Developing and implementing tools to advance precision psychiatry using electronic health records

In recent years, precision medicine has emerged as an avenue for risk stratification and the provision of highly individualised
care, through careful consideration of differences across individuals' characteristics, genes, environments, and lifestyles. Clinical
prediction modelling is the backbone of precision medicine, and can be utilised to predict the probability of a condition being
present (diagnostic models), its future course (prognostic models), and the response to associated interventions (predictive
models).

In several branches of medicine (e.g., cardiovascular disease, and cancer), a wide range of these models have been
implemented for routine usage in clinical practice. However, within psychiatry, whilst such models have been developed, they
have largely not been considered for implementation in clinical practice (Meehan et al., 2022; Salazar de Pablo et al., 2021). This
is largely due to the relative lack of testing the performance of models using wholly independent data in new, real-life settings
to assess their generalisability to related, but different settings (external validation). External validation, however, is costly and
time-consuming. Thus, novel approaches are required to assess the robustness of models across different settings in order to
improve the likelihood of models being implemented in clinical practice and at lower costs, and consequently progress precision
psychiatry. Monte Carlo simulations using information of existing prediction models assessed in different settings would help to
assess the expected robustness of a model in external settings. This would allow an assessment of the likelihood that a model
will exhibit good external validation without requiring the collection of data and aids in decision-making regarding whether
further costly and time-consuming external validation studies are warranted.

Further, one of the most frequently investigated conditions of existing individualised prediction models are psychotic disorders
(Meehan et al., 2022; Salazar de Pablo et al., 2021). Amongst other influences, this focus has been driven by a need to tackle the
sizeable health disparity and burden associated with psychosis, as well as to advance the detection of individuals at clinical high risk state for psychosis (CHR-P) and prognostication of their outcomes (Fusar-Poli et al., 2016, 2020). An individualised prognostic model for psychosis onset has previously been developed and extensively externally validated (Fusar-Poli et al.,
2017, 2019; Oliver et al., 2020; Puntis et al., 2021) and is ideal for testing novel approaches of simulating performance across
multiple settings.

Similar to psychotic disorders, personality disorders are associated with significant comorbidity, premature mortality, and high
societal and personal costs (Tyrer et al, 2015). Further, they predict poor adherence to and engagement in the treatment of
other comorbidities, and consequently can moderate transdiagnostic mental health and physical health outcomes. Despite this
evident disease burden, a recent review (Meehan et al., 2022) identified only two individualised prediction models focussed on
borderline personality disorder (BPD) and both were diagnostic, rather than prognostic. Given the extensive research into
subthreshold personality disorder pathology and the malleability of BPD traits in youth/early adolescence (Sharp & Fonagy,
2015; Chanen & McCutcheon, 2018), the development of a prognostic model for personality disorder would be well-placed to
help improve our understanding of emerging personality disorders; promote their prevention and early intervention; and
improve long-term outcomes.