Establishment and validation of intelligent review rules for blood cell analysis based on the gradient boosting...

To develop artificial intelligence (AI)-based intelligent review rules to accurately screen samples that require retests or blood smear microscopy examinations, thereby reducing the review rate, improving the work efficiency of haematology laboratories, and ensuring the quality of whole blood analysis.

A total of 10,212 EDTA-K2 venous blood samples from the clinical laboratories of 4 hospitals from May 2022 to April 2025 were collected. Among them, 9000 samples from the First People's Hospital of Foshan, the Zhongda Hospital affiliated to Southeast University, and the First Affiliated Hospital of Sun Yat-sen University were used to establish the rules, and 1212 samples from the Zhongshan Hospital affiliated to Fudan University were used for rule validation. The obtained rules were compared with the 41 review rules set by the International Consensus Group for Haematology Review (ICGHR). All samples were tested using a haematology analyser under the customized dilution ratio (CDR) mode and smear microscopy, if the microscopic examination results trigger the 10 microscopic anomalous rules established by the ICGHR, then the sample is deemed to be an anomalous sample. Intelligent review rules were established using the gradient boosting decision tree (GBDT) algorithm. Moreover, the false negative rate, false positive rate, and review rate of the validation set were statistically analysed.

(1) Among the 9000 samples in the establishment set, 4493 anomalous samples were identified by microscopic examination, and 26 intelligent review rules were established using the GBDT algorithm. (2) For the external validation set, 180 anomalous samples were identified by microscopic examination. Compared with those of the 41 international review rules, the false positive rate (14.60%) and review rate (27.55%) of the intelligent review rules were significantly lower (false positive rate: 30.19%, review rate: 44.14%). The false negative rate of the intelligent review rules (1.89%) was comparable to that of the 41 international review rules (0.90%) and met the ICGHR requirement of a false negative rate < 5%, with no critical hematologic cells missed.

Intelligent review rules generated using AI techniques effectively reduce the false positive rate and review rate and prevent missed diagnoses of hematologic cells while ensuring test quality. The review rules generated by AI methods significantly increase laboratory efficiency and are especially suitable for medical institutions.