Back to List View Graph View

Deep learning

Deep learning is a branch of machine learning in which multilayered neural networks learn hierarchical representations from data.

Rebuilt from PubMed 3 Sept 2026 · no new papers today

Where the papers sit

12 papers study deep learning directly. The themes below are drawn from those 12. 1 new direction follows.

  • Multimodal Clinical Prediction : Multimodal fusion combines imaging, body composition and clinical signals for cancer pain detection, concussion prediction, myocarditis severity assessment and gastric-cancer prognosis. Attention and explainability recur as the field moves toward clinically useful risk stratification. 4 papers · 33.3%

  • Medical Image Diagnosis : Automated image diagnosis covers meniscal tears, vitiligo, dental restorations and diabetic retinopathy. Clinical annotation, lesion-aware attention, adversarial augmentation, uncertainty estimates and interpretable outputs recur as models are pushed toward reliable deployment. 4 papers · 33.3%

  • Others — Broad Clinical AI Applications : AI is being applied to oral-care outcomes, antimicrobial resistance, ECG-based cardiovascular detection and HIV prognosis. The common aim is better clinical prediction and decision support, but the applications do not converge on a shared disease mechanism or endpoint. 4 papers · 33.3%

NEW DIRECTION

Deep learning is used to audit model stability rather than only to produce a clinical prediction

The cardiovascular-disease ECG-scan study treats deep learning models themselves as the object of investigation, rather than merely as tools for diagnosis. Whereas the other papers use deep learning to classify disease, estimate severity, or predict outcomes, this study deliberately perturbs ECG-printout images and measures how model predictions and explanations change, identifying vulnerabilities under discoloration, handwriting, and paper wrinkles. This gives deep learning a distinct role as an auditable system whose robustness and explanatory stability are clinical outcomes in their own right 42155349May.

Recent Findings on deep learning

These papers describe deep learning as a clinical support tool for prediction, detection, and treatment decisions, although only one directly addresses antimicrobial resistance. The AMR review links AI-based surveillance, antimicrobial susceptibility prediction, stewardship, drug discovery, and resistance-gene detection, while emphasizing data quality, bias, interoperability, privacy, and clinician adoption 41859321Mar. The oral-health bibliometric analysis reports increasing use of deep learning and frequently high Accuracy near or above dental professionals’ performance 42678726Sep. The ECG study shows that augmentation with image perturbations improves generalization, reaching an Area Under the Receiver Operating Characteristic Curve of 0.894 on contaminated data, while explanation stability remains limited 42155349May. The HIV systematic-review protocol shifts attention toward validating prognostic models for mortality, drug resistance, and other outcomes, but its quantitative findings remain unavailable 42612074Aug.

These studies use deep learning to extract clinically relevant signals from pathology images, synchronized sensor streams, speech, facial behavior, and text, rather than relying on a single clinical input. The myocarditis model combined multiple instance learning with a transformer to capture inflammatory infiltrates, myocyte injury, and architectural disruption, achieving an AUROC of 0.993 for severe disease 42585173Aug. The concussion and cancer-pain studies found that multimodal fusion can improve or balance prediction across modalities, although the cancer-pain study could not exclude subject-level overfitting and the concussion study requires larger clinical trials 42581081Aug42496586Jul. The gastric cancer paper extends this direction to body composition-based prognosis after neoadjuvant treatment, but its abstract reports only the model objective and no performance results 42068095May. Together, the studies favor attention-based fusion and explainability, while their controlled settings and incomplete validation limit conclusions about clinical deployment.

These validation studies combine clinically curated datasets, repeated or patient-wise cross-validation, explainability, and measures of uncertainty or generalization. The dental-restoration model improved small-target recall to 0.81, but precision varied substantially across categories and the dataset comparison was not directly controlled 42615416Aug. The diabetic-retinopathy framework combined lesion-aware attention, adversarial augmentation, Bayesian uncertainty, and semi-supervised learning, achieving 93.8% Accuracy on APTOS 2019, 91.2% on Messidor-2, and an Area Under the Curve of 0.963 for referable disease 42557287Aug. The vitiligo model reported Accuracy of 94.88%, F1 score of 94.86%, and an AUC of 0.9885 with patient-wise cross-validation and clinically meaningful explanations in 98.48% of reviewed images 42497363Jul. The meniscus study showed that repeated 10-run, 5-fold cross-validation produced lower but more reliable performance than a single run, supporting independent multicenter validation across imaging protocols 42483770Jul.