Convolutional neural network for red blood cell transfusion prediction: A machine learning based study

Subjective, inconsistent clinician judgement on red blood cell (RBC) transfusion demand causes unnecessary blood wastage, increased perioperative safety risks, and poorly balanced hospital blood inventory. To tackle these persistent clinical limitations, we developed a workflow-adapted clinical decision support tool: a one-dimensional convolutional neural network (1D CNN) trained on routinely acquired haematological and clinical variables to stratify inpatient RBC transfusion volume categories.

Using a retrospective cohort of 7,250 patients from Beijing Friendship Hospital (April 2024-June 2025), we trained a 1D CNN on 19 standardised pre-transfusion variables to predict transfusion volume categories (1, 2, 3, 4, ≥5 units). The model was optimised with Adam and evaluated on a 30% hold-out test set. Comprehensive characterisation of the cohort's baseline clinical features was also performed to clearly define the transfusion contexts and clinical indications of the study population.

On the 30% hold-out test set, the 1D CNN reached an overall accuracy of 0.7314 and converged without overfitting. The model accurately identified the predominant 2 units transfusion class (68.03% of cases, precision 0.7314). However, performance was limited in minority classes (≥5 units and 3 units), which showed low prevalence (< 5%) and overlapping haemoglobin values (60-80 g/L), indicating a critical safety limitation.

The 1D CNN reliably forecasts routine transfusion requirements in a single-center setting but under-detects high-volume transfusion events. External validation across diverse clinical centers, targeted data enrichment for minority classes, and development of indication-specific sub-models are mandatory to improve the model's generalizability and clinical utility.