The initial cohort comprised 177 individuals selected by simple random sampling from patients admitted to emergency-care wards of the “Prof. Dr. Alexandru Obregia” Clinical Psychiatry Hospital in Bucharest between September 2022 and January 2023. The full details of the initial cohort are described elsewhere [17]. For the present study, patients who were unable to be interviewed due to their cognitive status (e.g., patients with severe dementia or developmental disorders) and patients not residing in Bucharest were excluded, resulting in a study population of 144 patients. The rationale behind the last exclusion criteria was that Romania does not have a centralized digital database that tracks admissions, making the collecting of follow-up data on patients not residing in the hospital’s catchment area difficult and unreliable. A follow-up examination of patient digital records in the hospital database in December 2023, gathered data on the number of readmissions. Revolving-door patients were defined as those who had at least three admissions during the two years prior to follow-up. Patients who met this criterion constituted the study group, while the remainder were designated as controls.

Socio-demographic and clinical data were collected during the baseline visit by interview and from charts and were corroborated by the attending physician and the patient’s family when needed. Diagnoses on admission were grouped in one of the following categories: psychosis (including the International Statistical Classification of Diseases and Related Health Problems 10th Revision (ICD-10) diagnoses of schizophrenia, schizoaffective disorder, schizophreniform disorder and brief psychotic disorder), mania, depression (including both uni- and bipolar depression), substance use disorders and others (any diagnosis that did not belong in any of the previous categories). For patients diagnosed both with substance use disorder and another psychiatric condition, the attending physician was consulted and the one considered the primary reason for admission was chosen.

Disease severity was assessed using the Clinical Global Impression Severity Scale (CGI-S) [18], aggression was evaluated using the modified Overt Aggression Scale (mOAS) [19]. Treatment adherence, defined as taking 80–120% of prescribed psychiatric treatment dose in the four weeks before the current admission (or between the previous hospital discharge and present admission if they were less than four weeks apart), was evaluated only when at least two outside sources were available for corroboration (medical files, family, friends, attending physician, general practitioner).

All analyses were conducted using IBM SPSS Statistics software v26 (Armonk, NY, USA: IBM Corp). Descriptive analysis was used to evaluate social, demographic, and clinical characteristics. Continuous variables were assessed for normality using the Shapiro-Wilk test and analyzed using Mann-Whitney (for non-parametrical distributed data) or the Student t-test (for parametrically distributed data). Parametric data are presented with mean and standard deviation, while non-parametric data are given as median and interquartile range. Statistical significance was set at p = 0.05. Categorical variables with no more than two categories (e.g., treatment compliance yes/no) were analyzed using Chi-Square while categorical variables with more than two categories (e.g., mean of referral by: police, ambulance, family/friends, self) were individually analyzed using binary logistic regression. Variables that reached statistical significance at first analysis were entered into a multivariate logistic regression.

Sociodemographic characteristics are presented in Table 1. Of the 144 subjects, 38.9% (56) were RD patients. The two groups were similar in age, education, ethnicity, relationship status, housing and employment status. Most RD patients were male (64.3%), single (76.7%), received social benefits (50%) and shared a household with at least one other person (69.6%). Monthly income was lower in the RD group: 1450 $\pm$ 1440.9 RON vs 1892.7 $\pm$ 1980.5 RON (corresponding to 291 $\pm$ 289 EUR vs 381 $\pm$ 397 EUR), although the difference was not statistically significant.

Characterization of healthcare use is shown in Table 2. The involuntary admission at the initial visit resulted in a lower probability of RD status on follow-up (exp(B) = 0.392, 95% confidence interval (CI) = 0.155–0.991, p = 0.048). Number of mean lifetime admissions prior to baseline was significantly higher (p $\le$ 0.001) in the RD group (2.67 $\pm$ 2.91) than in the control group (0.49 $\pm$ 0.83). While the length of the initial hospitalization stay was similar, RD patients had less time between the initial admission and the next hospitalization (142.3 $\pm$ 143.6 days vs 324.8 $\pm$ 145.1 days for non-RD patients, p $\le$ 0.001). There were no significant differences in baseline method of admission (e.g., via police, ambulance, family, or self-referral).

The two groups were also similar in clinical characteristics at recruitment admission (Table 3), including diagnosis, disease severity, aggression and treatment compliance. Average disease duration was longer in RD patients (11.2 $\pm$ 10 years vs 8.1 $\pm$ 10.2 years). The difference was statistically significant in the univariate analysis (p = 0.012) but not in the multivariate analysis (Table 4), which controlled for other predictive factors (number of lifetime admissions and involuntary admission on initial visit).

Parameters associated with RD status by univariate analysis were introduced into a binary logistic regression model to account for potential reciprocal moderating effects. Of the four introduced in the model, three maintained statistical significance. Higher numbers of lifetime admissions were predictive of RD status (exp(B) = 3.956, p $\le$ 0.001), even after controlling for disease duration. Time between readmission after the initial visit remained significantly lower in the RD group (exp(B) = 0.991, p $\le$ 0.001). Involuntary status at the initial visit significantly decreased the probability of subsequent RD status (exp(B) = 0.188, p = 0.008) and the association was greater than in the univariate analysis.

In this study, no sociodemographic characteristic was associated with RD status. A potential explanation is that in Romania, while outpatient services are covered by medical insurance are limited to periodic checkups, more complex services such as psychotherapy, behavioral or occupational therapies and addiction centers are only available with payment and others such as halfway homes or mobile crisis units are not available at all. Thus, patients who cannot afford to access more complex aftercare services that might prevent readmission are forced to return to public hospitals. It is worth noting that monthly income in the current cohort was lower than the national minimum wage and less than half of the average national wage (1790 RON ~ 360 Euro vs. 2050 RON ~ 419 Euro vs. 4230 RON ~ 850 Euro). A recent systematic review [15] did not identify any consistent sociodemographic predictors; there is roughly an equal split between studies that found associations between RD and gender, age, ethnicity, education, marital status, employment status, income and urbanicity and those that did not.

The number of lifetime admissions before the baseline evaluation was significantly higher for the RD group studied here, even after controlling for disease duration, suggesting that the higher rate at which these patients access hospital care is not simply a function of longer periods of illness. As previous research has established, patients who enter a pattern of frequent hospitalizations tend to maintain it over time [23], further stressing the importance of prevention by ensuring access to high-quality outpatient care. Initial visit length was not predictive of subsequent RD status, although other studies found that shorter admissions were more likely to lead to rehospitalization [28]. The period of time immediately following discharge is known to carry the most risk of readmission [29] and RD patients in the current cohort had much shorter durations between the initial visit and readmission, than non-RD patients. In partial concordance with available data [27, 30, 31], involuntary admission at baseline emerged as a protective factor against RD status. This could be due to the alleviating effect of compulsory hospitalization and treatment on severely ill patients but could also reflect an increased reluctance to access mental healthcare services [28]. Conversely, other authors found a positive association between involuntary admission and RD status, attributing it to the greater severity of psychiatric symptoms in this category of patients [32].

Clinical and treatment characteristics, including diagnosis, disease severity, aggression and treatment compliance did not differ between the two groups. Disease duration was associated with RD status in the univariate analysis, but not after controlling for other predictors. While numerous studies cited psychosis and substance use disorders as risk factors [15], there is a significant minority in which there was no link between RD status and psychiatric diagnosis [33, 34]. Diverging from most existing literature, aggression, including self-harm, was not more common in the RD group studied. Nevertheless, this is not unprecedented, as [27] lower levels of lifetime physically violent behavior has been reported for RD patients. As previously speculated by the authors, with regard to sociodemographic variables, this is probably a consequence of the structural characteristics of the Romanian healthcare system, as patients who are not necessarily severely ill but cannot afford to access complex outpatient services after discharge tend to return to public hospitals. This effect appears to be significant enough to mask other variables with smaller effect sizes.

The main limitations of this study are the relatively small size of the sample which might not allow for the identification of risk factors with smaller effect sizes and the inclusion of patients only from one hospital’s catchment area, which might affect how representative the study is. However, the source hospital is by far the largest inpatient unit in the country, serving the metropolitan area of a capital city of around three million inhabitants. Potential changes over time in the baseline variables (e.g., losing treatment adherence between hospitalizations) were not accounted for and it was not possible to access information regarding quality of life outside the hospital, social support and access to aftercare services, all of which would have been relevant.