Cardiometabolic prediction models for young people with psychosis spectrum disorders in the UK (Psychosis Metabolic Risk Calculator, PsyMetRiC 2.0): development, refinement and external validation

Abstract

Background: young people with psychosis spectrum disorders are at high risk of cardiometabolic morbidity and subsequent premature mortality, but accurate clinic-ready prediction models for this group are lacking. We aimed to collaboratively refine, extend, and validate the Psychosis Metabolic Risk Calculator (PsyMetRiC) prediction models (comprising age, sex, ethnicity, body mass index, smoking status, antipsychotic prescription, high-density lipoprotein levels, triglyceride levels; originally developed to predict incident metabolic syndrome within six years of a first recorded psychosis spectrum disorder in people aged 16-35 years) for accuracy, clinical usefulness, and acceptability, and to translate the models into a regulated, clinically available medical device.

Methods: we used primary care (CPRD, QResearch) and secondary care (South London and Maudsley NHS Foundation Trust) datasets. Individuals from primary care sources were aged 16-35 when they received a first recorded diagnosis of a psychosis-spectrum disorder between Jan 1, 2005, and Dec 31, 2015, with follow-up to Dec 31, 2020; Individuals from the secondary care source were enrolled in the psychosis early intervention service between Jan 1, 2012, and Dec 31, 2024. We developed models for a binary outcome of metabolic syndrome within 1 to 6 years using logistic regression; a time-to-event outcome of type 2 diabetes within 10 years using Weibull regression; and a binary outcome of clinically-significant weight gain within 1 year using logistic regression. We revised existing predictors for finer detail (hereafter PsyMetRiC1 models) and added new ones: cardiometabolic disorder family history, antidepressant prescription, systolic blood pressure, glycated haemoglobin levels (hereafter PsyMetRiC2 models). Refinement and external validation were performed for metabolic syndrome models (PsyMetRiC1-MetS; PsyMetRiC2-MetS). Development and external validation were performed for type 2 diabetes models (PsyMetRiC2-T2D). Development and internal validation were performed for clinically-significant weight gain models (PsyMetRiC2-WG). “Partial” versions without biochemical results were also developed for weight gain and metabolic syndrome models. We conducted discrimination, calibration, and decision curve analyses (whole sample and by demographic subgroup); involved stakeholders; and implemented the models in a web application compliant with regulatory standards in Great Britain.

Findings: in total, we included n=25,850 individuals (male n=13,614 [52·7%]; female n=12,236 [47·3%]; non-White European n=9,405 [36·3%]; mean age = 26·7 years [SD=5·4]). From primary care, this comprised n=3,989 for development and n=4,347 for external validation of metabolic syndrome outcomes; and n=9,181 for development and n=7,487 for external validation of type 2 diabetes outcomes, representing 121,202 person-years of follow-up. From secondary care, this comprised n=846 for development and internal validation of weight gain outcomes. For metabolic syndrome, revision and extension improved discrimination performance by 4% at external validation compared with the original PsyMetRiC models (PsyMetRiC2-MetS full-model: C=0·81, 95% C.I., 0·77-0·82; partial-model: C=0·79, 95% C.I., 0·76-0·83). For type 2 diabetes, discriminative performance was excellent at internal (PsyMetRiC2-T2D full-model: C=0·86, 95% C.I., 0·78-0·94) and external validation (PsyMetRiC2-T2D full-model: C=0·81, 95% C.I., 0·75-0·87). For weight gain, discriminative performance was good at internal validation (PsyMetRiC2-WG full-model: C=0·77, 95% C.I., 0·73-0·82; partial-model: C=0·76, 95% C.I., 0·72-0·80). Calibration plots were acceptable for all models. All models displayed evidence of clinical usefulness at all plausible thresholds. Subgroup analysis revealed some accuracy differences that did not impair clinical usefulness. The web application (https://psymetric.app/) is available for clinical use in Great Britain.

Interpretation: we developed prediction models for incident cardiometabolic disorders in young people with psychosis. The PsyMetRiC models are among the first in psychiatry to be available for routine clinical use. PsyMetRiC can support a shift toward collaborative, preventive physical healthcare for young people with psychosis.

Funding: NIHR Advanced Fellowship (NIHR304365).

More information

Identifiers

ORCID for Richard I.G. Holt: orcid.org/0000-0001-8911-6744

Catalogue record

Export record

Altmetrics