Data source

The study used data from the Clinical Practice Research Databank (CPRD) GOLD and Aurum databases. ^{Reference Wolf, Dedman, Campbell, Booth, Lunn and Chapman16,Reference Herrett, Gallagher, Bhaskaran, Forbes, Mathur and Staa17} These two large research databases contain longitudinal, anonymised patient-level electronic health records (EHRs) from general practices (GPs) in the UK. Vision software was commonly used in the past but has reduced substantially in recent years in England. The GOLD database includes over 11.3 million patient data contributed by GPs that deploy the Vision software system, and Aurum captures 19 million patient data from GPs that encompass the EMIS-Optum Web software. ^{Reference Wolf, Dedman, Campbell, Booth, Lunn and Chapman16,Reference Herrett, Gallagher, Bhaskaran, Forbes, Mathur and Staa17} The patient-level data in CPRD GOLD and Aurum have been linked through a trusted third party to hospital admission data (Hospital Episode Statistics) in England, using unique patient identifiers. ^{Reference Herrett, Gallagher, Bhaskaran, Forbes, Mathur and Staa17} The hospital data contained information on the date of hospital admission and the clinical diagnoses established at and during admission and coded using the ICD-10.

Ethics and consent statement

The study protocol was approved by CPRD’s Independent Scientific Advisory Committee (references 20_150R and 24_004253). All data sent to the CPRD is anonymised and therefore consent is not required.

Participants

This study used a matched case–control study design. The study population comprised patients aged 65–100 years at any point during the observation period (1 January 2000 to 1 July 2020 for CPRD GOLD, or up to 1 September 2020 for Aurum), who were registered in practices from England and participated in record linkage. Patients were followed from the date they met all inclusion criteria (registration in a practice, attainment of age 65 years and prior history of registration in the practice of at least 3 years) until the date of death, list removal (e.g. because of moving away) or reaching the maximum follow-up age of 101 years. The follow-up of each patient was divided into 3-month intervals, with risk factors such as morbidity indicators assessed at each of these time intervals. These data were used in the matching process. At the start of each 3-month interval, the history of polypharmacy was defined as the prescribing of five or more distinct drug classes within the preceding 84 days, ^{Reference Guthrie, Makubate, Hernandez-Santiago and Dreischulte18} which was used to determine eligibility for both cases and controls. Non-pharmacological prescribing was excluded, following a previous study. ^{Reference Guthrie, Makubate, Hernandez-Santiago and Dreischulte18} Supplementary Fig. 1 (available at https://doi.org/10.1192/bjp.2026.10588) illustrates the timeline of eligibility criteria, follow-up and baseline characteristics.

Cases and controls

The index date for cases was defined as the date of hospital admission of interest, as specified in the outcome section. Cases were patients who experienced a first hospital admission of interest and had a history of polypharmacy at the start of the 3-month interval containing the index date. Cases were matched to up to six controls who had a history of polypharmacy but had not been admitted to a hospital in the year before the index date. The index date for matched controls was assigned to closely align with the case’s index date based on calendar time, ranging from within 3 months to a maximum difference of up to 5 years. Hospital admissions after the index date were permitted for controls.

The objective of the matching was to achieve comparability in the risk of hospital admission between cases and controls based on morbidity, as determined by disease (but not by treatments). A directed acyclic graph illustrating the selection and definition of risk factors for matching morbidity indicators is shown in Supplementary Fig. 2. Matching was done using propensity matching (using the QAdmission Score) as well as matching by age, gender, morbidity cluster, presence of frailty, practice coding level and calendar time. Supplementary Fig. 3 illustrates the matching process. Matching was done separately for CPRD GOLD and Aurum. The risk-set approach to control sampling was used (with control patients potentially included as controls for multiple cases, although only once for a particular case).

Variables

The variables extracted from CPRD databases included socioeconomic information, clinical diagnoses, symptoms, patients’ characteristics and lifestyle factors, prescriptions and referrals to secondary care. Deprivation score, chronic disease, age, gender, ethnicity and lifestyle factors (smoking, alcohol intake) were used to estimate the QAdmission score. ^{Reference Hippisley-Cox and Coupland19} Prescription information contained product name, formulation and route of product administration, British National Formulary (BNF) chapter, dosage instruction, quantity and strength. Daily dose and the number of tablets to be taken per day were calculated from the unstructured free text with an algorithm, as done by the data provider. ^{Reference Shah and Martinez20}

Outcome

The primary outcome was emergency hospital admission, defined as a hospital admission accompanied by an emergency department visit on the same day, following the methodology outlined by Budnitz et al. ^{Reference Budnitz, Lovegrove, Shehab and Richards21} The secondary outcome was ADR-related hospital admission, defined as a hospital admission with an ADR-related admission code recorded. Details on ADR-related hospital admission are available in Supplementary Table 1.

Exposure

The primary exposure of interest was phases of antidepressant treatment use. We extracted details of all antidepressant prescription history (section 4.3 of BNF) for up to 3 years before the index date. A total of 31 antidepressants listed in Supplementary Table 2 were included. If no antidepressant prescriptions appeared during the 3 years before the index date, patients were classified as no exposure. Otherwise, participants were regarded as users who were grouped into nine mutually exclusive phases of antidepressant treatment. We used clinically meaningful periods of antidepressant use to define the phases of antidepressant treatment, as adopted in Cochrane reviews, clinical guidelines and observation studies. ^{Reference Van Leeuwen, van Driel, Horowitz, Kendrick, Donald and De Sutter10,Reference Valuck, Orton and Libby22} Because our data cannot distinguish between intentional discontinuation and temporary interruptions, we use the neutral term ‘treatment gap or discontinuation’, indicating that not all prescription gaps necessarily reflect a true antidepressant withdrawal. The phases include initiation, continuation, early maintenance, long-term maintenance, abbreviated trial, short treatment gap or early discontinuation after short-term use, short treatment gap or early discontinuation after long-term maintenance, moderate treatment gap or late discontinuation, and past exposure. In our population, antidepressant prescribing often includes gaps between prescriptions (see Supplementary Table 3). Therefore, it was necessary to specify a prescription-free interval that distinguished temporary interruptions from true discontinuation. We applied a threshold of ≥180 antidepressant-free days for this purpose. NICE guidance recommends reviewing patients who continue with antidepressant medication at least every 6 months, including discussion about whether to continue or stop treatment. ²³ Accordingly, in routine primary care, patients who restart medication within this time frame are likely to be continuing the same course of therapy. A shorter gap (<180 days) may reflect tapering, delayed refills or unsuccessful withdrawal attempts. Longer gaps (e.g. 365 days) could lead to misclassifying a new antidepressant initiation after a period of discontinuation as ongoing maintenance therapy.

Supplementary Fig. 4 shows our definition and cut-off points of the phases. If a prescription was issued after a gap of ≥180 days, it was considered a new treatment initiation, and phase classification was recalculated from that start date. If such a gap does not exist in the prescription history, the first prescribing date in the observation window was considered the start date of treatment. The gap was defined as the interval between the expected end date of a prescription and the start date of the next prescription (when applicable). The expected end date was equal to the start date of the antidepressant prescription plus the duration of the prescription. The duration of the prescription was estimated by dividing the prescribed quantity by the daily dose. ^{Reference Shah and Martinez20} For patients who received multiple antidepressant prescriptions on the same date, we assumed patients would concurrently take these drugs, and the length of these concurrent prescriptions was based on the longest length. Multiple imputation was employed for prescriptions with incomplete duration data. Limited information was available on the missingness mechanism for antidepressant treatment quantity and dose, making the missing-at-random assumption not fully testable. To assess the plausibility of this assumption, we conducted analyses to examine the distribution of predictors in the multiple imputation model between patients with observed and missing treatment duration.

Statistical analysis

The propensity matching procedure used a caliper (prespecified maximum difference) of 0.25 of the logit of the QAdmission score. ^{Reference Austin24} We used predictive mean matching to impute daily dose and quantity in five separate imputed data-sets, and did five times iterations (imputation cycles) during the imputation process. Because the antidepressant phase variable was defined relative to the outcome-specific index date, multiple imputation was performed separately for each outcome data-set (emergency hospital admission and ADR-related hospital admission). Definitions and a list of the predictors are given in Supplementary Table 4.

We used conditional logistic regression to examine the association between phases of antidepressant use and two outcomes: emergency and ADR-related hospital admission. In the conditional logistic regression, the effects of different phases were based on comparing each case with matched controls, and odds ratios and 95% confidence intervals were estimated with Rubin’s rules across imputed data-sets. Potential confounders in the regression models included Index of Multiple Deprivation quintiles, disease history, smoking history, the quantity of co-medications in BNF Chapter 4, the quantity of co-medications not in BNF Chapter 4, ethnicity and Charlson Comorbidity Index group. ^{Reference Charlson, Wells, Ullman, King and Shmukler25} Disease history covered both potential indications for antidepressant use (depression, anxiety, mania or schizophrenia, dementia and Parkinson’s disease) and comorbidities that could influence both antidepressant prescribing and the risk of hospital admission (cardiovascular disease, atrial fibrillation, heart failure, chronic obstructive pulmonary disease or asthma, diabetes mellitus, renal disease, cancer and previous falls), with each condition represented as a separate variable. The quantities of co-medications were treated as continuous variables with a linear relationship to the outcome, and all other covariates were treated as categorical variables. Subgroup analyses stratified by age, gender and specific ICD-10 hospital admission codes were conducted using the same imputed data-sets as those generated for the corresponding main outcome. We displayed the resulting odds ratios and 95% confidence intervals in heatmaps, where red cells (print version: blue cells) indicate higher adjusted odds ratios and blue cells (print version: grey cells) indicate lower adjusted odds ratios, compared with the no exposure group. We conducted five sensitivity analyses to examine the robustness of our findings: (a) reclassifying antidepressant phase exposures: patients with ≤7-day gaps before the index date were reclassified into antidepressant current use phases; (b) stratifying phase exposures by classes (selective serotonin reuptake inhibitors (SSRIs), tricyclic antidepressants and other antidepressants); (c) repeating heatmap analysis stratified by the three antidepressant classes; (d) redefining antidepressant phase exposures and (e) resetting the antidepressant-free gap to 84 days as the threshold to distinguished temporary interruptions from true discontinuation. We applied a threshold of ≥180 antidepressant-free days for this purpose. All statistical and sensitivity analyses were performed using R software version 4.2.2 on the Windows platform (R Foundation for Statistical Computing, Vienna, Austria; https://www.r-project.org).