The SLaM BRC Case Register and CRIS application

Initial development of the SLaM BRC Case Register has been previously described in detail, as has SLaM as a provider (and see also http://www.slam.nhs.uk).7 In summary, the data are sourced from EHRs used by SLaM, which provides comprehensive mental health services to a geographic catchment of over 1.2 million residents in four south London boroughs—Croydon, Lambeth, Lewisham and Southwark—as well as some regional/national specialist services. SLaM catchment service provision is currently structured within the following specialty groupings: Addictions; Behavioural and Developmental Psychiatry; Child and Adolescent Mental Health Services; Mental Health of Older Adults and Dementia; Mood, Anxiety and Personality; Psychological Medicine; Psychosis. These are aligned with academic groupings at King's College London, reflecting the university–health service partnership enshrined within King's Health Partners Academic Health Sciences Centre (KHP AHSC; http://www.kingshealthpartners.org; also incorporating two major acute care providers). The groupings also encompass services delivered to all age groups, standard specialties such as Addictions, Eating Disorders and Learning Disabilities, as well as provision within Forensic and General Hospital Liaison settings. In addition, wider national provision by SLaM at the time of writing includes the following services: adult attention deficit hyperactivity disorder, adult personality disorder, affective disorders, anxiety disorders (residential), autism assessment and behavioural genetics, brain injury (outpatient and inpatient), anxiety disorders and trauma, chronic fatigue, eating disorders (day care, outpatients, inpatients), female hormone clinic, psychosis (inpatient, outpatient and specialist rehabilitation), mother and baby unit, autism, practitioner health, psychological interventions, psychosexual disorders, self-harm (outpatients) and traumatic stress. Finally, some SLaM services provide to a wider geographic catchment (eg, Addiction services to Bexley and Greenwich boroughs) and others are catchment independent (eg, General Hospital Liaison services are provided to the four Acute Trusts within the catchment regardless of individual patients’ areas of residence).

Clinical records have been fully electronic (ie, paperless) across all SLaM services since April 2006, using the bespoke Patient Journey System (PJS) which incorporated legacy data from earlier service-specific EHRs. The CRIS application was developed in 2007–2008 and consists of a series of data-processing pipelines which both structure and de-identify PJS fields, rendering effectively anonymised data from the full clinical record available at the researcher interface, with search and database assembly functionality facilitated by a front end designed for non-technical use. The anonymisation process and its effectiveness, including the de-identification of open-text fields and the generation of a pseudonymised identifier (CRIS ID), have been described in detail.8 The wider patient-led oversight and security model have also been previously described and have not changed significantly since the SLaM BRC Case Register was set up.7 ,8 Ethical approval as an anonymised database for secondary analysis was originally granted in 2008, and renewed for a further 5 years in 2013 (Oxford C Research Ethics Committee, reference 08/H0606/71+5). In terms of cohort coverage, all SLaM care is represented on CRIS. An opt-out model is in place for service users, and is advertised in all publicity material and initiatives; to date, only three people have requested this.

The SLaM BRC Case Register conforms to the WHO's formal description of a psychiatric case register—a ‘patient-centred longitudinal record of contacts with a defined set of psychiatric services originating from a defined population’,9 although its dynamic nature, updating against source files every 24 h, renders it distinct in some respects. The inclusion of both structured and unstructured (open-text) data in anonymised form, also variously distinguish the SLaM BRC Case Register from other local, regional and national case registries, including those extracted from EHRs such as the disease registries maintained by the US Veteran's Administration.10 ,11 Routinely collected data resources such as the Mental Health Minimum Dataset and Hospital Episode Statistics (HES) for England and Wales overlap with SLaM Case Register data but are limited to prespecified structured fields.

Early experience with CRIS and its influence on subsequent design

Developments in the technical architecture underlying CRIS are summarised in the online supplementary appendix and the current model is displayed in figure 1. Studies published to date using CRIS-derived data have generally fallen into two groups. The first have used a combination of open-text and structured data, with open-text data identified using search terms and then manually coded into numeric form for the purpose of analysis. Because of this, sample sizes have been limited to no more than several hundred. However, productive examples include one of the largest case series assembled of people with neuroleptic malignant syndrome, in order to evaluate the range of diagnostic criteria,12 and associations with antipsychotic exposure,13 as well as a study of factors associated with khat use in a comprehensive sample of Somali mental health service users.14 The second group of studies have used only structured data or have made very limited use of open-text data. These have typically analysed sample sizes of several thousand or more. Examples include studies of residential mobility and of homelessness among inpatients on mental health wards, and a series of investigations of mortality associated with mental disorder, described later.15 ,16

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

Diagram/map of CRIS technical architecture including natural language processing and data linkage. CRIS, Clinical Record Interactive Search; GATE, General Architecture for Text Engineering; SLaM, South London and Maudsley.

Important experiential learning occurred during the initial stages of CRIS use. First, we found that it was sometimes desirable to select and combine data from records in ways that were unsupported by the original CRIS interface (eg, because of complex temporal relationships required between fields). Second, it became clear that while being able to identify and retrieve open-text records according to the presence of prespecified search terms did achieve helpful economy of effort, it did not remove the work needed to generate quantitative data from open text. Indeed, for those projects dependent on the use of open text, the manual coding process placed important limitations on sample size and study duration. Finally, researchers began to develop ideas that required data in addition to those stored in the source EHR, such as data from primary care, acute care and outcomes such as mortality. In the succeeding sections, we set out how the SLaM BRC Case Register has evolved to respond to these challenges.

Handling open text

As outlined above, a priority for development has been to develop more efficient ways of using open-text data in the SLaM BRC Case Register. Early case register data collection included manually reading the de-identified text fields returned by CRIS, such as routine case notes, correspondence and medication notes. For example, one of the recent publications involved manually reading of 2860 records on CRIS of patients receiving acetylcholinesterase inhibitors in order to record their Mini-Mental State Examination (MMSE) scores and respective dates, and other medication prescribed.17 Through this process over 11 000 MMSE scores were ascertained; however, there were significant demands in terms of time and resources and the exercise was only possible as the focus of a PhD studentship. Beyond the efficiencies in manual coding gained by extracting only those records required for coding, through keyword searches and postsearch processing, further gains may be made by displaying text fields in ways that make text of interest easier to see, and by displaying data that are required to be reviewed together in close proximity, and away from other data. For example, in studies of homelessness and residential mobility among inpatients, 4485 admissions were selected according to defined criteria, and free-text records corresponding to these admissions were selected if they contained the terms ‘homeless’, ‘NFA’ or ‘no fixed abode’.15 ,16 The aim was to check structured data on homelessness against free-text data, and if necessary, to supplement the former. SAS was used to insert ‘tags’ that change font colour (red) and weight (bold) for the target words when the data are displayed in Excel, allowing around 2000 free-text progress notes to be coded as homeless/not homeless in less than a day. A SAS Enterprise Guide project developed in collaboration with Amadeus Software Ltd allows CRIS users to do this via a graphical user interface.

A more ambitious approach has been followed for an ongoing project to capture incident cases of psychosis, supported by another Enterprise Guide project developed in collaboration with Amadeus Software Ltd. First, a structured query language (SQL) query retrieves a selection of data for individuals not already present on a cumulative database of first-episode psychotic patients and not already diagnosed as having a psychotic disorder, and whose recent free-text entries contain particular words of interest such as ‘delusion’ or ‘hallucination’. Second, these data are imported into SAS and then automatically outputted in a format suitable for manual coding. This involves splitting data into a multiworksheet Excel workbook, such that each worksheet (tab) contains only data relating to a single person (in the case of our proposed project, each worksheet would similarly pertain to a single episode of care). Targeted words are displayed in colour and in bold.

In contrast to the facilitated, but still manual approaches described above, natural language processing (NLP) techniques have been evaluated and applied for extracting knowledge from unstructured text data. For our purposes, the key NLP technique has been information extraction (IE) where unstructured text is converted into structured tables.18 Such methods promise massive reductions in the time resource required by researchers to unlock information held in clinical notes that in turn may be connected to other parts of the structured record. It was therefore decided, early in the postdevelopment phase, to implement a text-mining capability in CRIS. This was to be generic, in that information to be extracted could not necessarily be foreseen in advance of the design of individual research studies. General Architecture for Text Engineering (GATE) was chosen as the core NLP infrastructure for CRIS.19 ,20 GATE is a widely used suite of open source software for text engineering that includes a workbench for developing applications, tools for distributing those applications on different computer hardware architectures, a quality assurance suite and facilities for manual preparation of example data.19–21 GATE’s origins are in clinical IE and it has been widely applied in this context.22 ,23 GATE includes a flexible architecture for IE and text mining, a large set of pluggable text processing components, and graphical tools for organising those components into new applications. The GATE suite also includes tools for text-mining workflow, distributed processing and visualisation. A variety of text processing tools and document formats may be plugged into this architecture, with individual tools being chained together into processing ‘pipelines’, and documents processed in series through these pipelines.

Two distinct shallow language processing methodologies have been adopted for CRIS development, in collaboration with University of Sheffield Department of Computer Science. The first may be described as rule-based pattern matching of key concepts. Sentences are first processed to find and create annotations based on simple surface linguistic information (such as words, sentences, etc). This step is then followed by the process of finding concept-specific keywords, which are used to recognise likely sentences of importance to the IE task. For example, in an application to determine the smoking status of a patient implied by texts, such a dictionary, might list the terms of common tobacco products and activities—‘cigarette’, ‘smoker’, etc. Finally, a set of patterns specific to the text-mining task are run over the previously generated annotations in order to create a final annotation containing all of the information required in a readily extractable format. The challenge of the pattern matching approach is that it is knowledge intensive. A successful series of patterns need to be developed in relation to a specific IE task (eg, to extract medications, educational level or particular test results). They have to be built manually by GATE users with language engineering skills, using definitions agreed with clinicians and epidemiologists. A sample of the output from an initial prototype application is then corrected by a clinician or epidemiologist, which in turn is used to stimulate discussion about requirements and to provide a basis for multiple iterations of development until performance requirements are met. An advantage of this IE approach is that it also allows researchers to combine information available from open text and structured fields available in CRIS, through SQL, thus combining multiple sources of information. At the postprocessing stage, we can further apply specific filtering criteria to data extraction, such as frequency and length of prescribing and number of concomitant drugs, thus identifying more complex patterns in the text, such as antipsychotic medication profiles (ie, antipsychotic polypharmacy).24

Because of the lengthy development cycles of building shallow parsing algorithms, a second IE methodology has also been evaluated. Here, support vector machines (SVMs) are used to rapidly achieve respectable results for certain types of IE problem. A SVM is a machine-learning technique where the intention is to represent instances of text as vectors in high dimensional space. With a training set of instances labelled as indicative of a desired class, the SVM implementation in GATE generates a hyperplane which can in turn be used to classify unseen instances pertaining to the described class in the training set. In practice, this primarily uses a technique known as ‘bag of words’, where the occurrence of single words within a sentence is the principal currency used to distinguish the various classes. The first part of the model construction requires an expert (eg, clinician) to review a set of documents and label sentences which are relevant to the concept in question, in much the same way that they might signal to a language engineer the relevance of a given sentence for a pattern-based approach. The combination of labelled and unlabelled sentences forms the training data, from which the SVM learns the classification function. This model is then applied to unseen data, and the model quality assessed by human review. If required, further training data can be supplied, which may involve an active learning-inspired approach. A limitation with SVMs applied in CRIS has been that they have limited suitability for complex data extraction problems; however, in scenarios where the assertion to be extracted is simple and tend to be restricted to a concise set of clinical language, performance has been found to be very good and IE applications with immediate utility can be rapidly developed.25 The TextHunter program was designed specifically to aid the process of clinical text annotation in CRIS, providing an easy-to-use interface for annotators with a focus on the sentence containing the word(s) of interest and immediately proximal text and functionality for rapid coding into discrete groups, typically comprising the following: (1) positive (ie, implying that the construct is present); (2) negative (ie, a statement indicating that the construct is absent); and (3) irrelevant text.26 Additional TextHunter functionality includes platforms for interannotator agreement testing, and the creation of gold standard and test annotation sets.

Whether rules-based or machine-learning approaches are used, separate training and test data sets are constructed. Standard metrics for evaluating IE application performance in the test data sets, at the level of the individual text annotation, comprise precision (equivalent to positive predictive value; the proportion of IE application ‘hits’ which are found to identify the genuine construct) and recall (equivalent to sensitivity; the proportions of instances of the genuine construct which are identified by the application). Employing text mining within the CRIS data set has involved a trade-off between the two. However, the longitudinal nature of EHR data means that there are generally multiple opportunities for an NLP application to capture a piece of information; therefore, suboptimal recall can be compensated for and the focus has been on maximising precision. For the purpose of precision and recall testing, there are two reportable outcomes. The first is ‘annotation level’, which is carried out across randomly selected documents and is an indicator of the base level of performance of the application. This figure is useful for developmental purposes, or, in the case of simple concepts that do not require postprocessing, for estimating the final performance of the algorithm. The second type of precision and recall are ‘currency level’, measuring performance after postprocessing.

The SLaM Clinical Data Linkage Service

SLaM comprises one part of the KHP AHSC (established with King's College London, Guy's and St Thomas’ and King's College Hospitals NHS Foundation Trusts) and received National Institute of Health Research (NIHR) funding to set up a service to meet the growing demand from SLaM and KHP researchers whose projects require linked data extracts. SLaM consequently established the Clinical Data Linkage Service (CDLS) as a trusted third party safe haven set up to enable safe and secure data processing services (linkage, and/or storage, and/or extraction) on distinct data sets for secondary research use. The two main methods of linkage have involved either (1) CDLS performing a secure linkage using deterministic or probabilistic matching if/as required or (2) CDLS supporting another trusted third party service to perform the linkage outside of the SLaM electronic firewall followed by CDLS receiving the linked data afterwards (eg, CRIS-HES linkage). Linked data are stored by CDLS in accordance with the SLaM ICT Security Policy and a set of standards contained in a CDLS Memorandum of Understanding completed by the data controllers providing data to individual projects, prior to undertaking any data processing for the project. Linked data are stored on a CDLS server within the SLaM firewall. To date, linkages have been successfully carried out between CRIS and a number of databases, described below.

Cohort characteristics

Initial descriptive data were assembled on the catchment area for SLaM (Croydon, Lambeth, Lewisham and Southwark) using publicly available sociodemographic information from ONS census data.30 Analyses of CRIS data used 31 December 2014 as a census date for descriptive statistics including sociodemographic and diagnostic profiles. ‘Active’ patients on this date were defined as those who had been referred to and accepted by SLaM and had not been discharged by 31 December 2014. ‘Inactive’ patients had a recorded activity date on or before 31 December 2014 and excluded referrals categorised as ‘rejected’ or ‘waiting’. On 31 December 2014, 223 224 patient records were available on CRIS, of which 31 961 described ‘active’ patients and 191 263 ‘inactive’. The remaining 21 882 records described referrals, which were either solely characterised as ‘rejected’ or ‘waiting’, and in which no team episode (for outpatients) or ward stay (for inpatients) was indicated. Descriptive data were further provided for key linked data sets at that time. In this respect, the most recent mortality date recorded in the linked ONS mortality data set was 16 December 2013; cancer registry data were linked up to 31 December 2008; HES data were available to 31 March 2013. For analyses of linked HES data, contacts with mental health services were excluded.

Descriptive data from the UK Census for the catchment populations served by SLaM are summarised in table 1 and contextualised with the same information for London as a whole and for England. There are slight differences in population structure between the four boroughs served, with Croydon having higher proportions of young children and older residents compared with London and the other three boroughs. Highest proportions in the young adult (20–39 year) age range were living in Lambeth and Southwark. As a whole, the SLaM catchment has a slightly higher predominance of working adults in the 20–59-year range compared with London, and shares with London lower proportions in older age ranges compared with England. The SLaM catchment has substantially higher proportions of residents from minority ethnic groups and/or born outside UK compared with England, whereas compared with London as a whole, there are higher proportions from black minority groups and lower proportions from Asian groups. In common with London as a whole, proportions are higher in both highest and lowest socioeconomic groups compared with England; proportions in unemployment are higher, but so are proportions with higher levels of education. Of the catchment boroughs, Lambeth, Southwark and Lewisham have higher levels of both in-migration and out-migration compared with Croydon. Based on the ratios between summed borough statistics and those for the catchment overall, 76.9% of inflow migration and 78.5% of outflow migration was from/to areas outside the catchment, rather than between catchment boroughs.

Geographic characteristics are summarised in figures 2⇓–4. Figure 2A visually contextualises deprivation levels in SLaM compared with other areas of London, and figure 2B summarises the most recently recorded residence of active SLaM patients. In the latter, most active SLaM patients were identified as residing within its geographic catchment, although appreciable numbers were drawn from a wider geography. Within the SLaM catchment, higher numbers of active patients were generally found in areas of higher deprivation, although several anomalous areas can be seen—for example, those with high deprivation and relatively low numbers of active patients (figure 3A, B). Figure 4 illustrates the most recent recorded residence of non-active patients in London (figure 4A) and specifically in SLaM's catchment (figure 4B). Outside SLaM's catchment, relatively high numbers of inactive patients were recorded as residing in neighbouring local authorities in South East London including Bexley, Greenwich and Bromley.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

Maps contextualising deprivation levels in the South London and Maudsley (SLaM) catchment compared with London as a whole, and illustrating the distribution of recorded residences for active patients (on 31 December 2014) within London.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3

(A, B) Maps showing distribution of deprivation levels in the four catchment boroughs served by South London and Maudsley (SLaM), the key hospital sites and the number of active patients (on 31 December 2014) across the same geography.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4

Maps illustrating the distribution of recorded residences for inactive patients (on 31 December 2014) within London and SLaM catchment area. LSOA, lower super output area; SLaM, South London and Maudsley.

Descriptive data are summarised in table 2 for all people who were represented on the SLaM BRC Case Register on 31 December 2014. Higher proportions of active patients were 80 years and older and in the 40–59 group compared with proportions in the four catchments. Compared with the catchment area characteristics described in table 1, active SLaM patients had a slightly higher male predominance, and there were higher proportions self-assigning as white, mixed or other ethnicity. Around 70% were single. Employment status data were available on less than 25% of the active sample, but of this group around 66% were unemployed. Of active SLaM patients on the census date, 6574 (20.9%) were either residing in boroughs outside London or living in London but outside SLaM's four catchment boroughs. Of these, 3385 (51.5%) were in contact with SLaM services that provided for other boroughs, 1941 (29.5%) were using one or more of SLaM's national services, 341 (5.2%) were in contact with General Hospital Liaison services covering one of the four Acute Trusts within the SLaM catchment, and 907 (13.8%) were previous catchment residents currently living outside the catchment (193 of whose addresses were recorded as temporary).

On the 31 December 2014 census date, there were nearly 32 000 active cases receiving care from SLaM services, with the largest numbers receiving care from Psychosis or Child and Adolescent Mental Health Services (table 3). A further 190 000 plus patients on the SLaM BRC Case Register were inactive to SLaM, nearly one-third of whom received care from Psychological Medicine services (which includes General Hospital Liaison services). Table 4 provides an additional description of overlap between services for active and inactive patients, with over 1000 active patients in contact with two or more specialties concurrently and over 15 000 inactive patients having received care from two or more specialties. Ever-recorded primary diagnoses are summarised in table 5. Of active patients, the most common mental disorder diagnoses ever recorded were schizophrenia (21.2%) and mood (19.0%) disorders, followed by organic (11.0%), substance use (11.7%) and neurotic (13.0%) disorders, and disorders of childhood and adolescence (11.3%). Sizes of data linkage samples are described in tables 6⇓–8. Nearly 85% of CRIS patients had records in HES (excluding mental health service data) and nearly 2% of CRIS patients had data linked to those from the cancer registry within the years of data availability (table 6). Distributions of underlying cause of death are summarised in table 9 for the linked sample with this information, and primary cancer diagnoses are similarly described in table 10.

Performance of NLP applications

Performances of IE applications to date are summarised for CRIS as a whole, supplementary to more detailed publications on some of these.31 ,32 ,33 ,34 The first NLP IE application to be developed was for the MMSE, a commonly used 0–30-point assessment of global cognitive function. The objective of the application was to ascertain both the numerator and denominator scores (because denominator scores of less than 30 are used where some items cannot be attempted because of, eg, sensory impairment), as well as the date implied for the assessment (because clinical text fields commonly refer to previous as well as current scores). Further rules for application postprocessing were that only MMSE scores with denominators over 25 were included (because scores below that level imply substantial missing data and a scale that was probably incompletely administered), and scores were excluded if two different numerators were assigned to the same date.34 The application for educational attainment sought to ascertain the numeric value associated with text commenting on school leaving age, whether the age itself or the year, and the application for ‘living alone’ simply sought to identify that phrase or equivalents applied to the patient. In developing the smoking application, authors extracted information from open-text fields, classifying patients as either ‘currently smoking’, ‘past smoker’ or ‘has never smoked’, with smoking of substances other than tobacco (eg, marijuana/cannabis and cocaine) specifically excluded.31 The methodology used an iterative process of manual ‘gold standard’ annotation of free-text documents, followed by comparison with the results generated by the application at each development stage, with analysis of this comparison feeding further development of the rules. The application for ‘diagnosis’ sought simply to extract any text strings associated with a diagnosis statement in order to supplement the existing structured (International Classification of Diseases (ICD)-10) fields. Its performance was evaluated formally in a random sample of 75 documents for ‘vascular dementia’,33 but is recommended for individual further evaluation in other conditions. The application for ascertaining pharmacotherapy was developed using a gazetteer of generic and commercial names for all medications in UK use in order to ascertain instances where the patient was reported as receiving these, with supplementary rules for ascertaining recorded dose, frequency/timing and starting/stopping statements. Its precision was first tested for clozapine receipt against a manual search of 279 documents, and recall was ascertained on a random set of 200 documents containing the word clozapine and scrutinised to ascertain an actual prescription.32 Finally, the validity of this application was recently further evaluated for six antipsychotic agents (amisulpiride, flupentixol, haloperidol, olanzapine, risperidone, zuclopenthixol) on instance level (ie, specific mentions in the text at individual points in time). To estimate precision and recall, the authors examined a subset of 20 patients for each medication, totalling 120 patients (the instances of antipsychotic prescribing varied from 328 to 1150 instances by antipsychotic agent) by running the NLP application over the set of unseen documents and comparing the results to the manual coding of the same data set.24 For all evaluations, an F-statistic was additionally calculated, representing the harmonic mean of precision and recall, and defined as: F=2×(precision×recall/(precision+recall)). As with the diagnosis application, further bespoke validation of the pharmacotherapy application is recommended for new medications or classes. Performance data are summarised for NLP IE applications in table 11, and table 12 describes the resulting additional structured data points generated across CRIS using these applications.

Findings to DATE

The SLaM BRC Case Register has been used for a wide range of research projects to date, as well as for key service evaluation and audit projects, and over 50 publications have arisen. Large-scale outcome studies supported by CRIS data have included those of residential mobility and of homelessness among inpatients on mental health wards.15 ,16 Evaluations of service interventions and other quality markers were also studied,35 ,36 and investigations are increasingly focusing on early symptoms and treatment pathways in psychosis.37 ,38 Keyword search functionality recently supported a large historic cohort study of service use and abuse experiences of trafficked people in contact with secondary mental health services.39

A particularly prominent theme has been the investigation of mortality and physical health outcomes in people with mental disorders. Initial reports highlighted the raised mortality and lower life expectancy of people in the most common disorder groups.40–43 More studies were carried out to attempt to profile those most at risk, which have indicated that disability and environmental circumstances appear to be more important than symptoms.44 ,45 This was supported by a study showing that, in those who received specific structured risk assessments, clinician-perceived risk of self-neglect was a strong and independent predictor of mortality, whereas clinician-perceived risks of suicide and/or violence were not predictive.46 In terms of mortality predictors in specific patient groups, the impact of psychiatric comorbidity and psychological health on all-cause and cause-specific mortality in opioid use disorder has been evaluated, highlighting the importance of personality disorder and comorbid alcohol use disorder.43 Similarly, the importance of alcohol and drug use, physical illness, and functional impairment as predictors of mortality in individuals with personality disorder has been demonstrated, a group with and life expectancies at birth reduced by 17–19 years compared with the general population in England and Wales.47 ,48 Mortality outcomes have been further evaluated in studies of cognitive impairment and delirium in older adults.34 ,49

Studies of pharmacotherapy profiles have continued investigations into mortality as an outcome, most notably in a report identifying a marked reduction in people using clozapine, not explained by a range of potential confounders including service use.32 Another study found that atypical antipsychotic agents were not associated with higher mortality in people with vascular dementia.33 Further work will examine antipsychotic polypharmacy in more detail, following recent successful development of algorithms to capture this.24 As described earlier, utilising the keyword search functionality in CRIS, exposure to non-pharmacological agents such as khat was investigated,14 and a large series of cases with suspected neuroleptic malignant syndrome were successfully identified which allowed a matched case–control study of antipsychotic exposures potentially responsible.12 ,13 The association between antidepressant use and risk of mania and bipolar disorder has also recently been investigated,50 as has antipsychotic use in children and adolescents with autistic spectrum disorder.51 Finally, the potential to use extensive routine data to monitor treatment response was exemplified in a recent study of people receiving acetylcholinesterase inhibitor treatments for Alzheimer's disease in which trajectories of cognitive function were plotted before and after treatment initiation in order to identify predictors of ‘response’—to our knowledge, the largest and most extensive cohort of its kind.17

Recent developments which are likely to generate substantial future output include the assembly of one of the largest cohorts to date of women with severe mental disorder who are followed from preconception and pregnancy to investigate medication use in relation to maternal and fetal outcomes.52 Supplementing CRIS-derived outcomes to large clinical research samples with genetic profiling has also begun to generate novel output, for example, indicating that a well-recognised genetic risk factor for schizophrenia may also be a risk factor for worse clinical outcomes after diagnosis.53 NLP applications have recently been extended to cover a range of affective and psychotic symptoms, allowing much more detailed phenotyping of large samples than a diagnosis alone provides,54 ,55 and a range of adverse drug events have also recently been successfully captured.56