Predicting 7‐year‐olds mental health in the perinatal period: Development and internal validation of a multivariable model using the prospective ALSPAC cohort

Abstract Background Mental health difficulties in childhood are increasing. Prevention is the only sustainable and ethical public health approach. However, predicting which children are most at‐risk of mental health difficulties prior to symptoms emerging remains elusive. Methods We developed and internally validated a perinatal multivariable model, predicting 7‐year‐olds mental health, using the Avon Longitudinal Study of Parents and Children ( N = 6021, 51.2% male, 98.6% White). Perinatal predictors were reported by the mother prospectively in pregnancy and the Strengths and Difficulties Questionnaire (SDQ) was completed by the mother at 7‐years‐old. This was dichotomised at recommended clinical cut‐off (total>16) Building on our previous model in a French cohort, 15 perinatal parameters spanning maternal pre‐pregnancy health, biological and psychosocial pregnancy‐specific‐experiences, maternal health behaviours in pregnancy and sociodemographic factors were entered into a logistic regression using the least absolute selection and shrinkage operator. Optimism‐adjusted estimates were achieved using bootstrapping. Model performance was stratified by sex, sociodemographic risk and admission to a special‐care baby unit. Results Combining eight variables predicted poor mental health, with a C‐statistic of 0.66; 95% Confidence‐Interval (0.64–0.68). It accurately predicted 85.6% of the participants mental health at 7‐years in the perinatal period. Model performance was similar across groups of interest. Applying this model leads to a higher benefit than serving ‘all’ or ‘no’ children, that is, using the model, 30.9% of children who later had poor mental health would have been identified in the perinatal period. Conclusion It is possible to predict childhood mental health at birth with moderate accuracy. Similar patterns of model performance were observed in this English cohort compared to a previous French cohort. At population‐level, the model is most useful for ruling‐out babies who are not predicted to be high‐risk. In addition to improving its positive predictive value and external validation, future research should examine the model's performance at service‐delivery level before implementation.