A retrospective analysis of determinants of involuntary psychiatric in-patient treatment

Cologne is the fourth largest city in Germany with a population of about one million inhabitants (exact population in the year 2011: 1,005,777 inhabitants). It is located in the densely populated and economically strong region of the Rhineland. Apart from a small area in in the northern part of the city, in-patient psychiatric care is provided by four hospitals and it is organized on a sectoral basis: each of the four hospitals provides care to the inhabitants of specific parts of the city ranging from approximately 501,000 to approximately 109,000 inhabitants depending on the size of the hospital. Apart from these four hospitals, there are only few small private psychiatric and psychosomatic units in and near Cologne. These are not responsible for emergency care, have no acute, closed wards and do not treat patients under the Mental Health Act. Out-patient care is mainly provided by a large number of psychiatrists and psychotherapists who work in practices spread across the city. In addition, out-patient units of the psychiatric hospitals provide complex multidisciplinary care to the most severely affected patient groups. In general, the services of the City of Cologne are organized in 9 distinct districts. Community-based social psychiatric services and centres exist in every district. They are coordinated by the Public Health Department and they all work in a similar way with similar resources.

A single Municipal Court is the deciding authority for all detentions in the four psychiatric hospitals of Cologne. The Mental Health Act of the federal state of North Rhine Westphalia (PsychKG NRW) is applicable to individuals who are mentally ill and therefore pose an immediate, severe threat to themselves or others. It is required that there is no other way of averting harm. In most cases, a preliminary detention for up to one day and night is initiated on the basis of a short report of a psychiatrist or a physician with sufficient experience in the field of psychiatry (this includes emergency physicians and staff of emergency units in general hospitals). After this, a court hearing will be held in the hospital on the same day or at the latest on the next day after admission. The court provides a legal representative for the patient. The court will either confirm detention for up to six weeks or it will lift detention immediately. There is no out-patient civil commitment in Germany.

We conducted a retrospective analysis of 5764 health records of inpatients in the four psychiatric hospitals of Cologne (Table 1). We included all patients who were treated under the PsychKG NRW in these hospitals in 2011 (n = 1773; including patients who were primarily involuntarily admitted and patients whose status changed from voluntary to involuntary during their in-patient stay in a psychiatric hospital). For feasibility reasons, we were not able to include all voluntary cases of the same year in our analysis (n = 8398). We included a large random sample of almost half of the entire voluntary population (n = 3991). The random samples of voluntary patients from the three largest hospitals were drawn by using every n^th patient based on lists of consecutive voluntary admissions (hospitals 1, 2 and 3, see Table 1). This procedure avoids potential sampling errors due to dates of admission. Regarding the smallest hospital with only 600 cases in 2011, we included all voluntary cases in the analysis.

Five trained assistant physicians extracted the clinical and sociodemographic data from the health records and administrative patient files. These assessors were trained to identify the data of interest in the hospital records and to enter them into a standardized form. We employed standardized training record sets in order to ensure a high degree of care in identifying the data of interest. Diagnoses were routinely assigned by the attending physicians according to the WHO ICD-10 classification. The assessors extracted both the main diagnosis (cause for admissions) and the psychiatric comorbidities (secondary diagnoses). A detailed description of the data set is given in Table 2.

We used χ²-tests for the analysis of categorical data (most demographic data, diagnoses and most data on other clinical features and treatment). Due to our sampling procedure, involuntary cases were overrepresented in our samples from hospitals 1–3 (see above, Study design and data sources, and Table 1). Therefore, we weighted the samples from hospitals 1–3 according to the true proportion of voluntary cases in the samples of these hospitals. As a result, the weighted sample matches the true distribution of voluntary and involuntary cases in the four participating hospitals. The analysis of categorical data was performed on the weighted sample (n = 10,171). Diagnoses were analysed both on the level of main and secondary diagnoses. In addition, we included the comorbidities addiction and psychosis (F1 and F2) and addiction and personality disorder (F1 and F6) in our analyses, because these were previously shown to be associated with increased risks for aggression, suicide and self-harm, involuntary admissions and containment measures [30‐32]. Metric data (age and length of in-patient stay) were analysed by means of ANOVA and t-tests on the unweighted sample (n = 5764). The level of significance was set at p ≤ 0.01. This conservative significance level was chosen in order to limit the number of potential false positives that could have occured in our analysis otherwise.

As a further means to identify predictor variables for involuntary hospital admissions, we used a statistical modelling approach employing the CHAID algorithm (Chi-Square Automatic Interaction Detector [29, 33]. CHAID is a decision tree-generating algorithm for analysing potentially interacting mixed categorical and continuous data providing a classification of predictive factors of an outcome of interest. Given the multitude of both potentially interacting categorical and continuous potential predictor variables in our study, CHAID was chosen because it provides a data-mining approach for a large dataset as in our study. CHAID is based on the null hypothesis of independence of the predictors and the outcome variable. It performs multiple stepwise χ² tests and provides statistically homogenous subgroups, which are used to generate a summary diagram (“decision tree”) of the relative importance of predictive factors. The decision trees represent both the hierarchy and the interactions of predictive factors for the outcome of interest. We preferred CHAID over multiple binomial logistic regression since decision tree analyses provide a hierarchical risk classification structure which immediately identifies the most promising areas of future preventive interventions. It has been argued that preventive tools based on main effects regression approaches do not adequately reflect the contingent nature of clinical risk assessment processes [34]. CHAID has been used previously in mental healthcare research for example to identify predictors of successful outcomes of methadone treatment [35] and vocational rehabilitation for patients with affective disorders [36]. In our study we carried out an Exhaustive CHAID on the weighted sample (n = 10,171). All clinical, treatment-related and sociodemographic items available for both voluntary and involuntary patients (see Table 2, Sections B-F) were included as possible predictor variables in relation to the dichotomous outcome “legal status of hospital treatment” (voluntary vs. involuntary). The model was restricted according to the following stopping criteria: minimum node size to be split n = 100, minimum leaf size to be created from a node n = 50, or a maximal depth of three. Missing values were treated as valid values, i.e. all cases were used for the computations of both the training and the test model. The model was built with a split-sample validation in a training sample (70%) and then tested with a test sample (30%). The level of significance was set at p ≤ 0.05. The results reported here refer to the test sample.

All statistical analyses were carried out using IBM SPSS Statistics Version 25.

Organic mental disorders (ICD10: F0) and psychotic disorders (ICD 10: F2) were overrepresented in the group of involuntary patients (F0: 18.6% vs. 3.9%, F2: 31.4% vs. 17.7%), whereas substance use disorders (ICD 10: F1) and affective disorders (ICD 10: F3) were more common among the group of voluntarily treated patients (F1: 37.9% vs. 23.2%, F3: 30.8% vs. 16.1%). These findings were robust when either the main diagnosis alone or the main and secondary diagnoses were taken into account for the analysis (Table 3). The comorbidity of addiction and psychosis (F1 and F2) was overrepresented among involuntary patients (12.0% vs. 8.4%). Prior to the admission, involuntary patients were more likely to be in touch with community psychosocial services (2.4% vs. 1.0%) and to be under legal guardianship (28.7% vs. 15.3%), and they were less likely to having received psychiatric treatment in an outpatient unit (31.1% vs. 33.1%). Involuntary patients were admitted for hospital treatment more often outside of regular service hours (62.9% vs. 38.9%). Finally, involuntary patients were admitted more often due to suicidal behaviour and self-harm (37.9% vs. 20.9%) and reported more often a history of suicidal attempts compared to voluntary patients (30.1% vs. 25.0%). There was an appreciable percentage of missing values for the items “previous suicide attempts” and “legal guardianship prior to admission” (18.1 and 15%, resp.), otherwise data quality was satisfactory (Table 3).

Patients treated under the Mental Health Act were older than voluntarily treated patients (48.0 vs. 44.0 years). They were more often retired (37.2% vs. 23.8%) including early retirement due to illness. They more often presented with a migratory background (31.0% vs. 27.7%). They lived more often in supervised accommodation (16.4% vs. 12.2%) and less often on their own (37.4% vs. 42.3%) (Table 3). Patients treated under the Mental Health Act differed from the voluntarily treated patients in terms of their professional training: they had received more often higher levels of education such as a secondary apprenticeship (9.3% vs. 6.1%) or a university degree (12.5% vs. 11.0%). However, they were less often in active employment (16.3% vs. 21.2%) and their main income source was more often a pension (39.0% vs. 25.4%). Despite efforts to extract all necessary data from the available medical records, the percentage of missing values on some sociodemographic data was high (e.g., school education 31.1%, professional education 20.2%, degree of employment 16.4% and main income source 16.9%) (Table 3).

The Exhaustive CHAID analysis yielded 16 statistically significant subgroups (terminal nodes). In the training sample, 93.1% of the involuntary group was predicted correctely, compared with 92.4% in the test sample. For both datasets, very similar areas under the curve (AUC) were found: Training AUC = 0.667; Test AUC = 0.665. These similarly correct classification rates indicate that the fitting process was performed adequately and satisfactorily, i.e. there is neither evidence for underfitting nor for overfitting. The main diagnosis was identified as the strongest predictor for involuntary treatment (χ²(3) = 573.603, corrected p < 0.001, first level of segmentation of the sample, see Fig. 1). Patients with organic mental disorders (ICD 10: F0) had the highest rate of involuntary admissions (51.7%); patients with psychotic disorders (F2) and patients with intellectual disabilities (F7) had the second highest rate (31.9%); patients with neurotic, stress-related and somatoform disorders (F4) and patients with personality disorders (F6) had a rate of involuntary admissions (19.0%) that was only slightly lower than the average rate of the entire study sample (19.7%). The group of patients with substance use disorders (F1) and affective disorders (F3) as well as the group of patients with behavioural and emotional disorders with onset in childhood or adolescence (F9) had the lowest rate of involuntary admissions (12.4%).

For the group of patients with organic mental disorders (ICD 10: F0), who were at the highest risk for involuntary treatment, admission outside of regular work hours (χ²(1) = 23.215, corrected p < 0.001), previous suicide attempts (χ²(1) = 21.577, corrected p < 0.001), and marital status (χ²(1) = 18.097, corrected p = 0.001), were identified as further critical variables (Fig. 2). Patients with organic mental disorders were more often subject to involuntary treatment if they were admitted outside regular working hours particularly if they were married or widowed (74.3% involuntary treatment in this highest risk group).

For the diagnostic groups with the second highest quota of involuntary treatment (ICD 10: F2 and F7, Fig. 3), lack of previous outpatient treatment (χ²(1) = 47.745 corrected p < 0.001), migratory background (χ²(1) = 11.787, corrected p = 0.002) and admission outside regular working hours (χ²(1) = 15.473, corredted p < 0.001) were found to be critical. Among patients who had had no regular outpatient psychiatric care, those with a migratory background were at highest risk (46.8% involuntary treatment). Patients with previous outpatient psychiatric care were more often admitted involuntarily (36.2% vs. 20.5%) if they attended an emergency service outside of regular working hours.

For the diagnostic groups with a quota of involuntary treatment around or below average (ICD 10: F4, F6, F1, F3, F9), suicidal and self-harm behaviour at admission were identified as most critical (F4, F6: χ²(1) = 33.221, corrected p < 0.001, Fig. 4, and F1, F3, F9: χ²(1) = 273.390, corrected p < 0.001, Fig. 5). In the next level of the CHAID decision tree, critical variables were admission out of regular working hours (χ²(1) = 84.027, corrected p < 0.001), the hospital in which the treatment took place (F4, F6: χ²(1) = 32.401, corrected p < 0.001; F1, F3, F9: χ²(1) = 54.588, corrected p < 0.001) and the patients’ level of professional training (χ²(1) = 17.053, corrected p < 0.001). When suicidal patients were admitted to either hospital 1 or 4, they were more likely to be treated under the German Mental Health Act (F4, F6: 54.5%; F1, F3, F9: 34.0%) compared to those admitted to hospitals 2 and 3 (F4, F6: 9.4%; F1, F3, F9: 10.6%). In search for an explanation for this finding, we carried out exploratory analyses and found that in the case of the two hospitals with higher quotas of involuntary treatment in suicidal cases, a higher proportion of Mental Health Act cases had been referrals from general hospitals (41.7 and 57.6% of the PsychKG cases in hospitals 1 and 4, compared to only 26.2 and 30.5% in hospitals 2 and 3 respectively). Non-suicidal patients with affective, substance related and behavioural and emotional disorders (F1, F3, F9) were more often admitted involuntarily (16.0% vs. 4.3%) if they attended an emergency service outside of regular working hours. Non-suicidal patients with neurotic and personality disorders (F4, F6) were at higher risk for treatment under the German Mental Health Act if they had no professional education (10.5% vs. 5.8%).

A major strength of our study is the detailed in-depth analysis of health records of a large sample of psychiatric in-patients representative for a complete metropolitan region of Germany. In addition to administrative and routinely available data, we screened all available data sources for a detailed list of sociodemographic and clinical items previously shown to be associated with detentions. We included all cases treated under the Mental Health Act in the City of Cologne over the course of one year (except from a defined small part of the northern city) and we also included a large sample of voluntary patients from the same hospitals and year who served as controls. We obtained data from all four psychiatric hospitals which serve this region and admit patients under the Mental Health Act, thus avoiding sampling bias and ensuring the generalizability of our findings. The City of Cologne comprises a wide range of sociodemographic neighbourhoods. Community-based social psychiatric services are organized in similar ways all over the region and a single Municipal Court is responsible for all detentions in the city. Hence, variation due to systemic factors such as differing jurisdictions and/or major differences in out-patient services was minimized.

In addition, our statistical analysis went beyond previous studies through the application of the CHAID decision tree analysis. For the purpose of our research questions, CHAID was superior to logistic regression [43] as it allowed us to identify interactions between different risk factors for detention and develop a classification of risks leading to tailored suggestions for preventive interventions. By using a CHAID algorithm with a maximum depth of three levels we developed a model which is easily understood by clinicians. The resulting model appears to be adequately accurate with no overfitting. Further improvements of the model fit may be accomplished by ensemble methods such as bagging, boosting or stacking, which we intend to use in a future project employing machine learning algorithms.

The shortcomings of our study are pertinent to its retrospective design. We collected clinical and administrative data from existing medical records and these data were incomplete for research purposes. Hence, we have no information on potentially important features that may drive involuntary hospitalization, such as symptom severity, level of psychosocial functioning, insight or perceived social support. Lack of this information prevents the development of a comprehensive risk model for involuntary psychiatric hospitalizations. Moreover, the retrospective nature of the study does not allow for making causal interpretations, as the observed relations may be influenced by unobserved third variables. In addition, although we searched all available data sources in-depth, the number of missing values was considerable for some sociodemographic variables. This may have reduced the power of our study to detect possible links between sociodemographic factors and the legal status of in-patient treatment. Finally, although there are many similarities between our findings and results from another recent study from Germany [11, 12] as well as several studies from metropolitan regions of other European countries [7, 9, 13, 16, 26, 44], we cannot be sure how far the generalizability of our findings goes. Differences between health systems across different countries and differences in availability and quality of health services depending on the degree of urbanicity may well lead to increased or decreased detention rates. Hence, results may be very different in other countries and in rural regions.

Patients with organic mental disorders had the highest likelihood of involuntary treatment. Support measures, e.g. specific training of relatives and professionals providing home care, may help manage crises in at-home situations and avoid hospital admissions, in particular at night or during weekends [45]. Measures for the prevention and management of mental crises in persons with intellectual disabilities should be established in a similar way [46].

There is also a great need for action in improving mental healthcare for migrants. Local networks of mental health care providers need to become fully accessible and offer linguistically and culturally appropriate services. Measures to be implemented may include training of staff in cultural sensitivity and the use of interpreters in diagnostic and treatment procedures [47].

We cannot be sure about the nature of the association between the time of admission in a psychiatric hospital and the legal status. It does seem plausible, however, that emergency hospital services with lower staffing levels may increase the risk of involuntary admission. In this case, community-based, outreach crisis intervention services may lower the need for hospital attendance and this may lower involuntary admission rates [48, 49]. Finally, it may be warranted to improve deescalation skills in general hospital emergency departments and other emergency services. Deescalation training of emergency professionals may be crucial in reducing involuntary psychiatric in-patient admissions. These factors appear to be not specific to the situation in Cologne or Germany.

In summary, a variety of interventions may help to improve health services and reduce involuntary treatment in different risk groups for involuntary admissions by addressing modifiable risk factors as identified in our analysis. Currently, only limited evidence is available for the efficacy of preventive measures targeted at certain risk groups, pointing to a need to develop, implement and evaluate such programs [48].