Surg Innov. 2025 Nov 19:15533506251400113. doi: 10.1177/15533506251400113. Online ahead of print.
ABSTRACT
BackgroundAccurate intraoperative blood loss estimation is crucial, particularly for pediatric and adolescent patients. Traditional methods (visual, gravimetric, formula-based) are often imprecise and unreliable, highlighting a need for objective and efficient approaches.MethodsThis study proposes a novel machine learning (ML) approach to predict blood loss from surgical sponges via a pixel-counting rate system. We utilized image preprocessing, including edge detection, and evaluated various ML algorithms with hyperparameter tuning. Details include deployment on commodity smartphones, a deterministic preprocessing pipeline, and app-level measures to mitigate folding, lighting, and motion blur. Model accuracy was assessed using RMSE and R2. A lightweight CNN baseline on raw images was also evaluated, positioning our system as a low-cost, mobile alternative to proprietary solutions.ResultsThe ML-based method significantly reduced reliance on subjective visual estimation. Gradient Boosting (GB) and Artificial Neural Networks (ANN) achieved a minimum RMSE of 0.91 and a maximum R2 of 0.93, outperforming traditional methods. The system demonstrates strong potential for integration into mobile phone applications for practical clinical use.ConclusionsOur machine learning-based pixel-counting approach offers a more accurate and efficient alternative for intraoperative blood loss estimation. Future research will expand the dataset, refine ML algorithms, and adapt the system for specialized equipment. While a CNN baseline approached GB performance, GB remained superior at the current data scale. A prospective paired-data plan is outlined for future benchmarking. Clinically, the per-sponge RMSE (∼0.9 cc) aligns with an acceptable error margin (0.5-1.0 cc) for decision support, noting that pediatric transfusion decisions rely on cumulative loss and multimodal monitoring.
PMID:41257324 | DOI:10.1177/15533506251400113
