Decision-making logic for samples of clinically significant EDTA-dependent pseudothrombocytopenia with machine learning.

OBJECTIVES

To establish practical, clinically meaningful criteria for identifying significant platelet aggregation in suspected samples, defined by a platelet-count deviation exceeding the reference change value (RCV).

METHODS

Suspected aggregated samples were collected. Venous blood anticoagulated with citrate/EDTA underwent platelet counting via impedance method (PLT-I) and blood smear analysis at 5-, 30-, 60-, and 120-min intervals. Manual microscopy (30-min smears) assessed aggregation cluster size/number, while digital cytomorphology (PLT-pro) quantified clusters at all time points. Using 5-min citrate platelet counts as reference, relative deviations (d%) of EDTA PLT-I were calculated. Clinically significant aggregation-positive samples were defined as d% exceeding the negative RCV (30.02% for PLT < 100 × 10(Zhang et al., 20259)/L; 20.92% for PLT ≥ 100 × 10(Zhang et al., 20259)/L). Samples were randomly split into training (70%) and test (30%) sets to develop logistic regression and decision-tree models.

RESULTS

A total of 86 patients contributed 338 time-point results; 183 were classified as positive and 155 as negative. Manual microscopy showed that counting clusters with ≥5 platelets (vs. ≥ 3) improved specificity (p = 0.083). A criterion of ≥3 clusters, each containing ≥5 platelets, achieved 100% sensitivity and 33.3% specificity. PLT-I and PLT-pro were key discriminators via the Boruta algorithm. In the test set, the decision-tree model outperformed logistic regression (sensitivity: 92.42% vs. 89.39%; specificity: 94.29% vs. 94.29%; accuracy: 93.07% vs. 91.09%).

CONCLUSION

Criteria were established for manual microscopy (≥ 3 clusters with ≥5 platelets) and digital analysis (PLT-pro 27-97 with PLT-I < 58 × 10(Zhang et al., 20259)/L, or PLT-pro ≥97 with PLT-I < 164 × 10(Zhang et al., 20259)/L). However, these thresholds are merely exploratory results, and independent validation must be performed in this laboratory before they can be used for clinical decision-making.