Abstract
Background: Neural networks (NNs) are increasingly applied for various healthcare practices. Their performance, however, depends on large datasets often unavailable in clinical contexts, necessitating innovative strategies for enabling NNs to function effectively with small or limited datasets. Thus, this study evaluated evidence on NN applications in healthcare with a focus on strategies for handling data scarcity.
Methods: PubMed, Scopus, Web of Science, Embase, and IEEE Xplore databases and Google Scholar were searched for related studies. Data were extracted on clinical domains, model architecture, dataset characteristics, limited data strategies, and performance metrics and synthesized into thematic categories.
Results: NNs were applied in radiology, pathology, cardiology, neurology, oncology, genomics, clinical natural language processing, decision support, and remote monitoring. Data scarcity was mitigated through data-centric and model-centric approaches and advanced paradigms. Imaging-focused applications heavily relied on augmentation and transfer learning, whereas text- and tabular-based tasks leveraged weak supervision and contrastive pretraining. Multimodal integration, combining imaging, genomics, and electronic health records (EHRs), emerged as a powerful strategy to overcome single-modality data constraints. Despite promising performance gains, external validation and generalizability remain limited across populations.
Conclusion: NNs demonstrate strong potential in healthcare, even when constrained by small datasets, provided that specialized strategies are employed. Transfer learning, data augmentation, generative modeling, multimodal integration, and meta-learning substantially enhance model performance and clinical applicability. However, challenges of reproducibility, bias, and scalability persist. Future research should prioritize federated learning, harmonized benchmarking, and explainability to ensure safe, equitable, and clinically trustworthy NN implementation in data-limited healthcare settings.