This study was approved by the Institutional Review Board and Ethics Committee of The University of Tokyo (Institutional Review Board number: 3501). The requirement for obtaining written informed consent was waived as the study was a secondary analysis of anonymous administrative data.

Patient data were extracted from the Japanese Diagnosis Procedure Combination database, a national database of administrative claims and discharge abstracts in Japan. Eighty-two university hospitals in Japan participate mandatorily in the database, whereas over 1600 community hospitals participate voluntarily. The database included the administrative data of 11 million inpatients in 2020, accounting for approximately 80% of all acute care inpatients in Japan [14]. The details of the database have been described previously. [15]

The database includes the following data: hospital information (hospital identifier, hospital type, and number of hospital beds), patient information (patient identifier, age, sex, height, body weight, and smoking status), information at admission (purpose of hospitalization, primary diagnosis, comorbidities, route of hospitalization, activities of daily living [ADL]), information regarding treatments (surgery, anesthesia, medication, and blood transfusion), information at discharge (discharge status and ADL), and costs. Diagnoses are recorded using the International Classification of Diseases, 10th Revision (ICD-10) codes and Japanese text. All medical procedures are encoded using the original Japanese medical procedure codes.

A previous validation study on the Japanese Diagnosis Procedure Combination database revealed that the recorded procedures and drugs had high sensitivity and specificity, whereas the recorded diagnoses of common diseases, including malignant tumors, cardiac diseases, renal diseases, and stroke, had moderate sensitivity and high specificity [15].

This study included hospitalized patients with cancer aged ≥ 20 years who received chemotherapy and PCT intervention within 2 days of admission between January 1, 2015, and December 31, 2020. The present study only included hospitalized patients with cancer receiving chemotherapy to obtain a relatively homogeneous population. PCT intervention was identified using the registry of the Japanese Medical Procedure Code A226-2. Patients with cancer who were concurrently diagnosed with schizophrenia (ICD-10 codes F20.x–F29.x) were excluded from this study as it was difficult to distinguish whether these patients were receiving antipsychotics for schizophrenia or delirium. Emergency hospitalized patients were also excluded because receiving chemotherapy within 2 days of emergency hospitalization was considered an exceptional case.

The exposure of interest was the hospital PCT intervention volume, which was defined as the annual number of new PCT interventions performed in a hospital. The PCT intervention volume was calculated using the hospital identifier, and the hospitals were categorized into low-, intermediate-, or high-volume groups according to their tertiles.

The primary outcome was the incidence of hyperactive delirium within 30 days of admission. Hyperactive delirium was identified by the administration of haloperidol or risperidone ≥ 2 days after admission. This delirium identification algorithm was adopted in this study as a similar identification algorithm achieved sufficient validity for use in claims-based databases [16]. The secondary outcomes were mortality within 30 days of admission and a decline in ADL at discharge. Decline in ADL was identified by comparing the Barthel Index at admission and discharge. Death was categorized as a decline in ADL.

The potential confounders used for the regression analyses included the demographic and hospital characteristics of the patients, comorbidities, and medications prescribed at admission. These variables were selected from the pretreatment factors that could be associated with the incidence of delirium, mortality, and decline in ADL, according to clinical judgment and the existing literature [17,18,19,20,21,22,23,24,25,26]. Demographic characteristics included the year of admission, sex, age, body mass index, smoking status, and Barthel Index of the patients. The Barthel Index is a measure of the functional independence of individuals in ADL [27]. The total score on the Barthel Index ranges from 0 to 100. Higher scores on the Barthel Index indicate greater functional independence, with a score of 100 indicating that the individual was completely independent in all ADLs. The hospital characteristics included the type of hospital (academic/community), number of hospital beds, and chemotherapy volume (i.e., the annual number of inpatient chemotherapies in a hospital). The hospitals were categorized into two groups based on the median number of hospital beds, whereas the chemotherapy volume was categorized into three groups according to tertiles. Comorbidities included the Charlson Comorbidity Index, dementia (ICD-10 codes F00.x–F03.x or use of anti-dementia agents), brain metastasis (C79.3), and type of cancer (lung cancer, C33.x-C39.x; lower gastrointestinal cancer, C17.x-C21.x; upper gastrointestinal cancer, C15.x-C16.x; leukemia and lymphoma, C81.x-C96.x; and other cancers). The Charlson Comorbidity Index, a score used to classify the comorbid conditions of a patient, was calculated using Quan’s algorithm [28, 29]. Medications included antibiotics, opioids, gabapentinoids, and hypnotics. Patients who received these medications within 2 days of admission were considered to be on medication.

Multivariate logistic regression analyses were performed to estimate the odds ratios for 30-day delirium, 30-day mortality, and decline in ADL using the hospital PCT intervention volume and potential confounders as independent variables. The odds ratios were calculated for the intermediate and high hospital PCT intervention volume groups using the low hospital PCT intervention volume group as the reference. Multicollinearity was assessed using the variance inflation factor, with a variance inflation factor larger than 10 indicating deleterious multicollinearity [30].

Restricted cubic spline regression analyses were conducted subsequently to evaluate the potential non-linear associations between continuous hospital PCT intervention volume and outcomes. Restricted cubic spline regression analysis is advantageous over standard categorical regression analysis in that it circumvents the power loss associated with categorization [31, 32]. Three hospital PCT intervention volume points (the 10th, 50th, and 90th percentiles) were used as knots in the restricted cubic spline regression analysis. The odds ratios for each value of hospital PCT intervention volume were calculated using the lowest hospital PCT intervention volume (1 case/year) as the reference value. The hospital PCT intervention volume and the same potential confounders used in the logistic regression described above were used as independent variables in the restricted cubic spline regression analyses.

Categorical variables are presented as numbers (percentages) and were compared using the Chi-square test. Odds ratios are presented as 95% confidence intervals (CI). All reported P values were two-sided, and P values of < 0.05 were considered statistically significant. All statistical analyses were performed using Stata/SE 17.0 (Stata Corp., College Station, Texas, USA). This manuscript adhered to the Strengthening the Reporting of Observational Studies in Epidemiology guidelines.