Data-derived subtypes of delirium during critical illness.

Background: To understand delirium heterogeneity, prior work relied on psychomotor symptoms or risk factors to identify subtypes. Data-driven approaches have used machine learning to identify biologically plausible, treatment-responsive subtypes of other acute illnesses but have not been used to examine delirium.
Methods: We conducted a secondary analysis of a large, multicenter prospective cohort study involving adults in medical or surgical ICUs with respiratory failure or shock who experienced delirium per the Confusion Assessment Method for the ICU. We used data collected before delirium diagnosis in an unsupervised latent class model to identify delirium subtypes and then compared demographics, clinical characteristics, and outcomes between subtypes in the final model.
Findings: The 731 patients who developed delirium during critical illness had a median age of 63 [IQR, 54-72] years, a median Sequential Organ Failure Assessment score of 8.0 [6.0-11.0] and 613 [83.4%] were mechanically ventilated at delirium identification. A four-class model best fit the data with 50% of patients in subtype (ST) 1, 18% in subtype 2, 17% in subtype 3, and 14% in subtype 4. Subtype 2-which had more shock and kidney impairment-had the highest mortality (33% [ST2] vs. 17% [ST1], 25% [ST3], and 17% [ST4], p = 0.003). Subtype 4-which received more benzodiazepines and opioids-had the longest duration of delirium (6 days [ST4] vs. 3 [ST1], 4 [ST2], and 3 days [ST3], p < 0.001) and coma (4 days [ST4] vs. 2 [ST1], 1 [ST2], and 2 days [ST3], p < 0.001). Each of the four data-derived delirium subtypes was observed within previously identified psychomotor and risk factor-based delirium subtypes. Clinically significant cognitive impairment affected all subtypes at follow-up, but its severity did not differ by subtype (3-month, p = 0.26; 12-month, p = 0.80).
Interpretation: The four data-derived delirium subtypes identified in this study should now be validated in independent cohorts, examined for differential treatment effects in trials, and inform mechanistic work evaluating treatment targets.
Funding: National Institutes of Health (T32HL007820, R01AG027472).
Competing Interests: Declaration of interests Kelly Potter acknowledges support from the National Institutes of Health (NIH). Jason Kennedy acknowledges support from the NIH. Niall Prendergast received funding from the NIH and receives authorship royalties from McGraw-Hill. Chukwudi Onyemekwu acknowledges support from the NIH. Wes Ely reports work with Eli Lily as an unfunded investigator on COVID-19 studies with Baricitinib and acknowledges VA and NIH grant support. Christopher Seymour receives research funding from the NIH, receives personal fees from Inotrem, Beckman–Coulter, and Octapharma, and is an Associate Editor for JAMA. Timothy Girard receives research funding from the NIH, the Department of Defense, and Ceribell and served previously on an advisory board for Lungpacer Medical Inc.
(Copyright © 2023 The Authors. Published by Elsevier B.V. All rights reserved.)