DepressNet: an interpretable machine learning framework for depression severity stratification using psychometric data

1 Introduction

Depression is a mood disorder characterized by persistent low mood, reduced interest in activities, and impairment in daily functioning. According to the World Health Organization (WHO), it is a leading cause of disability-adjusted life years (DALYs) worldwide (World Health Organization, 2025). Estimates suggest that approximately 332 million people globally are affected, with a lifetime prevalence of major depressive disorder (MDD) of around 20% among adults (Santomauro et al., 2021; Chen et al., 2025). The burden is especially pronounced in low- and middle-income countries, where over 75% of affected individuals do not receive adequate treatment due to limited clinical infrastructure (Thornicroft et al., 2017). As a result, understanding and modeling depression severity through scalable computational approaches remains an important research area. In computational research, structured self-report instruments such as the Patient Health Questionnaire-9 (PHQ-9) and the Depression Anxiety Stress Scales (DASS-42) provide standardized psychometric data that can be analyzed using machine learning techniques (Sharma and Verbeke, 2020). These datasets have enabled the development of predictive models that learn associations between questionnaire responses and symptom severity patterns. Ensemble methods, including Random Forest and XGBoost, have shown strong performance on tabular health-related data (Bader et al., 2024; Sharma and Verbeke, 2020; Chung and Teo, 2023). However, much of the existing literature formulates depression-related prediction as a binary classification problem (e.g., depressed vs. non-depressed). This simplification reduces the ordinal structure of severity levels into a single threshold, limiting the ability of computational models to reflect clinically defined gradations in symptom intensity (Nusrat et al., 2024; Torous et al., 2020). From a computational perspective, multi-class depression severity modeling introduces several challenges. The DASS-42 dataset used in this study exhibits a highly imbalanced distribution, where extremely severe cases constitute 68.2% of observations, while Mild cases account for only 4.4%. This imbalance arises from self-selection bias in online psychometric platforms (Kosolwattana et al., 2023). Models trained under such conditions often struggle to learn minority class boundaries effectively. In addition, while individual classifiers such as gradient boosting models perform well on structured data, they may not fully capture complex decision boundaries across multiple severity levels, motivating the use of ensemble learning strategies (Shatte et al., 2019; Salehi and Khedmati, 2024). Another limitation of many existing approaches is their limited interpretability, particularly in terms of identifying which questionnaire items most influence predictions for different severity levels (Shaha et al., 2024; Rudin, 2019). Stacking ensemble learning provides a framework in which out-of-fold predictions from heterogeneous base learners are used to train a higher-level meta-learner. This approach has demonstrated strong performance in tabular classification tasks, particularly in healthcare-related applications (Lundberg and Lee, 2017; Wolpert, 1992). Despite this, relatively limited work has explored stacking frameworks that combine neural meta-learners with GPU acceleration for psychometric data analysis. In parallel, explainable AI methods such as SHAP (SHapley Additive exPlanations) have been widely used to interpret machine learning models in mental health research, although detailed class-wise decomposition for severity-specific interpretation remains underexplored (Ahmed et al., 2025; Liu et al., 2026; Mimikou et al., 2025). Within this context, this study focuses on developing an interpretable machine learning framework for analyzing depression severity patterns in DASS-42 questionnaire data. DepressNet framework integrates XGBoost, Random Forest, LightGBM and calibrated SVM as base learners with a PyTorch-based multilayer perceptron meta-learner. The model is evaluated on the OpenPsychometrics DASS-42 dataset (N = 39,775) for five-level severity stratification. In addition, a dual explainability module based on SHAP is used to analyze both global and class-wise feature attribution patterns across severity levels. To our knowledge, this is the first study to apply class-wise SHAP decomposition across all five DASS-42 severity levels, providing structured symptom-response interpretation at each severity stratum.

This study makes four principal contributions:

• Stacking architecture (DepressNet): DepressNet is a hybrid ensemble framework that integrates four complementary base learners—XGBoost, Random Forest, LightGBM and calibrated SVM—with a PyTorch-based multilayer perceptron (MLP) serving as the meta-learner. The stacking procedure is implemented using out-of-fold (OOF) probability predictions to ensure proper separation between training and validation data and to prevent information leakage during meta-model training. This design allows the meta-learner to effectively learn from the probabilistic outputs of diverse base models while maintaining methodological rigor in the evaluation process.
• GPU-accelerated training pipeline: A training pipeline leveraging CUDA-compatible base learners and a PyTorch MLP meta-learner is developed, featuring cosine annealing learning rate scheduling, class-weighted loss, and best model checkpointing to support scalable experimentation on large psychometric datasets.
• Dual explainability module: The global SHAP attribution identifies the most influential DASS-42 items across the dataset, while class-wise decomposition enables interpretation of feature importance variations across different severity levels, revealing structured symptom attribution patterns.
• Comprehensive statistical validation: Statistical validation is performed using the Friedman test along with the Nemenyi post-hoc test, the Wilcoxon signed-rank test, and the McNemar test to assess performance differences between models and evaluate the relative improvement of the proposed framework over baseline classifiers.

In Section 2, the recent related works have been addressed along with finding their contributions and research gaps. The overall methodology of this research has been demonstrated in Section 3. The efficacy and experimental results are analyzed in Section 4. The interpretation of the model is also analyzed in the same section. The future direction and conclusion are discussed in Section 5.

2 Related works

2.1 Machine learning for depression detection

Machine learning techniques on a diverse range of data types and evaluation tools have been applied to detect depression. Sharma and Verbeke (2020) utilized the XGBoost algorithm with multiple resampling strategies on a big Dutch biomarker dataset (n = 11,081) for the classification of depression. Bader et al. (2024) implemented weighted Random Forest and SMOTE on the PHQ-9 data set that is composed of 830 samples. It achieved an AUC score of 0.91 for the multiclass severity classification based on oxidative stress biomarkers, enhancing that with sociodemographic variables. Chung and Teo (2023) empirically evaluated single classifier vs. ensemble machine learning approaches for mental health prediction using the OSMI Tech Survey dataset, finding that Gradient Boosting achieved the highest accuracy of 88.80%, with XGBoost achieving 87.20%, demonstrating the strong potential of boosting-based methods for mental health classification tasks.

Limited attention has been given to the multi-class case severity classification of depression. Although multi class severity classification of depression remains relatively underexplored, recent work has focused on handling class imbalance; for example, Kosolwattana et al. (2023) proposed an adaptive SMOTE variant (SASMOTE) to improve classification on highly imbalanced healthcare datasets, illustrating methods that could be applied to depression severity problems. In order to classify depression in students on an extensive dataset, GLNet is being introduced by Li et al. (2025). This model integrates Mamba and convolutional layers. The results achieved by this model were 88.84%. The authors identify academic pressure and financial stress as two vital reasons for their analysis. According to the ML-based PHQ-9 item pairing study by Glavin et al. (2023), item combinations (PHQ-2&4) were found to have an AUC of 0.954 for depression prescreening, highlighting the efficacy of item-level feature selection over aggregate scoring.

2.2 Ensemble methods in mental health classification

Ensemble learning has consistently outperformed individual classifiers in mental health prediction tasks (Shatte et al., 2019; Iyortsuun et al., 2023). Stacking–training a meta-learner on the out-of-fold predictions of diverse base learners has shown particular promise for tabular data by exploiting complementary learning biases of heterogeneous classifiers (Wolpert, 1992). Salehi and Khedmati (2024) proposed CSBBoost, a cluster-based SMOTE ensemble combining XGBoost, Random Forest, and bagging with a balanced meta-aggregation strategy, demonstrating consistent AUC improvements over individual models on 20 benchmark datasets. The systematic review by Shatte et al. (2019) identified ensemble methods as the most consistently high-performing approach across mental health classification tasks.

Neural meta-learners in stacking architectures have received limited exploration in mental health contexts. Most existing stacking approaches use logistic regression or linear models as meta-learners (Ahmmed et al., 2025), failing to capture non-linear interactions in the meta-feature probability space. Ahmmed et al. (2025) proposed a stacking ensemble with logistic regression mediation for depression prediction among professionals, achieving 98.75% test accuracy on a synthetic Kaggle dataset. However, the use of synthetic data and simple meta-learners limits the generalizability of these findings to real clinical populations.

2.3 Explainable AI in mental health

Doctors and clinicians need interpretable explanatory models for clinical deployment of ML models in their decision making. SHAP continues to be the most used and cited framework for post-hoc feature attribution in tabular ML models. It provides global population-level importance rankings, or global interpretability, and local instance-level explanations or local interpretability (Lundberg and Lee, 2017). Ahmed et al. (2025) utilized SHAP and LIME to explain an XGBoost model for depression detection and severity classification using wearable actigraphy signals for circadian disruption (i) Power spectral density and (ii) Autocorrelation found age as a major predictor. XGBoost obtained 85.91% multiclass accuracy comparable with our DepressNet.

In nutrition epidemiology, Liu et al. (2026) applied explainable AI methods, including SHAP and LIME, to interpret ML models predicting comorbid depression and stroke from daily dietary nutrient intake in a US population study based on NHANES data. Thus, their paper showed how combined global-local explainability may be useful to derive clinical insights. Mimikou et al. (2025) explored explainable machine learning approaches for depression prediction and reported strong performance for tree based models; they highlighted several demographic and psychosocial predictors as influential. Specific claims about SHAP combined with a genetic algorithm and the 97.83% accuracy should be supported only if drawn from the corresponding study that reports those methods and results. A recent systematic review (Trelles et al., 2025) analyzed explainable ML models for clinical depression detection; while some studies report strong performance for XGBoost, aggregate performance metrics (for example, mean F1 and AUC) should be cited only from the review that computed those summary statistics and attributed them to XGBoost. Seventy percent of the studies used SHAP because it is a well-known interpretability method. To the best of our knowledge, no paper has so far applied class-wise SHAP decomposition over five severity levels of depression for monitoring symptom importance evolution.

2.4 DASS-42 in computational research

Depression Anxiety Stress Scales (DASS-42) developed by Lovibond and Lovibond (1995) is a well-validated self-report tool. The DASS-42 provides standardized sub-scale scores for depression, anxiety, and stress with accepted clinical cutpoints. Sulak and Koklu (2024) utilized five data mining algorithms, namely Naive Bayes, Artificial Neural Networks (ANN), Logistic Regression (LR), Support Vector Machines (SVM), and Random Forests (RF), to the DASS-42 data collected from 680 students. It was found that ANN gave the highest accuracy of 99.26% for classification of depression. Nonetheless, this finding was achieved on a small, demographically homogeneous sample with features recording the subscale sum scores which get directly encoded into the target label, suggesting data leakage. To address this, we use raw item-level responses (Q1A-Q42A) from the DASS-42 obtained from 39,775 participants distributed globally with no mention of any computed subscale scores in the feature matrix to prevent leakage.

3 Dataset and feature engineering

3.1 Data source

The current study uses the DASS-42 dataset from OpenPsychometrics, which was collected through online administration of the Depression Anxiety Stress Scales on the OpenPsychometrics platform (openpsychometrics.org) (OpenPsychometrics, 2019) between 2017 and 2019. The unprocessed data contains 39,775 complete responses from those participants who agreed to the use of their responses for research purposes by stating their answer as accurate at the end of the survey. Global distribution of the participants brought demographic diversity across age, gender, education level, and culture, which cannot be achieved in a clinical sample from the same institution.

3.2 DASS-42 instrument

The DASS-42 contains 42 items rated on a four-point Likert scale (1 = Did not apply to me at all; 4 = Applied to me very much or most of the time), assessing the frequency of affective symptoms over the preceding week. Items are mapped to three subscales: Depression (14 items: Q3, Q5, Q10, Q13, Q16, Q17, Q21, Q24, Q26, Q31, Q34, Q37, Q38, and Q42), Anxiety (14 items), and Stress (14 items). Raw subscale scores range from 0 to 42 after recoding responses to a 0–3 scale (Lovibond and Lovibond, 1995). The instrument has demonstrated strong psychometric properties across diverse populations, with Cronbach's alpha values ranging from 0.83 to 0.97 across subscales and samples, with Depression and Stress subscales consistently exceeding 0.90, and good discriminant validity (Sulak and Koklu, 2024).

3.3 Target variable: depression severity

The Depression severity label was derived from the raw Depression subscale score, multiplied by two, and 194 classified using the standard DASS severity cutpoints (Lovibond and Lovibond, 1995). Class distribution and five-class encoding are shown in Table 1.

Table 1. Depression severity class distribution based on DASS-42 clinical cutpoints (N = 39, 775).

3.4 Feature set

The feature matrix comprises all 42 raw DASS item responses (Q1A-Q42A) plus 11 demographic variables (age, gender, education, urban background, English nativity, handedness, religion, sexual orientation, race, marital status, and family size), totaling 53 predictors. Raw Depression, Anxiety, and Stress subscale scores were explicitly excluded as they directly encode the target label (Kaufman et al., 2012). No missing values were present after initial filtering, yielding a complete-case dataset for all analyses.

4 Methodology

4.1 Preprocessing pipeline

Z-score normalization was applied to standardize continuous demographic features (age, family size) that were limited exclusively to the training partition to avoid leakage from the test. All 53 features were confirmed to be statistically significant (p < 0.05) in the classification task of depression severity, using ANOVA F-test feature selection. The dataset was split using stratified random splitting (80% train, 20% test) to preserve the class distribution over the splits.

SMOTETomek, a combined oversampling-undersampling technique was applied only on the training partition to correct the severe class imbalance (Batista et al., 2004). The SMOTETomek class generates synthetic instances belonging to the minority class using SMOTE interpolation. Moreover, it removes borderline instances of the majority class by using Tomek Links. It results in a training dataset with 108,535 instances (21,707 for each class). To ensure evaluation in conditions resembling real deployment, the held-out test set (7,955 samples) was not ever balanced, which means the class distribution was as originally created (Abdelhay et al., 2025).

4.2 DepressNet architecture

DepressNet is a two-level stacking ensemble designed for five-class depression severity classification. Figure 1 shows the entire architecture from raw input features over base learner training to out-of-fold meta-feature generation, MLP meta-learner training and final class probability output.

4.2.1 Base learners

Four complementary classifiers serve as base learners, selected for their complementary inductive biases and GPU acceleration capability:

• XGBoost (Chen and Guestrin, 2016): Extreme Gradient Boosting with CUDA acceleration (device = “cuda”), 300 estimators, learning rate 0.05, max depth 6, subsample ratio 0.8, column subsample ratio 0.8, and multi-class log-loss objective.
• Random Forest (Breiman, 2001): Bagged decision tree ensemble with 300 estimators, max depth 12, balanced class weights, and full CPU parallelization (n_jobs = –1).
• LightGBM (Ke et al., 2017): Leaf-wise gradient boosting with GPU acceleration (device = “gpu”), 300 estimators, 63 leaves, learning rate 0.05, and balanced class weights.
• Calibrated SVM: Linear support vector classifier (LinearSVC, C = 1.0, balanced class weights) wrapped in CalibratedClassifierCV for Platt-scaled probability outputs, enabling compatibility with the stacking protocol's probability-based meta-features.

4.2.2 Out-of-fold stacking protocol

To prevent data leakage from base learner training into the meta-learner—a critical methodological requirement in stacking (Wolpert, 1992)—out-of-fold (OOF) predictions were generated using five-fold stratified cross-validation on the balanced training set. For each fold, each base learner was trained on the remaining four folds and predicted class probabilities for the held-out fold. The resulting OOF probability vectors (shape: n train × 20, i.e., 4 models × 5 classes) constitute the meta-feature matrix. Test-set meta-features were computed as the mean prediction across all five folds for each base learner.

4.2.3 PyTorch MLP meta-learner

The meta-learner is a three-layer MLP implemented in PyTorch and trained on the Tesla T4 GPU (CUDA). The architecture (Figure 1, right panel) comprises two hidden layers (256 and 128 units) with ReLU activations, batch normalization, and dropout regularization (rates 0.4 and 0.3), followed by a five-unit linear output layer. Class-weighted cross-entropy loss was applied, with weights inversely proportional to class frequency, preserving sensitivity to minority severity classes. Training used the Adam optimizer (lr = 0.001, weight decay = 1e-4) with cosine annealing scheduling over 60 epochs and best-model checkpointing.

4.3 Baseline models

Six classifiers were established as baseline classifiers Logistic Regression, Decision Tree, Random Forest, XGBoost, LightGBM, MLP individual-MLP. To perform cross-validation, all baselines are wrapped in an ImbPipeline which encapsulates SMOTE oversampling in each CV fold. In this way, synthetic samples are created only from the training fold data to prevent leakage into the validation folds, as shown in Batista et al. (2004). This pipeline was applied to the raw (unbalanced) training partition with a five-fold stratified cross-validation. In order to carry out a final test-set evaluation of the different baselines, these were trained on the full SMOTETomek-balanced training set, and evaluated on the held-out test set, using exactly the same feature sets and metrics used to evaluate DepressNet.

4.4 Evaluation protocol

The performance of the models considered in this study was evaluated according to four complementary metrics for imbalanced multi-class classification, namely: (1) Test Accuracy, that is, the proportion of correctly classified samples; (2) Macro-averaged F1-score, where all classes are equally weighted regardless of their size, allowing for an unbiased evaluation in imbalanced settings; (3) Matthews Correlation Coefficient (MCC), which is a balanced measure that's robust to class imbalance and is guaranteed to yield high values only when the model's predictions are correct across all classes; and (4) Macro-averaged AUC-ROC (one-vs-rest), that is, the area under the receiver operating characteristic curve that quantifies discriminative ability across all severity classes simultaneously (Chicco and Jurman, 2020).

4.5 Statistical testing

Statistical significance of inter-model performance differences was assessed via four complementary tests: (1) Friedman test—a non-parametric equivalent of repeated-measures ANOVA for comparing multiple classifiers across multiple folds; (2) Nemenyi post-hoc test—pairwise comparison controlling family-wise error; (3) Wilcoxon signed-rank test—pairwise comparison of DepressNet against each base-line across CV folds; and (4) McNemar test—comparing per-sample correctness of DepressNet and the best baseline on the held-out test set (Chicco and Jurman, 2020).

4.6 Explainability framework

SHAP TreeExplainer was applied to the trained XGBoost base learner to compute feature attributions for all 7,955 test instances (Lundberg and Lee, 2017). For multi-class output, SHAP returns a three-dimensional array (samples × features × classes). Global feature importance was computed as mean absolute SHAP value averaged across samples and classes. Class-wise analysis averaged absolute SHAP values per severity class, identifying the most influential questionnaire items at each level. Figure 2 illustrates the explainability pipeline.

5 Results and discussion

5.1 Overall performance comparison

Table 2 presents the complete performance comparison of all baseline models and DepressNet on the held-out test set (N = 7,955), alongside five-fold cross-validation results on the training partition. Figure 3 provides a visual comparison of test accuracy and macro F1-score across all models.

Table 2. Performance comparison of all models on the DASS-42 depression severity classification task (five-class, N_test = 7, 955).

Best values in each column are highlighted in bold.

DepressNet achieves the best Test Accuracy (85.38%), Macro F1-score (59.01%), and MCC (0.7158) among all the evaluated models, achieving consistent empirical improvement over the strongest individual baseline, the MLP standalone (85.02%, 58.41%, 0.7095), under the adopted experimental protocol. The AUC-ROC of 0.9714 is slightly lesser than that of the standalone MLP (0.9727), indicating a trade-off between discriminative calibration and per-class sensitivity balancing. There is a big difference between the CV accuracy and the test accuracy across the tree-based models. This indicates that there is a domain shift between the balanced training distribution (after SMOTETomek) and the original heavily skewed test distribution. This particularly affects the minority severity classes.

5.2 Per-class analysis

The classification report per class for DepressNet on the held-out test set is shown in Table 3. The confusion matrix is shown in Figure 4.

Table 3. Per-class classification report for DepressNet on the held-out test set (N = 7, 955).

Performance across severity classes varies a lot due to the inherent difficulty of the classification task under natural class imbalance. The Extremely Severe category makes up 68.2% of the test set, resulting in a perfect score of F1 = 1.00 due to its proper representation and unique symptom profile. The model has a moderate precision of 0.56 and a high recall of 0.93 for Severe class, which indicates that the model adequately prioritized sensitivity for clinically important cases. Moderate shows balanced performance (F1 = 0.45) in accordance with its intermediate and ambiguous symptoms. The most difficult subclass to classify is the Mild class (4.4% of test set, F1 = 0.10). It has a considerable symptom overlap with the Normal and Moderate classes. This limitation has been observed in all models assessed and is discussed further in Section 5.5.

To further address the misclassification of the Mild class, the MLP meta-learner was retrained using Focal Loss (γ = 2.0) as a cost-sensitive alternative to the standard class-weighted cross-entropy loss. Focal Loss assigns greater emphasis to hard-to-classify minority samples by exponentially down-weighting well-classified easy examples. However, the experimental results indicate no meaningful performance improvement. The Mild class F1-score remains at 0.10, while the overall Macro F1-score decreases slightly from 59.01% to 58.39% (−0.62%). These findings suggest that the difficulty in classifying the Mild class is primarily structural rather than optimization-related. Specifically, the DASS-42 Mild severity band (scaled scores 10–13) spans a narrow four-point interval positioned between the Normal (0–9) and Moderate (14–20) classes, resulting in substantial overlap in symptom patterns and inherent ambiguity in item responses. Consequently, neither resampling nor cost-sensitive learning is sufficient to fully resolve this boundary ambiguity. Future work will investigate ordinal classification methods, hierarchical two-stage classification, and Mild-targeted boundary-aware oversampling techniques to improve discrimination of borderline cases.

5.3 Ablation study

Table 4 presents the ablation study results quantifying each base learner's individual contribution to DepressNet's performance. The ablation study was conducted using the three tree-based base learners (XGBoost, Random Forest, and LightGBM) to isolate their individual contributions. All configurations were trained and evaluated under uniform experimental conditions. The final DepressNet model reported in Table 2 additionally includes a Calibrated SVM base learner; consequently, the full configuration figures here differ slightly from those in Table 2. Figure 5 visualizes the ablation findings.

Table 4. Ablation study showing the performance impact of removing each tree-based base learner from the tree ensemble.

Bold values indicate the best performance in each column.

The ablation results indicate that each base learner adds incrementally to the complete ensemble. The performance drop caused by removing Random Forest is the largest, which results in MCC decreasing from 0.6994 to 0.6786, presenting a decrement of Δ = −3.1%. Thus, the most important base learner is confirmed to be Random Forest because of its complementary bagging-based feature sampling that covers different decision boundaries from the gradient boosting-based methods used in the other three base learners. Omitting XGBoost results in the second-highest decline (MCC: Δ = −1.3%). Eliminating LightGBM brings about the least single drop (Δ = −0.7%), yet having it in positively affects performance consistently. Full configurational settings of DepressNet outperforms all three metrics validating the design choice of combining all three learners.

5.4 Statistical significance analysis

The Friedman test analysis using five-fold cross-validation macro F1-scores of all seven models resulted in a test statistic of 21.34 (p = 0.0007), indicating significant performance difference. Nemenyi post-hoc analysis informs us of a significant pairwise difference found between DepressNet and MLP alone (p = 0.000343) and a MLP and Logistic Regression with p = 0.0166. Comparisons via Wilcoxon signed-rank tests against each single baseline reveal DepressNet yields p = 0.0625. This indicates almost consistent improvement toward alpha = 0.05 due to the few numbers of folds (5) used. The results of the McNemar test conducted on held-out test predictions (DepressNet vs. MLP standalone: statistic = 1.88, p = 0.170) confirmed that differences per sample are not statistically significant at the test level. This is because they reflect marginally but consistently improved accuracy across the entire test distribution rather than due to particular subpopulations benefiting the most.

Given that the McNemar test does not demonstrate statistically significant per-sample superiority (p = 0.170), claims of outperformance in this study should be interpreted as consistent empirical improvements under the adopted experimental protocol rather than as statistically demonstrated superiority at the individual prediction level. The Friedman test (p < 0.001) confirms significant differences among the full model set, and the Nemenyi post-hoc result (p = 0.0003 vs. MLP standalone) provides pairwise evidence of DepressNet's advantage at the cross-validation level.

5.5 Explainability analysis

Before presenting the SHAP results, an important methodological clarification is required. All SHAP values reported in this section represent model-derived feature attributions that reflect learned associations within the DepressNet framework. These values quantify each feature's contribution to the model's classification decisions and should not be interpreted as clinically validated causal mechanisms or as evidence that specific questionnaire items independently cause depression severity. The patterns described below are consistent with established clinical frameworks and are offered as model-level exploratory insights to support human understanding, rather than as causal clinical claims. Independent clinical validation is required before attributing causal significance to any identified feature patterns.

5.5.1 Global feature importance

Figure 6 presents the global SHAP feature importance bar chart. Table 5 lists the 10 most influential DASS-42 items with their mean absolute SHAP values and clinical descriptions.

Table 5. Top 10 most influential DASS-42 features ranked by global mean absolute SHAP value across all classes.

The 10 most influential predictors align precisely with the core symptom domains of Major Depressive Disorder as defined in the DSM-5 (APA, 2013): persistent sadness (Q13A, Q26A), worthlessness and negative self-evaluation (Q17A, Q34A), anhedonia and loss of interest (Q16A), hopelessness and negative future orientation (Q21A, Q10A, Q38A), and psychomotor retardation and avolition (Q5A, Q42A). This convergence with established clinical diagnostic criteria provides face validity for the model's learned representations, suggesting that the feature attribution patterns are clinically plausible. However, this alignment reflects the model's learned associations within the DASS-42 scoring structure and should not be interpreted as evidence that these items causally determine depression severity. These findings are consistent with SHAP-identified predictors in related mental health ML studies (Ahmed et al., 2025; Mimikou et al., 2025).

5.5.2 Class-wise symptom progression pathway

Figure 7 presents SHAP beeswarm plots for the Severe class. Table 6 summarizes the top five predictors per severity class based on class-wise SHAP decomposition.

Table 6. Class-wise SHAP top predictors and dominant clinical symptom domains for each depression severity level.

The SHAP decomposition for each Class reveals a clinically meaningful pathway for the progression of depression.

• Normal: Normal mood is defined by transient indicators of downheartedness, sadness, and low motivation with no psychiatric significance. This class is distinguished by an absence of worthlessness or of an existential sort.
• Mild: The initial appearance of worthlessness (Q21A: life not worthwhile; Q17A: not worth much) and decline of motivation (Q5A) suggesting the transition from subclinical to clinical MDD is mild, consistent with the cognitive onset model of MDD (APA, 2013).
• Moderate: Disruption of self-schema becomes dominant (Q34A, Q13A, Q16A), with emerging anhedonia and worthlessness continuously aligned with Beck's cognitive triad of negative self, world, and future.
• Severe: Stable negative self-schema with prevailing hopelessness characterize entrenched depressive cognition that requires clinical intervention (Q17A, Q21A, Q34A).
• Extremely severe: When existential collapse markers prevail—pervasive sadness (Q13A), hopelessness (Q10A, Q37A, Q38A) and motivational disintegration (Q42A)—it suggests cognitive shutdown and emotional flattening seen in severe MDD requiring urgent intervention.

The progression pathway provides a data-driven validation of existing frameworks and demonstrates that the predictions of DepressNet have clinical meaning beyond statistical accuracy. Similar SHAP class patterns further revealing severity of the symptoms was reported in a study on depression detection from wearable data (Ahmed et al., 2025) and then from nutritional covariates (Liu et al., 2026). This shows that such interpretability can be applied to various types of data.

5.6 Limitation of the study

A key limitation of this study arises from the construction of the target variable. The depression severity labels used in this work are derived directly from the Depression subscale of the DASS-42 instrument based on established clinical cut-off scores. As a result, the target labels are inherently dependent on a subset of the same item-level responses that are also used as input features in the predictive model. While the model is trained on raw item-level responses rather than aggregated subscale scores, this design introduces a degree of structural dependency between input features and the target definition. Therefore, the proposed DepressNet framework should not be interpreted as a replacement for standardized psychometric scoring or clinical diagnosis. Instead, it serves as a computational modeling approach that learns patterns of association within psychometric responses and provides an interpretable machine learning perspective on how individual symptom items collectively align with severity categories defined by the DASS-42 instrument. The observed predictive performance reflects the model's ability to approximate these structured scoring rules and capture latent relationships among self-reported symptoms, rather than independent clinical ground truth. Future work should validate the proposed framework on clinically diagnosed datasets and externally labeled cohorts to assess generalizability beyond questionnaire-derived severity labels. Incorporating clinician-assessed diagnostic outcomes as independent ground-truth labels would further strengthen the clinical validity, interpretability, and translational relevance of the proposed framework. A second specific limitation concerns the Mild severity class, which achieves an F1-score of only 0.10 across all evaluated models, including the proposed DepressNet, despite SMOTETomek resampling and Focal Loss training. This persistent underperformance reflects the inherent ambiguity of Mild depressive presentations—characterised by symptom profiles that substantially overlap with both Normal and Moderate categories—and the narrow 4-point severity band of the Mild stratum (scaled scores 10-13). Clinically, Mild depression is the most important class for early detection and intervention; its poor classification performance therefore represents the most consequential limitation for practical deployment of the current framework. A third limitation is the absence of external validation on independently clinically diagnosed cohorts. The DASS-42 OpenPsychometrics dataset, while large (N = 39,775) and demographically diverse, consists entirely of voluntary online self-report responses without independent clinical diagnosis or clinician assessment. The model has not been evaluated on hospital-based patient cohorts, PHQ-9 clinical datasets, or DASS-21 community samples with independent severity ratings. Validation on such datasets is a necessary condition before DepressNet can be considered for real-world clinical deployment. Results from this study should therefore be interpreted as a reproducible computational framework for psychometric data analysis rather than a clinically validated screening or diagnostic tool.

6 Conclusion

This study presented DepressNet, a stacking ensemble framework for five-class depression severity stratification using item-level responses from the DASS-42 psychometric instrument. The proposed model integrates XGBoost, Random Forest, LightGBM, and Calibrated SVM as heterogeneous base learners with a PyTorch-based multilayer perceptron meta-learner, trained on out-of-fold probability predictions to mitigate information leakage and improve generalization reliability. To address severe class imbalance, the training pipeline incorporates SMOTETomek resampling together with class-weighted cross-entropy loss. Experimental evaluation on a largescale real-world dataset (N = 39,775) demonstrates that DepressNet achieves competitive and stable performance under an imbalanced multi-class setting, attaining 85.38% accuracy, 59.01% macro F1-score, 0.7158 MCC, and 0.9714 AUC-ROC on a held-out test set. Across all evaluation 387 metrics, the proposed model demonstrates consistent empirical improvement over six baseline classifiers under identical experimental conditions. Statistical validation using the Friedman test (p < 0.001)and subsequent post-hoc analyses further confirms that observed performance differences are statistically significant. Beyond predictive performance, this study emphasizes model interpretability through a dual SHAP-based explainability framework, enabling both global feature attribution and class-wise decomposition across severity levels. The resulting analysis reveals clinically meaningful symptom-response structures aligned with established depression symptomatology, while ablation studies confirm the complementary contribution of three tree base learner, with Random Forest contributing the most significant marginal impact within the ensemble. Despite these advantages, a key limitation is the poor predictive performance on the Mild severity class (F1 = 0.10), primarily due to strong symptom overlap with adjacent categories and persistent class imbalance even after resampling. A structural dependency exists between the input features and target construction because the severity labels are derived from DASS-42 scoring rules, which may influence predictive behavior and limit generalizability beyond questionnaire-derived severity labels. Therefore, external validation on clinically diagnosed and independently labeled datasets is necessary before considering real-world deployment. Future work will address these limitations through ordinal and hierarchical classification strategies, advanced imbalance-learning methods tailored to minority severity classes, and validation on independently assessed cohorts, including hospital-based populations and datasets with clinically established outcomes. Overall, DepressNet provides a reproducible and interpretable machine learning framework for psychometric data analysis, contributing to severity-aware modeling and explainable AI approaches in computational mental health research.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Andrej Košir, University of Ljubljana, Slovenia

Reviewed by: Chaimae Taoussi, Université Sultan Moulay Slimane, Morocco

Md. Mortuza Ahmmed, American International University Bangladesh, Bangladesh

Footnotes

DASS, depression anxiety stress scales; MDD, major depressive disorder; SHAP, SHapley Additive exPlanations; MCC, Matthews Correlation Coefficient; OOF, out-of-fold; MLP, multilayer perceptron; XAI, explainable artificial intelligence.