Predicting the risk of delirium in older population: development and validation of clinical prediction models using big data

Introduction: Older people with multiple long-term conditions and polypharmacy are at increased risk of adverse health outcomes, including delirium. This study aimed to develop and validate prediction models for an individual’s risk of delirium within 1 and 5 years.

Method: We conducted a retrospective observational cohort study using the Clinical Practice Research Datalink Aurum database. Participants were aged ≥65 years, prescribed at least one chronic medication, registered with a UK general practice between 1/10/2006 and 1/3/2023, and had at least 1 year of records. The primary outcome was delirium leading to hospital admission or death within a year of the index date. Prespecified predictors included patient characteristics, comorbidities, and all medications other than vaccines, immunoglobulins and anaesthetics (199 drug classes). Variable selection was performed using LASSO Cox regression with Bayesian information criterion. The primary analysis used a Fine-Gray competing risks approach to account for death from other causes. Missing data were handled using multiple imputation.

Results: Model development cohort included 2,681,502 patients and external validation included 1,043,270 patients. Delirium occurred in 20,661 (0.77%) and 99,052 (3.7%) patients within 1 and 5 years, respectively, in the development cohort; and in 11,142 (1.07%) and 51,049 (4.9%) patients in the validation cohort. The 1-year and 5-year models included 76 and 129 predictors, respectively. Memory concerns, psychosis and bipolar disorder, and prescription of anti-epileptic drugs were the strongest predictors at both time points, whereas dementia was a strong predictor only in the 1-year model. External validation demonstrated good calibration (1-year O/E ratio=1.411, 95%CI 1.409-1.414) and discrimination (1-year C-statistic=0.848, 95%CI 0.845-0.851). 

Conclusions: We developed and externally validated two clinical prediction models for delirium risk that demonstrated good predictive performance. Neuropsychiatric conditions were among the strongest predictors. Despite the underprediction at lower risks and overprediction at the highest risks, decision curve analysis indicated potential clinical utility.