Predicting inpatient mortality in patients with inflammatory bowel disease: A machine learning approach

Abstract

Background and aim: Data are lacking on predicting inpatient mortality (IM) in patients admitted for inflammatory bowel disease (IBD). IM is a critical outcome; however, difficulty in its prediction exists due to infrequent occurrence. We assessed IM predictors and developed a predictive model for IM using machine-learning (ML).

Methods: Using the National Inpatient Sample (NIS) database (2005-2017), we extracted adults admitted for IBD. After ML-guided predictor selection, we trained and internally validated multiple algorithms, targeting minimum sensitivity and positive likelihood ratio (+LR) ≥ 80% and ≥ 3, respectively. Diagnostic odds ratio (DOR) compared algorithm performance. The best performing algorithm was additionally trained and validated for an IBD-related surgery sub-cohort. External validation was done using NIS 2018.

Results: In 398 426 adult IBD admissions, IM was 0.32% overall, and 0.87% among the surgical cohort (n = 40 784). Increasing age, ulcerative colitis, IBD-related surgery, pneumonia, chronic lung disease, acute kidney injury, malnutrition, frailty, heart failure, blood transfusion, sepsis/septic shock and thromboembolism were associated with increased IM. The QLattice algorithm, provided the highest performance model (+LR: 3.2, 95% CI 3.0-3.3; area-under-curve [AUC]:0.87, 85% sensitivity, 73% specificity), distinguishing IM patients by 15.6-fold when comparing high to low-risk patients. The surgical cohort model (+LR: 8.5, AUC: 0.94, 85% sensitivity, 90% specificity), distinguished IM patients by 49-fold. Both models performed excellently in external validation. An online calculator (https://clinicalc.ai/im-ibd/) was developed allowing bedside model predictions.

Conclusions: An online prediction-model calculator captured > 80% IM cases during IBD-related admissions, with high discriminatory effectiveness. This allows for risk stratification and provides a basis for assessing interventions to reduce mortality in high-risk patients.

Keywords: IBD; artificial intelligence; calculator; hospitalized patients; machine learning; prediction model.

Figure 1

Flowchart of data selection. NIS, National Inpatient Sample; IBD, inflammatory bowel disease; LOS, length of stay.

Figure 2

Model performance by sensitivity and positive likelihood ratio for inpatient mortality. White area is the acceptance area with > 80% sensitivity and positive likelihood ratio of > 3. DOR, diagnostic odds ratio.

Figure 3

QLattice model structure. Gray boxes are predictors, white boxes are interactions between each predictor. Each predictor carries a learned weight (coefficient) per category level, plus additional bias (estimated error). Moving from left to right, a probability is finally calculated and converted to a binary outcome prediction (1 for probability > 0.5, else 0). (a) All adults admitted for IBD. (b) All adults admitted for IBD who underwent IBD‐related surgery.