SIMPLE NOMOGRAM TO PREDICT 30-DAY READMISSION WITH RECURRENT VARICEAL BLEEDING IN LIVER CIRRHOSIS PATIENTS USING SUPERVISED MACHINE LEARNING FROM UNITED STATES POPULATION

Introduction:
Acute variceal bleeding (AVB) is a major cause of death in cirrhotic patients. We developed and internally validated a pragmatic model to predict the individualized risk of 30-day readmission with recurrent AVB in liver cirrhosis patients.

Methods:
Hospitalizations with a primary diagnosis of esophageal AVB were identified using the 2019's Nationwide Readmission Database (NRD). We utilized the NRD as it recognizes the same patient's index admissions and recognition of readmissions. Patients were excluded if aged <18 years, had non-AVB, end-stage renal disease, solid organ transplants, anticoagulation use, immunosuppression, para/quadriplegics, lymphomas/leukemias, or malignant tumors. We used supervised machine learning to input variables with increased association with recurrent AVB readmissions to undergo Least Absolute Shrinkage and Selection Operator (LASSO) penalized regression for selecting the best predictors. Receiver operating characteristic (ROC) curves assessed predictive power for each selected predictor and those with a poor threshold of discrimination [area under the curve (AUC <0.60)] were eliminated. The remaining predictors were utilized to develop a variceal AVB nomogram. The nomogram was internally validated using 10 fold cross validation, and ROC curves were generated along with bootstrapped Bias corrected (BC) 95% confidence intervals (CI) for the AUC. The Brier score was used to report measures of overall performance.

Results:
The prevalence of recurrent AVB readmissions was 2.65%. For the model predicting the risk, five predictors were included: Chronic pulmonary disease history (AUC 0.84± 0.003 ), Age >=50 years (AUC 0.61 ± 0.002), Transjugular intrahepatic portosystemic shunt (TIPS) during index hospitalization (AUC 0.89±0.003 ), Obesity (AUC 0.84± 0.002), history of cardiac arrhythmia (AUC 0.84± 0.002) (Figure 1). These variables were used to develop a nomogram that displayed outstanding discrimination AUC 0.91 (Bootstrap Bias Corrected 95%CI 0.90-0.92), correlating to a 91% probability of the model correctly assigning a higher score to patients at risk of recurrent variceal bleed readmission within 30 days of discharge (Figure 2A, 2B). Liu's index was used to determine the cut-off (21 points). Therefore patients with a score ≥ 21 were deemed at high risk for recurrent AVB readmission. For high-risk patients, with a sensitivity of 80.44% and a negative predictive value of 99.47%, the specificity was 100.00% with a positive predictive value of 100.00%. The Brier score was 0.005, indicating the good overall performance of the nomogram.

Conclusions:
The proposed nomogram score can be used to identify such patients with a risk for recurrent variceal bleed readmission within 30 days of discharge.