An evaluation of symptom domains in the 2 years before pregnancy as predictors of relapse in the perinatal period in women with severe mental illness

Abstract

Abstract Background Symptoms may be more useful prognostic markers for mental illness than diagnoses. We sought to investigate symptom domains in women with pre-existing severe mental illness (SMI; psychotic and bipolar disorder) as predictors of relapse risk during the perinatal period. Methods Data were obtained from electronic health records of 399 pregnant women with SMI diagnoses from a large south London mental healthcare provider. Symptoms within six domains characteristically associated with SMI (positive, negative, disorganization, mania, depression, and catatonia) recorded in clinical notes 2 years before pregnancy were identified with natural language processing algorithms to extract data from text, and associations investigated with hospitalization during pregnancy and 3 months postpartum. Results Seventy-six women (19%) relapsed during pregnancy and 107 (27%) relapsed postpartum. After adjusting for covariates, disorganization symptoms showed a positive association at borderline significance with relapse during pregnancy (adjusted odds ratio [aOR] = 1.36; 95% confidence interval [CI] = 0.99–1.87 per unit increase in number of symptoms) and depressive symptoms negatively with relapse postpartum (0.78; 0.62–0.98). Restricting the sample to women with at least one recorded symptom in any given domain, higher disorganization (1.84; 1.22–2.76), positive (1.50; 1.07–2.11), and manic (1.48; 1.03–2.11) symptoms were associated with relapse during pregnancy, and disorganization (1.54; 1.08–2.20) symptom domains were associated with relapse postpartum. Conclusions Positive, disorganization, and manic symptoms recorded in the 2 years before pregnancy were associated with increased risk of relapse during pregnancy and postpartum. The characterization of routine health records from text fields is relatively transferrable and could help inform predictive risk modelling.