Open access Protocol Catalonia Suicide Risk Code BMJ Open: first published as 10.1136/bmjopen-2020-037365 on 12 July 2020. Downloaded from http://bmjopen.bmj.com/ on February 25, 2022 by guest. Protected by copyright. Epidemiology (CSRC-­Epi) study: protocol for a population-­representative nested case–control study of suicide attempts in Catalonia, Spain Philippe Mortier ‍ ‍,1,2 Gemma Vilagut,1,2 Beatriz Puértolas Gracia,1,2 Ana De Inés Trujillo,1,3 Itxaso Alayo Bueno,1,2 Laura Ballester Coma,1,2,4 María Jesús Blasco Cubedo,1,2,5 Narcís Cardoner,6,7,8,9 Cristina Colls,10 Matilde Elices,8,11 Anna Garcia-­Altes,2,10,12 Manel Gené Badia,13 Javier Gómez Sánchez,1 Mario Martín Sánchez,14 Rosa Morros,15,16,17 Bibiana Prat Pubill,18 Ping Qin,19 Lars Mehlum,19 Ronald C Kessler,20 Diego Palao,6,7,8,9 Víctor Pérez Sola,7,8,11,21 Jordi Alonso,1,2,5 On behalf of the CODIRISC Epidemiology Study Group To cite: Mortier P, Vilagut G, Abstract Puértolas Gracia B, et al. Strengths and limitations of this study Introduction Suicide attempts represent an important Catalonia Suicide Risk Code public health burden. Centralised electronic health record Epidemiology (CSRC-­Epi) study: ►► Suicide attempt cases (estimated n~6000) in the (EHR) systems have high potential to provide suicide Catalonia Suicide Risk Code case registry are iden- protocol for a population-­ representative nested case– attempt surveillance, to inform public health action aimed tified through in-­depth psychiatric evaluation, which control study of suicide attempts at reducing risk for suicide attempt in the population, and allows to carefully differentiate between suicidal in Catalonia, Spain. BMJ Open to provide data-­driven clinical decision support for suicide and non-­suicidal self-­injurious behaviour. 2020;10:e037365. doi:10.1136/ risk assessment across healthcare settings. To exploit this ►► A wide range of predictor variables will be included, bmjopen-2020-037365 potential, we designed the Catalonia Suicide Risk Code taken from centralised electronic health record data ►► Prepublication history for Epidemiology (CSRC-­Epi) study. Using centralised EHR representing five clinical settings, that is, emergen- this paper is available online. data from the entire public healthcare system of Catalonia, cy care, primary care, outpatient mental healthcare, To view these files, please visit Spain, the CSRC-­Epi study aims to estimate reliable and general and psychiatric hospitalisations. the journal online (http://​dx.​doi.​ suicide attempt incidence rates, identify suicide attempt ►► A two-­ stage exposure-­ enriched nested case– org/​10.​1136/​bmjopen-​2020-​ risk factors and develop validated suicide attempt risk control study combined with the use of inverse 037365). prediction tools. probability weights will enable efficient and Methods and analysis The CSRC-­Epi study is registry-­ population-­representative estimations. Received 30 January 2020 Revised 02 May 2020 based study, specifically, a two-­stage exposure-­enriched ►► An unknown proportion of suicide attempt cases do Accepted 29 May 2020 nested case–control study of suicide attempts during the not contact healthcare services, and are therefore period 2014–2019 in Catalonia, Spain. The primary study not included in this study. outcome consists of first and repeat attempts during the ►► Limited information on history of suicide attempt observation period. Cases will come from a case register before the study observation period will be available. linked to a suicide attempt surveillance programme, which offers in-­depth psychiatric evaluations to all Catalan residents who present to clinical care with any suspected risk for Ethics and dissemination This protocol is approved by suicide. Predictor variables will come from centralised EHR the Parc de Salut Mar Clinical Research Ethics Committee © Author(s) (or their systems representing all relevant healthcare settings. The (2017/7431/I). Dissemination will include peer-­reviewed employer(s)) 2020. Re-­use study’s sampling frame will be constructed using population-­ scientific publications, scientific reports for hospital and permitted under CC BY-­NC. No representative administrative lists of Catalan residents. government authorities, and updated clinical guidelines. commercial re-­use. See rights Inverse probability weights will restore representativeness of Trial registration number NCT04235127. and permissions. Published by the original population. Analysis will include the calculation BMJ. of age-­standardised and sex-­standardised suicide attempt For numbered affiliations see incidence rates. Logistic regression will identify suicide Introduction end of article. Suicide attempts constitute a major public attempt risk factors on the individual level (ie, relative risk) Correspondence to and the population level (ie, population attributable risk health issue worldwide, despite the fact Dr Philippe Mortier; proportions). Machine learning techniques will be used to that prevention strategies have shown to ​pmortier@​imim.​es develop suicide attempt risk prediction tools. be effective in reducing attempt rates.1 Mortier P, et al. BMJ Open 2020;10:e037365. doi:10.1136/bmjopen-2020-037365 1 Open access BMJ Open: first published as 10.1136/bmjopen-2020-037365 on 12 July 2020. Downloaded from http://bmjopen.bmj.com/ on February 25, 2022 by guest. Protected by copyright. Population-­based surveys estimate the lifetime prevalence study rare events such as suicidal behaviour in an efficient of suicide attempts among adults at 2.7% (range 0.5%– and population-­representative way. 5.0%),2 while a recent meta-­ analysis among children From a clinical viewpoint, the use of centralised EHR and adolescents found a pooled estimate of 6.0% (range data has high potential to provide data-­driven clinical 0.5%–34.1%) for suicide attempts in early life.3 Suicide decision support when evaluating risk for future suicide attempts are related to subsequent suicide,4 which has attempts.29 Such support is highly needed, as suicide a worldwide mortality rate estimated at 11.6 per 100 000 attempts constitute a complex behavioural outcome, person-­years, representing an annual loss of 34.6 million determined by highly multifactorial population processes, years of life.5 Apart from death by suicide, suicide attempts interdependencies and multilevel causality.30 Clinicians are also markers for subsequent persistent physical and are prone to heuristic-­based decision making,31 that is, a mental health issues, repeat suicide attempt, psychiatric rapid problem-­solving approach, including subconscious hospitalisations, impaired academic performance, unem- cognitive shortcuts, linking a limited set of risk factors ployment, partner abuse victimisation and perpetration, directly to suicide attempt potential. This leads to failure having children removed by social services, loneliness, to detect real suicide potential, poor patient experience relationship difficulties, impaired social functioning and and ineffective clinical decision making. In recent years, low life satisfaction.6–11 a number of studies started implementing advanced Despite the considerable burden that suicide analytical techniques on EHR data—including machine attempts represent in our society, there is a lack of reli- learning techniques—to model the complex additive and able surveillance data on suicide attempts that could interactive effects between large numbers of predictor inform public health action.12 This is in contrast with variables, and to improve the classification accuracy of actual suicide rates, that are increasingly monitored in data-­driven suicide attempt risk prediction tools.32 When many countries worldwide.13 WHO advocates the use routinely implemented at the healthcare system level, of centralised electronic health record (EHR) systems such data-­driven decision support tools could guide the to develop national suicide attempt surveillance.14 adequate allocation of clinical resources, such as in-­depth However, currently used disease classification systems suicide risk assessments and tailored treatment interven- in EHR systems (eg, the International Classification of tions in multistage screening approaches.33 34 Diseases15) do not allow to distinguish between suicidal Here, we present the protocol for the Catalonia Suicide and non-­ suicidal intent of self-­injurious behaviour.16 Risk Code Epidemiology (CSRC-­Epi) study, a large epide- In addition, due to the often difficult ascertainment of miological study of suicide attempts occurring during self-­injurious and suicidal intent, misclassification with the period 2014–2019 in Catalonia, Spain. The primary regard to suicidal outcomes often occurs.17 Offering an study outcome is suicide attempt among Catalan resi- in-­depth psychiatric evaluation to each individual who dents during the period 2014–2019. Secondary analyses presents to clinical care with suspected suicide risk, will focus on actual suicide among those with previous followed by a standardised registration of this clinical suicide attempts. The CSRC study combines a nested evaluation in a centralised EHR register, may therefore case–control sampling design with the use of inverse substantially improve the accuracy of public health probability weighting to enable the efficient analysis of surveillance of suicide attempts. a large amount of centralised EHR data. A unique aspect Apart from surveillance, centralised EHR systems have of the study is that data for the suicide attempt cases high potential to be used in epidemiological studies on come from an especially designed suicide attempt surveil- suicidal behaviour in the population.18–20 Indeed, it has lance protocol that stipulates that every Catalan resident been estimated that up to 92% of individuals that eventu- presenting to clinical care with any suspected risk for ally die of suicide have some type of healthcare contact in suicide receives an in-­depth psychiatric evaluation. These the year prior to death,21 with rates ranging from 54% to evaluations will allow us to differentiate carefully between 80% for primary care contacts,21 22 31% to 66% for mental suicidal and non-­suicidal self-­injurious behaviour in the healthcare contacts,21 22 24% to 60% for emergency study outcomes. department visits21 23–25 and 21% for psychiatric hospi- talisations.21 Clinical data collected through centralised Study objectives EHR systems could therefore provide new insight in the The CSRC-­Epi study main objectives are: population distribution of suicide attempt risk factors, ►► To provide reliable incidence measures for suicide outline the different healthcare trajectories preceding attempts occurring in Catalonia during the period an attempt and provide estimates of potential reduc- 2014–2019. tions in suicide attempt cases when designing preven- ►► To identify risk factor constellations for suicide tion interventions. The need for nested case–control attempts, both on the individual level (ie, the extent studies using EHR data to investigate suicidal behaviour to which risk factors increase the risk for subsequent has been recently highlighted.26 Developments in statis- suicide attempt in an individual) and on the popu- tical methods, including the use of inverse probability lation level (ie, the proportion of the total cases of weighting in case–control studies27 as well as two-­stage suicide attempt that are potentially attributable to risk (exposure-­enriched) case–control designs,28 now allow to factors). 2 Mortier P, et al. BMJ Open 2020;10:e037365. doi:10.1136/bmjopen-2020-037365 Open access BMJ Open: first published as 10.1136/bmjopen-2020-037365 on 12 July 2020. Downloaded from http://bmjopen.bmj.com/ on February 25, 2022 by guest. Protected by copyright. ►► To develop suicide attempt risk prediction tools using The CSRC case register subsequently includes all indi- machine learning techniques, and test their predic- viduals with a suicide attempt during the observation tive accuracy. period 2014–2019, including the exact date of event. For each event included in the SRC case register, a range Methods and analysis of predictor variables for future suicidal behaviour are Data sources assessed: the Suicidal Scale of the Mini-­ International All data for this study will be obtained from the Health Neuropsychiatric Interview (MINI) 5.0.0,37 38 the type Evaluation and Quality Agency of Catalonia (AQuAS35), and lethality of the suicide attempt that warranted evalua- a public entity attached to the Catalan Health Depart- tion, presence and type of mental disorder, hopelessness, ment. As from 2017, AQuAS manages the Public Data impulsivity, aggressiveness, altered state of conscience, Analysis for Health Research and Innovation Programme use or dependence of alcohol, use or dependence of (PADRIS35) to provide researchers with access to large illicit drugs, serious somatic disease (including chronic amounts of centralised EHR data, and to foster innovative diseases, chronic pain and disabilities), living status, pres- health research. ence of family or social support, social problems, stressful life events, access to lethal means and family history of Administrative data suicide. These variables will be used as predictor variables A first source of data for the CSRC-­Epi study will consist in analyses predicting repetition of suicide attempt and of six population-­ representative administrative lists of suicide after a previous attempt. Catalan residents, one for each year in the 2014–2019 period. These lists constitute annual censuses of indi- EHR data viduals with access to public healthcare, which, by law, A third source of data will consist of centralised EHR includes every Catalan resident. These lists include socio- registers, one for each of five clinical healthcare settings, demographic variables (ie, sex, age, nationality, a range that is, emergency care, primary care, outpatient mental of socioeconomic indicators and healthcare catchment healthcare and general and psychiatric hospitalisations. region) as well as associated small area geocode data These registers include a wide range of relevant predictor (ie, data available on the healthcare catchment region variables for suicidal behaviour, that is, history of self-­ level),33 the dates of immigration and emigration in/ injurious behaviours; all types of somatic conditions; out of Catalonia, the date of death, as well as a range neurodevelopmental, mental, behavioural, personality of healthcare summary variables that are constructed and substance use disorders; all types of medical proce- to monitor healthcare needs in the Catalan population dures performed; and detailed information on the (ie, 12-­month depression, 12-­ month complex mental number and type of healthcare contacts. Diagnoses and disease and the number of 12-­month healthcare contacts procedures in the EHR data are coded using the Inter- for each healthcare setting). These lists will be used to national Classification of Diseases-­ 9th revision-­Clinical Modification (ICD-9-­ CM) and ICD-10-­ CM disease clas- construct a sampling frame when conducting the nested sification system. The year of inception of the different case–control sampling, as further explained below. registers is 2012 for emergency care and primary care and 2008 for the other registers. CSRC case register In 2014, the Catalan Health Department and the Catalan Pharmaceutical register Health Service structurally implemented the CSRC A fourth source of data will consist of a register containing surveillance programme36 in the Catalan Public Health- all prescription drugs that have been delivered by officially care system. The CSRC surveillance programme is a specif- recognised pharmacies, including the date of delivery. ically designed suicide attempt surveillance protocol that Note that this excludes prescribed medication that was stipulates that every Catalan resident presenting with any not collected at the pharmacy. This register will provide suspected risk for suicide in any public healthcare setting an additional range of predictor variables for suicidal receives a face-­ to-­ face in-­ depth psychiatric evaluation behaviour, that is, all prescriptions for psychopharmaco- at the nearest emergency department. This assessment logical products, as well as prescriptions for a wide range includes differentiating suicide attempts from non-­ of medication used to treat relevant somatic conditions or suicidal self-­injurious behaviour, or from adverse mental known to have psychotropic effects. health states without self-­injurious intention. Each indi- vidual deemed at high risk for (repeat) suicide attempt Mortality register is subsequently eligible for two brief follow-­up interven- Suicide cases among those with a suicide attempt during tions (ie, a mental healthcare visit within 10 days (within the study observation period will be identified using 72 hours when aged 17 or less), and a phone call after data from the mortality register, managed by the Catalan 30 days) to increase access to adequate mental healthcare Department of Forensic Medicine, which provides detailed use. Clinical data of all individuals that received a special- data on causes of death using the ICD system (9th or 10th ised assessment are registered centrally in the CSRC case revision). State-­of-­the art forensic techniques, including register. psychological autopsy by a multidisciplinary team, is used Mortier P, et al. BMJ Open 2020;10:e037365. doi:10.1136/bmjopen-2020-037365 3 Open access BMJ Open: first published as 10.1136/bmjopen-2020-037365 on 12 July 2020. Downloaded from http://bmjopen.bmj.com/ on February 25, 2022 by guest. Protected by copyright. Figure 1 Overview of the CSRC-­Epi study design. CSRC-­Epi, Catalonia Suicide Risk Code Epidemiology; EHR, Electronic healthcare record; M, million; W, weight. to determine death by suicide in the mortality register.39 ~0.8%, respectively.39 Based on these figures, we expect Nevertheless, forensic examination of suicidal intention a maximum of ~9.1 million individuals with Catalan is difficult, and misclassification may occur. residency status on at least one point in time over the 2014–2019 period. However, we expect suicide attempts Study design in Catalonia to be extremely rare before the age of 10, in Figure 1 shows an overview of the CSRC-­Epi study design. line with findings that self-­injurious behaviour generally The CSRC-­ Epi study is a register-­based study, that is, occurs as from the adolescent period.43 We will, therefore, a study that uses the exposure and outcome data from exclude cases and controls that have not reached age 10 registries,40 which in turn, are representative for the by the end of the 2014–2019 period (ie, ~10.9%), lowering target population. The target population consists of the the total expected target population to ~8.1 million. dynamic cohort of all Catalan residents during the 6-­year period, 2014–2019. As explained in detail below, we will Case selection conduct a two-­stage nested case–control study within this Suicide attempt cases between 2014 and 2019 will be dynamic cohort,41 42 which, in combination with the use identified using the CSRC case register. Based on prelim- of inverse probability weights, will allow us to construct a inary data exploration, we expect to include ~6000 cases dataset representative for the original cohort of Catalan of clinically confirmed suicide attempt by the end of residents, and analyse the data accordingly. 2019, of which ~8% (~480) will be repeat attempters. This Annual total population of Catalonia is between 7.5 and substantially exceeds the number of suicide attempt cases 7.6 million, with annual rates for immigration, emigra- included in previous register-­based studies (median=1562, tion, birth and death being ~2.5%, ~1.9%, ~0.9% and IQR=1562–325032). 4 Mortier P, et al. BMJ Open 2020;10:e037365. doi:10.1136/bmjopen-2020-037365 Open access BMJ Open: first published as 10.1136/bmjopen-2020-037365 on 12 July 2020. Downloaded from http://bmjopen.bmj.com/ on February 25, 2022 by guest. Protected by copyright. Table 1 ICD codes to identify potentially missed cases of suicide attempt ICD-9-­CM Description ICD-10-­CM Description Suicide attempts E950*-E959* Self-­poisoning, hanging, X71*–X83* Drowning and submersion, and intentional self-­ strangulation, suffocation, firearms, explosive or thermal injurious behaviour fire arms, jumping, others material, sharp or blunt objects, jumping from a high place, jumping or lying in front of a moving object, crashing of motor vehicle and other specified means     T36*-T50* with a 5/6th Drug poisoning (overdose) character of 2     T51*-T65* with a 5/6th Toxic effects of nonmedicinal character of 2 substances     T71* with a sixth Asphyxiation, suffocation, character of 2 strangulation     T14.91 Suicide attempt Self-­injurious E980*-E989* Self-­poisoning, hanging,     behaviour of strangulation, suffocation, undetermined fire arms, jumping, others intent (intentional/ (undetermined intent) accidental) 994.7 Asphyxiation/strangulation T71* with a sixth Asphyxiation, suffocation, character of 4 strangulation 881*, 903.2, 903.3, Open wound of elbow, S51.001-­S51.009, Open wound of elbow/forearm/ 903.4 forearm, and wrist; injury S51.801-­S51.809, wrist, injury of ulnar/radial artery radial/ulnar vessels; injury S55.0-­S55.199, at forearm/wrist/arm level palmar artery S61.5-­S61.509, S65.0-­S65.199 965*, 967*, 969* Poisoning by T36*-T50* with a 5/6th Drug poisoning (overdose) analgesics, antipyretics, character of 4 antirheumatics, sedatives and hypnotics     T51*-T65* with a 5/6th Toxic effects of non-­medicinal character of 4 substances ICD-9-­CM, International Classification of Diseases-­9th revision-­Clinical Modification; ICD-10-­CM, International Classification of Diseases-­10th revision-­Clinical Modification. One of the main objectives of the CSRC programme is attempt case detection using ICD codes.17 Therefore, to enable reliable surveillance of suicide attempt in the we intent to validate the range of ICD codes we identi- population, and to tackle under-­registration and misclassi- fied against golden standard identification methods (ie, fication using regular EHR systems.16 Nevertheless, failure manual review as well as text mining of clinical notes) to adhere to the CSRC programme protocol may result in to increase the accurate detection of potentially missed an unknown number of suicide attempt cases that remain cases for our study. undetected. Therefore, ICD disease classification codes in the five centralised EHR registers will be inspected Control selection to identify potentially missed cases of suicide attempt. In a first stage, we will select a 20% stratified random For that purpose, a wide range of ICD codes related to sample of the 2014–2019 dynamic cohort members, using suicide attempt (see table 1) was identified through an the six population-­ representative administrative lists extensive MEDLINE search, including a recent overview of Catalan residents described above. Constructing this article with recommendations on the use of ICD codes for preliminary 20% subsample is necessary for two reasons: the surveillance of self-­injurious behaviour.16 ICD codes (1) we need to reduce the amount of data AQuAS will are unable to determine suicidal intent, but do allow to need to handle when conducting the age-­matched and identify subjects with intentional self-­injurious behaviour sex-­matched incidence density sampling in the second and to differentiate them from subjects with self-­injurious stage; and (2) based on publicly available data, we esti- behaviour of undetermined intent (ie, intentional or mate the probability of selecting controls with 12-­month accidental self-­injurious behaviour).16 Outcome defini- healthcare contacts for mental disorders to be relatively tion algorithms (ie, predefined sets of ICD codes) have low (ie, ranging from ~17% for primary care visits to shown promising in increasing the accuracy of suicide ~0.1% for general hospitalisations). Therefore, in order Mortier P, et al. BMJ Open 2020;10:e037365. doi:10.1136/bmjopen-2020-037365 5 Open access BMJ Open: first published as 10.1136/bmjopen-2020-037365 on 12 July 2020. Downloaded from http://bmjopen.bmj.com/ on February 25, 2022 by guest. Protected by copyright. to enrich the data for relevant exposure information, In addition, we will conduct separate analyses focusing controls in the 20% subsample will be oversampled for on clinical severity of suicide attempts (eg, lethality and number and specific types of healthcare use, using the method of attempt). A secondary study outcome consists past year healthcare summary variables available in the of suicide among those with a suicide attempt during administrative lists. This will result in a higher number 2014–2019. Cases of suicide will be identified through the of controls with (mental) healthcare diagnoses eligible in mortality register, managed by the Catalan Department of the second stage. Forensic Medicine (see Mortality register). In a second stage, we will create 67 risk sets, one for each month in the 6-­year observation period (ie, June 2014 to Suicide attempt occurrence December 2019). Within each risk set, a number of 30 Population-­representative annual incidence rates will be age-­matched and sex-­matched controls will be randomly estimated by dividing annual number of cases by total selected for each case (ie, case or potentially missed case) annual sums of person-­years at risk, multiplied by 100 without replacement (incidence density sampling or risk 000, using the weighted nested case–control dataset. We set sampling41). Eligible controls will include future cases will calculate both crude as well as age-­standardised and and controls will be allowed to be selected multiple times sex-­ standardised incidence rates, stratified by relevant across risk sets. To allow for the joint and separate anal- sociodemographic variables (eg, socioeconomic status, yses of suicide attempt and potential suicide attempt (see healthcare region, etc). As incidence rates do not inform the Data analytical plan section), controls are selected for of the distribution of cases over time, we will also estimate first suicide attempt within individuals (eligible controls and visualise incidence proportions (or cumulative inci- including previous potential cases, but not previous dence) over time using the Kaplan-­Meier estimator to suicide attempt cases), and if applicable, also for the estimate one minus the survival function. first potential suicide attempt within individuals (eligible controls including those without previous suicide attempt Suicide attempt risk factor associations or potential suicide attempt only). After the final selection To estimate the individual-­ level associations between of controls, inverse probability weights will be constructed predictor variables and outcome variable, (conditional) to restore population-­representativeness of the original bivariable and multivariable binomial logistic regression dynamic population cohort. Weights will be equal to 1 will be applied. For a first suicide attempt during the for cases and potentially missed cases, (ie, all are selected observation period as the outcome variable, the weighted in both sampling stages); for controls, weights will reflect nested case–control dataset will be used. In order for the the selection probabilities at stage 1 (including the over- OR to be a valid estimation of the HR (and hence, relative sampling according to healthcare summary variables), as risk) multiple inputs for those controls selected multiple well as at stage 2 (including the age–sex matching and the times and for those cases also selected as controls will be total time at risk of each control during the observation included in the analyses.45 Time-­varying predictor vari- period).42 44 ables (eg, the exact dates of diagnoses or medical prescrip- A specific objective of the SRC-­ Epi project is the tions) will be recoded by categorising time-­to-­event into construction of suicide attempt risk prediction tools by discrete time intervals. Relevant time-­ to-­ event cut-­ offs healthcare setting. For this purpose, a separate series of for these intervals will be identified by examining the controls will be selected at the second sampling stage, this changes in odds across time-­to-­event for a high number of time matching by age, sex as well as type and timing of last short time intervals. Given the high amount of predictor healthcare contact. For example, for a suicide attempt variables under study, the Least Absolute Shrinkage and assessed at the emergency department at time y and a last Selection Operator46 method will be employed as to select healthcare visit at primary care at time x, 30 age-­matched a subset of predictor variables that predict the outcome and sex-­matched controls will be selected among those best while maintaining a good model fit. For suicide and individuals that have not committed a suicide attempt for repetition of suicide attempt as the outcome variable, up to time y, restricting to those controls with primary the analysis will be very similar, but will now reduce to healthcare visits around time x. a cohort analysis of those individuals with first suicide attempt during 2014–2019 (which are all selected), and Data analytical plan regression models will also include the clinical variables The primary outcome of this study is suicide attempt from the CRSC case register obtained though in-­depth during the period 2014–2019. We will both focus on first psychiatric evaluation. suicide attempt during the observation period, as well Population-­level effect sizes will be estimated by calcu- as on repetition of attempts, defined as suicide attempts lating bivariable and multivariable population-­attributable among those with a previous suicide attempt during the risk proportions (PARP47), based on summary measures observation period. As explained above, we will also iden- of individual predicted probabilities obtained from the tify potentially missed cases of suicide attempt, using logistic regression models described above, comparing the ICD codes identified in the literature (see table 1). original models versus models in which regression coef- Cases and potentially missed cases will be considered ficients of the predictor variable(s) under study are set both as joint as well as separate outcomes in the analyses. to zero. PARP subsequently provide an estimate of the 6 Mortier P, et al. BMJ Open 2020;10:e037365. doi:10.1136/bmjopen-2020-037365 Open access BMJ Open: first published as 10.1136/bmjopen-2020-037365 on 12 July 2020. Downloaded from http://bmjopen.bmj.com/ on February 25, 2022 by guest. Protected by copyright. proportion of cases that could potentially be prevented indirect information will be available, consisting of (1) if certain risk factor(s) in the population were to be the Suicidal Scale of the MINI (item 6) used in the CSRC eliminated, assuming a causal relationship between risk programme protocol, which assesses lifetime history of factor and outcome variable. PARP estimates are of high suicide attempt among the cases. The timing of these value for policy makers involved in suicide prevention, previous attempts will be unknown, and this information as they provide insight in the population-­level impact will be based on patients’ self-­reported information and of risk factor(s) with regard to suicide risk. Specifically, (2) the EHR registers, which include episodes of self-­ PARP take into account that high-­prevalence risk factors harm before 2014, among both cases and controls. It will, carrying low individual risk may be equally or even more however, not be possible to determine the suicidal intent important to consider than low- prevalence risk factors of these episodes, and EHR register data are only available carrying high risk for the affected individuals. Taking into as from 2008. A second main limitation of the CSRC-­Epi account such knowledge is important, as it is the combi- study is that, although the entire Catalan population has nation of both individual-­ level and population-­ level free access to public healthcare, about 20% of the popu- interventions that has shown to be successful in reducing lation opts for private coverage or uses both public and adverse outcomes with complex multicausal aetiologies.30 private healthcare systems. This limits the population-­ representativeness of the EHR data to an unknown extent. Suicide attempt risk predictions As mentioned above, a third limitation is that, due to vari- Risk prediction tools for (repetition of) suicide attempt able adherence to the CSRC surveillance programme, will be constructed using machine learning techniques. A suicide attempt cases may still go undetected. This will first series of algorithms will focus on estimating suicide be countered by identifying potentially missed cases of attempt risk for different prediction windows, that is, suicide attempt in the EHR registers. Related to this, it censoring data of both cases and controls at different should be acknowledged that an unknown proportion of time points relative to the time of event.48 A second series suicide attempt cases do not contact healthcare services, of algorithms will predict suicide attempt after specific and are therefore not included in this study. This limita- healthcare contacts using the separate series of controls tion is inherent to studies using EHR data, and points selected in the second sampling stage (see Control selec- to the need on complementing the knowledge that can tion). Machine learning techniques will include elastic be gained from registry-­based studies with findings from net penalised logistic regression,49 naïve Bayes classi- general population epidemiological survey research. fiers,50 multivariate adaptive regression splines,51 Bayesian additive regression trees,52 random forests,53 gradient Patient and public involvement boosting,54 k-­ nearest neighbour algorithms,55 support No patients involved. vector machines56 and artificial neural networks.57 Stacking ensemble techniques (super learning)58 will be implemented to further optimise prediction accuracy, by Ethics and dissemination using the predictions from the above mentioned algo- The protocol of this study has been approved by the rithms (the base learners) as input to train new models Parc de Salut Mar Clinical Research Ethics Committee (meta learners). To avoid model overfit, model develop- (CEIC protocol 2017/7431/I). The study is in line with ment and tuning will be conducted on a training dataset the principles established in the Declaration of Helsinki, using k-­fold cross-­validation, and model predictive accu- with the Charter of Fundamental Rights of the European racy will be evaluated in a separate validation dataset Union (2000/C 364/01), and the European Conven- using a recalibrated algorithm. Sample selection bias tion on Human Rights. All data for this study come from introduced by the two-­stage nested case–control design the PADRIS programme, and will involve processing of will be addressed using appropriately corrected classi- completely anonymised EHR data. For record linkage fiers.59 Model predictive accuracy will be evaluated by the activities, the Spanish Order SAS/3470/2009 for data area under the receiver operating characteristic curve, obtained in observational studies is followed. This is also as well as accuracy measures calculated for different in line with the Data Protection Directive of the European thresholds of the continuous predicted probability (ie, Union (Directive 95/46/EC). thresholds set to delineate the top x% at highest risk60), We aim to create awareness of the proposed action in including positive predictive values (precision, or the the general public, by providing comprehensive informa- probability that a predicted case is actually a true case), tion on the need for detecting new suicide attempt risk sensitivity values (recall, or the proportion of true cases factors constellations, on the need to improve suicide that has been predicted correctly), and F1-­scores (ie, the attempt risk estimation, and on the ongoing exploration harmonic mean of recall and precision). of clinical decision support for the improved assessment of suicide risk. The following communication measures Study limitations will be taken: (1) the design of a website providing clear A first major limitation of the CSRC-­Epi study is that any and balanced information on the project; (2) balanced history of suicide attempt before the study observation newspaper articles and interviews to the press and (3) period is unknown for both cases and controls. However, providing all healthcare settings with patient folders on Mortier P, et al. BMJ Open 2020;10:e037365. doi:10.1136/bmjopen-2020-037365 7 Open access BMJ Open: first published as 10.1136/bmjopen-2020-037365 on 12 July 2020. Downloaded from http://bmjopen.bmj.com/ on February 25, 2022 by guest. Protected by copyright. 19 the project, providing a clear and balanced summary of National Centre for Suicide Research and Prevention, Institute of Clinical Medicine, the project. University of Oslo, Oslo, Norway 20 Department of Health Care Policy, Harvard Medical School, Boston, MA, USA Communication with patients and with next of kin will 21 Institut de Neuropsiquiatria i Addiccions, Hospital del Mar, Barcelona, Spain be in lay terms only. Feasible formats are internet websites and patient forums, patient folders and carefully planned Collaborators The CODIRISC Epidemiology Study Group are: Jordi Alonso, Itxaso releases to the press. We will provide clear and balanced Alayo Bueno, Laura Ballester Coma, Jordi Blanch, María Jesús Blasco Cubedo, Ana information on the project, acknowledging the unique De Inés Trujillo, Maria Teresa Campillo Sanz, Narcís Cardoner, Anna Isabel Cebrià, Cristina Colls, Matilde Elices, Anna García-­Altés, Ricard Gavaldà, Manel Gené Badia, experience of each patient and stressing our final aim of Javier Gómez Sánchez, Ronald C. Kessler, Lars Mehlum, Cristina Molina Parilla, improving (not replacing) human clinical practice and Rosa Morros Pedrós, Philippe Mortier, Jordi Ortiz, Diego Palao, Rosa Maria Pérez care. Patient groups will also be involved in the design of Pérez, Víctor Pérez, Maria J. Portella, Bibiana Prat Pubill, Beatriz Puértolas Gracia, the overall communication strategy (codesign). Raquel Suárez Pérez, Ping Qin, and Gemma Vilagut. Targeted expert audiences will consist of those involved Contributors Initial draft of the protocol: PM, GV, BPG, IAB and JA. Critical review of in suicide research, as well as in general psychiatry, the protocol: LM, PQ, and RCK. Initial draft of the manuscript: PM, GV, BPG, IAB, and JA. Critical review of the manuscript: PM, GV, BPG, ADIT, IAB, LBC, MJBC, NC, CC, psychiatric epidemiology and translational psychiatry. ME, AG-­A, MGB, JGS, MMS, RM, BPP, PQ, RCK, DP, VPS and JA. All authors read and Scientific publications will be sent to peer-­reviewed jour- approved the final manuscript. nals through open-­access publishing, and added to the Funding This project was supported by ISCIII/FEDER PI17/00521, ISCIII/FEDER Pompeu Fabra University’s repository of open access arti- PI17/01205, and Generalitat de Catalunya (2017 SGR 452). The Catalonia Suicide cles.61 Further dissemination of results will be through Risk Code surveillance programme is an initiative of the Mental Health and Addictions Plan of the Department of Health of the Catalan Government. PM has a scientific conferences and workshops. Hospital and Sara Borrell research contract awarded by the ISCIII (CD18/00049). ME has a Juan government authorities involved with mental healthcare de la Cierva research contract awarded by the ISCIII (FJCI-2017–31738). VPS and will be informed of the study results through scientific ME want to thank unrestricted research funding from Secretaria d′Universitats i reports, which will present series of suicide intervention Recerca del Departament d′Economia i Coneixement (2017 SGR 134 to ‘Mental Health Research Group’), and Generalitat de Catalunya (Government of Catalonia). prevention frameworks. In addition, we will provide the BPG and ADIT received funding from ISCIII FI18/00012 and FPU2017-06447, Spanish and Catalan Department of Health with updated respectively. LBC received funding by Ministerio de Educación, Cultura y Deporte clinical guidelines for the assessment of suicide risk. Clini- (FPU15/05728). DP and JA received funding by ISCIII/FEDER PI17/01205. cians with mental healthcare expertise as well as emer- Competing interests DP has received grants and also served as consultant or gency department clinicians and general practitioners advisor for Angelini, Janssen, Lundbeck and Servier. The other authors have no will be informed of these recommendations through competing interests to declare. the project’s website and the professional associations’ Patient and public involvement Patients and/or the public were involved in the design, or conduct, or reporting, or dissemination plans of this research. Refer to websites, periodicals and meetings. the Methods section for further details. Author affiliations Patient consent for publication Not required. 1 Health Services Research Group, IMIM (Hospital del Mar Medical Research Provenance and peer review Not commissioned; externally peer reviewed. Institute), Barcelona, Spain Open access This is an open access article distributed in accordance with the 2 CIBER Epidemiología y Salud Pública (CIBERESP), Madrid, Spain Creative Commons Attribution Non Commercial (CC BY-­NC 4.0) license, which 3 Department of Social Psychology, Autonomous University of Barcelona (UAB), permits others to distribute, remix, adapt, build upon this work non-­commercially, Cerdanyola del Vallès, Barcelona, Spain and license their derivative works on different terms, provided the original work is 4 Department of Psychology, University of Girona (UdG), Girona, Spain properly cited, appropriate credit is given, any changes made indicated, and the use 5 Department of Health & Experimental Sciences, Pompeu Fabra University (UPF), is non-­commercial. See: http://​creativecommons.​org/​licenses/​by-​nc/​4.​0/. Barcelona, Spain 6 Depression and Anxiety Program, Department of Mental Health, Parc Taulí Sabadell, ORCID iD Hospital Universitari, Sabadell, Spain Philippe Mortier http://​orcid.​org/​0000-​0003-​2113-​6241 7 Department of Psychiatry and Legal Medicine, Universitat Autònoma de Barcelona (UAB), Cerdanyola Del Vallès, Barcelona, Spain 8 Centro de Investigación en Red de Salud Mental, CIBERSAM, Madrid, Spain 9 Institut d'Investigació i Innovació Parc Taulí (I3PT), Sabadell, Barcelona, Spain 10 Agència de Qualitat i Avaluació Sanitàries de Catalunya - Health Evaluation and References Quality Agency of Catalonia (AQuAS), Catalan Health Department, Barcelona, Spain 1 Lewitzka U, Sauer C, Bauer M, et al. Are national suicide prevention 11 Neurosciences Research Programme, IMIM (Hospital del Mar Medical Research programs effective? A comparison of 4 verum and 4 control countries over 30 years. BMC Psychiatry 2019;19:1–10. Institute), Barcelona, Spain 2 Nock MK, Borges G, Bromet EJ, et al. Cross-­National prevalence and 12 Institut d’Investigació Biomèdica (IIB Sant Pau), Barcelona, Spain risk factors for suicidal ideation, plans and attempts. Br J Psychiatry 13 Legal Medicine Unit, Faculty of Medicine, University of Barcelona, Barcelona, 2008;192:98–105. Spain 3 Lim K-­S, Wong CH, McIntyre RS, et al. Global lifetime and 12-­month 14 Preventive Medicine and Public Health Training Unit PSMar-­UPF-­ASPB, Parc prevalence of suicidal behavior, deliberate self-­harm and non-­ suicidal self-­injury in children and adolescents between 1989 and de Salut Mar, Agència de Salut Pública de Barcelona, Pompeu Fabra University, 2018: a meta-­analysis. Int J Environ Res Public Health 2019;16:ijer Barcelona, Spain ph16224581. 15 Fundació Institut Universitari per a la recerca a l'Atenció Primària de Salut Jordi 4 Ribeiro JD, Franklin JC, Fox KR, et al. Self-­injurious thoughts and Gol i Gurina (IDIAPJGol), Barcelona, Spain behaviors as risk factors for future suicide ideation, attempts, 16 Departament de Farmacologia, de Terapèutica i de Toxicologia, Universitat and death: a meta-­analysis of longitudinal studies. Psychol Med Autònoma de Barcelona, Barcelona, Spain 2016;46:225–36. 17 5 Naghavi M. Global burden of disease self-­harm Collaborators. global, Institut Català de la Salut (ICS), Metropolitana Nord, Barcelona, Spain regional, and national burden of suicide mortality 1990 to 2016: 18 Master Plan on Mental Health and Addictions, Ministry of Health, Catalan systematic analysis for the global burden of disease study 2016. Government, Barcelona, Spain BMJ 2019;364:l94. 8 Mortier P, et al. BMJ Open 2020;10:e037365. doi:10.1136/bmjopen-2020-037365 Open access BMJ Open: first published as 10.1136/bmjopen-2020-037365 on 12 July 2020. Downloaded from http://bmjopen.bmj.com/ on February 25, 2022 by guest. Protected by copyright. 6 Goldman-­Mellor SJ, Caspi A, Harrington H, et al. Suicide attempt in 33 Kessler RC, Bernecker S, Bossarte RM, et al. The role of big data young people: a signal for long-­term health care and social needs. analytics in predicting suicide. In: Personalized psychiatry-­Big data JAMA Psychiatry 2014;71:119–27. analytics in mental health, 2019. 7 Joffe BI, Van Lieshout RJ, Duncan L, et al. Suicidal ideation and 34 Kessler RC, Bossarte RM, Luedtke A, et al. Suicide prediction behavior in adolescents aged 12-16 years: a 17-­year follow-­up. models: a critical review of recent research with recommendations Suicide Life Threat Behav 2014;44:497–509. for the way forward. Mol Psychiatry 2020;25:168–79. 8 Mortier P, Demyttenaere K, Auerbach RP, et al. The impact of lifetime 35 Health Evaluation and Quality Agency of Catalonia (Aqua). Public suicidality on academic performance in college freshmen. J Affect data analysis for health research and innovation program (PADRIS), Disord 2015;186:254–60. 2019. Available: http://​aquas.​gencat.​cat/​en/​ambits/​analitica-​dades/​ 9 De Luca SM, Franklin C, Yueqi Y, et al. The relationship between padris/ [Accessed 1 Aug 2019]. suicide ideation, behavioral health, and College academic 36 Barcelona: Servei Catal de la Salut. Atenció a les persones en RISC performance. Community Ment Health J 2016;52:534–40. de suïcidi: codi RISC de suïcidi (CRS), 2015. Available: https://​ 10 Pajonk F-­G, Ruchholtz S, Waydhas C, et al. Long-­Term follow-­up scientiasalut.​gencat.​cat/​handle/​11351/​1654?​locale-​attribute=​es after severe suicide attempt by multiple blunt trauma. Eur psychiatr. 37 Lecrubier Y, Sheehan DV, Weiller E, et al. The mini international 2005;20:115–20. neuropsychiatric interview (mini). A short diagnostic structured 11 Stanford S, Jones MP, Loxton DJ. Understanding women who interview: reliability and validity according to the CIDI. Eur psychiatr. self-­harm: predictors and long-­term outcomes in a longitudinal 1997;12:224–31. community sample. Aust N Z J Psychiatry 2017;51:151–60. 38 Roaldset JO, Linaker OM, Bjørkly S. Predictive validity of the mini 12 Bachmann S. Epidemiology of suicide and the psychiatric suicidal scale for self-­harm in acute psychiatry: a prospective perspective. Int J Environ Res Public Health 2018;15:1425. study of the first year after discharge. Arch Suicide Res 13 World Health Organisation. Who | suicide data. who website, 2019. 2012;16:287–302. Available: https://www.​who.​int/​mental_​health/​prevention/​suicide/​ 39 Martin-­Fumadó C, Gómez-­Durán EL. Suicide investigation: suicideprevent/​en/ psychological autopsy. Span J Leg Med 2017;43:135–7. 14 WHO. Practice manual for establishing and maintaining surveillance 40 Mathes T, Pieper D. Study design classification of registry-­based systems for suicide attempts and self-­harm. WHO, 2016. studies in systematic reviews. J Clin Epidemiol 2018;93:84–7. 15 World Health Organization. The International classification of 41 Rothman KJ, Greenland S, Associate TLL. Modern epidemiology. 3rd diseases, 2019. Available: https://www.​who.​int/​classifications/​icd/​en/ Edition. Hastings Cent Rep, 2014. [Accessed 31 Jul 2019]. 42 Ø B, Breslow NE, Chatterjee N, et al. Handbook of statistical 16 Hedegaard H, Schoenbaum M, Claassen C, et al. Issues in methods for case-­control studies, 2018. developing a surveillance case definition for nonfatal suicide 43 Glenn CR, Lanzillo EC, Esposito EC, et al. Examining the course attempt and intentional self-­harm using International classification of suicidal and Nonsuicidal Self-­Injurious thoughts and behaviors of diseases, tenth revision, clinical modification (ICD-10-­CM) coded in outpatient and inpatient adolescents. J Abnorm Child Psychol data. Natl Health Stat Report 2018:1–19. 2017;45:971–83. 17 Swain RS, Taylor LG, Braver ER, et al. A systematic review of 44 Kim RS, Kaplan RC. Analysis of secondary outcomes in nested case-­ validated suicide outcome classification in observational studies. Int control study designs. Stat Med 2014;33:4215–26. J Epidemiol 2019;48:1636–49. 45 Kim RS. Analysis of nested case-­control study designs: revisiting the 18 Chute CG, Koo D. Public health, data Standards, and vocabulary: inverse probability weighting method. Commun Stat Appl Methods crucial infrastructure for reliable public health surveillance. J Public 2013;20:455–66. Health Manag Pract 2002;8:11–17. 46 Tibshirani R. Regression shrinkage and selection via the LASSO. J R 19 Kukafka R, Ancker JS, Chan C, et al. Redesigning electronic STAT SOC B 1996;58:267–88. health record systems to support public health. J Biomed Inform 47 Krysinska K, Martin G. The struggle to prevent and evaluate: 2007;40:398–409. application of population attributable risk and preventive fraction 20 Casey JA, Schwartz BS, Stewart WF, et al. Using electronic health to suicide prevention research. Suicide Life Threat Behav records for population health research: a review of methods and 2009;39:548–57. applications. Annu Rev Public Health 2016;37:61–81. 48 Walsh CG, Ribeiro JD, Franklin JC. Predicting risk of suicide 21 Schaffer A, Sinyor M, Kurdyak P, et al. Population-­Based analysis attempts over time through machine learning. Clin Psychol Sci of health care contacts among suicide decedents: identifying 2017;5:457–69. opportunities for more targeted suicide prevention strategies. World 49 Zou H, Hastie T. Regularization and variable selection via the elastic Psychiatry 2016;15:135–45. net. J Royal Statistical Soc B 2005;67:301–20. 22 Stene-­Larsen K, Reneflot A. Contact with primary and mental health 50 Rish I. An empirical study of the naive Bayes classifier. IJCAI care prior to suicide: a systematic review of the literature from 2000 workshop on empirical methods in artificial intelligence, 2001. to 2017. Scand J Public Health 2019;47:9–17. 51 Friedman JH. Multivariate adaptive regression splines. Ann. Statist. 23 Vasiliadis H-­M, Ngamini-­Ngui A, Lesage A. Factors associated with 1991;19:1–67. suicide in the month following contact with different types of health 52 Chipman HA, George EI, McCulloch RE. Bart: Bayesian additive services in Quebec. Psychiatr Serv 2015;66:121–6. regression trees. Ann Appl Stat 2010;4:266–98. 24 Morrison KB, Laing L. Adults’ use of health services in the year 53 Breiman L. Random forests. Mach Learn 2001;45:5–32. before death by suicide in Alberta. Heal Reports. 22, 2011. 54 Friedman JH. Greedy function approximation: a gradient boosting 25 Ahmedani BK, Simon GE, Stewart C, et al. Health care contacts in machine. Ann Stat 2001;29:1189–232. the year before suicide death. J Gen Intern Med 2014;29:870–7. 55 Cover T, Hart P. Nearest neighbor pattern classification. IEEE Trans 26 Pirkis J, Nicholas A, Gunnell D. The case for case–control studies in Inform Theory 1967;13:21–7. the field of suicide prevention. Epidemiol Psychiatr Sci 2020;29:e62. 56 Cortes C, Vapnik V. Support-­vector networks. Mach Learn 27 Kim RS. A new comparison of nested case-­control and case-­cohort 1995;20:273–97. designs and methods. Eur J Epidemiol 2015;30:197–207. 57 Ayat S, Farahani HA, Aghamohamadi M, et al. A comparison of 28 Huque MH, Carroll RJ, Diao N, et al. Exposure enriched case-­control artificial neural networks learning algorithms in predicting tendency (EECC) design for the assessment of gene-­environment interaction. for suicide. Neural Comput & Applic 2013;23:1381–6. Genet Epidemiol 2016;40:570–8. 58 van der Laan MJ, Polley EC, Hubbard AE. Super learner. Stat Appl 29 Sitapati A, Kim H, Berkovich B, et al. Integrated precision medicine: Genet Mol Biol 2007;6:Article25. the role of electronic health records in delivering personalized 59 Krautenbacher N, Theis FJ, Fuchs C. Correcting classifiers for treatment. Wiley Interdiscip Rev Syst Biol Med 2017;9:e1378. sample selection bias in two-­phase case-­control studies. Comput 30 Knox KL, Conwell Y, Caine ED. If suicide is a public health problem, Math Methods Med 2017;2017:1–18. what are we doing to prevent it? Am J Public Health 2004;94:37–45. 60 Kessler RC, Stein MB, Petukhova MV, et al. Predicting suicides 31 Horrocks M, Michail M, Aubeeluck A, et al. An electronic clinical after outpatient mental health visits in the army study to assess risk decision support system for the assessment and management of and resilience in Servicemembers (army STARRS). Mol Psychiatry suicidality in primary care: protocol for a mixed-­methods study. JMIR 2017;22:544–51. Res Protoc 2018;7:e11135. 61 Universitat Pompeu Fabra (UPF), Barcelona S. Upf open access 32 Belsher BE, Smolenski DJ, Pruitt LD, et al. Prediction models for publishing, 2019. Available: https://​guiesbibtic.​upf.​edu/​acces-​obert/​ suicide attempts and deaths. JAMA Psychiatry 2019;76:642–11. eng Mortier P, et al. BMJ Open 2020;10:e037365. doi:10.1136/bmjopen-2020-037365 9