Delirium misdiagnosis risk in psychiatry: a machine learning-logistic regression predictive algorithm

Abstract

Background

Delirium is a frequent diagnosis made by Consultation-Liaison Psychiatry (CLP). Numerous studies have demonstrated misdiagnosis prior to referral to CLP. Few studies have considered the factors underlying misdiagnosis using multivariate approaches.

Objectives

To determine the number of cases referred to CLP, which are misdiagnosed at time of referral, to build an accurate predictive classifier algorithm, using input variables related to delirium misdiagnosis.

Method

A retrospective observational study was conducted at Alfred Hospital in Melbourne, collecting data from a record of all patients seen by CLP for a period of 5 months. Data was collected pertaining to putative factors underlying misdiagnosis. A Machine Learning-Logistic Regression classifier model was built, to classify cases of accurate delirium diagnosis vs. misdiagnosis.

Results

Thirty five of 74 new cases referred were misdiagnosed. The proposed predictive algorithm achieved a mean Receiver Operating Characteristic (ROC) Area under the curve (AUC) of 79%, an average 72% classification accuracy, 77% sensitivity and 67% specificity.

CONCLUSIONS: Delirium is commonly misdiagnosed in hospital settings. Our findings support the potential application of Machine Leaning-logistic predictive classifier in health care settings.