The identification of predictors of treatment response is crucial for improving treatment outcome for children with anxiety disorders. Machine learning methods provide opportunities to identify combinations of factors that contribute to risk prediction models.

A machine learning approach was applied to predict anxiety disorder remission in a large sample of 2114 anxious youth (5–18 years). Potential predictors included demographic, clinical, parental, and treatment variables with data obtained pre-treatment, post-treatment, and at least one follow-up.

All machine learning models performed similarly for remission outcomes, with AUC between 0.67 and 0.69. There was significant alignment between the factors that contributed to the models predicting two target outcomes: remission of all anxiety disorders and the primary anxiety disorder. Children who were older, had multiple anxiety disorders, comorbid depression, comorbid externalising disorders, received group treatment and therapy delivered by a more experienced therapist, and who had a parent with higher anxiety and depression symptoms, were more likely than other children to still meet criteria for anxiety disorders at the completion of therapy. In both models, the absence of a social anxiety disorder and being treated by a therapist with less experience contributed to the model predicting a higher likelihood of remission.

These findings underscore the utility of prediction models that may indicate which children are more likely to remit or are more at risk of non-remission following CBT for childhood anxiety.

Elucidation of the interaction of biological and psychosocial/environmental factors on opioid dependence (OD) risk can inform our understanding of the etiology of OD. We examined the role of psychosocial/environmental factors in moderating polygenic risk for opioid use disorder (OUD).

Data from 1958 European ancestry adults who participated in the Yale-Penn 3 study were analyzed. Polygenic risk scores (PRS) were based on a large-scale multi-trait analysis of genome-wide association studies (MTAG) of OUD.

A total of 420 (21.1%) individuals had a lifetime diagnosis of OD. OUD PRS were positively associated with OD (odds ratio [OR] 1.42, 95% confidence interval [CI] 1.21–1.66). Household income and education were the strongest correlates of OD. Among individuals with higher OUD PRS, those with higher education level had lower odds of OD (OR 0.92, 95% CI 0.85–0.98); and those with posttraumatic stress disorder (PTSD) were more likely to have OD relative to those without PTSD (OR 1.56, 95% CI 1.04–2.35).

Results suggest an interplay between genetics and psychosocial environment in contributing to OD risk. While PRS alone do not yet have useful clinical predictive utility, psychosocial factors may help enhance prediction. These findings could inform more targeted clinical and policy interventions to help address this public health crisis.

Cost-effectiveness analysis needs to be considered when introducing new tools and treatments to clinical services. The number of new assessment tools in mental health has rapidly expanded, including suicide risk assessment. Such suicide-based assessments, when linked to preventative interventions, are integral to high-quality mental health care for people with severe mental illness (SMI). We examined the cost implications of implementing Oxford Mental Illness and Suicide (OxMIS), an evidence-based, scalable suicide risk assessment tool that provides probabilistic estimates of suicide risk over 12 months for people with SMI in England.

We developed a decision analytic model using secondary data to estimate the potential cost-effectiveness of incorporating OxMIS into clinical decision-making in secondary care as compared to usual care. Cost-effectiveness was measured in terms of costs per quality-adjusted life years (QALYs) gained. Uncertainty was addressed with deterministic and probabilistic sensitivity analysis.

Conducting suicide risk assessment with OxMIS was potentially cheaper than clinical risk assessment alone by £250 (95% confidence interval, −786;31) to £599 (−1,321;−156) (in 2020–2021 prices) per person with SMI and associated with a small increase in quality of life (0.01 [−0.03;0.05] to 0.01 QALY, [−0.04;0.07]). The estimated incremental cost-effectiveness ratio of implementing OxMIS was cost saving. Using probabilistic sensitivity analysis, 99.96% of 10,000 simulations remained cost saving.

Cost-effectiveness analysis can be conducted on risk prediction models. Implementing one such model that focuses on suicide risk in a high-risk population can lead to cost savings and improved health outcomes, especially if explicitly linked to preventative treatments.

23% of people experiencing a first episode of psychosis (FEP) develop treatment resistant schizophrenia (TRS). At present, there are no established methods to accurately identify who will develop TRS from baseline.

In this study we used patient data from three UK early intervention services (EIS) to investigate the predictive potential of routinely recorded sociodemographic, lifestyle and biological data at FEP baseline for the risk of TRS up to six years later.

We developed two risk prediction algorithms to predict the risk of TRS at 2-8 years from FEP onset using commonly recorded information at baseline. Using the forced-entry method, we created a model including age, sex, ethnicity, triglycerides, alkaline phosphatase levels and lymphocyte counts. We also produced a machine-learning-based model, including an additional four variables. The models were developed using data from two and externally validated in another UK psychosis EIS.

The development samples included 785 patients, and 1,110 were included in the validation sample. The models discriminated TRS well at internal validation (forced-entry: C 0.70, 95%CI 0.63-0.76; LASSO: C 0.69, 95%CI 0.63-0.77). At external validation, discrimination performance attenuated (forced-entry: C 0.63, 0.58-0.69; LASSO: C 0.64, 0.58-0.69) but recovered for the forced entry model after recalibration and revision of the lymphocyte predictor (C: 0.67, 0.62-0.73).

The use of commonly recorded clinical information including biomarkers taken at FEP onset could help to predict TRS. These measures should be considered in future studies modelling psychiatric outcomes.

No significant relationships.

The comorbidity between cardiometabolic and psychotic disorders develops early. This is a crucial window of opportunity to reduce excess morbidity and mortality. Recently, a cardiometabolic risk prediction algorithm for young people with psychosis, the psychosis metabolic risk calculator (PsyMetRiC) was developed and externally validated in the UK. However, its international transportability is unknown.

We performed the first international validation study of PsyMetRiC in Lausanne, Switzerland, and examined whether additional variables (clinical and/or genetic) may improve the predictive performance of the algorithm

We included people aged 16-35y with psychosis from the PsyMetab cohort, who did not have MetS at baseline, and who had 1-6y follow-up data. The PsyMetRiC partial (age, sex, ethnicity, body mass index, smoking status, and prescription of a metabolically-active antipsychotic) and full (also including high-density lipoprotein and triglycerides) algorithms were applied. Predictive performance was assessed using measures of discrimination (C-statistic) and calibration (calibration plots). Recalibration steps included refitting the intercept and/or slope if necessary. Additional variables (e.g. speed of weight gain, polygenic risk scores) were added to the model and predictive performance was reassessed.

We included 545 participants. The discrimination performance of both PsyMetRiC algorithms was good (C>0.75), and calibration plots showed good agreement between observed and predicted risk. Additional analyses to be conducted.

PsyMetRiC is likely to be generalizable for use in Switzerland, suggesting that PsyMetRiC may also be suitable for use in other European populations. While additional international validations are required, these findings are an encouraging step toward an international cardiometabolic risk prediction algorithm for young people with psychosis.

No significant relationships.

During adolescence and young adulthood people appear to be more prone to violent behaviour. A greater tendency to violent behaviour appears to be associated with hyperactivity, impulsivity and low tolerance for frustration and provocation in social settings.

This prospective cohort study aimed to evaluate rates of violent behaviour among young people with mental disorders, compared with older age groups.

A total of 340 individuals with severe mental disorders (125 living in residential facilities and 215 out-patients) were evaluated at baseline with the SCID-I and II, Brief Psychiatric Rating Scale, Specific Level of Functioning scale, Brown–Goodwin Lifetime History of Aggression scale, Buss–Durkee Hostility Inventory, Barratt Impulsiveness Scale and State–Trait Anger Expression Inventory-2. Aggressive behaviour was rated every 15 days with the Modified Overt Aggression Scale (MOAS).

The sample comprised 28 individuals aged 18–29 years, 202 aged 30–49 and 110 aged 50 and over. Younger age was associated with a personality disorder diagnosis, substance use disorder, being single and employed. These results were confirmed even controlling for the gender effect. The patterns of the cumulative MOAS mean scores showed that younger (18–29 years old) individuals were significantly more aggressive than older (≥50) ones (P < 0.001).

This study highlights how young age in people with severe mental disorders is correlated with higher levels of impulsivity, anger and hostility, confirming previous analyses. Our results may assist clinicians in implementing early interventions to improve anger and impulsivity control to reduce the risk of future aggressive behaviours.

Predictors of new-onset bipolar disorder (BD) or psychotic disorder (PD) have been proposed on the basis of retrospective or prospective studies of ‘at-risk’ cohorts. Few studies have compared concurrently or longitudinally factors associated with the onset of BD or PDs in youth presenting to early intervention services. We aimed to identify clinical predictors of the onset of full-threshold (FT) BD or PD in this population.

Multi-state Markov modelling was used to assess the relationships between baseline characteristics and the likelihood of the onset of FT BD or PD in youth (aged 12–30) presenting to mental health services.

Of 2330 individuals assessed longitudinally, 4.3% (n = 100) met criteria for new-onset FT BD and 2.2% (n = 51) met criteria for a new-onset FT PD. The emergence of FT BD was associated with older age, lower social and occupational functioning, mania-like experiences (MLE), suicide attempts, reduced incidence of physical illness, childhood-onset depression, and childhood-onset anxiety. The emergence of a PD was associated with older age, male sex, psychosis-like experiences (PLE), suicide attempts, stimulant use, and childhood-onset depression.

Identifying risk factors for the onset of either BD or PDs in young people presenting to early intervention services is assisted not only by the increased focus on MLE and PLE, but also by recognising the predictive significance of poorer social function, childhood-onset anxiety and mood disorders, and suicide attempts prior to the time of entry to services. Secondary prevention may be enhanced by greater attention to those risk factors that are modifiable or shared by both illness trajectories.

Current approaches for early identification of individuals at high risk for autism spectrum disorder (ASD) in the general population are limited, and most ASD patients are not identified until after the age of 4. This is despite substantial evidence suggesting that early diagnosis and intervention improves developmental course and outcome. The aim of the current study was to test the ability of machine learning (ML) models applied to electronic medical records (EMRs) to predict ASD early in life, in a general population sample.

We used EMR data from a single Israeli Health Maintenance Organization, including EMR information for parents of 1,397 ASD children (ICD-9/10) and 94,741 non-ASD children born between January 1st, 1997 and December 31st, 2008. Routinely available parental sociodemographic information, parental medical histories, and prescribed medications data were used to generate features to train various ML algorithms, including multivariate logistic regression, artificial neural networks, and random forest. Prediction performance was evaluated with 10-fold cross-validation by computing the area under the receiver operating characteristic curve (AUC; C-statistic), sensitivity, specificity, accuracy, false positive rate, and precision (positive predictive value [PPV]).

All ML models tested had similar performance. The average performance across all models had C-statistic of 0.709, sensitivity of 29.93%, specificity of 98.18%, accuracy of 95.62%, false positive rate of 1.81%, and PPV of 43.35% for predicting ASD in this dataset.

We conclude that ML algorithms combined with EMR capture early life ASD risk as well as reveal previously unknown features to be associated with ASD-risk. Such approaches may be able to enhance the ability for accurate and efficient early detection of ASD in large populations of children.

Risk prediction algorithms have long been used in health research and practice (e.g. prediction of cardiovascular disease and diabetes). However, similar tools have not been developed for mental health. For example, for psychotic disorders, attempts to sum environmental risk are rare, unsystematic and dictated by available data. In light of this, we sought to develop a valid, easy to use measure of the aggregate environmental risk score (ERS) for psychotic disorders.

We reviewed the literature to identify well-replicated and validated environmental risk factors for psychosis that combine a significant effect and large-enough prevalence. Pooled estimates of relative risks were taken from the largest available meta-analyses. We devised a method of scoring the level of exposure to each risk factor to estimate ERS. Relative risks were rounded as, due to the heterogeneity of the original studies, risk effects are imprecisely measured.

Six risk factors (ethnic minority status, urbanicity, high paternal age, obstetric complications, cannabis use and childhood adversity) were used to generate the ERS. A distribution for different levels of risk based on simulated data showed that most of the population would be at low/moderate risk with a small minority at increased environmental risk for psychosis.

This is the first systematic approach to develop an aggregate measure of environmental risk for psychoses in asymptomatic individuals. This can be used as a continuous measure of liability to disease; mostly relevant to areas where the original studies took place. Its predictive ability will improve with the collection of additional, population-specific data.