Machine learning methods and schizophrenia spectrum disorders: a scoping review

Abstract

Background

Schizophrenia spectrum disorders (SSDs) are chronic conditions associated with profound functional impairment, disability and economic burden. Traditional approaches to diagnosis and management often struggle to capture the heterogeneity of SSD. In recent years, machine learning (ML) has expanded rapidly in mental health research, offering ways to analyse complex datasets, uncover subtle patterns and generate predictive models that could inform more individualised interventions.

Objectives

This scoping review aimed to map how ML has been applied to SSD research with emphasis on clinical and functional outcome trajectories.

Eligibility criteria

Peer-reviewed studies investigating any use (eg, predictive tool, treatment algorithm) of ML in determining outcomes in SSD were eligible. Non-English language studies, grey literature, systematic reviews and conference abstracts were excluded.

Sources of evidence

Following Arksey and O’Malley’s framework and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) extension for scoping reviews, we searched three bibliographic databases (MEDLINE, APA PsycInfo, Embase) with input from a health sciences librarian. The initial search was completed on 28 July 2023 and re-run on 10 March 2025.

Charting methods

Two independent reviewers screened studies at title/abstract and full-text stages and extracted data using a standardised table in Covidence, capturing study design, ML techniques, data sources and outcomes. Methodological and reporting quality of included studies was appraised using the APPRAISE-AI instrument.

Results

88 studies met inclusion criteria. For included studies, n=62 were cross-sectional, n=19 were cohort studies, n=5 were case-controls, n=1 was an evaluation study and n=1 carried out a secondary analysis of data from a randomised controlled trial. Overall, methodological and reporting quality were moderate (mean APPRAISE-AI score 47.4/100, range 31–65). ML applications were grouped into five domains: clinical, treatment, functional, behavioural and provider outcomes. Clinical outcome studies used electronic health records, clinical assessments, neuroimaging and mobile sensing to predict relapse, remission, symptom severity and suicide risk. Treatment studies modelled treatment response, resistance and adverse effects, with attention to clozapine and electroconvulsive therapy. Functional studies employed deep learning to analyse speech, social cognition and real-world functioning. Behavioural outcomes included prediction of aggression, self-harm and offending. Provider outcomes examined prescribing patterns using supervised learning. Across domains, integration of multimodal data and adoption of advanced approaches such as deep neural networks were evident.

Conclusions

ML in SSD research shows significant promise for advancing diagnosis, prognosis and treatment planning. However, methodological rigour, interpretability, fairness and clinical validation remain critical for future translation into practice.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• To our knowledge, this is the first review to map machine learning applications in schizophrenia spectrum disorders across a breadth of clinical, treatment, functional, behavioural and provider outcomes, rather than a single outcome domain such as diagnostic classification.

• A comprehensive, librarian-designed search strategy was applied across seven bibliographic databases and re-run to extend coverage to March 2025.

• Screening and data extraction were performed independently by two reviewers with a third reviewer resolving disagreements and reliability was calibrated using a 5% sample.

• Methodological and reporting quality was formally appraised using APPRAISE-AI, an instrument developed for artificial intelligence studies supporting clinical decision-making; most included studies were of moderate quality, with reproducibility and methodological conduct the weakest domains.

• Restriction to English-language, peer-reviewed publications may have omitted relevant evidence, and heterogeneity in outcomes, data types and reported performance metrics precluded quantitative synthesis of model performance.

Introduction

Schizophrenia spectrum disorders (SSDs) are associated with significant functional impairment, disability and low rates of personal recovery.^{1 2} These disorders pose an immense burden in terms of economic costs, generally due to lost productivity and premature mortality.^{3 4} These challenges are compounded by societal stigma, leading to low rates of personal recovery and integration into the community.^{5 6} The heterogeneous nature of SSD, characterised by varying symptom profiles and outcomes, underscores the urgent need to better understand long-term functional outcomes of people affected by SSDs.^{7 8}

The use of artificial intelligence, specifically the application of machine learning (ML), has become increasingly popular in mental health research and clinical practice, particularly concerning schizophrenia.^{9 10} ML, a subset of artificial intelligence, involves the development of algorithms that can learn from and make predictions based on data.¹¹ In the context of schizophrenia, ML techniques have been employed to analyse complex datasets, ranging from neuroimaging scans to genetic information and electronic health records.^12–14 These advanced analytical methods enable the identification of patterns and correlations that may not be evident through traditional statistical approaches.

For instance, ML algorithms can predict disease onset, progression and treatment response by integrating diverse data sources, potentially leading to more personalised and effective interventions for individuals with SSD.¹⁵ Recent studies have demonstrated the potential of ML in early detection of psychosis risk,¹⁴ prediction of treatment outcomes¹⁶ and identification of neuroimaging biomarkers associated with SSD.¹⁷ Moreover, ML approaches have shown promise in analysing language patterns and social media data to detect early signs of relapse or symptom exacerbation.^{18 19}

Prior meta-analyses and reviews have examined ML applications in SSD. Wang et al examined task-based fMRI diagnostic classification; Di Camillo et al focused on structural and functional MRI-based classification; and Parsaei et al addressed prediction of violent behaviour. These examples each focused on a single outcome domain.^20–22 To our knowledge, no prior review has mapped ML applications across the breadth of clinical, functional, behavioural, treatment and provider outcome domains in SSD.

This scoping review aims to characterise how ML has been used for determining clinical and functional outcome trajectories in individuals with SSD, including psychotic affective disorders. By synthesising the current state of ML applications in SSD research, this review seeks to identify promising avenues for future investigation and explore the potential impact of ML on improving outcomes for individuals living with SSD.

Methods

Given that ML applications within mental health are a novel and rapidly evolving field, few primary studies were expected. For this reason, a scoping review was chosen from the available methodologies to gain a comprehensive sense of existing evidence and to identify key knowledge gaps. Typically, scoping review stages involve determination of the research question(s); consultation with a librarian to identify key search terms; screening and abstracting identified articles; discussing and synthesising the results; and finally creating a scientific manuscript. The proposed scoping review was conducted in five stages which have been adapted from Arksey and O’Malley as described below. Additionally, the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) extension for scoping reviews was used to ensure reporting and synthesis of evidence met current standards for reporting on knowledge synthesis research.

The primary question is: ‘How is ML being used for individuals with schizophrenia spectrum disorders toward determining treatment and functional outcome trajectories?’

Search strategy

A comprehensive search strategy was designed by a health sciences librarian in consultation with the research team. The core strategy was designed, tested, and finalised in MEDLINE (Ovid), and then translated and run in APA PsycInfo and Embase bibliographic databases. The search strategies used database-specific subject headings, keywords in natural language and advanced search operators to capture two concepts: ‘artificial intelligence’ and ‘schizophrenia spectrum disorders’. Publication year limit for all original searches was 2023, with no earlier year limit; the initial search was completed 28 July 2023, and re-run on 10 March 2025. The following publication types were removed using the screen when possible: book chapters, dissertations, editorials, letters, commentaries and conference abstracts. To supplement the database searches, reference lists of included studies were scanned for additional eligible articles. The full search strategies are available as supplementary files (online supplemental table S1).

Screening and selection

The authors discussed and determined the eligibility criteria which included: peer-reviewed studies investigating any use (eg, predictive tool, treatment algorithm) of ML in determining outcomes in psychotic disorders. Exclusion criteria involved the following characteristics that disqualify prospective articles from being included: non-English studies, grey literature, systematic reviews and conference abstracts. Two levels of screening took place: a title and abstract review for articles that fit predefined criteria and a full-text review of articles that meet those criteria. Two investigators reviewed each article and extracted data using Covidence, a systematic review management software. A 5% random sample of the results from the search strategy was used to calibrate and establish the reliability of the screening process applied. The review team met to compare their assessments and to modify the eligibility criteria as needed. There were two levels of screening: a title and abstract review to identify articles meeting criteria and a full-text review of those passing the first screen. For both levels, two independent reviewers (LAR, KAC-D) performed the screen, met, and compared results, and, where needed, discussed with a third reviewer to resolve any differences.

Data extraction and synthesis

As with screening, two independent investigators (LAR, KAC-D) reviewed and extracted data from the selected articles. Data extraction was carried out using an extraction table developed by the authors on Covidence. Information extracted from studies included: author, year of study publication, study population and size, ML techniques used, study objectives and summarised key findings. The literature was extracted by LAR and KAC-D and then analysed based on the study outcomes of focus.

Quality appraisal

The methodological and reporting quality of the included studies was assessed using the APPRAISE-AI instrument.²³ Full-text articles were evaluated by two reviewers using the instruments’ item-level scoring criteria. The maximum score is 100. Overall scores were categorised as very low quality (0–19), low quality (20–39), moderate quality (40–59), high quality (60–79) or very high quality (80–100).

Patient and public involvement

No patients or members of the public were involved in the design, conduct, reporting or dissemination plans of this scoping review.

Results

Our scoping review revealed a diverse landscape of ML applications in the study of SSD. In total, 88 studies were analysed and included in the final study. Of these, 62 (70.5%) were cross-sectional, 19 (21.6%) were cohort studies, 5 (5.7%) were case-control studies, 1 (1.1%) was an evaluation study and 1 (1.1%) was a secondary analysis of a randomised controlled trial. In total, 20 studies (22.7%) used a longitudinal design capable of modelling change over time. Sample sizes ranged between n=18 and n=24 614 study participants. Figure 1 illustrates the complete study PRISMA diagram. Further detailed descriptive information about included articles is contained in a supplementary table (online supplemental table S2).

Figure 1. Preferred Reporting Items for Systematic Reviews and Meta-Analyses diagram.AI, Artificial Intelligence.

Of the 88 included studies, 55 (62.5%) drew on two or more data modalities and 33 (37.5%) relied on a single modality. Clinical assessment data were most frequent (n=71), followed by electronic medical record/health administrative data (n=46), neuroimaging or neurophysiological data (n=16), speech or audiovisual data (n=7), laboratory and other biological measures (n=6) and mobile sensing data (n=3). Multimodal studies most commonly combined electronic medical record data with clinical rating scales, or neuroimaging with clinical rating scales.

The research landscape can be categorised into five primary domains based on the outcomes of focus: clinical, treatment, functional, behavioural and provider outcomes (table 1).

Table 1. Summary of types of outcomes and corresponding data types elicited via scoping review for individuals with SSD.

1. Clinical outcomes Symptom burden Depression Psychosis Obsessive-Compulsive Disorder Sleep Symptom severity Symptom remission Relapse prediction Suicide risk Disease progression Hospital readmission Length of hospitalisation Laboratory values	Data types: Electronic medical records Clinical assessment scales (eg, PANSS, BPRS, CGI) Neuroimaging data (MRI, EEG) Speech and language data Mobile sensing data Machine learning methods: Supervised learning: Support Vector Machines (SVM), random forests, gradient boosting, logistic regression Unsupervised learning: topic modelling, latent class growth analysis Deep learning: neural networks Regularisation methods: LASSO
2. Treatment outcomes Treatment response Clozapine Electroconvulsive Therapy Treatment resistance Treatment discontinuation Treatment adherence Adverse treatment effects including toxicity	Data types: Electronic medical records Clinical assessment scales EEG data Machine learning methods: Supervised learning: support vector machines, random forests, artificial neural networks, kernel partial least squares regression
3. Functional outcomes Social cognition Social functioning Social integration Social isolation Educational/academic functioning Quality of life Resilience	Data types: Clinical assessment scales Speech data Mobile sensing data Machine learning methods: Supervised learning: random forests deep learning: neural network analysis, Long Short-Term Memory (LSTM) neural networks Probabilistic methods: Bayes network
4. Behavioural outcomes Self-injurious behaviour Aggressive behaviours Violent offending Completed homicide Recidivism	Data types: Electronic medical records Clinical assessment scales Neuroimaging data Machine learning methods: Supervised machine learning using various algorithms, support vector machines, random forests
5. Provider outcomes Prescribing patterns	Data types: Electronic medical records Machine learning methods: Supervised machine learning: random forests, support vector machines, ordinary least square regression

BPRS, Brief Psychiatric Rating Scale; CGI, Clinical Global Impression; PANSS, Positive and Negative Syndrome Scale; SSD, Schizophrenia spectrum disorders.

Studies were assigned to a single outcome domain based on the primary analytic objective of the reported ML model, that is, what the model was built to predict or classify, rather than by symptom domain alone. For example, suicide risk was retained under clinical outcomes since studies operationalised it as a risk-assessment outcome derived from clinical and health record data, consistent with how it is assessed in routine clinical care. Self-injurious behaviour was retained under behavioural outcomes because it was analysed as an enacted behavioural event alongside aggression, offending and recidivism.

Clinical outcomes

A wide range of data sources were used to investigate clinical outcomes. Electronic medical records with both structured and unstructured data have been a rich source of information,^{24 25} complemented by standardised clinical assessment scales such as Positive and Negative Syndrome Scale (PANSS), Brief Psychiatric Rating Scale and Clinical Global Impression.^{26 27} Additionally, researchers have used neuroimaging data from MRI,^28–30 EEG, speech data and data from mobile sensing technologies such as smartphones and videos.^31–33 These diverse sources have enabled the application of sophisticated machine learning techniques, ranging from traditional supervised learning methods like Support Vector Machines^{19 29 34} and Random Forests^{35 36} as well as regularisation techniques like LASSO,^{33 37} to more advanced approaches such as deep learning. These tools have been applied to a broad spectrum of clinical challenges, including predicting symptom burden across various domains (psychotic symptoms, psychotic affective including depression symptoms, obsessive-compulsive disorder symptoms, isolated sleep disturbances), assessment and diagnostic aims, assessing symptom severity and remission, forecasting relapses, evaluating suicide risk, modelling disease progression and predicting hospital readmissions and length of stay.^38–41

Treatment outcomes

There has been application of machine learning, primarily drawing from electronic medical records, clinical assessment scales and EEG data on treatment outcomes. Researchers have employed supervised learning techniques, including Support Vector Machines, Random Forests, Artificial Neural Networks and Kernel Partial Least Squares Regression. These methods have been used to predict treatment response, with particular attention to specific interventions such as clozapine pharmacotherapy^42–44 and electroconvulsive therapy.^{45 46} Other important areas of investigation include treatment resistance, factors influencing treatment discontinuation and adherence and the prediction of adverse treatment effects.^47–50

Functional outcomes

A combination of clinical assessment scales, speech data and mobile sensing technologies have been used to investigate functional outcomes. This domain has an emphasis on advanced ML techniques, including deep learning approaches such as neural network analysis and Long Short-Term Memory (LSTM) neural networks, as well as probabilistic methods like Bayes Networks. These tools have been applied to study various aspects of social and cognitive functioning, including social cognition, social integration, educational and academic performance, and overall quality of life.^{33 51–53}

Behavioural outcomes

Researchers have relied on electronic medical records, clinical assessment scales and neuroimaging data to study behavioural outcomes. The ML approaches in this domain have largely focused on supervised learning techniques, including Support Vector Machines and Random Forests. These methods have been applied to study and predict a range of challenging behaviours associated with SSD, including self-injurious behaviour,^{36 54 55} aggression, violent offending, homicide and recidivism.^{56 57}

Provider outcomes

A small body of research has focused on provider outcomes, particularly prescribing patterns.^{58 59} This work has primarily used patient data from hospital visits and healthcare encounters via electronic medical records and employed supervised machine learning methods such as Random Forests, Support Vector Machines and Ordinary Least Square Regression to analyse and predict physician behaviour in managing SSD.

Reported model performance

Performance was reported heterogeneously across studies, expressed variously as accuracy, balanced accuracy, area under the receiver operating curve (AUC), sensitivity and specificity and correlation between predicted and observed values. Consistent with scoping review methodology, estimates were not pooled; representative results are described below to illustrate the range of performance reported. Within clinical outcomes, Kottaram et al reported dynamic resting-state connectivity predicted change in positive, negative and overall symptom severity with accuracies of 87%, 83% and 77% respectively.³⁰ A digital phenotyping model classified disease status with AUC of 0.94,⁶⁰ while a speech-derived model predicted PANSS with Pearson correlations of 0.51 and 0.77.⁶¹ For treatment outcomes, gradient boosting predicted adverse treatment courses with a balanced accuracy of 67% and an AUC of 0.72.⁵⁸ Within functional outcomes, regularised regression predicted employment or education status with an AUC of 0.88, while the same modelling approach predicted remission less well (AUC 0.63).⁶² A speech-derived model predicted social skills and functional competency with correlations of 0.79 and 0.62.⁶¹ Within behavioural outcomes, gradient boosting distinguished offender from non-offender patients with a balanced accuracy of 79.9% and AUC of 0.87.⁵⁶ Studies of provider outcomes did not report comparable metrics.

Most studies either did not address generalisability at all or only mentioned it as a limitation without substantive discussion. When generalisability was actively addressed, researchers typically used cross-validation techniques, particularly leave-one-site-out and 10-fold cross-validation; testing models on unseen data/test sets; and geographical validation across different recruiting centres. Examples include nested leave-one-out cross-validation across multiple recruiting centres,⁶² leave-one-out cross-validation,^{48 50} cross-validation frameworks applied within single-site samples²⁹ and replication of a finding in an independent validation sample.⁵¹ Among studies that discussed generalisability in their limitations, the recurrent theme was acknowledgement of uncertainty about model performance in broader populations or different clinical settings, most often attributed to single-site recruitment, modest sample size or the absence of external validation.

Quality appraisal

All 88 included studies underwent quality appraisal. The mean APPRAISE-AI score was 47.4 out of 100 with a median of 17 and a range of 31–65. Most studies were classified as moderate quality (n=72). 11 studies were classified as low quality and five as high quality. No studies were classified as very low quality or very high quality. Among the lowest scoring domains for included studies were ‘robustness of results’ (mean 9.1/20) and ‘methodological conduct’ (mean 6.2/20). Complete domain-level scores and total scores are available in online supplemental table S3.

Discussion

Our scoping review has revealed a diverse and rapidly evolving landscape of ML applications in SSD research. The main areas where ML is being applied can be categorised into five domains: clinical outcomes; treatment outcomes; functional outcomes; behavioural outcomes; and provider outcomes. Across these categories, we observed a trend towards integrating multiple data types, including electronic medical records, neuroimaging data, clinical assessment scales and novel data streams such as mobile sensing and speech analysis. The most common ML methods used include supervised learning techniques like Support Vector Machines (SVM) and Random Forests, as well as more advanced approaches such as deep learning neural networks. We also noted a lack of standardisation in how generalisability is addressed in this field.

Across the included studies, Support Vector Machines and Random Forests were the most frequently applied algorithms overall, appearing across all five outcome domains, reflecting their relative interpretability and reasonable performance on the small/moderate sample sizes. Deep learning approaches, including neural networks and LSTM architectures, were concentrated in functional outcome studies and in clinical outcome studies using neuroimaging or mobile sensing data. Regularisation methods such as LASSO appeared primarily in clinical outcome studies.

Regarding the suitability of various methods to a given dataset, our work suggests that model choice broadly tracked with data characteristics: traditional supervised approaches such as SVM and Random Forests were applied to structured, lower-dimensional clinical and electronic health record data, where interpretability and smaller sample sizes favour these methods. Deep learning approaches were concentrated in high-dimensional, unstructured data such as neuroimaging and speech data. We note, though, that few included studies directly benchmarked multiple ML approaches head-to-head on the same dataset and outcome type, so these represent review-level observations about how methods have been applied in practice, rather than a validated recommendation for which method is preferred or superior for a given data type.

The use of multiple data types and application of sophisticated ML techniques reflects the complex, multifaceted nature of SSDs and highlights the potential of ML to uncover novel insights and improve patient outcomes. Notably, there has been a shift towards more sophisticated ML techniques, particularly in functional and clinical outcome prediction, with deep learning approaches showing promise in analysing complex, high-dimensional data such as neuroimaging and speech patterns.⁶³

The application of ML in SSD research holds potential for improving clinical practice across several domains. Regarding diagnosis, ML algorithms have demonstrated the ability to identify subtle patterns in text, neuroimaging and clinical data that may aid in early detection and differential diagnosis of SSD.^{12 14} Predictive models developed using ML techniques have shown promise in forecasting disease trajectories, relapse risk and long-term functional outcomes.¹⁶ In the realm of treatment planning, ML-based approaches have been used to predict treatment response and adverse effects, potentially enabling more personalised and effective treatment strategies.¹⁵ These advancements offer opportunities for personalised medicine in SSD, moving towards a more individualised approach to patient care. For example, ML models could potentially be used to identify patients at high risk of treatment resistance, allowing for earlier intervention with clozapine or other targeted therapies. Similarly, ML-driven prediction of functional outcomes could inform rehabilitation strategies and social support interventions.

While ML approaches offer powerful tools for analysing complex datasets in SSD research, several methodological considerations must be addressed. The field remains predominantly at the model development stage, with robustness and methodological conduct being the lowest-scoring domains when the included articles in this review were subject to APPRAISE-AI quality appraisal. The performance of ML models is heavily dependent on the quality and quantity of available data. Issues such as missing data, small sample sizes and lack of standardisation across datasets pose significant challenges.⁶⁴ Many advanced ML techniques, particularly deep learning models, operate as ‘black boxes,’ making it difficult to understand the reasoning behind their predictions. This lack of interpretability can be a significant barrier to clinical adoption.⁶⁵ Finally, generalisability remains problematic; models developed on specific patient populations or datasets may not transfer well to other contexts or populations.

These generalisability concerns take on particular urgency given equity considerations. Performance disparities in ML predictions for socially marginalised populations have emerged which raises concern that ML could inadvertently amplify health inequity. This risk is particularly significant for patients with SSD, where many individuals already face a great deal of social disadvantage and barriers to accessing care. Biases in training data could disproportionately harm specific populations, particularly regarding behavioural outcome predictions such as violence or recidivism.

Several limitations of this scoping review should be noted. Given the fast-evolving nature of ML research, some very recent developments, such as advances in generative AI applications, may not be captured. This review focused on research applications with limited insight into real-world clinical implementation. Additionally, publication bias, English language restriction, and the absence of grey literature may overrepresent positive findings. A further issue is the heterogeneity across the datasets used in this review. Included studies varied widely in sample size (ranging from n=18 to n=24 614). Neuroimaging studies varied in acquisition protocols, and speech, text and social-media-biased studies varied in language and cultural context. This variance likely contributes to the inconsistent and possibly site-specific performance of ML models noted throughout this review and complicates efforts to pool or directly compare findings across studies.

Beyond bias and equity, broader ethical considerations warrant attention as ML moves closer to clinical use for patients with SSD. These include obtaining informed consent for the secondary use of clinical data in model development, ensuring appropriate data privacy and security safeguards given the sensitivity of medical records, and establishing clear accountability when ML-informed decisions contribute to clinical decision-making. Meaningfully involving patients, clinicians and people with lived experience of severe mental illnesses in the design, evaluation and governance of ML tools, consistent with the Patient and Public Involvement principles noted in our Methods, is likely to be important in developing applications that are clinically useful and trusted by the individuals they are intended to benefit.

Based on our review, there are several promising areas for future research and development in ML applications for SSD. This includes potential advancements from combining diverse data types (eg, genomics, neuroimaging, clinical assessments and real-world functioning data) to provide a more comprehensive understanding of SSD.⁶⁶ Developing ML approaches that can effectively model the dynamic, time-varying nature of SSD symptoms and functioning can also be clinically useful.⁹ Finally, advancing techniques for making complex ML models more accessible and user-friendly could assist in promoting more inclusive and interdisciplinary research as well as trust among care providers.

ML applications in SSD research offer promising avenues for advancing our understanding of these complex disorders and improving patient outcomes. The integration of diverse data types and increasingly sophisticated ML techniques has the potential to enable more personalised and effective approaches to diagnosis, prognosis and treatment in SSD. However, realising this potential will require addressing significant methodological, ethical and implementation challenges. As the field continues to evolve, close collaboration between data scientists, clinicians, ethicists and individuals with lived experience of SSD will be crucial to ensure that ML-driven innovations translate into meaningful improvements in clinical care and quality of life for individuals with SSD.

Data availability statement

Data are available upon reasonable request.