Article Text
PDF
Abstract
Objectives The overall objective was to analyse service-related factors involved in the complex processes that precede suicide in order to identify potential targets for intervention.
Design and setting Explorative network analysis study of post-suicide root cause analysis data from Swedish primary and secondary healthcare.
Participants 217 suicide cases reported to the Swedish national root cause analysis database between 2012 and 2017.
Primary and secondary outcome measures A total of 961 reported incidents were included. Demographic data and frequencies of reported deficiencies were registered. Topology, centrality indices and communities were explored for three networks. All networks have been tested for robustness and accuracy.
Results Lack of follow-up, evaluations and insufficient documentation issues emerged as central in the network of major themes, as did the contributing factors representing organisational problems, failing procedures and miscommunication. When analysing the subthemes of deficiencies more closely, disrupted treatments and staffing issues emerged as prominent features. The network covering the subthemes of contributing factors also highlighted discontinuity, fragile work structures, inadequate routines, and lack of resources and relevant competence as potential triggers. However, as the correlation stability coefficients for this network were low, the results need further investigation. Four communities were detected covering nodes for follow-up, evaluation, cooperation, and procedures; communication, documentation and organisation; assessments of suicide risk and psychiatric status; and staffing, missed appointments and declined treatment.
Conclusion The results of this study suggest that healthcare providers may improve patient safety in suicide preventive pathways by taking active measures to provide regular follow-ups to patients with elevated suicide risk. In some cases, declined or cancelled appointments could be a warning sign. Tentative results show organisational instability, in terms of work structure, resources and staffing, as a potential target for intervention, although this must be more extensively explored in the future.
- Suicide & self-harm
- PSYCHIATRY
- Risk management
- Quality in health care
- Health & safety
- Organisational development
Data availability statement
Data are available in a public, open access repository. Copies of used R scripts and data sets are available at Open Science Framework: DOI 10.17605/OSF.IO/3DHTF.
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
- Suicide & self-harm
- PSYCHIATRY
- Risk management
- Quality in health care
- Health & safety
- Organisational development
Strengths and limitations of this study
The data source was based on standardised reports performed by trained healthcare teams.
The data were examined and categorised by four professionals, experienced in performing and peer-reviewing root cause analysis reports.
In addition to analysis of reported frequencies of adverse events, network analysis was applied.
Each network has been tested for robustness and accuracy.
The main limitation is that a relatively small proportion of all suicides was submitted to the national database.
Background
Suicide is one of the leading causes of death worldwide, affecting people of all ages, socioeconomic groups, and cultures. More than 700 000 suicides (1.3% of all deaths) occur globally each year, which exceeds the deaths due to malaria, HIV/AIDS, breast cancer, war and homicide.1 For every completed suicide, there are indications of more than 20 other attempts.2 Considering this and that the rate is markedly higher in people with psychiatric illness,3–5 preventing suicide is a general priority in mental healthcare. Due to the complex and heterogenous nature of suicidal behaviour, fluctuating levels of suicide intent and the lack of reliable assessment tools, suicide preventive decision making is difficult.6–8 Clinical actions depend not only on the competence of individual healthcare professionals, but also on patient safety management on a structural level.9
Postsuicide reviews commonly use root cause analysis (RCA) to identify service-related risks.10 11 In Sweden, RCA has been a widespread method for investigating adverse events in healthcare for more than 15 years. The analyses are performed according to a standardised protocol by trained teams.12 The RCA procedure has been exhaustively described elsewhere,11 13 14 and the workflow of the Swedish RCA teams is identical to the steps listed there.12 A short summary of the history of incident reporting and patient safety legislation in Sweden can be found in Fröding et al.15
Network analysis is an approach to statistically analyse and visualise core elements of a data set. Application spans from mathematics and physics to social sciences and psychology. The method is useful for modelling complex patterns among correlated variables.16 17 Over the last decade, a wide array of studies within the field of personality, psychopathology, and comorbidity has taken place.16 18–37
Previous research
Previous healthcare research on suicide prevention has focused mainly on single risk factors.38 39 Besides highlighting the importance of providing treatment to underlying illness, it stresses the reduction of accessibility to lethal means,40–45 combining immediate and long-term multilevel interventions,42 43 46 47 building trustful staff–patient relationships and involving relatives,48–50 conducting regular assessments in outpatient settings,49 51 52 designing safer environments for inpatients,53–56 following up earlier and maintaining closer supervision in the post-discharge period.57–66 To reduce organisational risk factors, better communication among professionals, proper education and provision of adequate guidelines have been suggested.15 49 52
Previous studies based exclusively on postsuicide RCA material, including systematic reviews, meta-analyses and observational studies from inpatient45 67 and outpatient settings,62 67 Veterans Health administration facilities68–72 and nursing homes67 68 report inadequacies in cooperation,62 68–70 72 accessibility to care,45 69 assessments of suicidal risk68–72 and follow-up67 as the main deficiencies in suicide prevention.
Network analysis used in suicidal behaviour modelling suggests an association between suboptimal treatment of psychiatric illness and increased levels of suicidal ideation.73–80 Feelings of thwarted belonginess,74 75 entrapment, hopelessness and perceived burdensomeness75 77 79 81 are also core phenomena of self-harm and suicidality. Physical illness, trauma, harassment and acute life stress due to economic or relational circumstances are examples of external individual factors associated with suicidal ideation.76 81 82 Personalising treatment strategies, for instance, by using electronic devices for repeated ecological momentary assessment, has been suggested as an application of these findings.73 75 83
Network studies on service-related risk factors for suicides among persons in contact with health services are lacking. Therefore, this study aims to explore relationships of common deficiencies in healthcare preceding suicide, identify potential targets for clinical intervention and generate hypotheses for future research.
Methods
This study followed the guidelines of the Strengthening the Reporting of Observational Studies in Epidemiology checklist for reporting cross sectional studies (online supplemental file A).
Supplemental material
Material
The analysed material consists of 217 RCA reports concerning patient suicides uploaded to the Swedish national database for RCA (Nationellt IT-stöd för HändelseAnalyser - NITHA) from 2012 to 2017. The search criteria were: ‘Type of consequence: suicide/suicide attempt’; ‘death: yes’.84 Information in NITHA is anonymised, so we could not link any information to actual patient records. The reports were produced by RCA teams from 12 of Sweden’s 21 regions. The teams consisted of 3–4 investigators trained in RCA methodology. who were responsible for data collection, identifying deficiencies, listing possible contributing factors and proposing and evaluating adequate actions to avoid future recurrences. The data were collected from all data sources available to the team at the time of the investigation, including medical records, information from booking systems, data from external service settings, and qualitative data, such as interviews with healthcare professionals and interviews with relatives (64%, n=139). The final reports varied in terms of scope and content. In some cases, particular facts about the medical condition or specific circumstances were omitted. Although we do not know the exact background to this, it may have been done to protect the integrity of those deceased. As we only had access to the final RCA reports, we have not been able to scrutinise how the RCA teams processed the original raw data.
Suicide reports in the NITHA database
Table 1 shows the distribution of demographic characteristics among the included 217 patients, as stated by the RCA teams. Approximately half were between the ages of 18 and 49, and half were in contact with psychiatric services at time of death. Men were slightly over-represented. For two persons, gender was reported as ‘other’. Mood disorders were recorded as the most common type of primary diagnosis for both sexes.
Patient demographics as reported to NITHA
Data extraction and processing
A data coding tool was developed to organise the data into inductively constructed categories. The original protocol was tested by two teams (CBC, MD and EvH, MR), and refined until consistent themes and subthemes had been identified. The team members had different professional backgrounds (two psychiatrists, one psychiatric nurse and one psychologist) and were experienced in performing and peer-reviewing RCA reports at their own clinic. The teams worked independently but had regular meetings to discuss the data coding tool, which was audited by external reviewers and revised several times to cover all areas of interest in the RCAs. Every modification prompted a second review of previously reviewed cases. The final version of the data coding tool was used to derive data from all 217 cases. Data were double-checked for discrepancies; none were found in the final version of the dataset.
The two teams reviewed and coded all 217 RCA reports. The extracted raw data underwent a keyword-based sorting of text strings in Microsoft Excel 2016 before classification, resulting in 499 registered deficiencies and 462 underlying, contributing factors. In the original RCA terminology, deficiencies are termed adverse events, and contributing factors root causes. Examples of typical cases are reported for each category in table 2. While some minor misclassifications in the original data were noted by the research teams, terminology used in the original NITHA reports (including definition of missing data) was retained.
Frequencies and percentages of reported variables
In the original RCA reports each item could be reported multiple times. The range for some items varied from 1 to 6, depending on whether the RCA team had registered deficiencies in a merged or split form. To avoid skewed results, all observations were binarised (using the simple algorithm ‘IF count value ≥0, THEN 1, ELSE 0’) before being entered into the network model.
Data analysis
The synthesised network model contains two elements: nodes (sometimes called vertices), representing variables and edges (also called links) which represent pairwise association among nodes.16 81 The network can be either directed, displaying the influential effect from one node to another, or undirected, where mutual influences are indicated by a line between two nodes without any direction.16 Centrality indices, such as strength, betweenness and expected influence (EI), are employed to evaluate the network.16 85 An overview of different types of networks and applicable models has been published by Hevey, 2018.23 For a further discussion on psychometrics and network estimation, we refer to previous researchers in this field.31 86–90
Three networks were produced: one giving an overview of the major themes (figures 1 and 2), another showing subthemes of deficiencies (figures 3 and 4) and a third covering subthemes of contributing factors (online supplemental file B). Frequencies and percentages are reported for each variable (table 2), alongside the centrality indices and stability measures (figures 1–4, table 3). Data were analysed using IBM Statistical Package for the Social Sciences V.25 and R V.3.5.0 (bootnet package V.1.4.3, ggplot2 V.3.3.5, igraph package V.1.2.6, qgraph package V.1.6.9, IsingFit package V. 0.3.1).91–97 To visualise the dependencies, we used an undirected network (formally called a pairwise Markov random field).28 87 98 Relevant relationships among nodes were estimated using IsingFit package which uses an enhanced least absolute shrinkage and selection operator, based on the Ising Model. The operator reduces spurious edges by suppressing minimal connections to exactly zero. Selection is performed by combining logistic regression (l1-regularised) and a model selection based on the Extended Bayesian Information Criterion (EBIC).96 Since the network structures were sparse, the EBIC hyperparameter (γ) was adjusted to 0 after careful consideration and comparisons of different settings. A γ set to 0 (can vary from 0 to 1, default is 0.25) results in a lower shrinkage of estimated connections. As simulation studies have shown, the likelihood of false positives is low and the specificity will still be higher compared with a non-regularised partial correlation network.23 88 99 The estimated networks were then bootstrapped for accuracy and stability using the bootnet function, which performs a non-parametric bootstrap to calculate the 95% bootstrapped CIs for the edges by resampling the data with replacement 2500 times per network. The networks were visualised with the plotting tools in qgraph, using the force-directed layout ‘spring’, which employs the Fruchterman-Reingold algorithm and draws nodes with higher centrality towards the centre.93 100 Lastly, network communities were calculated using the walktrap algorithm and plotted with igraph, qgraph and ggplot2 plotting tools.92 94 97 The data and R code necessary to reproduce our results can be found on The Open Science Framework repository.101–104
Supplemental material
Major network: black fields show significant differences (alpha=0.05) of edges.
Major network: standardised centrality index and significant differences (alpha=0.05) of node strength (black fields).
Deficiencies network: black fields show significant differences (alpha=0.05) of edges.
Deficiencies network: standardised centrality index and significant differences (alpha=0.05) of node strength (black fields).
CS-coefficients for each network (cut-off=0.25)
Patient and public involvement
Neither patients nor the public were involved in this study.
Results
Frequencies and percentages of reported variables
Frequencies and percentages for identified categories of deficiencies and contributing factors are reported in table 2.
Deficiencies (499 in total) were identified and classified under six major themes. The three most frequently reported categories concerned psychiatric evaluation, follow-up and cooperation. Typical cases involved patients who were referred from inpatient to outpatient services or changes in primary clinical contacts, both of which could result in missed appointments or incomplete assessments of health status. In 28% of the cases involving follow-up, healthcare planning that could have provided a framework for treatment during the transition was also lacking. Seventy-seven per cent of the deficiencies categorised as problems in cooperation were linked to unclear delimitation of responsibility. Lack of adequate information was also a relatively common explanatory factor and was identified in 16% of all cases. In contrast, deficiencies concerning safety and relatives were rare.
In line with the structure of the RCA protocol, the 462 contributing factors formed five major themes (table 2). Nearly half of the factors pointed towards failing procedures, routines, or guidelines as contributors. Examples included poor compliance to, or insufficient knowledge about an existing policy, or lack of guidelines that could be applied in a specific context. Suboptimal work structures, communication problems and insufficient competence regarding medical, juridical or organisational matters were also reported as common.
Network stability
Correlation stability coefficients (CS-coefficients) denote the estimated maximum number of cases that can be dropped from the data to retain a correlation of at least 0.7 between statistics, based on the original network data and statistics computed with fewer cases (with 95% probability). The coefficient should not be below 0.25 and is preferably above 0.5.87 105 The CS-coefficients for each of the three networks (the major network, deficiencies network and contributing factors network) are shown in table 3. As CS-coefficients for the Contributing factors network were below the cut-off value, indicating instability, further investigations are required before any final conclusions can be drawn.87 The visualisation of this network is included in online supplemental file B, along with the centrality indices calculated for this subset.
Central and peripheral nodes in the network
The centrality indices node strength and edge strength were included to quantify impact on each network structure. Node strength is defined as the total sum of the magnitude of each of its edges. Edge strength in a partial correlation or regularised network reflects the magnitude of the pairwise relationship between two nodes, while controlling for indirect influences via other nodes.23 88 The centrality indices closeness, betweenness and EI were examined, but excluded from the main section of this paper as the CS-coefficients for closeness and betweenness were below cut-off and EI did not add anything to the interpretation that was not already explained by node strength. Calculated values for these indices are included in online supplemental files C–E.
Supplemental material
Supplemental material
Supplemental material
The major network and significant differences of edges are shown in figure 1.
As shown in both figures 1 and 2, nodes representing documentation, communication, organisation, follow-up, procedures and psychiatric evaluation were central, compared with nodes related to safety, competence, contact with relatives, technical issues and cooperation. Although the nodes involved in this subset all scored high in strength, two non-significant, negative connections were found: (1) between organisation and communication, and (2) between psychiatric evaluation and procedures, which may reflect how data were registered by the RCA teams.
In the Deficiencies network (figure 3), missed appointments, particularly the absence of booked follow-ups but also cancellations made by the patient, scored high in node strength. Consequently, missed assessments of suicide risk and continuous re-evaluation of the psychiatric status, were also central, along with the node representing shortages in staff.
In relation to these nodes, the nodes representing administrative problems, such as missed referrals or other types of transferred information, safety issues, suboptimal contact with relatives, healthcare plan being either absent or incomplete, and assessment not being recorded were more peripheral (figure 4).
The third network, representing contributing factors, was too instable to estimate. Although the nodes for work structure, resources, competence and continuity had the highest node strength centrality, the differences were not significant. Our recommendations are to examine these more thoroughly in a future study with a larger sample. The topology and centrality indices for the Contributing factors network are shown in online supplemental file B.
Detected communities
Communities were detected using the walktrap algorithm.92 The nodes belonging to a community are colour marked in the visualisations of the networks in figures 1 and 2.
Two communities were present in the major network (figure 1):
The nodes for the deficiencies psychiatric evaluation, follow-up, and the contributing factor for procedures, routines and policies.
The nodes for the deficiency communication and the nodes representing the contributing factors organisation and communication.
Analysis of the deficiencies network (figure 2) resulted in two detected communities. The first included the nodes representing understaffing, declined/missed appointments, and cases where future appointments had not been booked. The second covered the nodes representing assessments of suicide risk and of the overall mental condition.
Discussion
The results of this study suggest that reported adversities are linked to a group of activities, rather than to single mistakes. Providing suicidal patients with regular assessments, for instance, and proposing adequate actions depends not only on the personal conditions of the evaluating clinician and the patient being assessed, but also on proper work structures, good intrateam communication, adequate routines and well-known procedures, and sufficient documentation of planned and performed activities.
There are three main findings of this study. First, missed and declined appointments are central features when examining elements occurring prior to the suicide. Together they account for more than a fifth of the total amount of deficiencies. We have not examined the positive effects of feedback loop systems which enhances the ability for healthcare providers to react when a patient does not turn up on scheduled meetings. Nor have we investigated cases with negative correlations between treatment cancellations and suicide. However, one hypothesis drawn from our results and extrapolated conclusions from previous studies,51 52 57 58 60–63 65–72 is that any disruption in treatment is negative, and cancellations made by the patient could be an early warning sign of an ongoing exacerbation of the suicidal process. During phases of acute suicidality or in the early stages of recovery from a suicide attempt, the well-being of the patient is frail, and the suicide risk may fluctuate rapidly.58 61 64–66 Establishing a backup system, which safeguards follow-up plans and alerts healthcare staff when patients cancel planned appointments, could help improve patient safety. Second, many nodes are still disconnected. Even if it is likely that there is an underlying covariance, the correlation is not independently significant. The sparsity of the networks could be explained by the estimation procedure. Each network has been regularised to reduce false positive connections and produce parsimonious graphs. When comparing them with networks based on partial correlation matrices, many edges have been omitted due to the penalisation. It is therefore likely that other patterns would appear if more data were entered. Third, the nodes representing security, technical issues and contact with relatives have both low frequencies and low centrality. This means that adversities related to these areas are rarely reported. One reason for this can be the very nature of the type of failures that can occur in these areas. Denied access to an important medical record system at a specific time rarely affects more than one or a few team members at a time. Ligature points, once removed, do not reappear at the exact same location. Establishing and maintaining stable work conditions, on the other hand, is more elusive. The concept of organisational prerequisites to provide safe interventions to suicidal patients is subjective which could lead to a higher rate of recurrences of management related issues. While adverse events concerning security at the inpatient facilities were rare, the transition to outpatient services was frequently mentioned in the post-mortem audits. Transitions imply a change in primary caregiver and a shift from short-term to long-term treatment goals. A connection to elevated risk levels could be expected, although the direct relationship has not been investigated in this study. To gain more knowledge about the mechanisms involved, network studies covering these steps of the process are needed. Even though interviews with relatives were included in 64% of the reports, their perspective were only reflected in 1% of the deficiencies (table 2). This situation has been previously described by Bouwman et al.50 . After examining policies from 15 healthcare organisations and spoken to 35 stakeholders (including patient, families and their counsellors, national regulators and professionals) they concluded that involvement by relatives, insofar they had been involved, rarely extended beyond aftercare and information provision.50 With this in mind, studies based on the narratives of relatives would probably complement and enrich our results.
We acknowledge that from a general point of view, some of our findings are similar to the conclusions drawn by our colleagues in the same field. Suicide risk is multifactorial, and decisions about appropriate safety measures are dependent on factors on both individual and structural levels.5 6 10 38 40 42 44 46 48 49 51 53–55 57 58 60–64 66 68–70 However, following the argumentation of Fried and Robinaugh (2021) on complexity, adverse events cannot be prevented by understanding the single components alone, neglecting the interactions among them.106 If the value of a unique node is determined not only by the intrinsic properties of the node itself, but by its relations to other objects, the study of single factors will not yield any ultimate answers about how to prevent undesired events. To gain more knowledge, we must first examine the dynamics of the systems from which adversity arises.
Conclusion
Network analysis adds to previous research in patient safety by elucidating patterns which may be unclear if only incident rate is considered. The results show that failed assessments and cancelled treatments during follow-up are both frequent and have a high centrality, thus functioning as a warning sign for exacerbation. Organisational instability, in terms of understaffing, shortages of resources and suboptimal work procedures are also prominent features of the networks. Although comparative studies are needed before any final conclusions can be drawn, focusing on these areas may improve patient safety in suicide prevention.
Strengths and limitations
Strengths of the study include data collected from NITHA, the only open national resource in Sweden for the dissemination of RCA reports. These reports were produced in a standardised manner by trained RCA teams. The data were examined and categorised by four professionals, all experienced in performing and peer-reviewing RCA reports. Considering the dynamic nature of deficiencies in healthcare where underlying factors are rarely sharply outlined, but rather multilayered, network analysis can bring new and valuable insights of risk-prone areas.
The study also has several limitations. This was a cross-sectional study, limiting the capacity to identify the directions of effects. Since we obtained our data exclusively through the NITHA system, other postsuicide investigations were not included. Because regional institutional praxis concerning submission to the NITHA database varied, RCA reports cannot be considered representative for the country of Sweden. A relatively small proportion of all suicides were submitted to the database, and therefore, selection bias cannot be ruled out. The RCA methodology is designed to scrutinise organisations and detect possible causes for systematic negative output. Consequently, the reported findings may focus on incidental discoveries, rather than some latent factor which lies beyond the scope of the protocol. Moreover, since RCA aims to identify organisational vulnerabilities, the reports lack certain details concerning the patients themselves. As we did not have access to original records, we have not been able to verify the accuracy of the content in the RCA reports. Therefore, our findings will reflect any misclassification done by the RCA teams during the initial investigation process. Lastly, the classification tool used by the auditing teams has not been validated by independent reviewers. The data were qualitatively categorised and could have been organised differently.
Future research
Based on the findings of this study, we suggest further research on security systems which help healthcare providers to react when patients drop out of treatment. Considering the relatively low number of observations, we also recommend future network studies based on a larger sample. To gain more insights into the perspectives of patients and relatives, network studies based on their experiences would be a fruitful approach.
Data availability statement
Data are available in a public, open access repository. Copies of used R scripts and data sets are available at Open Science Framework: DOI 10.17605/OSF.IO/3DHTF.
Ethics statements
Patient consent for publication
Acknowledgments
The authors would like to express their gratitude to Cecilia Boldt-Christmas (CBC), Marzia Dellepiane (MD), and Evalena von Hausswolff (EvH) who participated in the data extraction and coding.
References
Supplementary materials
Supplementary Data
This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.
Footnotes
Twitter @lilasali
Contributors MR (guarantor) designed the study, collected and registered the data, performed the analyses, and drafted the manuscript. EC, LA, MW and TB contributed to the study design, analyses and manuscript revision. All authors read and approved the final manuscript.
Funding This work was supported by the Affective Clinic, Sahlgrenska University Hospital, Gothenburg, Region Västra Götaland (no specific grant number).
Competing interests None declared.
Patient and public involvement Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.
Provenance and peer review Not commissioned; externally peer reviewed.
Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.