We partner with a secure submission system to handle manuscript submissions.
Please note:
You will need an account for the submission system, which is separate to your Cambridge Core account. For login and submission support, please visit the
submission and support pages.
Please review this journal's author instructions, particularly the
preparing your materials
page, before submitting your manuscript.
Click Proceed to submission system to continue to our partner's website.
To save this undefined to your undefined account, please select one or more formats and confirm that you agree to abide by our usage policies. If this is the first time you used this feature, you will be asked to authorise Cambridge Core to connect with your undefined account.
Find out more about saving content to .
To send this article to your Kindle, first ensure no-reply@cambridge.org is added to your Approved Personal Document E-mail List under your Personal Document Settings on the Manage Your Content and Devices page of your Amazon account. Then enter the ‘name’ part of your Kindle email address below. Find out more about sending to your Kindle.
Find out more about saving to your Kindle.
Note you can select to save to either the @free.kindle.com or @kindle.com variations. ‘@free.kindle.com’ emails are free but can only be saved to your device when it is connected to wi-fi. ‘@kindle.com’ emails can be delivered even when you are not connected to wi-fi, but note that service fees apply.
Challenging macrotrends in recent decades have characterised population mental health across the globe, with apparent deterioration in adolescent mental health, as well as underfunded, inequitable, and inefficient psychiatric care systems [1]. Unnecessary delays to treatment are widespread, with systemic difficulties, such as clinician burnout, a dearth of funding, and staff shortages, negatively impacting mental health care across Europe and beyond [2]. With limited clinician time and resources available, new methods for improving mental health care are an urgent priority. Digital technologies, and particularly recent advances in the capabilities of artificial intelligence (AI), offer potential to address some of these challenges, thereby transforming approaches to mental health care.
The application of artificial intelligence (AI) to resting-state electroencephalography (EEG) data represents a burgeoning frontier in schizophrenia diagnostics. However, the inferential validity of findings is frequently compromised by methodological heterogeneity. This systematic review rigorously evaluates the diagnostic efficacy and methodological robustness of AI-generated EEG biomarkers by implementing the Prediction model Risk Of Bias Assessment Tool for Artificial Intelligence (PROBAST+AI).
Methods
Following PRISMA guidelines, a systematic search of PubMed, Scopus, and Web of Science identified 43 primary studies encompassing 3,420 patients with schizophrenia or first-episode psychosis and 3,164 healthy controls (total N = 6,584). Methodological quality and validation strategies were assessed using the 34 signaling questions of the PROBAST+AI framework.
Results
The aggregate mean diagnostic accuracy across the literature was 93.38%, with deep learning architectures demonstrating superior metrics relative to classical machine learning (96.83% vs. 87.68%). Critically, external validation was restricted to 4.7% of the included studies. A pronounced performance discrepancy was observed between subject-wise (88.22%) and segment-wise (98.14%) validation protocols, indicating pervasive data leakage. Overall, 76.7% of the analyzed studies were characterized by a high risk of bias, predominantly stemming from suboptimal data partitioning and inadequate independent verification.
Conclusions
Contemporary diagnostic accuracies appear to be artifactually elevated due to structural methodological deficiencies and data leakage. Consequently, current evidence does not support the immediate clinical implementation of AI-EEG biomarkers in schizophrenia. To facilitate translational progress, we propose a standardized methodological framework optimizing validation strategies.
Brain-structural and language abnormalities in people with psychotic symptoms (PSx) have long been reported. A key question is whether language alterations are a potential behavioral readout of brain cortical gray matter changes. Adding to previous evidence for this in the cases of cortical thinning and gyrification, we here used structural MRI (sMRI) to relate a complexity-theoretical metric, the Largest Lyapunov Exponent (lambda [λ]), to computational linguistic complexity metrics extracted from spontaneous speech. This study aimed to identify psychosis-related regional changes in structural complexity using λ and an explainable AI (XAI) approach. We hypothesized that altered structural complexity will relate to computational language features that have previously been shown to track brain dysconnectivity in schizophrenia in fMRI.
Methods
MRI data were acquired from 92 patients with PSx and 38 healthy controls (HCs) from the TOPSY study. Nonlinear analysis of gray matter distribution was performed by extracting the λ for gray matter voxels. XAI was employed to identify the voxels contributing significantly to classifying PSx against HCs. Finally, the brain voxels’ contribution was tested for associations with word perplexity and syntactic and semantic complexity metrics, from spontaneous speech as obtained from picture descriptions, and with clinical symptom severity (SOFAS and PANSS scores).
Results
The classification framework resulted in a balanced accuracy of 75.9%. The voxels contributing most to the classification decision were located in the temporal lobe, cingulum, angular, lingual, calcarine, occipital, fusiform, and parietal cortices, and parts of the cerebellum and vermis. Structural complexity in these regions was significantly associated with clinical variables (SOFAS and PANSS), word perplexity, and lexical diversity.
Conclusions
We present a new analytical framework using spatial series extracted from sMRI, demonstrating that alterations in structural complexity are an identifiable feature of psychosis, which co-vary with probabilistic and structural changes in speech as a behavioral readout.
Attention deficit hyperactivity disorder (ADHD) affects 5–7.2% of children and 2.5% of adults. Despite its prevalence, ADHD remains underdiagnosed and undertreated, leading to significant challenges for affected individuals. Early diagnosis and intervention can prevent adverse outcomes and improve quality of life.
Methods
We developed a predictive model to identify adults with ADHD using electronic health records. The dataset comprised 2,973 adult patients (aged 18 years and above) diagnosed with ADHD and a control group of 4,447 adults referred to psychologists with no ADHD diagnosis. A transformer-based architecture was implemented, utilizing only clinical codes and gender as input features. Fivefold cross-validation was adopted, and model performance was evaluated on held-out test data consisting of 800 patients, 400 of whom had an ADHD diagnosis.
Results
Our study demonstrated the ability to predict adult ADHD using clinical data, with a 6-month model achieving an area under the receiver operating characteristic curve (AUC) of 0.79 (95% confidence interval: 0.76–0.81), F1-score of 0.79, sensitivity of 0.80, and specificity of 0.77. Shapley Additive Explanations identified key contributing codes, including F158 and Y903, consistent with known associations between ADHD and substance use.
Conclusions
Our findings show that machine learning can effectively use clinical codes and demographic data from routine EHRs to support early, cost-efficient diagnosis of adult ADHD, paving the way for earlier intervention and improved outcomes.
Suicide is a major public health challenge requiring early detection of suicidal ideation (SI). Traditional direct questioning methods suffer from stigma and disclosure bias, failing to identify many at-risk individuals. While machine learning (ML) models show promise, most lack external validation. Indirect screening, using psychosocial data rather than direct SI questions, offers a scalable alternative. This study aimed to externally validate an indirect, ML-based SI screening tool. We tested if a model trained on a Slovenian general population sample retained predictive accuracy when applied to an independent Croatian sample during a period of societal stress (pandemic and earthquakes), assessing performance across age and gender subgroups.
Methods
A logistic regression model was trained on a Slovenian sample (N = 2,989) and validated on a Croatian sample (N = 2,364). The model used only indirect predictors, including sociodemographics, life satisfaction, behavioral changes, and Brief COPE subscales. The target outcome was the presence of SI (SIDAS score > 0). Performance was measured by the area under the receiver operating characteristic curve (AUROC).
Results
The model demonstrated strong external validity on the entire Croatian sample, achieving an AUROC of 0.80. Performance remained robust across subgroups: males (AUROC = 0.83), females (AUROC = 0.79), younger adults (AUROC = 0.77), and older adults (AUROC = 0.81). Self-blame, behavioral disengagement, and relationship dissatisfaction were key predictors.
Conclusions
An indirect, ML-based screening tool can reliably identify SI risk in the general population. The model demonstrated strong cross-national transferability and resilience during a societal crisis, proving it is a feasible and valid strategy for population-level prevention.
Melanie, a mother of two, has struggled with severe contamination-related obsessive-compulsive disorder (OCD) for years, spending hours each day cleaning and avoiding activities that feel “unsafe,” which causes constant anxiety. Her symptoms make daily tasks difficult, as she cannot play with her children outside, do groceries without taking drastic measures, or visit friends and family without intense fear of contamination. Her rigid routines and compulsions take a toll on her mental and physical health and her children’s, leaving her feeling isolated and overwhelmed. Melanie knows she needs specialized treatment. However, living in a rural area, the nearest specialized clinic is too far away for a day treatment program, inpatient care is not an option due to childcare responsibilities, and outpatient treatment is inaccessible due to long waiting lists and insufficient intensity for her needs. As a result, she remains stuck, unable to access the intensive therapy she requires.
Cannabis use is elevated in youth with depression and attention-deficit/hyperactivity disorder (ADHD), but drivers of this increase remain underexplored. The self-medication hypothesis suggests cannabis is used by patients for mood regulation, a common difficulty in ADHD and depression. This study aimed to examine associations between mood instability and cannabis use in a large, representative clinical cohort of adolescents diagnosed with ADHD and/or depression.
Methods
Natural language processing (NLP) approaches were utilised to identify references to mood instability and cannabis use in the electronic health records of adolescents (aged 11–18 years) with primary diagnoses of ADHD (n = 7,985) or depression (n = 5,738). Logistic regression was used to examine mood instability as the main exposure for cannabis use in models stratified by ADHD and depression.
Results
Mood instability was associated with a 25% higher probability of cannabis use in adolescents with ADHD compared to those with depression. Following adjustment for available sociodemographic and clinical covariates, mood instability was associated with increased cannabis use in both ADHD (aOR: 1.61 [95% CI: 1.41–1.84]) and depression (aOR: 1.38 [95% CI: 1.21–1.57]) groups.
Conclusions
This was the first study to explore the differential impact of mood instability on adolescent cannabis use across distinct diagnostic profiles. NLP analysis proved an efficient tool for examining large populations of adolescents accessing psychiatric services and provided preliminary evidence of a link between mood instability and cannabis use in ADHD and depression. Longitudinal studies using direct measures or tailored NLP techniques can further establish the directionality of these associations.
The use of artificial intelligence (AI) in psychiatry holds promise for diagnosis, therapy, and the categorization of mental disorders. At the same time, it raises significant theoretical and ethical concerns. The debate appears polarized, with proponents and critics seemingly irreconcilably opposed. On the one hand, AI is heralded as a transformative force poised to revolutionize psychiatric research and practice. On the other hand, it is depicted as a harbinger of dehumanization. To better understand this dichotomy, it is essential to identify and critically examine the underlying arguments. To what extent does the use of AI challenge the theoretical assumptions of psychiatric diagnostics? What implications does it have for patient care, and how does it influence the professional self-concept of psychiatrists?
Methods
To explore these questions, we conducted 15 semi-structured interviews with experts from psychiatry, computer science, and philosophy. The findings were analyzed using a structuring qualitative content analysis.
Results
The analysis focuses on the significance of AI for psychiatric diagnosis and care, as well as on its implications for the identity of psychiatry. We identified different lines of argument suggesting that expert views on AI in psychiatry hinge on the types of data considered relevant and on whether core human capacities in diagnosis and treatment are viewed as replicable by AI.
Conclusions
The results provide a mapping of diverse perspectives, offering a basis for more detailed analysis of theoretical and ethical issues of AI in psychiatry, as well as for the adaptation of psychiatric education.
Artificial intelligence (AI) is transforming nearly every domain of science, and scholarly publishing is no exception. From automated language editing to machine-assisted peer review and large-scale content analysis, AI tools are increasingly embedded in scientific writing and publishing. The response from the scientific community has ranged from cautious optimism to outright skepticism. This viewpoint aims to articulate an editorial perspective on the integration of AI into scientific writing and publishing, evaluating both the opportunities and tensions that arise, and offering principles for navigating the road ahead.
Improving media adherence to World Health Organization (WHO) guidelines is crucial for preventing suicidal behaviors in the general population. However, there is currently no valid, rapid, and effective method to evaluate the adherence to these guidelines.
Methods
This comparative effectiveness study (January–August 2024) evaluated the ability of two artificial intelligence (AI) models (Claude Opus 3 and GPT-4O) to assess the adherence of media reports to WHO suicide-reporting guidelines. A total of 120 suicide-related articles (40 in English, 40 in Hebrew, and 40 in French) published within the past 5 years were sourced from prominent newspapers. Six trained human raters (two per language) independently evaluated articles based on a WHO guideline-based questionnaire addressing aspects, such as prominence, sensationalism, and prevention. The same articles were also processed using AI models. Intraclass correlation coefficients (ICCs) and Spearman correlations were calculated to assess agreement between human raters and AI models.
Results
Overall adherence to WHO guidelines was ~50% across all languages. Both AI models demonstrated strong agreement with human raters, with GPT-4O showing the highest agreement (ICC = 0.793 [0.702; 0.855]). The combined evaluations of GPT-4O and Claude Opus 3 yielded the highest reliability (ICC = 0.812 [0.731; 0.869]).
Conclusions
AI models can replicate human judgment in evaluating media adherence to WHO guidelines. However, they have limitations and should be used alongside human oversight. These findings may suggest that AI tools have the potential to enhance and promote responsible reporting practices among journalists and, thus, may support suicide prevention efforts globally.
Nowadays, both researchers and clinicians alike have to deal with increasingly larger datasets, specifically also in the context of mental health data. Sophisticated tools for dataset visualization of information from various item-based instruments, such as questionnaire data or data from digital applications or clinical documentations, are still lacking, specifically for an integration at multiple levels and for use in both data organization and appropriate construction for its valid use in subsequent analyses.
Methods
Here, we introduce ItemComplex, a Python-based framework for ex-post visualization of large datasets. The method exploits the comprehensive recognition of instrument alignments and the identification of new content networks and graphs based on item similarities and shared versus differential conceptual bases within and across data and studies.
Results
The ItemComplex framework was evaluated using four existing large datasets from four different cohort studies and demonstrated successful data visualization across multi-item instruments within and across studies. ItemComplex enables researchers and clinicians to navigate through big datasets reliably, informatively, and quickly. Moreover, it facilitates the extraction of new insights into construct representations and concept identifications within the data.
Conclusions
The ItemComplex app is an efficient tool in the field of big data management and analysis addressing the growing complexity of modern datasets to harness the potential hidden within these extensive collections of information. It is also easily adjustable for individual datasets and user preferences, both in the research and clinical field.
Recent advances in natural language processing (NLP), particularly in language processing methods, have opened new avenues in semantic data analysis. A promising application of NLP is data harmonization in questionnaire-based cohort studies, where it can be used as an additional method, specifically when only different instruments are available for one construct as well as for the evaluation of potentially new construct-constellations. The present article therefore explores embedding models’ potential to detect opportunities for semantic harmonization.
Methods
Using models like SBERT and OpenAI’s ADA, we developed a prototype application (“Semantic Search Helper”) to facilitate the harmonization process of detecting semantically similar items within extensive health-related datasets. The approach’s feasibility and applicability were evaluated through a use case analysis involving data from four large cohort studies with heterogeneous data obtained with a different set of instruments for common constructs.
Results
With the prototype, we effectively identified potential harmonization pairs, which significantly reduced manual evaluation efforts. Expert ratings of semantic similarity candidates showed high agreement with model-generated pairs, confirming the validity of our approach.
Conclusions
This study demonstrates the potential of embeddings in matching semantic similarity as a promising add-on tool to assist harmonization processes of multiplex data sets and instruments but with similar content, within and across studies.
Cardiovascular disease (CVD) is twice as prevalent among individuals with mental illness compared to the general population. Prevention strategies exist but require accurate risk prediction. This study aimed to develop and validate a machine learning model for predicting incident CVD among patients with mental illness using routine clinical data from electronic health records.
Methods
A cohort study was conducted using data from 74,880 patients with 1.6 million psychiatric service contacts in the Central Denmark Region from 2013 to 2021. Two machine learning models (XGBoost and regularised logistic regression) were trained on 85% of the data from six hospitals using 234 potential predictors. The best-performing model was externally validated on the remaining 15% of patients from another three hospitals. CVD was defined as myocardial infarction, stroke, or peripheral arterial disease.
Results
The best-performing model (hyperparameter-tuned XGBoost) demonstrated acceptable discrimination, with an area under the receiver operating characteristic curve of 0.84 on the training set and 0.74 on the validation set. It identified high-risk individuals 2.5 years before CVD events. For the psychiatric service contacts in the top 5% of predicted risk, the positive predictive value was 5%, and the negative predictive value was 99%. The model issued at least one positive prediction for 39% of patients who developed CVD.
Conclusions
A machine learning model can accurately predict CVD risk among patients with mental illness using routinely collected electronic health record data. A decision support system building on this approach may aid primary CVD prevention in this high-risk population.