Anxiety and depressive disorders are the most prevalent mental disorders, with 301 million and 280 million individuals affected worldwide, respectively (1, 2). Mental disorders can develop at any age, where early years of life like childhood and adolescence are more vulnerable due to being turbulent and dynamic stages of growth and development (3). A meta-analysis published in 2015 estimates the global prevalence at 6.5% and 2.6%, respectively (4). Similar reviews conducted after COVID-19 reveal that the prevalence of childhood anxiety and depression has increased to 20.5% 25.2%, respectively (5). The presentation of these disorders at a young age is often undermined as transient mood lability and emotional bursts during puberty. However, it is estimated that almost 50% of mental disorders diagnosed in adulthood actually begin to develop during childhood and adolescence (6, 7). If these disorders remain untreated, they can become quite debilitating and affect an individual’s quality of life by reducing productivity and socialization. Early detection and timely treatment of these disorders can lead to better outcomes and improved prognosis as compared to later treatment in adulthood (8).

Numerous screening tools are used in clinical practice to readily detect the presence of mental disorders. A few tools used for screening anxiety and depressive disorders in young populations include the Revised Child Anxiety and Depression Scale (RCADS), Screen for Child Anxiety Related Disorders (SCARED), and Spence Children’s Anxiety Scale (SCAS) (9–11). While these tools are convenient to quickly detect the presence or risk of potential issues, they are limited to broadly label patients as “normal”, “borderline”, “at risk”, and/or “clinical”. The specific behaviors or symptoms that predominantly contribute to the development or progression of a disorder cannot be elucidated through these screening measures alone. Such evaluation and symptom profiling is critical for the optimization and personalization of therapeutic strategies to improve patient outcomes (12). However, repeated sessions with patients are required to pinpoint critical behaviors. This process is also dependent on the clinician’s experience and can be subject to clinical bias. An objective and methodical approach is required to highlight dominant symptoms and behaviors early in the screening process to readily formulate personalized treatment plans. The advent of Artificial Intelligence (AI) has introduced computational processes that are able to “learn” from patterns in data and extract critical information based on its mathematical significance or “contribution” in the process. Each variable is assigned a mathematical weight or importance metric depending on its effect upon the final outcome. This technique can be applied on screening tools to identify which specific questions of the tool (and subsequently, the associated symptoms) contribute significantly to disease progression.

Machine Learning (ML) is a field of AI which focuses on the development of computational techniques inspired by the human brain’s ability to learn, adapt, and improve (13). Over the past decade, the predictive capabilities of ML have been extensively researched in the field of healthcare for the development of fast and effective screening tools that can serve as clinical decision supports for healthcare professionals (14, 15). Coupled with clinician’s insights, such ML-based approaches can also prove as beneficial assistants for the prediction of mental disorders (16). Many studies have examined various machine learning algorithms on multidimensional data (demographics, socio-economic information, results of screening and diagnostic tests) for the early detection of depression and anxiety. Haque et al. developed models for the prediction of depression in Australian children using Random Forests with an accuracy of 95% (17), and another research from the same authors reported models based on Gaussian Naïve Bayes and Random Forest to be the best predictors of Obsessive-Compulsive Disorder, Separation Anxiety Disorder, and Attention-Deficit/Hyperactivity Disorder with accuracies ranging between 79-91% (18). A study on Palestinian adolescents used Support Vector Machines to develop models that can predict depression and anxiety at accuracies above 92% (19). Nemesure et al. used biometric and demographic data of university students to develop predictive models using an ensemble approach, where the models provided accuracies of 0.73 for Generalized Anxiety Disorder and 0.67 for Major Depressive Disorder (20). The primary objective of these studies was to develop a quick computational process for the prediction of mental disorders rather than focus on dominant behaviors or patterns for symptom profiling. There are a few studies that have utilized ensemble ML methods for the characterization of social risk factors that affect overall mental well-being of susceptible populations like children exposed to politically volatile circumstances (21). However, the utilization of such techniques for symptom profiling for anxiety and depressive disorders has not been reported thus far.

Moreover, there are no published studies that have analyzed these techniques on a Pakistani cohort. In addition to individual characteristics, environmental and social factors like poverty and inequality can also have detrimental effects on an individual’s mental well-being (3). Such problems are quite prevalent in low-middle income countries (LMICs) like Pakistan, where mental health stigma and uninformed cultural beliefs exacerbate the burden of mental disorders. Mental health conditions are frequently interpreted through moral or religious lenses, often perceived as personal weakness or insufficient faith, which discourages help-seeking behavior and contributes to underdiagnoses (22). The prevalence of collectivist family structures also tends to impose conformity, emotional restraint, and prioritization of family honor over individual psychological needs. Additionally, socioeconomic stressors, political instability, and limited access to mental health services interact with these cultural dynamics, worsening the prevalence and persistence of depression and anxiety in the Pakistani context. Therefore, determining the cultural significance of individual interrogatories or items of screening tools is also required for the optimization of mental healthcare.

The following two aims were central to this study:

All statistical and computational processes have been implemented in the programming language Python, version 3.10, using Colab notebooks.

The top 3 features elucidated by the models show that in case of Total Internalizing Issues, two items pertaining to unexplainable anxiety and one item regarding issues with appetite are the most dominant. For Total Anxiety, accelerated heartbeat due to anxiety, being worried about poor performance, and a general fear about something bad happening to oneself are the top 3 symptoms. For MDD, the top 3 features relate to the depressive symptoms of lethargy and lack of interest. In case of GAD, general fears regarding bad things happening to oneself or a loved one dominate, and for SAD, being away from home, avoiding or being anxious about going to school, and sleeping alone appear on top. For SP, fears related to how perceive oneself, presenting in front of an audience, and appearing foolish are the top 3 symptoms. In case of PD, the three features are related to unexplainable increase in fear, trembling, and faster heartrate. For OCD, obsessive thoughts are the dominant features as opposed to compulsive acts. For each individual, RF models can highlight a list of the most significant features or symptoms, which can be used to formulate personalized therapies for each individual, where the most critical behaviors are addressed expeditiously.

Similar studies on the predictive modelling or classification of anxiety and depressive disorders either utilize tree-based ensemble ML algorithms like RF and XGBoost, or report them as the best performing model from a pool of different classical and ensemble algorithms (17, 18, 20, 34). Similar results were also obtained in the present research, where RF outperformed other algorithms. This is attributed to the ensemble nature of the algorithm which optimizes output by the aggregation of multiple decision trees, and its inherent ability to deal with potential non-linear relationships among features (35). This may also be attributed to the nature of the data and classification task, which is fairly straightforward as the numerical value of a few features elucidate how each instance will be classified in to the three classes of Normal, Borderline, and Clinical. However, it is worth mentioning that the performance of SVM is comparable to RF in some cases. Qasrawi et al. also report better performance of SVM for the classification of depression and anxiety in their research (19). SVM is also a powerful classification algorithm that performs well on the type of dataset used in this research. However, the reason for preference of RF over SVM in this regard is not merely informed by the slightly better performance of RF and the recurrence of the algorithm in quoted literature. Random Forests are more robust and scalable algorithms which can handle non-linear data without the requirement of pre-standardization and deal with missing information via imputation (35). Support Vector Machines require complete and scaled data in order to perform efficiently (36). As the wholeness of medical data is not guaranteed in real-world application, Random Forests might prove to be a more practical choice as an efficient decision support system for the screening of mental disorders.

In the final models selected for each scale of RCADS, Gender has been removed as a predictive feature for Total Internalizing Issues, Total Anxiety, MDD, GAD, and OCD based on the performance metrics. This is interesting as the original T-scoring of RCADS is dependent on the gender of the patient, and research also explains a distinction between psychopathology of males and females (37). In contrast, removing the attribute of gender from the models for SAD, SP, and PD reduces their performance. Similar results are obtained for the attribute of grade level, where its removal from the models of Total Anxiety, MDD, and GAD improves model performance, while the metrics of models for SAD, SP, and PD reduce. Additionally, in case of Total Internalizing Issues, the removal of Items 5 and 9 improve model performance. These items of RCADS pertain to the specific behaviors associated with SAD, therefore, they may not be considered the primary contributors or predictors of general internalizing behavior. However, as these items are retained in the specific model for SAD, their omission from the scale of Total Internalizing Issues seems counter-intuitive. For a more robust comparison, the final RF models were further evaluated on the same test set, and Cohen’s kappa was calculated for models trained on all features and the χ²-subsets. This additional testing also revealed contrasting results, where values of κ either increased, decreased, or remained unchanged for some scales. For scales, where any difference in the value of κ was observed, McNemar’s test confirmed that the differences were not statistically significant. Due to the conflicting nature of the outcomes and the proof that any difference in model performance has occurred by chance, it is concluded that no changes need to be made in the composition of RCADS and its interrogatories based on this portion of the research alone. All items within RCADS are culturally significant in a Pakistani cohort, which is to be expected from a widely-used and validated psychometric scale curated by experienced professionals. However, the lower Cronbach’s alpha values obtained for the scales of SAD and OCD indicate that further studies regarding the composition of these subscales with respect to Pakistani populations are required.

It is important to discern that the aim of AI-driven decision support systems is not to provide a definitive diagnosis and, ultimately, a “replacement” for traditional clinical practices. The role of these proposed tools is the enhancement of the existing clinical process. In resource-limited areas like Pakistan, psychiatrists are greatly outnumbered by potential patients. Reportedly, there is only one fully trained mental health professional for every 360,000 patients in Pakistan (38). The statistics become even more drastic for younger patients, with only 1% of the outpatient mental health facilities specializing in child and adolescent psychiatry (22). The implementation of an additional screening aid in the clinical workflow could alleviate the burden of mental health professionals. The speed of AI-driven tools enables them to screen multiple individuals simultaneously, and their efficiency can allow the prioritization of patients that require urgent review from a professional. As the influx of patients in child and adolescent psychiatry departments is quite low, it would be more beneficial to equip general or pediatric outpatient departments with these decision support systems. In addition to conventional check-ups, if parents or frontline healthcare providers are concerned about the mental well-being of their child or patient, they can quickly screen them using the proposed predictive models. This preliminary screening can then inform if the patient should be referred to a mental health specialist, and the symptom profiling can facilitate the specialist in forming patient-centered therapeutic regimes.

While the study establishes the importance of each item of RCADS and provides RF algorithms that can screen mental disorders efficiently, there are certain limitations that may affect the generalizability of the proposed models. Firstly, the sample size is quite small. Although similar sample sizes have been reported in related literature (24–28), ML algorithms require extensive data for optimal performance. While the implementation of SMOTE-N slightly increases the dataset, this is an artificial or synthetic inflation of data, which is not comparable to real instances that are collected organically. Secondly, the gender distribution is quite imbalanced which can be attributed to the sampling methods used for data collection. Both consecutive and purposive sampling are non-probability sampling techniques which can induce bias. However, these methods were employed for convenience, as data collection for mental health research poses some hurdles due to limited clinical cases and socio-cultural implications. Additionally, the reported predisposition of affective disorders in females may also contribute to the gender imbalance (37). Thirdly, the data features only consist of gender, grade, and the 47 items of RCADS. Socio-cultural attributes like family structure, socioeconomic status, and parental education levels are primary determinants of a child’s cognitive development (39, 40). Therefore, their inclusion with questions regarding behaviors and personality traits will result in inclusive and comprehensive screening tools.

In terms of future prospects of this study, the proposed models should be further optimized using a larger and uniform dataset. Socio-demographic and biometric indicators of patients should also be incorporated for more robust cultural adaptation of proposed decision support systems. This study serves as a foundation for the predictive modelling and symptom profiling of depressive and anxiety disorders in at-risk children and adolescents using ubiquitous and standard screening practices. It shows that simple computational approaches can provide efficient decision support systems for healthcare professionals to alleviate their workload and optimize patient outcomes. However, it is appreciated that the implementation of such technological interventions requires extensive validation on external datasets that the models have not been exposed to during initial training. This external validation will investigate the generalizability and reproducibility of the models on varying cohorts.