Respiratory diseases represent a persistent global health challenge, underscoring the need for intelligent, accurate, and personalized diagnostic and therapeutic systems. Existing methods frequently suffer from limitations in diagnostic precision, lack of individualized treatment, and constrained adaptability to complex clinical scenarios. To address these challenges, our study introduces a modular AI-powered framework that integrates an audio-based disease classification model with simulated molecular biomarker profiles to evaluate the feasibility of future multimodal diagnostic extensions, alongside a synthetic-data-driven prescription recommendation engine. The disease classification model analyzes respiratory sound recordings and accurately distinguishes among eight clinical classes: bronchiectasis, pneumonia, upper respiratory tract infection (URTI), lower respiratory tract infection (LRTI), asthma, chronic obstructive pulmonary disease (COPD), bronchiolitis, and healthy respiratory state. The proposed model achieved a classification accuracy of 99.99% on a holdout test set, including 94.2% accuracy on pediatric samples. In parallel, the prescription module provides individualized treatment recommendations comprising drug, dosage, and frequency trained on a carefully constructed synthetic dataset designed to emulate real-world prescribing logic.The model achieved over 99% accuracy in medication prediction tasks, outperforming baseline models such as those discussed in research. Minimal misclassification in the confusion matrix and strong clinician agreement on 200 prescriptions (Cohen’s $\kappa$ = 0.91 [0.87–0.94] for drug selection, 0.78 [0.74–0.81] for dosage, 0.96 [0.93–0.98] for frequency) further affirm the system’s reliability. Adjusted clinician disagreement rates were 2.7% (drug), 6.4% (dosage), and 1.5% (frequency). SHAP analysis identified age and smoking as key predictors, enhancing model explainability. Dosage accuracy was 91.3%, and most disagreements occurred in renal-impaired and pediatric cases. However, our study is presented strictly as a proof-of-concept. The use of synthetic data and the absence of access to real patient records constitute key limitations. A trialed clinical deployment was conducted under a controlled environment with a positive rate of satisfaction from experts and users, but the proposed system must undergo extensive validation with de-identified electronic medical records (EMRs) and regulatory scrutiny before it can be considered for practical application. Nonetheless, the findings offer a promising foundation for the future development of clinically viable AI-assisted respiratory care tools.

Detecting Traumatic Brain Injuries (TBI) through imaging remains challenging due to limited sensitivity in current methods. This study addresses the gap by proposing a novel approach integrating deep-learning algorithms and advanced image-fusion techniques to enhance detection accuracy. The method combines contextual and visual models to effectively assess injury status. Using a dataset of repeat mild TBI (mTBI) cases, we compared various image-fusion algorithms: PCA (89.5%), SWT (89.69%), DCT (89.08%), HIS (83.3%), and averaging (80.99%). Our proposed hybrid model achieved a significantly higher accuracy of 98.78%, demonstrating superior performance. Metrics including Dice coefficient (98%), sensitivity (97%), and specificity (98%) verified that the strategy is efficient in improving image quality and feature extraction. Additional validations with “entropy”, “average pixel intensity”, “standard deviation”, “correlation coefficient”, and “edge similarity measure” confirmed the robustness of the fused images. The hybrid CNN-ViT model, integrating curvelet transform features, was trained and validated on a comprehensive dataset of 24 types of brain injuries. The overall accuracy was 99.8%, with precision, recall, and F1-score of 99.8%. The “average PSNR” was 39.0 dB, “SSIM” was 0.99, and MI was 1.0. Cross-validation across five folds proved the model’s “dependability” and “generalizability”. In conclusion, this study introduces a promising method for TBI detection, leveraging advanced image-fusion and deep-learning techniques, significantly enhancing medical imaging and diagnostic capabilities for brain injuries.

Cortisol and testosterone are key digital biomarkers reflecting neuroendocrine activity across the hypothalamic–pituitary–adrenal (HPA) and hypothalamic–pituitary–gonadal (HPG) axes, encoding stress adaptation and behavioral regulation. Continuous real-world monitoring remains challenging due to the sparsity of sensing and the complexity of multimodal data. This study introduces a synthetic sensor-driven computational framework that models hormone variability through data-driven simulation and predictive learning, eliminating the need for continuous biosensor input. A hybrid deep ensemble integrates biological, behavioral, and contextual data using bidirectional multitask learning with one-dimensional convolutional neural network (1D-CNN) and long short-term memory (LSTM) branches, meta-gated expert fusion, Bayesian variational layers with Monte Carlo Dropout, and adversarial debiasing. Synthetically derived longitudinal hormone profiles that were validated by Kolmogorov–Smirnov (KS), Wasserstein, maximum mean discrepancy (MMD), and dynamic time warping (DTW) metrics account for class imbalance and temporal sparsity. Our framework achieved up to 99.99% macro F1-score on augmented samples and more than 97% for unseen data with ECE below 0.001. Selective prediction further maximized the convergence of predictions for low-confidence cases, achieving 99.9992–99.9998% accuracy on 99.5% of samples, which were smaller than 5 MB in size so that they can be employed in real time when mounted on wearable devices. Explainability investigations revealed the most important features on both the physiological and behavioral levels, demonstrating framework capabilities for adaptive clinical or organizational stress monitoring.

Despite the numerous advancements that have been made in the treatment and management of breast cancer, it continues to be a source of mortality in millions of female patients across the world each year; thus, there is a need for proper and reliable diagnostic assistance tools that are quite effective in the prediction of the disease in its early stages. In our research, in addition to the proposed framework, a comprehensive comparative assessment of traditional machine learning, deep learning, and transformer-based models has been performed to predict breast cancer in a multi-dataset environment. For the purpose of improving diversity and reducing any possible biases in the datasets, our research combined three datasets: breast cancer biopsy morphological (WDBC), biochemical and metabolic properties (Coimbra), and cytological attributes (WBCO), intended to expose the model to heterogeneous feature domains and evaluate robustness under distributional variation. Based on the thorough process conducted in our research involving traditional machine learning models, deep learning models, and transformers, a proposed hybrid architecture referred to as the FT-Transformer-Attention-LSTM-SVM framework has been designed and developed in our research that is compatible and well-suited for the processing and analysis of the given tabular biomedical datasets. The proposed design in the research has an effective performance of 99.90% accuracy in the primary test environment, an average mean accuracy of 99.56% in the 10-fold cross-validation process, and an accuracy of 98.50% in the WBCO test environment, with a considerable margin of significance less than 0.0001 in the paired two-sample t-test comparison process. In our research, we have performed the importance assessment in conjunction with the SHAP and LIME techniques and have demonstrated that its decisions are based upon important attributes such as the values of the attributes of radius, concavity, perimeter, compactness, and texture. Additionally, the research has conducted the ablation test and has proved the importance of the designed FT-Transformer.

Accurately identifying drug off-targets is essential for reducing toxicity and improving the success rate of pharmaceutical discovery pipelines. However, current deep learning approaches often struggle to fuse chemical structure, protein biology, and multi-target context. Here, we introduce HDPC-LGT (Hybrid Dual-Prompt Cross-Attention Ligand–Protein Graph Transformer), a framework designed to predict ligand binding across sixteen human translation-related proteins clinically associated with antibiotic toxicity. HDPC-LGT combines graph-based chemical reasoning with protein language model embeddings and structural priors to capture biologically meaningful ligand–protein interactions. The model was trained on 216,482 experimentally validated ligand–protein pairs from the Chemical Database of Bioactive Molecules (ChEMBL) and the Protein–Ligand Binding Database (BindingDB) and evaluated using scaffold-level, protein-level, and combined holdout strategies. HDPC-LGT achieves a macro receiver operating characteristic–area under the curve (macro ROC–AUC) of 0.996 and a micro F1-score (micro F1) of 0.989, outperforming Deep Drug–Target Affinity Model (DeepDTA), Graph-based Drug–Target Affinity Model (GraphDTA), Molecule–Protein Interaction Transformer (MolTrans), Cross-Attention Transformer for Drug–Target Interaction (CAT–DTI), and Heterogeneous Graph Transformer for Drug–Target Affinity (HGT–DTA) by 3–7%. External validation using the Papyrus universal bioactivity resource (Papyrus), the Protein Data Bank binding subset (PDBbind), and the benchmark Yamanishi dataset confirms strong generalisation to unseen chemotypes and proteins. HDPC-LGT also provides biologically interpretable outputs: cross-attention maps, Integrated Gradients (IG), and Gradient-weighted Class Activation Mapping (Grad-CAM) highlight catalytic residues in aminoacyl-tRNA synthetases (aaRSs), ribosomal tunnel regions, and pharmacophoric interaction patterns, aligning with known biochemical mechanisms. By integrating multimodal biochemical information with deep learning, HDPC-LGT offers a practical tool for off-target toxicity prediction, structure-based lead optimisation, and polypharmacology research, with potential applications in antibiotic development, safety profiling, and rational compound redesign.

Automated electrocardiogram (ECG) arrhythmia classification remains challenging due to signal noise, inter-patient variability, and limited annotated data, which constrain the generalization of supervised learning approaches. This study presents a self-supervised ECG representation learning framework that combines contrastive pretraining with ensemble-based supervised classification. A signal-to-noise ratio criterion is applied during self-supervised pretraining to stabilize contrastive optimization, while all extracted ECG beats, including noisy segments, are retained during downstream evaluation. The learned representations are classified using a hybrid ensemble composed of convolutional encoders and tree-based models. Model evaluation follows strict patient-level partitioning with stratified 10-fold cross-validation and bootstrap-based uncertainty estimation on a held-out test set. Under this evaluation protocol, the framework achieved high beat-level performance on curated datasets (internal and external). Class-wise performance shows precision and recall values between 0.99 and 0.999 across normal, supraventricular, ventricular, fusion, and paced beat categories. External validation is conducted on independent ECG cohorts, including PTB-XL, Chapman–Shaoxing, and INCART 12-lead datasets. On these datasets, the hybrid model attains macro-F₁ scores ranging from 0.91 to 0.94, compared with standalone convolutional and handcrafted feature-based Random Forest classifiers evaluated under identical conditions. These results characterize the behavior of the proposed representation learning framework across heterogeneous patient populations and recording configurations.