Journal: iScience

Article Title: Ensemble transformer with post-hoc explanations for depression emotion and severity detection

doi: 10.1016/j.isci.2025.114605

Figure Lengend Snippet: Class distribution of DepressionEmo and MDSD dataset

Article Snippet: DepressionEmo dataset consists of 6,037 Reddit posts that represent a wide range of emotions associated with depression.

Techniques:

Journal: iScience

Article Title: Ensemble transformer with post-hoc explanations for depression emotion and severity detection

doi: 10.1016/j.isci.2025.114605

Figure Lengend Snippet: Pearson correlation between emotion labels in the multi-label DepressionEmo dataset Warmer colors indicate higher co-occurrence across posts (e.g., worthlessness-hopelessness and loneliness-emptiness co-occur frequently), while Anger shows weaker correlations with inward-facing emotions.

Article Snippet: DepressionEmo dataset consists of 6,037 Reddit posts that represent a wide range of emotions associated with depression.

Techniques:

Journal: iScience

Article Title: Ensemble transformer with post-hoc explanations for depression emotion and severity detection

doi: 10.1016/j.isci.2025.114605

Figure Lengend Snippet: Wordcloud representation of the DepressionEmo dataset after performing preprocessing steps

Article Snippet: DepressionEmo dataset consists of 6,037 Reddit posts that represent a wide range of emotions associated with depression.

Techniques:

Journal: iScience

Article Title: Ensemble transformer with post-hoc explanations for depression emotion and severity detection

doi: 10.1016/j.isci.2025.114605

Figure Lengend Snippet: Micro- and macro-averaged precision, recall, and F1-score comparison across all models on DepressionEmo and DSL (A, C, and E) show Precision, Recall, and F1 scores for DepressionEmo, while panels (B, D, and F) display the same metrics for DSL. The bars represent mean test performance (blue for micro, orange for macro), and the models are ranked by macro-F1 for each dataset. The proposed DepTformer-XAI-SV is marked with hatched bars.

Article Snippet: DepressionEmo dataset consists of 6,037 Reddit posts that represent a wide range of emotions associated with depression.

Techniques: Comparison

Journal: iScience

Article Title: Ensemble transformer with post-hoc explanations for depression emotion and severity detection

doi: 10.1016/j.isci.2025.114605

Figure Lengend Snippet: Comparison of model performance across datasets using macro-averaged F1 scores (A) and (B) present the generalization trend of classifiers trained on the DepressionEmo and evaluated on the DSL dataset. Each line corresponds to a model family—transformers (blue), deep learning (green), and classical machine learning (gray)—with the proposed DepTformer-XAI-SV model (orange) highlighted. Upward slopes indicate improved generalization to DSL, whereas flatter or downward trends reflect limited transferability across datasets.

Article Snippet: DepressionEmo dataset consists of 6,037 Reddit posts that represent a wide range of emotions associated with depression.

Techniques: Comparison

Journal: iScience

Article Title: Ensemble transformer with post-hoc explanations for depression emotion and severity detection

doi: 10.1016/j.isci.2025.114605

Figure Lengend Snippet: Ablation and imbalance influence landscape across backbone ensembles and data-level interventions (A) Ablation robustness map: absolute change ( Δ ) in macro-F1 versus the full fusion, showing how each training constraint impacts both emotion (circle markers) and severity (triangle markers) encoders. Bars indicate absolute Δ macro-F1; marker color encodes minority-macro-F1, and whiskers denote 95% confidence intervals. Fusion- and threshold-level ablations (orange region) cause larger stability loss than data-level ones (blue region). (B) Backbone influence landscape (LOBO analysis): absolute macro-F1 loss when each backbone is removed from the ensemble. Blue (DepressionEmo) and orange (DSL) bars reflect distinct minority sensitivities. The right inset shows the trade-off correlation ( Δ macro-F1 vs. minority-F1), where stronger ensembles retain minority balance. (C) Imbalance sensitivity landscape: macro-averaged metrics under progressive imbalance corrections. Solid lines (DepressionEmo) and dashed lines (DSL) show F1, precision, and recall trends under class-weighting and oversampling caps. The shaded region marks DSL non-applicability. The inset (bottom-right) traces precision-recall trade-offs across oversampling ratios, revealing recall-driven F1 gains beyond r = 0.25 .

Article Snippet: DepressionEmo dataset consists of 6,037 Reddit posts that represent a wide range of emotions associated with depression.

Techniques: Marker

Journal: iScience

Article Title: Ensemble transformer with post-hoc explanations for depression emotion and severity detection

doi: 10.1016/j.isci.2025.114605

Figure Lengend Snippet: Learning curves of the proposed model on DepressionEmo and DSL (A) DepressionEmo and (B) DSL report epoch-wise trends for training and validation loss, accuracy, recall, and precision, illustrating stable convergence and consistent generalization across datasets.

Article Snippet: DepressionEmo dataset consists of 6,037 Reddit posts that represent a wide range of emotions associated with depression.

Techniques: Biomarker Discovery