Generalisable samples

The volume of data in electronic health records and population registries and their coverage of populations has enabled access to representative samples of women with schizophrenia. However, data such as physician- and hospital-based diagnoses are not collected for research purposes, and can be subject to data entry errors. To generate cohorts of women with schizophrenia who have a pregnancy, utilising these data-sets has required substantial data cleaning and development of algorithms to define and validate health indicators.

In Ontario, studies have included all women in the province of Ontario^{Reference Vigod, Kurdyak, Dennis, Gruneir, Newman and Seeman29} and researchers have developed and validated an algorithm to identify individuals with schizophrenia in health administrative data^{Reference Kurdyak, Lin, Green and Vigod30} using data from two sources – hospital admissions data and out-patient data. Researchers used ICD diagnostic codes for schizophrenia and schizoaffective disorder arising from a hospital admission or three out-patient visits within 3 years. This definition has a sensitivity of 90.1, specificity of 68.0; a positive predictive value of 77.1 and a negative predictive value of 88.7 compared to a clinical record.

Research from Denmark has included all women with past admissions for psychiatric disorder across the whole country. Studies have used ICD diagnoses from the Danish Psychiatric Central Research Register to identify schizophrenia.^{Reference Mors, Perto and Mortensen31} Early data used hospital admissions only but since 1995 the register has contained both hospital admissions and out-patient data. Validation of ICD-10 first time diagnoses of schizophrenia in the register of in-patients and out-patients against clinical case notes has been reported as having a positive predictive value of 89.7–97.5%.^{Reference Uggerby, Ostergaard, Roge, Correll and Nielsen32}

In the UK, research in women with schizophrenia and pregnancy initially used routinely collected primary care data, which gave rise to fairly small samples.^{Reference Howard, Goss, Leese and Thornicroft33,Reference Howard, Kumar, Leese and Thornicroft34} The secondary mental health data from CRIS has allowed research to be carried out on representative samples of women managed in secondary care, but will miss those few women managed solely in primary care.^{Reference Taylor, Stewart, Ogden, Broadbent, Pasupathy and Howard35} In CRIS, natural language processing applications have been developed for text mining, using General Architecture for Text Engineering^{Reference Cunningham, Tablan, Roberts and Bontcheva36} to derive structured data from free-text fields, taking into account the textual context of words/phrases of interest (for example distinguishing between positive and negative statements or past and present occurrences). These applications are developed using an iterative process and validated against ‘gold standard’ manual annotations of documents by calculating precision (positive predictive value) and recall (sensitivity).^{Reference Cunningham, Tablan, Roberts and Bontcheva36} Research focused on pregnant women with a history of severe mental illness has used an application developed to extract diagnoses of severe mental illness prior to pregnancy, reading free-text notes to resolve any ambiguities.^{Reference Taylor, Stewart, Ogden, Broadbent, Pasupathy and Howard35,Reference Taylor, Broadbent, Khondoker, Stewart and Howard37,Reference Taylor, Stewart and Howard38}

Thus, schizophrenia can be validly and reliably identified in routinely collected or ‘real-world’ health administrative and clinical data, opening up multiple opportunities for our own teams and for researchers around the world to rapidly study this population using these or similar data sources.

Longitudinal follow-up

As a result of the rapid accumulation of large volumes of data occurring over long periods of time, secondary analysis of healthcare data allows for efficient measurement of trends over time, as well as long-term follow-up of mothers and their children. An important application of this in the area of schizophrenia in pregnancy is research on time trends in fertility rates, as well as characterisation of the epidemiology of the course of psychiatric illness in relation with pregnancy.

Using population registry data, birth rates among women with schizophrenia have been measured by capturing records of births recorded in hospital data^{Reference Vigod, Seeman, Ray, Anderson, Dennis and Grigoriadis39} or civil registration data.^{Reference Laursen and Munk-Olsen40} Early research from the UK using primary care data showed lower birth rates in women with psychotic disorders than who were matched population controls.^{Reference Howard, Kumar, Leese and Thornicroft34} Research from Denmark showed that birth rates were lower in women with schizophrenia, than the general population and women with other mental disorders even when accounting for higher rates of terminations in this group of women.^{Reference Laursen and Munk-Olsen40} Research using the ICES data-sets in Ontario, Canada examined trends in birth rates in women with schizophrenia over time and found increases of 16% from 1996–1998 to 2007–2009.^{Reference Vigod, Seeman, Ray, Anderson, Dennis and Grigoriadis39} By the end of the study period, birth rates were similar between women with and without schizophrenia in the under 25 age group. These findings have shown that over time, more women with schizophrenia are becoming mothers, and have supported the argument for urgent research to determine how best to support this vulnerable group in their health and the mental and physical health of their children.^{Reference Jones, Chandra, Dazzan and Howard41}

Our research teams have also focused on gaining a better understanding of the course of a woman's psychiatric illness across pregnancy and postpartum. Course of illness in these large data-sets has generally been measured by identifying psychiatric hospital admissions to determine severe psychiatric relapses. Data from Canada has reported admissions during pregnancy at 12–23%^{Reference Vigod, Kurdyak, Dennis, Gruneir, Newman and Seeman29,Reference Rochon-Terry, Gruneir, Seeman, Ray, Rochon and Dennis42} in women with a history of schizophrenia and data from the UK has reported admissions during pregnancy at 24% in women with a history of non-affective psychosis who were in contact with secondary mental healthcare.

In the first 3 months postpartum, UK data reported 31% of women with non-affective psychosis were admitted.^{Reference Taylor, Stewart and Howard38} The figures reported for the first year postpartum were 19% and 16% in data from Canada and Denmark, respectively,^{Reference Rochon-Terry, Gruneir, Seeman, Ray, Rochon and Dennis42,Reference Munk-Olsen, Laursen, Mendelson, Pedersen, Mors and Mortensen43} possibly reflecting the different population in the UK – a secondary care cohort – whereas those for Denmark and Canada are from an epidemiologically representative primary and secondary care population cohort. Although there is some variability across data-sets, these data have consistently documented that women with schizophrenia are at high risk of admission in pregnancy and postpartum and that early postpartum is the highest risk for admission across the whole perinatal period.^{Reference Rochon-Terry, Gruneir, Seeman, Ray, Rochon and Dennis42,Reference Munk-Olsen, Laursen, Mendelson, Pedersen, Mors and Mortensen43}

In Canada, Denmark and the UK, our teams have also shown that markers of the severity of the psychiatric disorder such as number and recency of previous admissions are predictive of admissions in the perinatal period.^{Reference Taylor, Broadbent, Khondoker, Stewart and Howard37,Reference Munk-Olsen, Laursen, Mendelson, Pedersen, Mors and Mortensen43,Reference Vigod, Rochon-Terry, Fung, Gruneir, Dennis and Grigoriadis44} Interestingly, in a Canadian study, having a consistent mental healthcare provider in pregnancy (either a primary care provider or a psychiatrist) was associated with a lower likelihood of requiring psychiatric admission in the postpartum. Other predictors of postpartum psychiatric admission were age >34, low income, not having an ultrasound before 20 weeks’ gestation as well as psychiatric comorbidity and prior psychiatric hospitalizations.^{Reference Vigod, Rochon-Terry, Fung, Gruneir, Dennis and Grigoriadis44}

Conducting longitudinal research relying on clinical studies prospectively collecting data across years would be very costly and at high risk of substantial loss to follow-up across the course of pregnancy and postpartum. These resources of prospective routinely collected healthcare data include more women with longer follow-up periods, minimising attrition, and can include otherwise hard to reach women too unwell to take part in a clinical study. Measuring psychiatric outcomes using hospital admission data may not be as sensitive an indicator of a woman's mental health as a clinical assessment. However, these findings inform clinicians that women with schizophrenia are at risk of a severe relapse in the perinatal period and require close monitoring during this period to prevent relapse, particularly in the early postpartum where admission might involve separating mother and baby, as well as determining which women may be at particular risk and hence require extra support.

Carrying out life-course research, including following women for their reproductive trajectories, and intergenerational research (i.e. research on outcomes of mothers and outcomes of their children) in healthy populations by recruiting mothers and babies has proved increasingly difficult. In the UK, a new large cohort study (the LIFE study) could not recruit to target and was stopped (www.lifestudy.ac.uk). Conducting clinical life-course and intergenerational research for women with schizophrenia would be an even greater challenge. More recently, routinely collected healthcare data has increasingly integrated data from across generations, different phases of life and critical periods of development, and research has begun to address longer-term follow-up of both women and children. For example, Canadian data have shown that women with schizophrenia are at higher risk of repeat pregnancy within 12 months of an index live birth than women without schizophrenia.^{Reference Gupta, Brown, Barker, Dennis and Vigod45} Danish data through linkage to social care data and other family members, has also shown high risk of out-of-home placement in children of women with severe mental disorders^{Reference Vigod, Laursen, Ranning, Nordentoft and Munk-Olsen46,Reference Ranning, Munk Laursen, Thorup, Hjorthoj and Nordentoft47} and long-term follow-up of mental health records have shown children of parents with severe mental disorders have higher incidence of child and adolescent mental disorders up to age 17 compared with children of parents without severe mental disorders.^{Reference Thorup, Laursen, Munk-Olsen, Ranning, Mortensen and Plessen48}

Rare outcomes

Many obstetric and neonatal complications are relatively uncommon in the general population. For example, the rate of stillbirths in 2015 was 18.4 per 1000 live births globally and 3.4 in high-income countries/developed regions^{Reference Blencowe, Cousens, Jassir, Say, Chou and Mathers49} and in 2017, globally, neonatal death was reported at 18 per 1000 live births and 3 per 1000 live births in high-income countries.^{Reference Hug, Alexander, You and Alkema50} Pre-eclampsia has been estimated at 4.6%^{Reference Abalos, Cuesta, Grosso, Chou and Say51} and rates of preterm birth have been reported at 8.7% in Europe and 11.2% in North America.^{Reference Chawanpaiboon, Vogel, Moller, Lumbiganon, Petzold and Hogan52} Given that schizophrenia is already a rare disorder, detecting such outcomes in these women requires very large samples. Using population-level data and including a variety of sources of linked data, research from Canada and Denmark has shown that compared with women in the population who do not have a history of schizophrenia, women with schizophrenia tend to have higher risk outcomes including pre-eclampsia, venous thromboembolism, placental complications and haemorrhage,^{Reference Vigod, Kurdyak, Dennis, Gruneir, Newman and Seeman29} assisted deliveries such as caesarean sections, inductions of labour and maternal intensive care admissions.^{Reference Vigod, Kurdyak, Dennis, Gruneir, Newman and Seeman29,Reference Bennedsen, Mortensen, Olesen, Henriksen and Frydenberg53}

Regarding neonates; data from Canada have also found increased risk for preterm birth, small for gestational age, large for gestational age and low birth weight and an increased risk for other neonatal morbidities such as seizures, sepsis, and neonatal abstinence syndrome,^{Reference Vigod, Kurdyak, Dennis, Gruneir, Newman and Seeman29} and from Denmark; poorer Apgar scores.^{Reference Bennedsen, Mortensen, Olesen, Henriksen and Frydenberg53} Earlier research from the UK using primary care data has shown that women with history of psychotic disorder, compared with matched controls had more caesarean sections and higher proportion of stillbirths and neonatal deaths.^{Reference Howard, Goss, Leese and Thornicroft33}

Recent research from Canada using clinical birth registry data (BORN) to look at mediators of elevated risk, found that smoking in women with schizophrenia explained the greatest proportion of elevated risks for preterm, small-for-gestational-age scores and Apgar scores. Other relevant mediators were illicit substance use and reproductive history including prior preterm birth, caesarean section, surgical abortion and assisted reproduction.^{Reference Vigod, Fung, Amartey, Bartsch, Felemban and Saunders54}

Regarding rarer psychiatric outcomes, research in Canada and the UK has looked at self-harm in pregnancy, detecting rates of 1% using emergency department visits^{Reference Rochon-Terry, Gruneir, Seeman, Ray, Rochon and Dennis42} and 11.4% in women with non-affective psychosis in secondary mental healthcare using clinical notes.^{Reference Taylor, van Ravesteyn, Lambregtse van denBerg, Stewart and Howard55} Research using the Danish population (1 202 294 women) identified 1554 women with a severe first-onset postpartum mental disorder, of whom 64 had a hospital record of self-harm following diagnosis. Of these, 6 (9.4%) women had a diagnosis of schizophrenia or related disorder.^{Reference Johannsen, Larsen, Laursen, Bergink, Meltzer-Brody and Munk-Olsen56} Canadian research has also looked at suicide among women over a 15-year period in the province of Ontario identified using death records. Of 1914 suicides identified, 51 occurred during the perinatal period, but among women with psychotic disorders in the previous 5 years, numbers were too small to report,^{Reference Grigoriadis, Wilton, Kurdyak, Rhodes, VonderPorten and Levitt57} indicating that suicide among women with schizophrenia in the perinatal period is very rare.

Challenges with using routinely collected healthcare data to identify rare outcomes include the risk of false-positive findings when making multiple comparisons, statistical differences that may not be clinically meaningful and residual confounding. Important covariates such as partner abuse or childhood maltreatment are not usually recorded in administrative data, but some of these are potentially extractable from comprehensive electronic health records such as CRIS, and with linkage of clinical registry data, such as with BORN in Ontario. Despite these limitations, these data allow us to determine what is happening on a ‘real-world’ level and no other method of research would provide sufficient data to detect these outcomes with population-level coverage and with such a high degree of external validity. These data give us information at the public health level regarding adverse health outcomes for mother and baby in women with schizophrenia enabling us to develop targeted interventions to address identified health disparities in this group of women and their babies.

Comparator groups

Another and often overlooked advantage of using population-based data is they have allowed for novel and creative ways of identifying multiple and varying control or comparison groups. Schizophrenia is primarily managed with antipsychotic medication and one of the biggest clinical concerns for patients and healthcare providers is how to weigh the risks and benefits of medication use in pregnancy. Applying modern statistical methods, such as propensity score matching, to the question of the safety of antipsychotic medication use in pregnancy has led to advances in knowledge about the risk–benefit ratio of the use of these medications in pregnancy.

For example, research from Canada has investigated maternal physical health and perinatal outcomes in women exposed and unexposed to antipsychotic medication in pregnancy.^{Reference Vigod, Gomes, Wilton, Taylor and Ray58} In an unmatched comparison, the antipsychotic-exposed women had higher risk of gestational diabetes and hypertensive disorders, and higher risks of adverse labour-, delivery- and neonatal-related outcomes. These differences disappeared when the antipsychotic-exposed group was compared to a propensity-matched unexposed group, indicating that factors other than medication may account for these differences.

Prescribing and exposure to psychotropic medication in pregnancy remains an area of contention for women and clinicians. The rich source of clinical psychiatric records and use of natural language processing applications to extract medication information in UK data from CRIS has enabled us to look in detail at changes in exposure status to medication in pregnancy in women with schizophrenia, including stops and switches in medications. Changes may reflect concerns by clinicians and women regarding effects of these medications during pregnancy, suggesting need for more evidence on the risks and benefits of medications in pregnancy. Around 40% of women with a history of non-affective psychosis stopped or switched a medication in the first trimester.^{Reference Taylor, Stewart, Ogden, Broadbent, Pasupathy and Howard35} We also determined the reasons for changes and whether stops in medication were a woman or a clinician's decision and whether medication was discontinued gradually or abruptly. Given there were many complex changes in medication throughout pregnancy in some women we were unable to report by pregnancy whether medication discontinuation was a woman or clinician's decision. However, when women stopped medications themselves, all stopped abruptly. Reasons for stopping or switching medications were primarily as a result of becoming pregnant. Other reasons included side-effects, improved symptoms or ‘non-compliance’.^{Reference Taylor, Stewart, Ogden, Broadbent, Pasupathy and Howard35} We have also investigated medication exposure as a predictor of psychiatric relapse in pregnancy and postpartum. Comparing within women with a history of psychotic disorders, these data have shown that women on no medication tended to have greater risk of relapse in pregnancy than women exposed to medication.^{Reference Taylor, Broadbent, Khondoker, Stewart and Howard37}

In Denmark, researchers are also utilising new and creative methods for investigating psychotropic medication in pregnancy and have attempted to account for confounding by indication among antidepressant users for outcomes in offspring by comparing unexposed women, women exposed in pregnancy only, women exposed before pregnancy only, as well as before and during pregnancy. To account for any specific effects of intrauterine influence of medication use on child health, fathers were also used as negative controls.^{Reference Liu, Agerbo, Ingstrup, Musliner, Meltzer-Brody and Bergink59} Note, this approach has not yet been carried out with respect to antipsychotic medications.

There are limitations to using routinely collected healthcare data for medication research. Adherence is largely unknown, given that much of the data is prescription related, leading to potential misclassification of medication exposures. Use of clinical psychiatric records in UK data have been able to give a clearer picture regarding whether women in secondary care are taking medication, although the quality of clinical notes varies, which may depend again on women's engagement with services and the severity of their psychiatric illness. However, in the absence of randomised controlled trial data, utilisation of population scale records and detailed clinical data, as well as newer analytic techniques are increasing the likely validity of these findings.