A Time-Series Feature-Based Nomogram for the Prediction of Severe Acute Pancreatitis

Background The annual incidence of acute pancreatitis is approximately 30 per 100,000, with 20% progressing to severe acute pancreatitis and a mortality rate of 20%-40%. Traditional scoring models suffer from data lag or insufficient accuracy, while existing machine learning models mostly overlook the dynamic characteristics of vital signs. Methods Vital signs, laboratory and imaging indices within 24 hours of admission were collected. First, a bidirectional long short-term memory network model was constructed using time-series data. Then,key indices from laboratory and imaging data were screened by LASSO. Eight machine learning models were constructed and compared. Finally, a predictive nomogram was developed based on the Random Forest model and SHAP values. Result After propensity score matching, among 193 patients, there were 124 cases in the MSAP group and 69 cases in the SAP group, with no significant differences in baseline characteristics between the two groups. The BiLSTM model showed an average AUC of 0.9551, accuracy of 0.9222, F1-score of 0.8956, training loss of 0.2992 ± 0.0328, and validation loss of 0.4132 ± 0.0651 in 10-fold cross-validation. Features including Rmax, Pdiff_mean, and Tdiff_std extracted from time-series data, together with those screened by LASSO (PE, Neu, HCT, Ca, TG, AMY, and CRP), were used to construct 8 ML models. The Random Forest model demonstrated the best comprehensive performance, with an accuracy of 0.8793, ROC-AUC of 0.9588. SHAP value analysis identified key features as Rmax, Pdiff_mean, HCT, Tdiff_std, PE, Neu, and serum calcium. The nomogram constructed based on these features achieved AUC values of 0.969 and 0.964 in the training and test sets, respectively. Decision curve analysis showed that the net benefit exceeded 0.2 at high-risk thresholds (0.2–0.8), outperforming both the "treat all" and "treat none" strategies. Conclusion The BiLSTM-RF model constructed in this study improves the accuracy of SAP prediction by extracting time-series features of vital signs. The nomogram built based on key features demonstrates good clinical practicability, providing a visual tool for the early assessment of SAP.