An artificial intelligence-based semi-quantitative diagnostic model for intra-abdominal hemorrhage based on focused...

To develop and validate an artificial intelligence model based on focused assessment with sonography for trauma (FAST) for the semi-quantitative grading of intra-abdominal hemorrhage resulting from blunt abdominal trauma, particularly for use in prehospital or resource-limited settings.

Nine Bama miniature pigs, mean weight (31.46 ± 3.73) kg were enrolled. Graded hemorrhage from 0 to 1000 mL was simulated by infusing 100 mL of autologous arterial blood into the peritoneal cavity at each step. The hemorrhage volume was mapped to 3 grades based on total blood volume (estimated at 65 mL/kg): Grade I (< 15%), Grade II (15% - 30%), and Grade III (> 30%). FAST ultrasound videos were acquired from 6 standard sites: right upper quadrant-1, right upper quadrant-2, left upper quadrant-1, left upper quadrant-2, right pelvic cavity, and left pelvic cavity. The pixel area of hemorrhage was obtained by manually segmenting the frame with the largest fluid collection using ITK-SNAP, and the corresponding scanning depth was recorded. A linear mixed-effects model was used to assess the impact of scanning depth on pixel area. A deep neural network, incorporating class weighting and dynamic probability threshold optimization, was constructed using a multimodal feature set including animal weight, pixel areas and scanning depths from each site, and the total pixel area. A 3-grade classification was performed. The model's performance was evaluated using leave-one-out cross-validation on an animal basis and compared with logistic regression, random forest, gradient boosting decision tree, and support vector machine.

A total of 797 raw videos were acquired, with 522 videos comprising 87 data groups (each covering 6 sites) included after screening. As hemorrhage volume increased, heart rate and shock index rose, while systolic blood pressure decreased; at 800 mL of hemorrhage, the shock index was 2.31 ± 0.38. The mixed-effects model revealed a significant negative correlation between scanning depth and pixel area (β = -2099.00, SE = 1041.13, z = -2.02, p = 0.044). The proposed model achieved an overall accuracy of 81.19%, outperforming support vector machine (73.77%), gradient boosting decision tree (70.63%), random forest (69.52%), and logistic regression (65.99%).

In a porcine model of blunt abdominal trauma, a multimodal artificial intelligence approach based on FAST multi-site pixel area features, combined with a deep neural network optimized by class weighting and dynamic probability thresholds, can achieve semi-quantitative grading of intra-abdominal hemorrhage.