Study: Machine learning quadruples delirium detection in hospitalized older adults

Editor's Note

A machine learning (ML) model that integrates clinical data with natural language processing significantly improved detection and management of hospital delirium in older adults. Results were published May 7 in JAMA Network Open.

Conducted at Mount Sinai Hospital, the quality improvement study evaluated the association of an ML-based model designed to stratify risk of non-ICU delirium in patients aged 60 and older. The model combined structured electronic medical record (EMR) data with natural language features extracted from clinical notes using an NLP pipeline. Trained using admissions data from 2016 to 2020 and validated in live clinical use between March 2023 and March 2024, the model was assessed against outcomes including detection rates, medication use, and hospital length of stay.

This model improved monthly detection rates from 4.42% to 17.17%, a fourfold improvement. In addition, researchers write, patients received significantly lower daily doses of benzodiazepines and Olanzapine. However, a higher percentage of patients in the post-deployment cohort received antipsychotics and opiates overall. Median hospital stay rose from 6.78 to 13.11 days in the post-ML cohort. According to researchers, the rise likely reflects higher acuity, as indicated by comorbidity scores and delirium prevalence. 

Researchers attribute the model’s success to a vertically integrated development approach and the inclusion of diverse clinical inputs, rather than relying solely on structured EMR data. Although results are promising, the study’s single-site deployment and reliance on an existing delirium care program may limit generalizability. Future plans include expanding use to affiliated hospitals and conducting external validations.

As detailed in a May 7 article in TechTarget, this is only the latest example of a machine learning tool for predicting delirium, citing 2023 work by Johns Hopkins University researchers.