# A clinical prediction model for delirium tremens: development and validation in alcohol-dependent patients using multivariable logistic regression

## TL;DR

This study created a model to predict delirium tremens in alcohol-dependent patients using lab tests and medical history.

## Contribution

A new clinical prediction model for delirium tremens using multivariable logistic regression and routine biomarkers.

## Key findings

- The model achieved high accuracy (AUC of 0.9881 in training and 0.9599 in validation) for predicting delirium tremens.
- Key predictors included history of DT, ammonia, creatinine, albumin, and thyroid hormone levels.
- The model showed high net benefit in decision curve analysis.

## Abstract

Delirium tremens (DT) is a severe complication of alcohol withdrawal. This study aimed to develop and validate a prediction model for DT risk in hospitalized patients with alcohol dependence, using routine laboratory indicators.

We retrospectively analyzed 347 patients with alcohol dependence admitted to the Addiction Medicine Department of a tertiary psychiatric hospital from 2020 to 2024. The primary outcome was DT occurrence. A prediction model was constructed using logistic regression, with data split into training (70%) and validation (30%) sets by random sampling. Model performance was evaluated via the area under the receiver operating characteristic curve (AUC), calibration plots, and decision curve analysis (DCA).

Of 347 patients, 118 (34%) developed DT. LASSO regression identified 11 predictors: history of DT, ammonia, creatinine, uric acid, total bilirubin (Tbiliary), albumin (ALB), gamma-glutamyl transferase (GGT), chloride (Cl), free triiodothyronine (Free_T3), free thyroxine (Free_T4), neutrophil percentage (NEU%), and red blood cell (RBC) count. Logistic regression confirmed that history of DT, ammonia, creatinine, ALB, Free_T3, NEU%, and RBC were independent risk factors (P< 0.05). The model demonstrated robust performance: AUC = 0.9881 [95% CI: 0.9794–0.9967] in the training set and 0.9599 [95% CI: 0.9142–1.0000] in the validation set, with high net benefit in DCA.

This model, incorporating readily available biomarkers and clinical history, effectively predicts DT risk. Limitations include its retrospective design (potential selection bias) and exclusion of clinical scales (e.g., CIWA-Ar). Prospective multicenter studies are needed to validate its generalizability.