Table A1.

Overview of included studies.

References	Clinical research question(s)	Data source and type	Number of patients and/or documents included in analysis	Free text information extraction method and purpose	Summary of clinical findings	Statement of potential clinical impact	Theme of benefit
Anwar et al. (40)	To investigate if a decision support system be produced using EHRs to predict type of hearing aid prescription.	Hospital patient records; including audiograms categorical data and free text notes	Over 23,000 patients Over 180,000 documents.	Manual extraction of notes made by hearing aid technicians.	The decision support system was able to replicate the decisions of audiologists whether to fit an in the ear or a behind the ear hearing aid with high precision.	A decision support system was produced to predict prescription of different hearing aids alongside an explanation demonstrating how that decision was arrived at. Authors conclude this will provide a useful “second opinion” for audiologists.	Clinical decision support
Bogdanowicz et al. (66)	To investigate clustering of all-cause and overdose deaths after a transfer of patients to alternative treatment provider and after the end of opioid substitution therapy (OST) in opioid-dependent individuals in specialist addiction treatment.	South London and Maudsley (SLaM) Biomedical Research Center (BRC) Case Register using Clinical Records Interactive Search (CRIS)	5,335 patients and 216 documents.	NLP algorithms to extract diagnoses, start and end dates of treatment episodes, medication and reasons for treatment cessation from text.	Opioid-dependent people who are transferred to an alternative treatment provider for continuation of their opioid substitution therapy experience high overdose mortality rates, with substantially higher rates during the first month.	Any transfer of patients whether due to escalation of treatment (e.g., to an in-patient unit) or as part of successful recovery (e.g., from an in-patient unit to rehabilitation care) needs to be undertaken with caution.	Service improvement
Cea-Soriano et al. (41)	To investigate if there is an association between the use of non-insulin antidiabetics in early pregnancy and the risk of miscarriages, stillbirths and major structural malformations.	The Health Improvement Network (THIN)	1,511 pregnancies in women with pre-gestational diabetes linked to livebirths; one full record per patient.	Manual extraction of free text to identify infant malformations	In pregnant woman with diabetes, use of non-insulin antidiabetic agents in early pregnancy had no link with greater risks of fetal losses or major malformations than insulin.	Adverse outcomes to infants born to women with type 1 diabetes can be predicted and prevented by the analysis of EHRs including free text.	Drug prescribing safety
Chang et al. (54)	To assess how intellectual disability (ID) affects the risks of severe respiratory diseases leading to hospital admission and how much worse their treatment outcomes are by comparing critical indicators for service usage	SLaM BRC CRIS Case Register linked to Hospital Episodes Statistics (HES)	3,138 patients; one full record per patient.	Generalized Architecture for Text Engineering (GATE) for NLP to supplement diagnostic codes for case identification.	Respiratory system disease admissions in adults with ID are more frequent, of longer duration and have a higher likelihood of recurring.	More intervention studies are needed to generate prevention strategies specifically for respiratory system diseases in primary and secondary care for adults with ID.	Risk factors for disease & outcomes
Chang et al. (55)	To assess the stage at cancer diagnosis and survival after cancer diagnosis among people served by secondary mental health services, compared with other local people.	SLaM BRC CRIS Case Register linked to Thames Cancer Registry	28,477 patients; one full record per patient	GATE for NLP to supplement diagnostic codes for case identification.	No link was found between mental disorder diagnoses and spread of cancer at presentation. However, people with severe mental disorders, depression, dementia and substance use disorders had significantly worse survival after cancer diagnosis, independent of cancer.	EHRs can be used in observational research to identify outcomes of individuals with mental disorders and other co-morbidities, such as cancer.	Risk factors for disease & outcomes
Colling et al. (79)	To investigate what proportion of patients with psychosis received CBT as compared to published audits and explore whether demographics predicted receiving CBT.	SLaM BRC CRIS case register	2,579 and 2,308 service users from two published audits, one full record per patient.	NLP techniques to identify episodes of CBT for psychosis.	Younger, white patients with a diagnosis of other schizophrenia spectrum and schizoaffective disorder when compared with schizophrenia were significantly more likely to have received CBT. Free text analysis improved accuracy of identifying patients.	The method provides a useful way for evaluating delivery of CBT in persons with psychosis on a large scale. It provides quality and safety service improvement by providing more scope for routine monitoring, and the promotion of equitable access between different demographic groups.	Service improvement
Colling et al. (67)	To investigate whether EHRs predict priority service outcomes in a mental health setting and whether NLP enhance these predictions, through the evaluation of three outcomes: (1) inpatient duration (2) readmission following inpatient care (3) high-intensity service use following first referral	SLaM BRC CRIS case register	(1) 808 patients (2)1,650 patients (3) 4,494 patients, one full record per patient.	NLP using TextHunter to develop algorithms on GATE for symptoms and medications	For extended duration of hospital admission and readmission following hospital admission (outcomes 1 and 2), predictive models did not achieve adequate levels of performance. Predicting high-intensity service use following a first referral showed promisingly sustained performance from development to evaluation data. NLP improved the predictions	EHRs can lead to quality service improvement by improving routine clinical predictions through utilizing previously inaccessible data.	Service improvement
Das-Munshi et al. (56)	To estimate excess mortality for people with severe mental illness across different ethnic groups, and to assess the association of ethnicity with mortality risk.	SLaM BRC CRIS case register with linked mortality data from Office of National Statistics (ONS)	18,201 patients; one full record per patient.	NLP techniques using GATE to supplement diagnosis codes for case identification.	People with severe mental illness have a higher excess mortality relative to the general population regardless of ethnicity. Among those with severe mental illness, some ethnic minorities have lower mortality than the white British group, for which the reasons deserve further investigation.	EHRs can be used in observational research to investigate exposure-outcome association in large mental healthcare setting.	Risk factors for disease & outcomes
Downs et al. (57)	To investigate the prevalence of negative symptoms (NS) at first episode of early-onset psychosis (EOP), and their effect on psychosis prognosis.	SLaM BRC CRIS case register	1,033 patients, 618 records	NLP techniques using GATE to supplement diagnostic codes for case identification, identify antipsychotic prescriptions, identification of multiple treatment failure.	Findings support the hypothesis that presence of these symptoms around the first stages of the illness identify a subset of children and adolescents who may be at higher risk of responding poorly to antipsychotics, both through refractory symptoms and high sensitivity to side-effects	Optimization of current pharmacological and non-pharmacological strategies for these patients, and further research involving agents that better target NS are warranted.	Risk factors for disease & outcomes
Fernandes et al. (68)	To investigate the demographic and clinical factors associated with antidepressant use for depressive disorder in a psychiatric healthcare setting	SLaM BRC CRIS case register	1,561 patients; one full record per patient.	NLP techniques using GATE to identify medications	Age, past medication and/or psychotherapy receipt use and symptom profiles in the past 12 months were associated with antidepressant receipt in secondary mental healthcare.	Continual monitoring of treatment choices in this cohort may contribute to providing optimal care for secondary care patients	Drug prescribing safety
Fusar-Poli et al. (42)	To compare the clinical outcomes in first episode of psychosis (FEP) patients who presented to either high risk or conventional mental health services.	SLaM BRC CRIS case register	2,943 patients; one full record per patient.	Manual review of free text for case identification.	FEP patients who had presented to a high-risk service spent fewer days in hospital, had a shorter referral to-diagnosis time, a lower frequency of admission and a lower likelihood of compulsory admission in the 24 months following referral, as compared to FEP patients who were first diagnosed at conventional services.	This study provides the first evidence that services designed for people at high risk of psychosis may be associated with better outcomes in patients who are already psychotic, but were referred because they were thought to be at high risk.	Service improvement
Jewell et al. (83)	To determine whether dynamic factors related to behavior, cooperation with treatment, and activities on the ward are associated with outcome at Mental Health Review Tribunals (MHRTs) in forensic psychiatric patients.	SLaM BRC CRIS case register	79 patients; one full record per patient.	NLP techniques using GATE to identify violence risk assessment score.	Unescorted community leave, responsible clinician's recommendation of discharge, and restricted Mental Health Act section were associated were greater likelihood of discharge at tribunal. Inpatient aggression, recent episode of acute illness, higher scores on violence risk assessment and agitated behavior were all negatively associated with discharge at tribunal. Historical Clinical Risk scores and attempted or actual physical violence uniquely predicted outcome after controlling for other dynamic variables.	By identifying dynamic factors in EHRs associated with discharge at tribunal, the results have important implications for forensic psychiatric patients and their clinical teams to improve their chances of discharge at a MHRT	Service improvement
Kadra et al. (69)	This study aimed to investigate socio-demographic, socioeconomic, clinical, and service-use predictors of long-term	SLaM BRC CRIS case register	6,857 patients; one full record per patient.	NLP techniques using GATE to identify medication prescribing and supplement diagnosis codes.	Findings highlight that certain patient groups are at an increased risk for long-term APP initiation.	Identifying the patients who are at an increased risk for long-term APP treatment earlier in their treatment could encourage clinicians to employ a broader range of interventions in addition to pharmacotherapy to reduce the risk of APP prescribing.	Service improvement
		APP initiation in serious mental illness (SMI).
Kadra et al. (70)	The aim of this study was to determine if there was an association between being discharged on antipsychotic polypharmacy (APP) and risk of readmission into secondary mental health care.	SLaM BRC CRIS case register	5,523 patients; one full record per patient.	NLP techniques using GATE to supplement antipsychotic medications and diagnostic data and to identify medication non adherence.	Being discharged on APP was associated with a significantly increased risk of readmission, in comparison to patients discharged on monotherapy. Patients receiving clozapine antipsychotic polypharmacy (APP) were at a significantly increased risk for readmission in comparison to patients on clozapine monotherapy.	Patients discharged on APP are more likely to be readmitted into hospital within 6 months in comparison to those discharged on monotherapy. This needs to be considered in treatment decisions and the reasons for the association clarified.	Drug prescribing safety
Kadra et al. (71)	To investigate the association between long-term antipsychotic polypharmacy use and mortality; and determine whether this risk varies by cause of death and antipsychotic dose	SLaM BRC CRIS case register linked with ONS mortality data	10,945 patients; one full record per patient.	NLP techniques using GATE to identify medication data including dose, and comorbid diagnoses	Patients on long-term antipsychotic polypharmacy (APP) had a small elevated risk of mortality, which was significant in some but not all models.	The findings suggest that the effect of long-term APP on mortality is not clear-cut, with limited evidence to indicate an association, even after controlling for the effect of dose	Drug prescribing safety
Knapp et al. (58)	To explore the relationships between clinical and other characteristics of people with Alzheimer's disease living in the community and likelihood of care home or hospital admission, and associated costs.	SLaM BRC CRIS case register linked to HES	3,075 patients; one full record per patient.	NLP techniques using GATE to supplement diagnosis codes for case identification, identify MMSE score and care home admission.	The authors were able to predict probability of care home or hospital admission and/or associated costs over 6 months.	EHRs can be used in observational research to predict exposures, such as clinical and demographic characteristics with outcomes, such as destination of care, in a mental healthcare setting.	Risk factors for disease & outcomes
Legge et al. (76)	To investigate the risk factors, reasons and timing of clozapine discontinuation in patients with treatment-resistant schizophrenia (TRS).	SLaM BRC CRIS case register	316 patients; one full record per patient.	NLP techniques using GATE to identify clozapine use and reasons for discontinuation	Adverse drug reactions (ADRs) accounted for over half of clozapine discontinuation. Sedation was the most common ADR cited as a reason for discontinuation and the risk of discontinuation due to ADRs was highest in the first few months of clozapine treatment. High levels of deprivation in the neighborhood where the patient lived were associated with increased risk of clozapine discontinuation.	It is important that clinicians identify and treat ADRs attributed to clozapine, particularly in the first few months after treatment onset, before they lead to discontinuation. Patients who live in an area of high deprivation are at an increased risk of discontinuing clozapine and may need additional support to maintain engagement with treatment.	Drug prescribing safety
Leniz et al. (59)	To investigate determinants of end-of-life hospital transitions, and association with healthcare use, among people with dementia.	SLaM BRC CRIS case register linked to ONS mortality data	8,800 patients; one full record per patient.	NLP techniques using GATE to supplement dementia identification and MMSE score.	Early transitions were associated with more hospital admissions throughout the last year of life compared to those with late and no transitions.	In contrast to late transitions, early transitions are associated with higher healthcare use and characteristics that are predictable, indicating potential for prevention.	Service improvement
Maguire et al. (47)	To evaluate whether primary care EHRs from patients with severe asthma can be used to identify allergic bronchopulmonary aspergillosis (ABPA) cases.	Clinical Practice Research Datalink (CPRD)	21,054 patients; one full record per patient.	Keyword search and manual check of free text to identify ABPA diagnosis, symptoms tests and treatment.	From the observed concurrence of keywords, the authors were able to devise an algorithm to identify cases of ABPA with varying degrees of specificity. Inclusion of free text improved identification of ABPA.	Clinical Decision Support tools can be produced from EHRs to identify ABPA cases in severe asthmatics. This tool could be extrapolated identify other rare conditions and to quantify their potential burden.	Clinical decision support
Mansour et al. (60)	To compare symptoms and types of treatment between ethnic groups in patients with late-life depression.	SLaM BRC CRIS case register	5,546 patients; one full record per patient.	NLP techniques using GATE to supplement diagnosis codes for case identification, identify depressive symptoms, CBT and medications.	Black Africans and Black Caribbeans more frequently presented with psychotic problems and were significantly less likely to have anti-depressant treatment prescribed post-diagnosis compared to White British.	Ethnic minority elders have significantly different presentations and undertake different types of treatment both across groups and relative to their White British counterparts. These differences need to be taken into consideration to optimize pathways into care and to personalize treatment.	Risk factors for disease & outcomes
Mueller et al. (61)	To investigate associations between acetylcholinesterase inhibitors (AChEI) prescription and mortality in patients with Alzheimer's dementia (AD) in a naturalistic setting.	SLaM BRC CRIS case register linked to HES and ONS mortality data	3,199 patients; 2,464 records.	NLP techniques using GATE to supplement diagnosis codes for case identification, medication prescriptions, MMSE score	A strong association between AChEI receipt and lower mortality was observed.	In a large cohort of patients with AD, AChEI prescription was associated with reduced risk of death by more than 20% in adjusted models. This has implications for individual care planning and service development.	Risk factors for disease & outcomes
Mueller et al. (62)	To describe the risk and duration of hospital admissions in patients with dementia with Lewy bodies (DLB), and compare these to those in Alzheimer's disease (AD) and the general population	SLaM BRC CRIS case register linked to HES data	10,159 patients; 970 records.	NLP techniques using GATE to supplement diagnosis codes for case identification.	Patients with DLB are more frequently admitted to general hospitals and utilize inpatient care to a substantially higher degree than patients with AD or the general elderly population.	This data highlights an opportunity to reduce hospital days by identifying DLB earlier and providing more targeted care.	Service Improvement
Mueller et al. (63)	To investigate associations between polypharmacy and emergency department attendance, any and unplanned hospitalization, and mortality in patients with dementia.	SLaM BRC CRIS Case Register linked to HES and ONS mortality data.	4,668 patients; one full record per patient.	NLP techniques using GATE to supplement dementia diagnosis codes for case identification and identify polypharmacy and MMSE score.	Polypharmacy is associated with an increased risk of emergency department attendance, and both any and unplanned hospitalization, as well as mortality. There is also a dose-response relationship between number of medications prescribed and negative outcomes.	Increased risks associated with polypharmacy may have considerable public health and healthcare delivery consequences.	Risk factors for disease & outcomes
Mueller et al. (64)	To investigate the association between use of antidepressants and mortality in people diagnosed Alzheimer's disease (AD).	SLaM BRC CRIS Case Register linked to HES and ONS mortality data.	5,473 patients; one full record per patient.	NLP techniques using GATE to supplement diagnosis codes for case identification, medication and MMSE score	Prescription of an antidepressant, before or after dementia diagnosis, was associated with higher mortality. Risks remained significant in patients without neuropsychiatric symptoms.	Prescription of antidepressants around time of dementia diagnosis may be a risk factor for mortality.	Drug prescribing safety
Mukadam et al. (77)	To investigate interethnic differences in cognitive scores and age at dementia diagnosis.	SLaM BRC CRIS case register	13,789 patients; one full record per patient.	NLP techniques using GATE to supplement MMSE score	In general, ethnic group distributions in referrals did not differ substantially from those expected in the catchments	People from black and Asian groups were younger at dementia diagnosis and had lower MMSE scores than white referrals.	Risk factors for disease & outcomes
Patel et al. (74)	To identify negative symptoms of patients with schizophrenia and assess their relationship with clinical outcomes.	SLaM BRC CRIS case register	7,678 patients; one full record per patient.	NLP techniques using GATE to identify negative symptoms.	Negative symptoms were common and associated with adverse clinical outcomes, consistent with evidence that these symptoms account for much of the disability associated with schizophrenia.	Observational research can be conducted using EHRs in large representative samples of patients, using data recorded during routine clinical practice.	Risk factors for disease & outcomes
Patel et al. (75)	To assess the impact of mood instability on clinical outcomes in patients with a psychotic, affective or personality disorder.	SLaM BRC CRIS case register	27,704 patients; one full record per patient.	NLP techniques using GATE & TextHunter to identify mood instability.	Mood instability occurs in a wide range of mental disorders and is not limited to affective disorders. Mood instability is associated with poor clinical outcomes. Clinicians should screen for mood instability across all common mental health disorders.	Observational research can be conducted in a mental healthcare setting using EHR's to determine the impact of mood instability on patient outcomes.	Risk factors for disease & outcomes
Patel et al. (84)	To investigate factors associated with the delay before diagnosis of bipolar disorder and the onset of treatment in secondary mental healthcare	SLaM BRC CRIS case register	1,364 patients; one full record per patient	NLP techniques using GATE to supplement diagnosis and medication information.	Some individuals experienced a significant delay in diagnosis and treatment of bipolar disorder after initiation of specialist mental healthcare, particularly those who had prior diagnoses of alcohol and substance misuse disorders.	These findings highlight a need for further study on strategies to better identify underlying symptoms and offer appropriate treatment sooner in order to facilitate improved clinical outcomes, such as developing specialist early intervention services to identify and treat people with bipolar disorder.	Service improvement
Patel et al., (80)	To investigate whether cannabis use is associated with increased risk of relapse, and whether antipsychotic treatment failure may mediate this effect in patients with first episode psychosis (FEP).	SLaM BRC CRIS case register	2,026 patients; one full record per patient.	NLP techniques using GATE & TextHunter to identify cannabis use	The use of cannabis in patients with FEP was associated with an increased likelihood of hospital admission and antipsychotic treatment failure did not mediate this.	EHRs can be used to conduct observational research in a mental healthcare capacity, for example to determine that cannabis use (exposure) is associated with various outcomes (relapse risk and antipsychotic treatment failure).	Risk factors for disease & outcomes
Price et al., (46)	To estimate alarm symptoms for cancer recorded in free text.	Clinical Practice Research Datalink (CPRD)	4,915 bladder and 3,635 pancreatic cancer cases; 16,459 control group EHRs; one full record per patient.	Keyword search for symptoms followed by algorithm run in Stata.	Omission of text records from CPRD studies introduces bias that inflates outcome measures for recognized alarm symptoms.	The addition of free text when analyzing EHRs improves accuracy and therefore results in potential improved understanding of symptoms and improvement in recognition of cancer symptoms in general practice.	Service improvement
Ramu et al. (72)	To investigate recorded poor insight in relation to mental health and service use outcomes in a cohort with first-episode psychosis.	SLaM BRC CRIS case register	2,026 patients; one full record per patient.	NLP techniques using GATE & TextHunter to identify poor insight, cannabis use, and antipsychotic prescription.	Recorded poor insight in people with recent onset psychosis predicted subsequent legally enforced hospitalizations and higher number of hospital admissions, number of unique antipsychotics prescribed and days spent hospitalized.	Improving insight might benefit patients' course of illness as well as reduce mental health service use	Risk factors for disease & outcomes
Roberts et al. (65)	To investigate mortality in individuals diagnosed with chronic fatigue syndrome (CFS).	SLaM BRC CRIS case register linked ONS mortality data.	2,147 patients; one full record per patient	NLP techniques using GATE & TextHunter to supplement CFS diagnosis codes for case identification	All-cause and cancer-specific mortality of patients with chronic fatigue syndrome in specialist care is not significantly different to that of the general population, however the risk of suicide is higher	Clinicians need to be aware of the increased risk of completed suicide and to assess suicidality adequately in patients with chronic fatigue syndrome.	Risk factors for disease & outcomes
Ruigómez et al. (82)	To evaluate the validity of recorded diagnoses of ischemic cerebrovascular events requiring hospitalization	The Health Improvement Network (THIN)	4,239 patients; one full record per patient.	Manual extraction of free text to validate codes for cerebrovascular events.	THIN demonstrates a high validity for the study of ischemic cerebrovascular events when reviewing computer records with additional free text comments.	EHRs, especially with the addition of free text comments, are an efficient approach to ascertain the accuracy and validity of computerized data and lead to service improvement.	Service improvement
Shah et al. (50)	To describe the contribution of free text in primary care to the recording of information about myocardial infarction (MI), including subtype, left ventricular function, laboratory results and symptoms; and recording of cause of death.	Clinical Practice Research Datalink (CPRD)	2,000 patients with MI and 1,800 deaths; one full record per patient.	Freetext Matching Algorithm (FMA) NLP software to determine subtype of MI, left ventricular function, angiogram results and symptoms.	Inclusion of free text increased the recorded proportion of patients with chest pain in the week prior to MI from 19 to 27%, and differentiated between MI subtypes in a quarter more patients than structured data alone.	Natural language processing to convert this information into a structured form can enrich primary care data at scale for research, and yield population-based insights into early presentations of disease.	Case finding for clinical trials
Sharma et al. (78)	Investigate predictors of falls and fractures leading to hospitalization in a large cohort of people with dementia.	SLaM BRC CRIS case register linked to HES	8,036 patients; one full record per patient	NLP techniques using GATE to supplement medications and MMSE score	Medications (including psychotropic and antipsychotics), neuropsychiatric symptoms, cognitive (Mini-Mental State Examination scores), or functional problems did not predict hospitalized falls.	Clinicians should consider that besides established demographic and physical health related factors, the risk of hospitalization due to a fall or fractures in dementia is largely determined by environmental and socioeconomic factors.	Risk factors for disease & outcomes
Soysal et al. (85)	To test associations between polypharmacy and both short-term (6 months) and long-term (3 years) cognitive trajectories in patients with incident dementia.	SLaM BRC CRIS case register	12,148 patients; one full record per patient.	NLP techniques using GATE to identify dementia date of diagnosis and number of medications.	No significant differences to the control group were found in relation to polypharmacy or excessive polypharmacy, neither in the initial cognitive improvement nor long-term decline.	Polypharmacy defined by the number of drugs does not appear to predict cognitive decline in a naturalistic cohort of patients with dementia.	Risk factors for disease & outcomes
Tate et al. (48)	To investigate how much information on ovarian cancer diagnosis is ‘hidden' in the free text and the time lag between a diagnosis being described in the text or in a hospital letter and the patient being given a Read code for that diagnosis.	Clinical Practice Research Datalink (CPRD)	344 patients; one full record per patient.	Manual extraction of information on date of diagnosis	Free text contains a significant amount of extra information. Free text analysis showed that GPs do not always code a definite diagnosis on the date that it is confirmed. For diseases which rely on hospital consultants for diagnosis, free text (particularly letters) is invaluable for accurate dating of diagnosis and referrals.	EHRs using free text can be used to improve the amount of information available to clinicians, leading to improved quality of care.	Service improvement
Taylor et al. (43)	To investigate the socio-demographic and clinical characteristics of an epidemiologically representative cohort of pregnant women with affective and non-affective severe mental illness.	SLaM BRC CRIS case register linked with HES	456 patients; one full record per patient.	NLP techniques using GATE to supplement diagnosis codes for case identification, number of days in acute mental health care. Manual searches for expected delivery dates, number of other children, partner status, smoking, alcohol, drug use, maternal history of abuse, deliberate self-harm.	A significant proportion of women, particularly those with non-affective psychoses, have modifiable risk factors requiring tailored care to optimize pregnancy outcomes. Mental health professionals need to be mindful of the possibility of pregnancy in women of childbearing age and prescribe and address modifiable risk factors accordingly.	EHRs can be used for observational research by predicting risk factors (exposures) for pregnancy outcomes (outcome).	Risk factors for disease & outcomes
Taylor et al. (44)	To investigate risk factors for postpartum relapse, particularly the potential prophylactic effects of medication in women with non-affective or affective psychoses.	SLaM BRC CRIS case register linked to HES	452 patients; one full record per patient.	NLP techniques using GATE to supplement diagnosis codes for case identification, and medication exposure. Manual searches for expected delivery dates, number of other children, partner status, smoking, maternal history of abuse, family history of psychosis, smoking, drug and alcohol use.	There was no evidence of a prophylactic effect of medication in women with non-affective or affective psychoses	Recent relapse increases the risk of relapse in the postpartum period so women with severe illnesses with a recent history of relapse should be warned pre-conception about the high risk of relapse.	Risk factors for disease & outcomes
Tulloch et al. (45)	To investigate the associations of homelessness and residential mobility with length of stay after acute psychiatric admission	SLaM BRC CRIS case register	4,885 patients; one full record per patient.	Housing data supplemented by manual review of free text for homelessness status.	Homelessness and, especially, residential mobility account for a significant part of variation in length of stay despite affecting a minority of psychiatric inpatients; for these people, the effect on length of stay is marked.	Residential mobility and, to a lesser extent, homelessness appeared to have very significant effects on length of stay which should inform policy change. The effects are clearly detectable within the whole sample, but apply especially to the minority of individuals who are directly affected.	Risk factors for disease & outcomes
Tulloch et al. (49)	To investigate the clinical and demographic associations of khat use in a sample of Somali users of mental health service users in South London.	SLaM BRC CRIS case register	240 patients; one full record per patient	Keyword search plus manual coding of Somali origin, Khat, cannabis and alcohol use, and diagnoses.	Khat use was very strongly associated with a primary diagnosis of schizophrenia, psychosis or drug and alcohol disorder (compared to stress-related disorders and other non-psychotic disorders), male gender, harmful or dependent use of alcohol, and detention under the Mental Health Act.	Documentation of khat use should be improved, especially for patients of Somali, Yemeni and other relevant ethnicity.	Risk factors for disease & outcomes
Woodhead et al. (73)	To examine if breast and cancer screening uptake is lower among women with serious mental illness (SMI) and to identify variation in screening uptake by illness/treatment factors, and primary care consultation frequency.	SLaM BRC CRIS case register linked to Lambeth DataNet (LDN);	Patients with breast and cervical screening receipt among linked eligible SMI patients (n = 625 and n = 1,393), to those without SMI known only to primary care (n = 106,554 and n =25,385); one full record per patient	NLP techniques using GATE to identify antipsychotic use, and risk and severity variables.	Women with SMI are less likely to receive breast and cervical cancer screening than comparable women without SMI.	To tackle health disparities linked to SMI, efforts at increasing screening uptake are key and should be targeted at women with other markers of illness severity or risk, beyond SMI status alone.	Service improvement
Wu et al. (81)	To investigate the prevalence and correlates of EHR-derived current smoking in people with severe mental illness.	SLaM BRC CRIS case register	1,555 patients; one full record per patient.	NLP techniques using GATE to identify smoking status	Proportions of patients with recorded smoking status increased from 11.6 to 64.0% when using CRIS data. After adjustment, younger age (below 65 years), male sex, and non-cohabiting status were associated with current smoking status.	A natural language processing application substantially improved routine EHR data on smoking status above structured fields alone and could thus be helpful in improving monitoring of this lifestyle behavior.	Service improvement