BioMedInformatics

Background: Early and accurate detection of depression from multimodal data is a critical yet challenging task. Many existing models either rely on a single modality or exhibit limited robustness across datasets, while publicly available datasets are often small and highly imbalanced. Methods: In this study, we propose a robust multimodal fusion framework that leverages bidirectional cross-modal attention to effectively integrate audio and text features, allowing the model to learn complementary information from both modalities. To address class imbalance and limited data, we employ SMOTE-based oversampling. Our model further incorporates validation-based adaptive thresholding and probability ensembling to enhance decision robustness and generalization. Experiments are conducted on two benchmark datasets, EATD-Corpus (Chinese; 162 participants) and DAIC-WOZ (English; 189 usable interviews), and ablation experiments compare cross-modal attention settings. Results: On EATD-Corpus, our fusion model achieves strong performance (F1: 0.820, Recall: 0.934, Precision: 0.762), achieving higher mean performance than the text-only and audio-only baselines evaluated under the same participant-level five-fold protocol. On DAIC-WOZ, the framework remains competitive (F1: 0.836, Recall: 0.925, Precision: 0.775). The ablation results confirm that bidirectional cross-modal interaction yields the most reliable trade-off between Precision and Recall. Conclusions: This paper proposes a deep learning framework for depression screening that jointly leverages audio and text modalities, demonstrating robust performance on small, class-imbalanced datasets and architectural generalizability across two corpora with different languages and interview settings.

Cardiovascular risk assessment remains a central challenge in both population-based prevention and high-risk clinical settings. We benchmarked a previously developed mechanistic model of cardiovascular ageing against machine-learning baselines across two distinct cohorts and evaluated both discrimination and probability calibration. The mechanistic model was based on an ordinary differential equation (ODE) framework and adapted to the observable interaction structure of two independent public datasets: the Framingham Heart Study cardiovascular risk dataset, with 10-year coronary heart disease as the outcome, and the Heart Failure Clinical Records dataset, with mortality as the outcome. ElasticNet logistic regression and XGBoost were evaluated as cohort-specific machine-learning benchmarks. Predictive performance was assessed using repeated stratified five-fold cross-validation with three repeats. In the Framingham cohort (n = 4240; 644 events), ElasticNet achieved ROC-AUC = 0.727 (95% CI 0.705–0.747), PR-AUC = 0.343, and Brier score = 0.116, while XGBoost achieved ROC-AUC = 0.715 (95% CI 0.694–0.737), PR-AUC = 0.325, and Brier score = 0.118. The observable mechanistic score showed weaker discrimination, with ROC-AUC = 0.547 and PR-AUC = 0.202. In the Heart Failure cohort (n = 299; 96 deaths), ElasticNet and XGBoost achieved ROC-AUC values of 0.770 and 0.771, respectively, whereas the mechanistic score achieved ROC-AUC = 0.534. Post hoc calibration improved probability scaling of the mechanistic score but did not restore discriminative performance. Overall, cohort-specific machine-learning models demonstrated stronger discrimination, while the external applicability of the mechanistic framework depended on the alignment of available predictors and clinical endpoints in the target cohort.

Rapid and accurate diagnosis is crucial for the early detection of malaria caused by Plasmodium parasites. This study primarily focuses on identifying a suitable deep learning model for classifying malaria parasites. In an ablation process, we started with single-branch Convolutional Neural Network (SB-CNN) and dual-branch CNN (DB-CNN) architectures enhanced with attention mechanisms to improve feature representation. Specifically, we incorporate the Convolutional Block Attention Module (CBAM) to focus on both channel-wise and spatially important features. We also use the Efficient Channel Attention (ECA) module to capture local inter-channel relationships, while the Squeeze-and-Excitation (SE) block emphasizes globally significant feature maps, further improving the model’s ability to distinguish between classes. To identify the most effective model, we experimented with EfficientNet-B0 and EfficientNet-B3 as backbone networks and conducted an ablation study by integrating various attention modules, resulting in three DB-CNN variants. We applied t-Distributed Stochastic Neighbor Embedding (t-SNE) to visualize the high-dimensional feature space between infected and uninfected samples. Utilizing the 5-fold cross-validation method on the Thick dataset, the best-performing models included architectures such as SB EfficientNet-B3, SB EfficientNet-B3 combined with CBAM, and DB EfficientNet-B3 across both branches, with CBAM in the first and SE in the second.