Partitioned Feature Selection Identifies Key Depression Predictors While Preserving Clinical Domains

Computer scientists and mental health researchers have introduced a domain-partitioned feature selection framework designed to extract informative depression indicators from questionnaire responses while preserving their distinct clinical groupings. In high-dimensional psychological assessment data, traditional machine learning workflows typically apply conventional feature selection methods that evaluate all survey items across a single, global feature space using an overarching predictive or statistical criterion. This universal ranking strategy disregards the recognized psychiatric reality that depressive symptoms do not exist as independent, homogeneous variables. Instead, psychological distress manifests across interrelated yet distinct clinical domains, such as cognitive disturbances, emotional despair, functional impairments, somatic complaints, and suicidal tendencies. By evaluating every questionnaire item within an undifferentiated pool, conventional feature selection risks discarding domain-specific nuances and obscuring the clinical logic behind algorithmic risk stratification. To resolve this structural limitation, the investigative team proposed partitioned feature selection, a multi-stage machine learning methodology that coordinates domain-based feature grouping, group-specific local target generation, domain-wise feature selection, and local-to-global importance integration.

The conceptual foundation of the partitioned feature selection architecture rests on the structure of established clinical assessment instruments, specifically the twenty-five-item Burns Depression Checklist. Depressive symptoms inherently cluster into recognizable dimensions that psychiatric theory and diagnostic guidelines, such as the Diagnostic and Statistical Manual of Mental Disorders, treat as coordinated constructs rather than disconnected observations. Cognitive frameworks emphasize negative automatic thoughts, self-reproach, pessimism, and hopelessness as a unified dimension, while behavioral formulations highlight social withdrawal, reduced activity, and amotivation. Similarly, vegetative symptoms such as sleep disruption, appetite alterations, and fatigue constitute a separate physical domain, while self-harm ideation and intent form an urgent clinical axis. Rather than treating each questionnaire item as an isolated statistical predictor, the framework organizes features according to two formal criteria: semantic consistency and clinical relevance. Semantic consistency mandates that items placed in the same partition represent closely related underlying psychological concepts based on survey phrasing and expert review. Clinical relevance ensures that every resulting partition directly corresponds to an established psychosocial construct documented in clinical literature, facilitating transparent interpretation for healthcare decision support.

Under these design principles, the twenty-five items of the Burns Depression Checklist were partitioned into four structured clinical domains. The thoughts and feelings group encompasses the first ten survey items, capturing core cognitive and emotional symptoms including depressed mood, feelings of worthlessness, demoralization, guilt, and impaired decision-making capacity. The activities and personal relationships group spans seven items that assess functional and interpersonal dimensions, such as loss of motivation, reduced engagement in work, social withdrawal, and loneliness. The physical symptoms group comprises five items evaluating somatic manifestations, including persistent fatigue, sleep disturbances, appetite changes, reduced sexual drive, and somatic health worries. The suicidal urges group contains three items dedicated to assessing suicidal thoughts, the desire to end life, and active self-harm planning. The researchers emphasize that these four symptom partitions are not intended to represent separate, definitive diagnostic categories or an exhaustive psychiatric taxonomy; rather, they serve to maintain a coherent semantic structure that grounds machine learning feature reduction in clinically intelligible domains.

Once the symptom variables are organized into their respective domains, the framework constructs group-specific local severity targets to evaluate features within their immediate clinical contexts. In typical survey-based machine learning pipelines, models are trained directly against an overall severity score calculated across all questionnaire items. In contrast, partitioned feature selection calculates a local score for each domain by summing the numerical weights of the features belonging exclusively to that group. Because the four domains contain varying quantities of items—ten, seven, five, and three questions, respectively—the global checklist scoring thresholds cannot be applied directly. The framework resolves this by proportionally scaling the original Burns severity cutoffs according to the maximum achievable score for each partition, mapping local sums into six standardized severity tiers ranging from no depression to extreme depression. This transformation yields four distinct local target variables, allowing local feature selectors to estimate the relative predictive utility of each individual symptom against the severity of its specific clinical domain rather than against a generalized global outcome.

To estimate local feature importance objectively, the framework implements locally uniform feature selection, which deploys four complementary feature selection algorithms independently across each clinical partition. Rather than relying on a single ranking criterion, locally uniform feature selection combines recursive feature elimination, SelectKBest with the chi-square statistic, univariate feature selection using the analysis of variance F-test, and mutual information feature selection. Recursive feature elimination operates as a wrapper method using a random forest classifier to account for complex nonlinear interactions among symptoms. The chi-square statistic measures statistical dependence by comparing observed and expected class frequencies across non-negative ordinal responses. The analysis of variance F-statistic quantifies linear class discriminative capability, while mutual information captures nonlinear and heterogeneous dependencies without assuming linear distributions. The researchers explicitly avoided stochastic meta-heuristic search algorithms, such as genetic algorithms or particle swarm optimization, because those techniques produce solution-dependent feature subsets and stochastic run-to-run variability rather than reproducible, deterministic importance scores for individual survey features.

Because the four local feature selection techniques generate importance metrics based on fundamentally divergent mathematical principles and numerical scales, raw scores cannot be combined directly without introducing systematic bias. The framework applies min-max normalization to map all importance scores onto a uniform interval between zero and one, preserving the internal relative ranking established by each method while rendering wrapper, filter, and information-theoretic metrics directly comparable. These normalized scores are averaged to yield a balanced consensus score for each feature within its group. To account for varying domain-level relationships with overall depression severity, the framework computes normalized group weights using the absolute Spearman rank correlation between each domain’s local target and the global severity target. In the primary experimental dataset, the activities and personal relationships group yielded the highest association weight at 0.2912, followed by thoughts and feelings at 0.2841, physical symptoms at 0.2574, and suicidal urges at 0.1673. In the second evaluation dataset, thoughts and feelings contributed the highest weight at 0.2746, followed closely by activities and personal relationships at 0.2681, suicidal urges at 0.2344, and physical symptoms at 0.2229. Multiplying consensus feature importance by these domain weights produces the integrated group feature importance metric.

Validation of the partitioned feature selection architecture was carried out using two independent Burns Depression Checklist datasets. The first dataset consisted of 1,226 responses gathered from undergraduate university students in Bangladesh, providing a structured cohort for initial methodological development and evaluation. The second cohort comprised a publicly accessible benchmark dataset containing 2,500 survey respondents, allowing investigators to assess the pipeline across distinct sample sizes and populations. Both datasets exhibited an imbalanced multiclass distribution across six severity tiers: moderate depression represented the largest proportion of responses, followed by mild depression, while normal but unhappy, no depression, and extreme depression contained comparatively few instances. To preserve these natural class proportions, both datasets were partitioned into stratified training, validation, and test splits containing 70%, 15%, and 15% of the observations, respectively. In the first dataset, two missing values identified in the health worry feature within the training partition were imputed using the rounded training mean of 1.0, ensuring that no validation or test information leaked into data preparation.

To establish the final predictor sets, the investigators applied a validation-guided retention procedure that ranked features in descending order of their integrated group feature importance scores and evaluated retention thresholds of 60%, 70%, 80%, 85%, and 90%. Sixteen distinct supervised classifiers—spanning tree-based models, instance-based algorithms, kernel-based estimators, naive Bayes formulations, gradient boosting variants, bagging architectures, and stacking ensembles—were trained on the training partition and scored against the validation partition. For the first dataset, both the 85% and 90% retention levels achieved an identical peak validation accuracy of 90.76%. Applying a predefined tie-breaking rule that selects the more parsimonious model, the 85% retention level was chosen, retaining 21 features and eliminating four lower-ranked variables: avoiding work, spending less time with family or friends, self-criticism, and possessing a specific plan for self-harm. For the second dataset, validation accuracy scaled progressively up to the 90% retention level, reaching a peak of 91.47% with 23 features, while excluding loss of pleasure in life and difficulty making decisions.

Model configuration was finalized by evaluating the sixteen classifiers under two distinct training conditions: baseline training without synthetic oversampling and training with the synthetic minority over-sampling technique, which interpolates synthetic observations exclusively within the minority classes of the training data. Across both datasets, an untuned support vector classifier operating without synthetic oversampling achieved the highest validation accuracy, reaching 90.76% in the first dataset and 91.47% in the second. Incorporating synthetic oversampling reduced peak validation accuracy to 85.87% and 86.67%, respectively. Subsequent hyperparameter optimization using five-fold cross-validated grid search failed to improve upon the baseline support vector machine configurations, leading researchers to retain the untuned models. When evaluated on the held-out test partitions, the untuned support vector classifier achieved a test accuracy of 95.65%, a macro-averaged F1 score of 95.43%, and a multiclass area under the receiver operating characteristic curve of 99.72% on the first dataset. On the second dataset, the model produced a test accuracy of 89.60%, a macro-averaged F1 score of 92.37%, and an area under the curve of 99.07%.

To assess whether domain-aware partitioning offered measurable predictive advantages over standard techniques, the authors conducted direct comparisons against conventional recursive feature elimination, an exploratory paired Wilcoxon signed-rank analysis, and a random regrouping ablation study. Compared to conventional recursive feature elimination evaluated on matched feature counts, partitioned feature selection improved peak validation accuracy by 1.09 percentage points in the first dataset and by 3.73 percentage points in the second dataset, while raising mean accuracy across all sixteen classifiers. In the exploratory Wilcoxon analysis across the sixteen paired classifiers, the performance advantage of partitioned feature selection without oversampling over conventional recursive feature elimination remained statistically significant after Holm correction in the first dataset, with a mean difference of 1.7323 percentage points, but did not reach significance after correction in the second dataset, where the mean difference was 0.8500 percentage points. Furthermore, when ten alternative partitions containing identical group sizes were constructed through random feature assignment, the clinically structured grouping exceeded the random mean accuracy in both datasets, though three random configurations in the first cohort achieved higher peak validation accuracy, indicating that the predictive benefit of clinical partitioning varies between populations.

Model interpretability was independently evaluated by comparing the integrated group feature importance rankings against post-hoc Shapley additive explanations computed on the validation partition, alongside execution time profiling. The Shapley attribution analysis demonstrated notable qualitative consistency with the integrated rankings: cognitive symptoms such as hopelessness and worthlessness, functional deficits such as loss of motivation, physical complaints including sleep and appetite disturbances, and suicidal thoughts consistently emerged as influential drivers across both analytical frameworks. Computational timing benchmarks demonstrated that partitioned feature selection completed its local selection phase in 10.059 seconds for the first dataset and 12.564 seconds for the second, representing speed improvements of 3.7-fold and 4.2-fold over global recursive feature elimination, which required 37.597 seconds and 52.858 seconds, respectively. The investigators explicitly emphasized that their models estimate questionnaire-derived depression severity tiers rather than independently confirmed clinical diagnoses, noting that identified predictors reflect statistical symptom associations rather than established causal risk factors.

Subject of Research: Domain-partitioned feature selection framework for questionnaire-based depression severity estimation

Article Title: Partitioned feature selection: locally uniform feature selection with structured feature grouping for explainable predictor identification

Article References: Shaha, T. R., Al Farid, F., Begum, M., Al Mamun, A., Nazneen, T., Uddin, J., & Abdul Karim, H. B. (2026). Partitioned feature selection: locally uniform feature selection with structured feature grouping for explainable predictor identification. Frontiers in Artificial Intelligence, 9, Article 1879422. https://doi.org/10.3389/frai.2026.1879422

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1879422

Keywords: Partitioned feature selection, Burns Depression Checklist, Depression severity estimation, Clinical symptom domains, Locally uniform feature selection, Integrated group feature importance, Support vector classifier, Recursive feature elimination, Explainable artificial intelligence, SHAP analysis, Multiclass severity classification, Random regrouping ablation, Supervised learning, Mutual information feature selection, Univariate feature selection, SelectKBest chi-square, Class imbalance, Mental health machine learning