The institutional review board of Yokohama Rosai Hospital approved the study (No. 31-42) and waived the requirement for informed consent owing to the retrospective design.
Medical record review
This retrospective cohort study was conducted at Yokohama Rosai Hospital, a tertiary medical center in Japan. We reviewed the medical records of patients (a) who had been suspected to have any psychiatric disorder by physicians in departments other than the Department of Psychosomatic Medicine, (b) who had been referred to the Department of Psychosomatic Medicine from March 2019 to October 2019, (c) who had been diagnosed with delirium, and (d) who had not been under continuous sedation.
The diagnoses was performed based on the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5)  by senior residents (post-graduate year of 3–5) initially and then supervised by a specialist certified by the Japanese Society of Psychiatric and Neurology (Y.F.).
We defined the first 24 hours from 9:00 p.m. on the referral date as day 1, and every subsequent period of the following 24 hours as day N. We reviewed the medical charts to determine whether patients had presented with delirium on day 3, which was denoted as the outcome. Day 3 was selected because a previous study suggested that an outcome evaluation on this day could be used to estimate the treatment effectiveness , and this outcome has been widely adopted [15,16]. Two authors (K.K. and D.M.; senior residents) reviewed the medical records independently referring to the prepared manual. The manual included (a) confirmation of the baseline (before referring) level of consciousness, (b) confirmation of the satisfaction for the diagnostic criteria of DSM-5 , and (c) discussion for disagreement in evaluation of each item of the diagnostic criteria. The manual was supervised by a board member of the Japanese Society of Psychosomatic Medicine (K.Y.). The reproducibility of the evaluation (before discussion for disagreement) was analyzed using the kappa statistic.
The number of drugs prescribed on the referral date was counted, excluding general infusion fluids, nutritional or electrolytic products, and psychotropic agents.
We used the t-test to compare the means of continuous variables, and Fisher’s exact test or the chi-squared test to compare the proportions of categorical variables in patients with and without delirium symptoms on day 3.
We performed the following analyses to determine the cutoff for polypharmacy [17,18]. First, we developed a classification and regression tree (CART) model in which the maximum tree depth was fixed at one and the Gini index was used as the splitting metric. We considered the splitting value of the root node to be the cutoff value. We also generated a receiver operating characteristic (ROC) curve and calculated the Youden index for each cutoff point. The cutoff that maximized the Youden index was considered the optimal cutoff value.
We calculated the odds ratio (OR) of polypharmacy for the persistence of delirium on day 3 using a logistic regression model with the propensity score as a covariate. Propensity scores were calculated using a logistic regression model based on the following variables: age, sex, the presence of 17 diseases in the updated Charlson comorbidity index (an index to predict mortality based on the presence of comorbid conditions) , the Karnofsky Performance Status (an index of physical performance ability)  score on admission, and psychotropics (five categories: antipsychotics, benzodiazepine or Z-drugs, ramelteon or suvorexant, antidepressants, and others). The Karnofsky Performance Status was used as an index of the severity of patients. Psychotropics were included because of their potential effects on delirium [1,21]. To confirm the sufficiency of the sample size, we performed a power analysis of the univariate logistic regression model.
All analyses were conducted using R version 4.0.0 (R Foundation for Statistical Computing, Vienna, Austria, 2019) with the packages ‘powerMeditation’ (version 0.3.2), ‘ROCR’ (version 1.0-11), ‘rpart’ (version 4.1-15), and ‘vcd’ (version 1.4-7). A P-value < 0.05 was considered statistically significant.