For this observational study, we used administrative, sociodemographic, and clinical data from the electronic health record database of the Paris Psychiatry Hospital Group, which serves a population of approximately 1.6 million people and covers three-quarters of the city of Paris. Our source database contained 700,000 patient electronic health records. We leveraged structured and unstructured data (e.g., discharge summaries or medical observation reports). We selected patient stays corresponding to a full-time psychiatric hospitalization for patients ≥15 years old, with a discharge in 2022 (n=9,697 admissions). As detailed in Figure 1 , we excluded short hospitalizations (≤2 days) and iterative hospitalizations when they were dedicated to specific therapies such as ECT in case of treatment-resistant depression or long-acting neuroleptic injection for patients with psychoses. This study was approved by the Paris Psychiatry Hospital Group Research Ethics Committee (accreditation no. 2024-CER-A-007).

We constructed a database of 8,870 full-time hospital stays (6,216 patients) ( Figure 1 ). The median LOS was 14 days (IQR, 5–31; mean 27 ± 47.8 days) (see distribution plot in Figure 2 ). The median age of patients was 38 (IQR, 27–53; mean age 41 ± 16 years); 44.9% of long stays were for female patients ( Table 2 ). Overall, 25.5% of the stays were in a psychiatric ward that was different from the patient’s psychiatric sector of origin, 10.0% were during a travel-related psychiatric episode, and 68.9% were after a visit to a psychiatric emergency unit. We found that 25.1% of the admissions were related to mandatory care at the request of a third party, and 18.7% to imminent danger. In all, 9.1% of stays were associated with another type of mandatory care (by decision of the state representative, of detained persons, after a decision of criminal irresponsibility, or within the framework of a temporary placement order). The two most common diagnoses were schizophrenia and mood disorders, accounting for 52.4% and 40.3% of the 8,870 hospitalizations, respectively. A total of 20.9% of stays featured suicidal risk and 53.5% featured somatic comorbidity.

Factors that increased the odds of a long LOS were age >55 years, admission from outside the patient’s sectorization zone, admission via a psychiatric emergency unit, mandatory care, seclusion and physical restraint, diagnosis of schizophrenia and psychosis, treatment resistance, and somatic comorbidity for female sex. Factors that reduced the odds of a long LOS were being in a relationship, admission during a travel-related psychiatric episode, and a diagnosis of a personality disorder.

Our results agree with the French and international literature. Among patient characteristics previously validated as risk factors, we first identified age >55 years. We found a positive association between age and a long LOS in previous studies (13, 23, 24). The positive association of outcome with a diagnosis of schizophrenia has been widely confirmed (3, 9, 13, 14, 19, 23, 25–28). We also highlighted features related to clinical severity, reported in previous studies: treatment resistance (29, 30), mandatory care (3), and seclusion and mechanical restraint (3) (25). We noted some previous contradictory findings (31) on the effect of coercive treatment on LOS (alternatively positive, negative, or null, depending on the study). However, these results were obtained in different foreign countries with their own mandatory care regulations, but our conclusions were consistent with a recent French study (3). We found another significant result for the severity of seclusion and mechanical restraint: a multiplication factor of 3 for the risk of LOS between the first and second variable level, depending on the duration of the measure. These findings suggest that the use of mandatory care and seclusion and mechanical restraint may correspond to severe and difficult-to-treat symptoms, materialized by patient aggressiveness, dangerousness, a lack of illness awareness, social withdrawal, functional impairment, and a low adherence to care. This observation could explain the need for complex treatment, difficulties in discharge, and thus a long stay (3, 9, 32). The literature suggests that particular attention should be paid to the severity of illness (13).

Another notable finding of our study was the positive association of somatic comorbidity with a long LOS for only women. Somatic illness has been described in the literature as a potential “focus of care” (33) and a risk factor for long hospital stays (13, 34, 35). However, we could not find other examples of this interaction between sex and somatic comorbidity in the literature. To interpret our result, because women are more likely to report physical symptoms than men (36), somatic symptoms may be easier to treat for women within a psychiatry unit, in contrast to men, who have to be transferred to the general medical ward and are therefore discharged from psychiatry care earlier.

Multivariate analysis of imputed data revealed a diagnosis of organic mental disorders as a risk factor for a long LOS. Considering the specific low prevalence of positive cases for this feature, with more statistical power, this characteristic might have initially appeared among the significant predictors. This result was also confirmed with sensitivity analysis with cutoffs of 20, 60, and 90 days. In addition, homelessness became positively and significantly associated with the highest cutoff of LOS (90 days). This finding agreed with previous research reporting the impact of homelessness on LOS (21, 37).

We found a protective effect of non-single marital status on a long LOS, in line with the literature describing it as a proxy “for functional impairment” (13) and a factor “reflecting resilience” (18). This characteristic is related to human relationships and family connection providing moral support to the patient and is consistent with studies showing that family visits reduce the LOS (38). We found no association between the diagnosis of mood disorders and a long stay in our multivariate analysis. Yet, significance would have been in accordance with previous French studies (3) but not the international literature (8, 29, 37). A diagnosis of a personality disorder was a negative predictor, reported in previous research (21, 28, 39). This observation was consistent with international guidelines (40), recommending limited extended hospitalization in this case. Moreover, it was also consistent with the fact that personality disorders are more often a comorbidity rather than a primary cause of full-time hospitalization (for 2019–2022, the unpublished GHU Paris Psychiatry Hospital annual report highlights that 58% of personality disorder diagnoses related to full-time hospitalization in the Paris Psychiatry Hospital Group were “associated” and not “primary” diagnoses).

An interesting result of our exploration was the variables related to the patient pathway. We highlighted two new significant features. The first was admission after a visit via a psychiatry emergency unit: its positive effect on a long LOS could be interpreted as a marker of the clinical severity of the disease and need for more complex care. The second, the protective effect of a long LOS during a travel-related psychiatric episode, was in line with the profile of these patients (41). Such patients are mostly traveling from France (Paris suburbs or other regions) (41). They are out of their sectorization zone, psychotic, and characterized by a delusional state or disorganization of thought. Their vulnerability and lack of control justifies rapid hospitalization to protect themselves and others, but as soon as possible, health professionals organize their return to their original psychiatric sector to ensure better follow-up. In addition, we found admission from outside the sectorization zone positively associated with a long LOS. Sensitivity analysis with a lower cutoff point for LOS (20 days) confirmed this result. This finding was consistent with those from previous studies in the UK (28, 42). Out-of-area patients may be mostly new and unknown patients for psychiatric wards and thus require a long stay. In addition, out-of-area placements might penalize continuity of care and thus increase the LOS (28).

Our current results suggest that these characteristics related to the patient and the journey should be systematically assessed by healthcare professionals and taken into account during clinical decision-making. Knowledge of such determinants of LOS could help clinicians identify and focus on patients with high clinical severity and a risk of a prolonged stay. We suggest that long-stay patients require a more complex behavioral management and thus a higher level of staff in dedicated units, for specialized acute care. The aim would be to provide them tailored care with a multidisciplinary team, to optimize treatment, discharge planning, and use intensive case management for patients with disruptive behaviors (43, 44). Our results also suggested the benefits of early treatment resistance detection (45) and optimization of somatic comorbidity care. Moreover, the protective effect of non-single marital status on a long LOS emphasized the importance for health care professionals to favor any opportunity for patients to have constructive interactions with families, and also with the social environment. This confirmed the benefits of an approach integrating mental health and social care, through rehabilitation programs, assertive community treatment, and care provided in community-based settings (43).

Our findings should also be contextualized in regard to the variety of international results on the same subject. The mean LOS in this study (27 days) approximates the mean LOS highlighted in the international literature for the USA (24.9 days) (13) and Quebec (25.3 days) (43). It is above the mean LOS found for Italy (17.9 days) (10) and India (5.7 days) (46), and below the mean LOS found for other European countries (10) such as Germany (37 days), Poland (33.4 days), the United Kingdom (46.2 days), and Belgium (55.1 days). Previous studies conducted for Germany, Poland, the UK, Italy, and Belgium (10) highlighted features such as poor social functioning, psychotic disorder, and high clinical severity (in line with our findings) that predicted LOS regardless of the country. However, this study argued that the specific effect of variables might be different from one country to another, and that national context characteristics (cultural and organizational) might influence LOS and its determinants. A previous study comparing the USA and Germany (47) suggested that a greater number of social workers and private psychiatrists per inhabitant in the USA might ease a faster connection to after care, and thus partly explain a lower mean LOS in the USA. We suggest that French sectorization might also influence our LOS in the same way: our system provides specific pathways such as alternatives to full-time hospitalization (23) and rehabilitation programs (improving continuity of care, patient autonomy, and reintegration into the community) that might lower our mean LOS, compared with other European countries. A perspective of future studies would be to extend that kind of comparative study (10) to a greater number of countries, including France, with a rigorous approach that takes national healthcare system characteristics into account.

Although many studies have explored factors relating to LOS, there are few up-to-date studies conducted in France. We confirmed and broadened results from French studies related to LOS (3, 15, 23, 25). This study had several methodological assets. Our sample size (n=8,870) aligned with literature recommendations (at least 3,000 observations) (13), which ensured that our study was not underpowered, taking the number of covariables we used for modeling into account (48). We used data from the Paris Psychiatry Hospital Group electronic health records. This healthcare institution, consisting of three different hospitals, provides mental healthcare for much of the city of Paris, offering a representative target population of approximately 1.6 million people. Therefore, our sample was characterized by different types of patients with very different sociodemographic features. It was a representative sample of the Paris population and allowed us to minimize selection bias. To our knowledge, this study is the first to compute true LOS with a specific algorithm. With our computation method, we avoided a bias while evaluating LOS: we took into account the potential splitting of the healthcare period with a spurious administrative discharge, normally supposed to be a temporary absence. We performed a comprehensive analysis by including several categories of predictors (demographic, clinical, and related to care pathway) in our final model, selected after several rigorous statistical steps. As far as we know, this research was among the first in France to analyze the effect of factors related to the patient’s pathway, which are important understudied characteristics. We compensated for the lack of structured data for somatic comorbidity and suicide risk by implementing natural language processing methods on raw medical narratives and leveraged comprehensive lexical dictionaries assessed by different raters. We checked that ignoring sector effects did not bias our results significantly (see Supplementary Material ). Our sensitivity multivariate analysis results were consistent in terms of age, psychotic diagnosis, and clinical severity factors, whatever the cutoff used to define a long LOS. A bivariate comparison of included versus excluded observations revealed the latter observations were significantly associated with older male patients with a long LOS. This finding suggested that removing observations with missing values (necessary before logistic regression implementation) could lead to a potential bias selection, resulting in a minimization of the effect size of predictors in logistic regression modeling. However, we ran consistent sensitivity multivariate analyses on imputed data (n=8,870) and the dataset excluding variable marital status, accountable for most of the missing information (n=8,522). The missing data exclusion did not significantly bias our main findings in multivariate models. These methodological steps strengthened the robustness of our final conclusions.

Our results should be interpreted in light of several methodological limitations. The first are those related to observational and retrospective studies (49, 50). We could not assess any causal relationship between explanatory variables and long-stay outcome. Our results were at the association level only. Future research on a national level will be needed to ensure better generalizability and address these questions about sector-level LOS determinants more accurately. For instance, it would be relevant to take hospital and ward policies about coercive measures into account, specifically when we investigate the impact of seclusion and restraint on length of stay. There is important literature about programs for reducing seclusion and restraint (51, 52). De-escalation techniques for psychosis-induced aggression or agitation (53), crisis prevention plans (54), tools like advanced directives in psychiatry (55), the use of “sensory rooms” (55), and a focus on a secure therapeutic relationship (56, 57) can help limit tension and the frequency of conflict. It would be interesting in a future study to investigate how these hospital-level and ward-level strategies might contribute to reducing length of stay.

Another limitation was the lack of data availability for some patient characteristics. We were not able to include, among explanatory variables, unemployment, a “proxy for current functional impairment” (21, 58), dependency (3, 25), symptom measurements, which could be used to assess clinical severity more accurately (9, 18, 28, 31), and biochemical markers obtained at the beginning of hospitalization (18). Nor could we evaluate the effect of the treatments received. To deal with polypharmacy impact, suggested strategies might be to stratify patients starting a treatment at the point of inclusion in the study and those already receiving treatment (59), grouping patients depending on the kind of treatment they receive (18). For future research, the medical record could include a traceability of precise reasons preventing patient discharge, whether clinical or organizational (e.g., no place available in a medico-social institution, a lack of rehabilitative programs at discharge, and no possibility of rehousing) (3). Future studies need to include all these factors. Our exploration was a proof of concept for other studies based on extended and national-level data.