were applied to recognize crucial mortality predictors, demonstrating moderate to high importance in XGBoost and low importance in random forest.

Owing to the inconsistencies observed between the two methods, we opted to select the most pertinent features for training the models using the chi-squared test. Consequently, we developed three distinct datasets to train and validate the machine learning models. These datasets included a dataset with features selected using the chi-squared test, a dataset with features chosen by two pediatricians, and a dataset with all 37 features, according to Table 1, except for the target variable. Our objective was to determine the dataset that yielded the most favorable results.

The dataset containing characteristics chosen by pediatricians comprised 17 features: sex, age, ethnicity, region, virus strain, dyspnea, fever, cough, odynophagia, abdominal pain, ageusia, anosmia, respiratory distress, reduced oxygen saturation, total comorbidities, vaccine doses, and nosocomial. The dataset selected by the chi-squared test comprised 24 features: age, ethnicity, region, viral strain, dyspnea, cough, respiratory distress, reduced oxygen saturation, cardiology, pulmonary disease, hypertension, immunosuppression, renal disease, asthma, total comorbidities, comorbidities, dichotomous comorbidities, time for outcome, vaccine doses, hematology, neurology, oncology, Down syndrome, and nosocomial infection.

For the purpose to conducting feature selection calculations using the chi-square test, XGBoost, and random forest, the Scikit-learn library in its version 1.3.1 was used. The Pycaret library version 3.1.0 was employed for training and validating the models. Statistical significance was set at P < 0.01.

Outcomes

The primary endpoint was COVID-19-related death. Additionally, we assessed the severity of the disease, including hospitalization, need for respiratory support (none, non-invasive oxygen support, and mechanical ventilation), and admission to the intensive care unit (ICU).

Model development

In this study, a total of twelve machine learning algorithms were employed to develop predictive models. These algorithms included gradient boosting (GB), AdaBoost (ADA), CatBoost (Cat), random forest (RF), extreme gradient boosting (XGBoost), extra trees (ET), logistic regression (LR), linear discriminant analysis (LDA), decision tree (DT), naïve Bayes (NB), k-nearest neighbors (KNN), and Quadratic Discriminant Analysis (QDA) (Dorn et al., 2021).

These models were selected due to their superior performance compared to deep learning algorithms. Recent studies indicate that for tabular data, ML algorithms such as XGBoost, CatBoost, logistic regression, and decision tree family algorithms exhibit better performance than neural networks. Notwithstanding ongoing research efforts, neural networks have demonstrated limited efficacy in the processing of tabular data (Hwanga & Jongwoo, 2023; Shmuel, Glickman & Lazebnik, 2024; Shwartz-Ziv & Armon, 2022; Sivapathasundaram & Poravi, 2021).

The evaluation process involved the use of k-fold cross-validation, which is known to have low bias and variation. The optimized hyperparameters for the machine learning algorithms are provided in Table 2, with constant values maintained across the three variations of the SIVEP-Kids dataset.

Assessment metrics

The performance of the predictive model was evaluated using various metrics, such as accuracy, precision, sensitivity, F1 score, and area under the ROC curve (AUC). A comprehensive analysis was conducted across all 12 machine learning algorithms to determine the best model for predicting mortality in COVID-19 patients (Fawcett, 2006; Powers, 2011; Sokolova & Lapalme, 2009). Other performance metrics information will be detailed in the results section.

• Accuracy: Measures the proportion of correctly classified instances. $$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

• where TP = true positives, TN = true negatives, FP = false positives, and FN = false negatives (Sokolova & Lapalme, 2009)

• Precision: Assesses how many of the predicted positive instances are actually correct. $$Precision={\displaystyle \frac{TP}{TP+FP}.}$$

• Precision is particularly important in applications where false positives must be minimized (Powers, 2011)

• Recall (Sensitivity): Measures the proportion of actual positives that were correctly identified. $$Recall={\displaystyle \frac{TP}{TP+FN}.}$$

• This metric is crucial in applications where false negatives are costly (e.g., medical diagnosis) (Sokolova & Lapalme, 2009).

• F1-score: A harmonic mean of precision and recall, balancing both metrics. $$F1=2\times {\displaystyle \frac{Precision\times Recall}{Precision+Recall}.}$$

• F1-score is useful when dealing with imbalanced datasets.

• AUC-ROC: Measures the ability of the model to distinguish between classes by plotting the true positive rate against the false positive rate. $$AUC={\int }_{{\{}_0}^{1}TPR\left(FPR\right)d\left(FPR\right)$$

• where:

• FPR–false positive rate

• TPR–true positive rate

d (FPR) represents the infinitesimal variation in the false positive rate (FPR).

In practice, AUC is numerically calculated as the sum of the areas under the ROC curve, approximating the integral by summing small rectangular or trapezoidal regions along the curve.

AUC-ROC is widely used in binary classification problems to assess model discrimination power (Fawcett, 2006).

We employed SHAP summary and force plots to elucidate the decision-making processes of the models. Given its superior performance across all datasets, the gradient boosting classifier (GBC) was selected for in-depth analysis. SHAP summary plots visualize feature importance by mapping the impact of each feature on the model output to a dot on the horizontal axis. The position of the dot represents the SHAP value, which quantifies the contribution of the feature to the prediction. Feature values were color-coded (red: high, blue: low) to reveal the direction and magnitude of their influence. For detailed individual predictions, please refer to the force plots in the Supplemental Material.

Feature selection

Twenty-four features, comprising demographic and clinical factors, were identified as the most relevant predictors using the chi-square independence test (Table 3). Additionally, Table 3 shows mean decreases in impurity and the importance scores of these variables calculated using the XGBoost and random forest algorithms. The descriptive statistics of these features are summarized in Table 4.

Table 3 indicates that age, cardiovascular disease, decreased oxygen saturation upon admission, total comorbidities, and time to outcome were significantly associated with patient outcomes, as determined using the chi-square test. These factors demonstrated a strong predictive power in distinguishing between fatal and discharged cases. This statistical significance is also apparent in the developed models and was of paramount importance in the training process.

In contrast, odynophagia, abdominal pain, fever, vaccination, transplant, diabetes mellitus, vomiting, other syndromes, sex, diarrhea, and ageusia were less predictive of COVID-19 mortality. Despite their clinical importance in treatment and mortality risk assessment, many of these factors could be excluded from our machine learning models without compromising predictive accuracy. This demonstrates the potential of simplifying mortality prediction while maintaining effective outcomes.

Assessment of the developed models

In this study, COVID-19 mortality prediction models were developed using 12 ML algorithms, namely, GBC, ADA, CatBoost, RF, XGBoost, ET, LR, LDA, DT, NB, KNN, and QDA. These models were trained on three feature datasets: Dataset 1, containing all features; Dataset 2, with features selected by pediatricians; and Dataset 3, with features selected by the chi-squared independence test. The performance evaluation metrics used were accuracy, AUC, recall, precision, and sensitivity. The results are shown in Fig. 1.

In general, most of the models demonstrated comparable levels of accuracy, displaying good to excellent performance across all three datasets. More specifically, numerically, the models performed best when trained on Dataset 3, which was selected using the chi-square method, followed by Datasets 2 and 1. However, Dataset 1 still exhibited commendable performance even when all features were included. For Dataset 3, the highest accuracies were achieved by LR (92.53%), GBC (92.34%), and ADA (92.19%). For Dataset 2, GB (92.08%), ADA (91.92%), and LR (91.73%) achieved the highest accuracy. For Dataset 1, GBC (91.41%), ADA (90.32%), and CatBoost (90.01%) were the best-performing models in terms of accuracy. Among the 12 algorithms analyzed, QDA consistently displayed the lowest performance across all datasets. Detailed comparison of the AUC for the top three models trained on Dataset 3, which achieved better results, is provided in Fig. 2. Considering the reliability of the AUC metric for imbalanced datasets, particularly relevant in our study despite using SMOTE for balancing, is crucial. The AUC results are nearly identical across all three datasets, with a notable emphasis on Dataset 1 containing all features.

Model interpretation

Figure 3 presents a SHAP summary plot that visualizes the impact of each feature on the model’s predictions for individual data points. Each line represents a data point, with the points distributed along the feature axis indicating their corresponding values. A wider spread of points for a given feature suggests a stronger influence on the model’s output. Among the features, “oxygen saturation reduced” demonstrates the most significant impact on predictions. Blue points, representing normal oxygen saturation levels, are associated with favorable outcomes (patient discharge), whereas red points (low oxygen saturation) correlate with unfavorable outcomes (death). The comorbidity variable (categorical) also had a notable influence. Higher values of this feature, indicating a greater number of comorbidities (ranging from 0 to ≥ 3), are linked to an increased likelihood of predicting death. Similar trends were observed for “dyspnea”, “respiratory distress”, “total comorbidities”, and “comorbidities”.

Key points

This study aimed to develop and evaluate machine learning (ML) models for predicting COVID-19 mortality risk in Brazilian pediatric patients using a large public dataset. We analyzed demographic and clinical data to identify key mortality predictors. The ML models were trained using three datasets: (i) all available features; (ii) features selected by pediatricians; and (iii) statistically relevant features. Although all models demonstrated robustness, our findings suggest that feature selection significantly enhances model performance. The model trained on the statistically relevant feature set (Dataset 3, 24 features) achieved the highest accuracy, followed by the model trained on pediatrician-selected features (Dataset 2, 17 features). The model using all features (Dataset 1) showed lower performance and may not generalize to other datasets. Our results indicate that simpler models with fewer features, such as those based on datasets 3 and 2, are preferable for clinical use as they require less input while maintaining high predictive accuracy. Consistent across all models, older age, low initial oxygen saturation, and pre-existing chronic conditions emerged as the strongest predictors of COVID-19 mortality in children and adolescents.

Comparative analysis

We evaluated 12 ML algorithms for predicting mortality in hospitalized pediatric COVID-19 patients. LR demonstrated superior performance, achieving 92.5% accuracy, 98.11% sensitivity, 94.13% precision, 96.07% F1-score, and 80.15% AUC. GBC and ADA also yielded strong results (AUC ≥ 79.6%). Other models showed acceptable performance (AUC 80.1–81.6%), while DT and quadratic discriminant analysis (QDA) exhibited weaker results (AUC = 62.9%, accuracy 7.9–24.3%). To identify key predictors, we employed the XGBoost, random forest, and chi-squared tests. SHAP analysis revealed that reduced oxygen saturation, comorbidities, and older age were the most critical factors. These, along with 23 additional statistically significant features, enhanced ML model performance. Our findings align with previous studies that have reported some important clinical predictors for COVID-19 patient mortality, the most relevant features included age (Moulaei et al., 2021, 2022; Wu et al., 2020; Yadaw et al., 2020; Zakariaee et al., 2023a, 2023b), ethnicity (Baqui et al., 2020, 2021), geographic region (Baqui et al., 2020, 2021), dyspnea (Shi et al., 2020), cough (Assaf et al., 2020; Das, Mishra & Saraswathy Gopalan, 2020; Gao et al., 2020; Moulaei et al., 2021, 2022; Zakariaee et al., 2023a, 2023b), reduced oxygen saturation (Assaf et al., 2020; Banoei et al., 2021; Kar et al., 2021), cardiology disease (Allenbach et al., 2020; Assaf et al., 2020; Baqui et al., 2020, 2021; Das, Mishra & Saraswathy Gopalan, 2020; Yadaw et al., 2020; Zakariaee et al., 2023a, 2023b), pulmonary disease (Banoei et al., 2021; Zakariaee et al., 2023a, 2023b), immunosuppression (Baqui et al., 2020, 2021; Gao et al., 2021; Xu et al., 2021), renal (Baqui et al., 2020, 2021; Shi et al., 2020), asthma (Aktar et al., 2021; An et al., 2020; Chimbunde et al., 2023), total comorbidities (Aktar et al., 2021; Banoei et al., 2021), hematology disease (Huyut, Velichko & Belyaev, 2022; Kamel et al., 2023), neurology disease (Baqui et al., 2020, 2021; Moulaei et al., 2021, 2022; Zakariaee et al., 2023a, 2023b), oncology disease (Assaf et al., 2020; Chin et al., 2020; Hu, Yao & Qiu, 2020; Zakariaee et al., 2023a, 2023b), hypertension (Assaf et al., 2020; Das, Mishra & Saraswathy Gopalan, 2020; Yadaw et al., 2020; Zakariaee et al., 2023a, 2023b), and chromosomal abnormalities (Landes et al., 2021).

Below, we detail how each of these best models selected on tests works technically.

Gradient boosting classifier

Gradient boosting is a machine learning method based on sequential decision trees. It minimizes a loss function using gradient descent.

Gradient boosting steps:

• •

  Start with an initial prediction, typically the mean of the target values.

• •

  Train a decision tree $ht\left(X\right)$ to minimize the residuals of the previous prediction. $$Rt=Y-Ft-1\left(X\right)$$

• •

  The new prediction is updated by adding the weighted tree output: $$Ft\left(X\right)=Ft-1\left(X\right)+\eta ht\left(X\right)$$

• •

  Repeat the process until a stopping criterion is met, such as a maximum number of iterations or minimum error.

• The loss function depends on the task:

• •

  Binary classification: Log-loss $$L\left(y,\hat{y}\right)=-\sum _{i=1}^m\left[y_i\log \widehat{y_i}+\left(1-y_i\right)\log \left(1-\widehat{y_i}\right)\right]$$

• •

  Regression: mean squared error (MSE) $$L\left(y,\hat{y}\right)={\displaystyle \frac{1}{m}\sum _{i=1}^m{\left(y_i-\widehat{y_i}\right)}^2}$$

• Popular GBC-based models include XGBoost, LightGBM, and CatBoost, widely used in machine learning applications for structured data.

Machine learning models and larges scale datasets

The machine learning algorithms presented in our work have proven to be feasible for large databases. Specifically, models such as logistic regression, AdaBoost, and gradient boosting demonstrate significant utility due to their inherent scalability and computational efficiency relative to their predictive power. Logistic regression, particularly when implemented with stochastic gradient descent (SGD) or its variants, can process vast amounts of data incrementally without requiring the entire dataset to reside in memory, making it suitable for streaming or disk-based learning scenarios. Ensemble methods like AdaBoost and gradient boosting, while potentially more computationally intensive per iteration, derive their strength from building sequences of weak learners; this iterative nature allows for potential parallelization and, crucially, they often achieve high accuracy with relatively shallow trees as base learners, mitigating the complexity explosion sometimes seen in other non-linear methods. Their proven effectiveness across diverse large-scale benchmarks underscores their suitability for modern data-rich environments (Hastie, Tibshirani & Friedman, 2009).

Furthermore, the specific characteristics of these algorithms lend themselves well to the challenges posed by large databases, such as high dimensionality and the presence of complex, non-linear relationships. Logistic regression, often combined with L1 or L2 regularization, provides a robust linear baseline capable of handling sparse, high-dimensional feature spaces commonly encountered in large datasets, while also offering interpretable coefficients. Boosting algorithms, particularly gradient boosting, excel at capturing intricate patterns and interactions within the data by sequentially fitting models to the residuals of prior iterations. This adaptive fitting process allows them to model complex functions effectively without necessarily overfitting, especially when parameters like learning rate, tree depth, and subsampling are carefully tuned. The capacity of gradient boosting frameworks to optimize arbitrary differentiable loss functions further enhances their flexibility for diverse large-scale prediction tasks (Friedman, 2001).

Contrasting these machine learning approaches with traditional statistical methods reveals critical differences when applied to large databases. While traditional methods, such as ordinary least squares regression or maximum likelihood estimation for generalized linear models, provide rigorous inferential frameworks, they often rely on assumptions (e.g., normality of errors, specific distributional forms) that may be violated in massive, heterogeneous datasets. Moreover, many classical techniques involve computations, like matrix inversion, that scale poorly with the number of samples or features, rendering them computationally infeasible for very large n or p without specialized implementations. Machine learning models, particularly those discussed, often prioritize predictive accuracy and computational scalability, frequently employing iterative optimization techniques and making fewer stringent assumptions about data generation processes. This distinction, emphasizing prediction over parameter inference or exact distributional modeling, aligns well with the practical demands of extracting actionable insights from large-scale data repositories (Breiman, 2001).

Strengths of the study

This study leveraged a nationwide database to provide a comprehensive overview of COVID-19 in Brazilian pediatric patients. The large sample size of laboratory-confirmed cases enabled the rigorous evaluation of multiple ML algorithms. Our findings suggest that ML models are robust when applied to extensive specifically designed datasets, offering the potential for future public health applications.

Limitations of the study

This study has several limitations. First, the SIVEP-Gripe database, which focuses on hospitalized patients, restricts the generalizability of the findings to a broader pediatric population. Second, we were unable to conduct an external validation of our models, an important step in the development of clinical prediction models (Steyerberg & Harrell, 2016). This involves evaluating the performance of a model on an independent dataset that was not used during the model’s training or internal validation phases (Collins et al., 2024a; Collins & Moons, 2019; Collins et al., 2024b). This process addresses critical issues such as overfitting and bias, while ensuring that the model generalizes well to new, unseen data, which is essential for its real-world applicability. Nevertheless, many of the available datasets are too small to provide reliable answers (Riley et al., 2024a, 2024b). To address this pivotal issue, we are currently integrating data from all official Brazilian databases, including non-hospitalized and hospitalized patients, across the country. Therefore, we believe that with this updated dataset, encompassing more than two million pediatric cases, we will be able to use some modern recommended techniques, such as the use of resampling methods for internal validation, to evaluate model performance and generalizability across clusters. Third, the administrative nature of the database hinders the assessment of certain clinical management details. Additionally, missing data, a common challenge in such registries, was mitigated through meticulous manual review of case records, including in-depth analysis of the "clinical observation" field in SIVEP-Gripe database. Finally, the absence of a national audit system for the SIVEP Gripe database is a notable limitation. However, our extensive analysis of these data since the onset of the pandemic has yielded consistent results comparable to those from other low-to middle-income countries using conventional statistical techniques (Nachega et al., 2022).

Clinical and policy implications

We believe that the utilization of ML models to predict outcomes in pediatric COVID-19 cases and other public health threats presents a transformative potential for healthcare systems. The application of ML models in predicting mortality in children with COVID-19 may have significant clinical, research, and policy implications. Our findings indicate that ML models can assist in accurately identifying high-risk children at an early stage, enabling healthcare systems to allocate resources (e.g., ICU beds, ventilators, and medications) more effectively for those at the highest risk of severe outcomes. ML models can inform individualized treatment plans by identifying risk factors specific to pediatric populations, potentially leading to the development of tailored clinical guidelines (Chumachenko et al., 2024; Collins et al., 2024b; Singh, 2019; Wynants et al., 2020). Predictive models can help policymakers prioritize vaccination strategies for children at elevated risk of severe outcomes, particularly in resource-limited settings. Furthermore, ML models can be integrated into public health surveillance systems to monitor trends in pediatric COVID-19 mortality and to inform targeted interventions (Bragazzi et al., 2020). However, policymakers must address ethical, equity, and regulatory challenges to ensure the effective implementation of these tools. Collaboration among researchers, clinicians, and policymakers is essential to maximize the benefits of ML in pediatric care (Malhotra et al., 2023).