Source-linked AI summary
SMOTE-VAR: An Uncertainty-Aware Oversampling Method for Predicting Depression Remission in University Students
Dang Nguyen, Arun Kumar A, Taylor A. Braund, Wu Yi Zheng, Debopriyo Bal, Leonard Hoon, Jill Newby, Helen Christensen, Svetha Venkatesh, Alexis Whitton, Sunil Gupta
TL;DR
Predicting depression remission in university students is hindered by class imbalance and SMOTE’s tendency to generate invalid minority samples. The paper introduces SMOTE-VAR, which uses Gaussian-process variance to filter uncertain synthetic samples, and reports improved remission prediction on an Australian student dataset. Its scope is constrained by computational challenges for high-dimensional or large minority-class datasets and by outcome assessment after only two weeks.
Problem
Class imbalance and SMOTE-generated false-positive minority samples complicate reliable machine-learning prediction of depression remission in university students.
Method
SMOTE-VAR uses a Gaussian-process variance score to estimate synthetic-sample uncertainty and filter samples above a preset threshold.
Results
SMOTE-VAR achieved the highest average balanced accuracy across five classifiers, improving by 1% over TVAE, 2% over SMOTE, and 7% over Imbalance.
Takeaways & Limitations
SMOTE-VAR provides a computational approach for reducing false positives while predicting treatment remission in a real-world cohort of Australian university students.
Takeaways & Limitations
Exact Gaussian-process inference scales at O(N^3), and remission was assessed after a 2-week intervention rather than the 12- to 16-week period when lifestyle benefits typically manifest.
Abstract
from arXiv · showhide
University students experience disproportionately high rates of common mental health conditions, such as depression, which can impair learning, social functioning, and overall well-being. Although lifestyle interventions such as mindfulness and physical activity can reduce the symptoms, many do not achieve symptomatic remission. Developing new approaches to identify students with poor outcomes could enable earlier and more targeted intervention. Machine learning (ML) methods have increasingly been used to predict remission in depressive patients. However, these ML models often suffer from class imbalance, where there may be an unequal proportion of people in the remitted group relative to the non-remitted group. This imbalance can reduce model accuracy and bias predictions. To address this, studies commonly employ the popular oversampling strategy SMOTE. However, SMOTE has a notable limitation: it may generate invalid synthetic minority samples. In a clinical context, these false positives can lead to incorrect risk stratification, potentially delaying necessary escalated care for patients unlikely to remit. In this paper, we introduce a novel and effective oversampling method that addresses this shortcoming. Our approach leverages the variance function of a Gaussian process to estimate the uncertainty of generated minority samples to reduce false positives. We validate our method on a depression dataset collected from university students and demonstrate that it is better than existing oversampling approaches in predicting remission (i.e., treatment outcome). By improving the reliable identification of non-responders, our method provides a robust computational tool to help clinicians rapidly pivot to adjunctive therapies, thereby personalizing and optimizing mental health care pathways.
1 Introduction
Depression remission prediction in university students is complicated by high non-remission rates and imbalanced clinical data. The paper introduces SMOTE-VAR, which uses Gaussian-process uncertainty to filter unreliable synthetic minority samples and evaluates it on Australian student data.
- Non-remission rates remain 40% to 52% despite lifestyle interventions, motivating early identification of students unlikely to achieve remission.
- SMOTE rebalances training data by linearly interpolating between real minority samples but can generate invalid samples when minority regions are non-convex.
- SMOTE-VAR assigns Gaussian-process variance scores to synthetic samples and filters those exceeding a preset threshold to reduce false positives.
- The evaluation uses 784 Australian university students with DASS-21 measures, treatment assignments, and mobility-pattern features, with 64% in remission and 36% in non-remission.
- The introduction identifies SMOTE-VAR as a proposed oversampling method and its application with standard ML classifiers as a real-world clinical contribution.
- SVM with SMOTE-VAR achieved a Balanced Accuracy (bACC) of 0.73 (±0.02), a 22% improvement over a baseline SVM without oversampling.
2 Related Works
Related work addresses remission prediction through non-oversampling and oversampling strategies. Existing approaches face limited performance on imbalanced clinical data, especially because deep learning methods suit larger datasets while SMOTE-based methods can generate false positives.
- Non-oversampling approaches train directly on imbalanced datasets and typically achieve a bACC of approximately 0.66–0.68.
- Oversampling methods address class imbalance by generating or rebalancing minority-class instances, with many established techniques extending SMOTE.
- Deep learning-based oversampling methods often underperform in mental health because clinical datasets commonly have small sample sizes.
- Traditional SMOTE-based methods remain inadequate at regulating false-positive generation, motivating a method specifically designed to address this weakness.
3 Framework
SMOTE-VAR augments standard SMOTE with Gaussian-process uncertainty scoring to filter unsupported synthetic minority samples while retaining broader interpolation coverage.
- Oversampling workflow: Oversampling learns from an imbalanced training set, generates synthetic minority samples, and evaluates classifiers trained on the rebalanced set using held-out-test bACC.The workflow constructs ˆDtrain from majority and synthetic minority samples before evaluation.
- Standard SMOTE: Standard SMOTE linearly interpolates between a minority sample and a randomly selected minority-class neighbor using a random λ ∈ (0, 1).The neighbor is selected from the k-nearest minority-class neighbors.
- SMOTE limitation: SMOTE can generate erroneous minority samples when interpolation crosses majority-sample regions because it implicitly assumes a convex minority-class region.Clinical data distributions may instead be non-convex, producing false-positive synthetic samples.
- SMOTE-VAR: Unlike distance-based filtering, SMOTE-VAR retains distant neighbors and selectively removes high-variance samples, avoiding localized over-dense synthetic clusters.Its GP variance incorporates spatial correlations and minority-class structure beyond pairwise distance.
- SMOTE-VAR: SMOTE-VAR assigns each synthetic sample a Gaussian-process variance score and rejects samples whose uncertainty exceeds a prespecified threshold.The GP estimates support from true minority observations rather than predicting class labels.
4 Experiments
This section describes the empirical evaluation of SMOTE-VAR, including the clinical dataset, preprocessing, experimental configurations, predictive-performance analysis, and ablation studies.
- The evaluation covers the clinical dataset, preprocessing pipeline, experimental configurations, predictive performance, and ablation studies.
4.1 Dataset and Feature Engineering
The study refines GPS and clinical data from university students into a balanced-feature prediction dataset, retaining 482 participants with substantial remission-class imbalance.
- The Vibe-Up dataset began with 1,282 participants, of whom 784 provided concurrent GPS mobility data.
- Strict GPS-quality filters and a minimum two-day mobility requirement reduced the analytical cohort to 482 students.
- The refined cohort contained 308 students in remission (64%) and 174 classified as non-remission (36%).
- Features combined DASS baseline and pre-treatment scores, treatment assignments, and GPS-derived daily locations visited and distance traveled.
- The illustrative dataset table pairs student DASS scores, treatment types, and outcomes with timestamped latitude-and-longitude locations.
- Linear imputation standardized variable-length mobility trajectories to a uniform 15-day observation window.
4.2 Experimental Settings
The experiments benchmarked SMOTE-VAR against ten oversampling or non-oversampling methods using five classifiers and repeated randomized validation.
- SMOTE-VAR was compared with Imbalance, interpolation methods, SMOTE variants, CTGAN, TVAE, and ImbLLM.
- The variance threshold for SMOTE-VAR was set to ν = 0.001.
- Five classifiers—kNN, SVM, DT, RF, and XGB—were trained on rebalanced datasets.
- Hyperparameters were tuned with five-fold cross-validation, while performance used randomized 90/10 train-test splits repeated across ten seeds.
- Balanced Accuracy was the primary reported metric, supplemented by F1-score, AUC, sensitivity, specificity, and ROC curves.
4.3 Results and Discussions
SMOTE-VAR achieved the strongest average balanced accuracy across evaluated classifiers and outperformed the non-oversampling baseline, although differences from the strongest oversampling baselines were not statistically significant.
- SMOTE-VAR achieved the highest average balanced accuracy, ranking first with four classifiers and second with one.
- Its average improvement was 1% over TVAE, 2% over SMOTE, and 7% over Imbalance.
- Paired Wilcoxon tests found no statistically significant differences from TVAE or SMOTE (p > 0.05), likely reflecting 1–2% gaps and limited repeated splits.
- Traditional interpolation methods outperformed CTGAN, while TVAE and ImbLLM were the closest competitors.
- Among SMOTE variants, SMOTE-VAR maintained the highest overall predictive performance and SMOTE-NC ranked second.
4.4 Ablation studies
Ablation studies show that SMOTE-VAR is robust to imputation-window choices but sensitive to its variance threshold, while clinical symptom features drive performance more than GPS mobility data. Multiple metrics and ROC analyses further assess predictive stability.
- Temporal robustness: bACC remained above 0.72 when the imputed observation window varied from 2 to 30 days.The study reports negligible fluctuations in predictive accuracy across these windows.
- Imputation methods: Linear and s-linear imputation achieved the highest efficacy among the evaluated imputation methods.Other imputation methods also performed well.
- Variance threshold: Optimal bACC occurred at variance thresholds ν ∈[0.001, 0.01], whereas thresholds above 0.05 or below 0.0005 reduced performance.Higher thresholds admitted false positives, while overly restrictive thresholds discarded valid synthetic samples and reduced training volume.
- Feature importance: Removing DASS pre-treatment and baseline features reduced bACC by approximately 5% to 6%, compared with a 2% decline after removing GPS mobility features.The results identify direct clinical symptom measures as primary drivers and GPS mobility as supplementary predictors.
- Feature importance: DASS pre-treatment had the highest individual feature-importance score, followed by DASS baseline.This individual-feature analysis agreed with the feature-removal results.
- Other performance metrics: Across 10 independent test runs, the SVM with SMOTE-VAR achieved mean AUC 0.74, mean F1-score 0.66, mean Specificity 0.82, and mean Sensitivity 0.64.The ROC curves clustered tightly around the mean curve, with AUC 0.74±0.04.
5 Conclusion
The study combines clinical survey and GPS location data with a probabilistic oversampling method for remission prediction in depressive university students. On an Australian student dataset, SMOTE-VAR consistently achieved the highest average predictive performance among evaluated oversampling methods.
- Conclusion: SMOTE-VAR assigns uncertainty scores to synthetic minority samples and filters high-variance samples to reduce false positives.The method adds a probabilistic component to standard SMOTE for imbalanced remission-prediction datasets.
- Conclusion: Clinical survey data and GPS location patterns effectively predicted treatment remission in depressive students.The evaluation used a real-world depression dataset collected from Australian students.
- Conclusion: SMOTE-VAR consistently achieved the highest average predictive performance among the evaluated oversampling methods.The conclusion reports this result for the Australian student depression dataset.
6 Limitations and Future Directions
The study’s main limitations concern computational scaling to complex datasets, short-term assessment of remission, and limited predictive contribution from GPS mobility data. Future work targets more scalable Gaussian-process methods, longer follow-up, and behavior-aware imputation.
- Methodological limitations: Scaling SMOTE-VAR to highly complex, high-dimensional datasets presents computational challenges.Exact Gaussian-process inference scales at O(N^3), while high dimensionality can reduce covariance-function discriminatory power.
- Clinical study design: Remission was assessed immediately after a 2-week digital intervention, which may capture early rather than long-term remission.The paper notes that lifestyle-intervention benefits typically manifest over 12 to 16 weeks and recommends extended longitudinal follow-up.
- Data processing: Removing GPS mobility features reduced bACC by only 2%, and linear imputation may have smoothed clinically relevant nonlinear behavioral spikes.Future digital-phenotyping studies should explore nonlinear or behaviorally aware imputation methods.