# Development and evaluation of a machine learning model to predict unplanned readmission risk in patients with ulcerative colitis

## TL;DR

A machine learning model was developed to predict the risk of unplanned hospital readmission in patients with ulcerative colitis, offering a useful tool for personalized patient care.

## Contribution

The study introduces a novel random forest-based machine learning model with high accuracy for predicting unplanned readmission risk in UC patients.

## Key findings

- The random forest model achieved an AUC of 0.936 in training and 0.813 in external validation for predicting readmission risk.
- Five key predictors were identified: C-reactive protein, erythrocyte sedimentation rate, red blood cell count, bowel movement frequency, and platelet count.
- An online risk prediction platform was developed based on the model for clinical use.

## Abstract

Ulcerative colitis (UC), a chronic inflammatory bowel disease marked by recurrent flares and remissions, often necessitates repeated hospitalization owing to disease variability. However, commonly used risk-scoring systems have limited predictive accuracy for hospital readmission. This study aimed to develop and validate a machine learning (ML)-based model to predict the risk of unplanned readmission within 1 year in patients with UC.

Unplanned readmission within 1 year was defined as an endpoint event, and a predictive model was developed using a retrospective cohort (n = 324) and externally validated using an independent prospective cohort (n = 137). Demographic characteristics, medical history, medication use, clinical symptoms, laboratory findings, and endoscopic data were integrated as input variables. The optimal feature subset was selected using Recursive Feature Elimination (RFE), and eight ML models were constructed. All models were optimized via five-fold cross-validation, and the best-performing model was selected as the final predictive tool and was subjected to external validation. Shapley additive explanation plots were used to interpret the predictive model.

The RFE algorithm identified five critical predictors: C-reactive protein, erythrocyte sedimentation rate, red blood cell count, increased frequency of bowel movements, and platelet count. All ML models achieved an AUC above 0.75 in the training cohort, demonstrating their robust predictive capability. The random forest (RF) model consistently outperformed the others across the training, internal validation, and external validation cohorts, with AUCs of 0.936, 0.815, and 0.813, respectively, reflecting excellent stability and generalization. Building upon the RF model, an online risk prediction platform was developed to estimate the probability of unplanned readmission in patients with UC.

The RF-based model showed strong predictive accuracy for assessing the 1-year risk of unplanned readmission in UC patients. The corresponding web-based risk calculator offers clinicians a valuable tool for personalized risk evaluation and enhanced patient management.