[Establishment and validation of serum quality identification model]

Objective: To construct an artificial intelligence-based recognition model for serum quality using machine learning, enabling the accurate identification of three types of abnormal specimens (hemolysis, lipemia, and jaundice) as well as their severity levels. Methods: Serum sample images were collected from outpatients, inpatients, and physical examinees in the Department of Clinical Laboratory of the First Medical Center of Chinese PLA General Hospital from November to December 2024. After double-blind labeling by five senior laboratory technicians, image preprocessing and data augmentation were performed. The dataset was split into training and validation sets in a 7∶3 ratio. A residual convolutional neural network was used to extract features, which were then used to construct a fully connected model. Model performance was evaluated using precision, recall, area under the receiver operating characteristic curve (AUC), and other indicators. A total of 2 218 samples randomly collected on September 2, 2025, were used for double-blind test validation. Results: Among the 38 278 serum samples, the AUC values of the model for identifying hemolysis, chyle, and icterus were 0.997, 0.998, and 0.967, respectively, and the AUC values for severity grading identification were all greater than 0.963. In the double-blind test, the accuracy for serum window recognition, normal serum, and abnormal serum was 99.37%, 93.27%, and 90.91%, respectively, and the accuracy for identifying the three types of abnormal specimens was 90.11%, 85.21%, and 71.70%, respectively. The overall recognition accuracy for severe abnormalities was higher than that for mild abnormalities. The difference was statistically significant for hemolysis (92.11% vs.75.47%, χ²=4.23, P=0.040), but not statistically significant for chyle (88.24% vs.78.40%, P=0.420) or icterus (85.71% vs. 69.57%, P=0.630). Conclusion: The deep learning model constructed in this study can accurately identify abnormal serum quality types and their severity grades, and shows high recognition accuracy in double-blind tests.