Development and validation of risk prediction models for high-risk patients with non-traumatic acute abdominal pain: a prospective observational study

Comprehensive Summary

Li et al. developed and validated a machine learning model to detect high-risk patients presenting to the emergency department (ED) with non-traumatic acute abdominal pain (NAAP). Data from patients at two centers were prospectively collected and analyzed to identify predictive risk factors. Seven early risk warning models were constructed and externally validated. Of 3,090 patients with NAAP, 14.49% were identified as high-risk. The ten strongest predictive factors included age, mode of admission, history of heart disease, history of tumor, MEWS score ≥5, post-coital trigger, knife-like pain, abdominal distension and fullness, tenderness, and muscle tension. The seven models had good predictive performance, with the random forest model demonstrating the best overall performance (Area under the Curve (AUC) = 0.786). Upon external validation, logistic regression using a nomogram had the strongest clinical application and generalizability.

Outcomes and Implications

Quality assessment of ED patients presenting with NAAP directly determines the efficiency of medical procedures and patient outcomes. This study found that the random forest machine-learning model had the strongest performance in predicting high-risk patients with NAAP. A series of independent risk factors were identified, including age, mode of admission as well as history of heart disease and tumors. These results suggest that such models can help nurses improve triage accuracy, implement timely interventions, and optimize emergency resource allocation. The study was limited by its reliance on convenience sampling from only two medical centers, the lack of external validation across different hospital levels, and the exclusion of pregnant patients, which may restrict generalizability. Future research is warranted to externally validate the models across various clinical settings and test dynamic risk warning approaches.