Background - Whether the metabolomic profile of stored platelets can predict storage duration across donors under donor-wise validation is unknown. We assessed a multi-task learning framework.
Materials and methods - Utilizing the public metabolomic dataset ST003937 (239 samples, 24 donors, 111 metabolites), a dual-task machine learning framework was constructed to predict storage duration (Task A) and biotinylation status (Task B). Models were evaluated by donor-wise 5-fold cross-validation with random-split comparison. Inter-donor distances were quantified and SHAP (SHapley Additive exPlanations) identified key metabolites. Multi-task learning was also performed.
Results - Task A failed to generalize, with a mean area under the curve (AUC) of 0.46, improving to 0.59 under random split, whereas Task B was robust (AUC 0.77). Median inter-donor distance (6.35) far exceeded the storage group distance (1.73). Multi-task learning marginally improved Task A by 0.03. SHAP and Gini importance identified metabolites related to energy metabolism and oxidative stress as key features for Task B.
Discussion - Inter-donor metabolic variability exceeded storage-induced changes in magnitude, explaining why cross-donor prediction of storage duration was unsuccessful despite documented group-level metabolic differences. The robust AUC gap between validation strategies confirms a "donor fingerprint" effect. Donor-specific baselines may be useful for future assessment of platelet quality. Multi-task learning provided marginal improvement. Task B's robust performance validates the analytical framework.
pdf