JMIR Publications

ABSTRACT

Background:

Background:

Intraoperative bleeding (IOB) critically affects surgical safety and patient outcomes. Machine learning (ML) offers predictive potential; however, evidence regarding model characteristics and clinical translation remains fragmented.

Objective:

Objective:

 This scoping review aims to synthesise the methodologies, performance, and challenges of clinical translation for ML-based IOB prediction models to guide future research and implementation in clinical settings.

Methods:

Methods:

Following the PRISMA-ScR guidelines, we systematically searched PubMed, Web of Science, Embase, CNKI, and Wanfang from inception to April 2025. Two reviewers independently screened studies and extracted data using the CHARMS checklist, including only original studies that evaluated IOB prediction in surgical patients.

Results:

Results:

Out of 2,651 screened records, 23 studies met the inclusion criteria. Key findings include that Tree-Based methods (RF/XGBoost) were predominant, accounting for 69.6% of the models, with AUC values ranging from 0.809 to 0.933 in orthopaedics and oncology. In contrast, deep learning models excelled in obstetrics, particularly with multimodal data, achieving an AUC of up to 0.920. Sample sizes varied from 48 to 48,543, with electronic health records (EHR) and electronic medical records (EMR) serving as primary data sources; notably, 47.8% of the studies integrated imaging data. Validation gaps were evident, as only 30.4% of the studies implemented external validation, and 52.2% utilised cross-validation.

Conclusions:

Conclusions:

In conclusion, while ML enhances the accuracy of IOB prediction, barriers to clinical adoption persist due to data heterogeneity and insufficient validation. Future research should prioritise the development of real-time prediction systems and prospective multicenter validation. Clinical Trial: no