The field of NLP emerged in the 1950s, a product of the confluence of AI, linguistics, and mathematics. It is an analysis technique that enables the automatic decoding and understanding of language with the support of machines [24]. The field is subdivided into two distinct domains: the first is natural language understanding (NLU), which translates natural language into a machine-understandable language [25]; the second is natural language generation (NLG), which produces a natural language output from structured data (e.g., text to text, text to other and vice versa) [26]. The process used by NLP consists of four parts: pre-processing of the text, representation of the text, training of the model, and evaluation of the model. In the pre-processing stage, the text is simplified and corrected to enhance its accuracy and efficiency in the subsequent steps. Subsequently, the text is converted into numerical values, which are then used to train the algorithm. The efficacy of the algorithm is then evaluated in real-world settings [27].

The algorithms encompassed within the domain of ML are based on the principle of “learning”, understood as the process of acquiring information through observation and experience, which is subsequently used to infer generalizable patterns or rules and apply them to new data.

Machine learning can be categorized into three distinct models [9]: Supervised Learning, where the machine has access to all the data and is tasked with the construction of a function that elucidates the underlying phenomenon; Reinforcement Learning, where the machine is provided with data sequentially; and Unsupervised Learning, where the machine is endowed with all the input and output data, yet it is not subdivided. In addition to these three categories, there are also systems known as Semi-Supervised Learning (a combination of Supervised and Unsupervised) [28] and Deep Learning.

Deep Learning is a sub-type of Machine Learning that utilizes a system of operation based on the structure of neural networks (neural networks), wherein learning occurs through continuous “training”. The term “deep neural networks” (DNN) refers to the architecture of these networks, which is based on simple units called “artificial neurons” (ANN). These artificial neurons are combined in interconnected layers (typically three or more, although recent algorithms can have hundreds of layers). The term “deep” in “deep neural networks” refers to the depth of this network architecture. This phenomenon mirrors the process of learning in the human nervous system, wherein the strength of connections between neurons, or “weights”, is adjusted, allowing for the establishment of more functional connections that are prioritized [29]. Artificial neural networks process input data in a non-linear manner, exhibiting characteristics like those of brain neurons, such as activation only when a threshold potential is reached, which is determined by the weighted sum of the inputs, and subsequent output generation.

It is imperative to underscore the fundamental distinction between the operational principles of deep learning and those of non-deep learning machine learning algorithms. The former is predicated on the processing of unlabeled (“raw”) data, whereas the latter is designed to extrapolate and select data based on the inputs provided by the programmer. A critical aspect of deep learning is the inherent processing and extrapolation of data by the algorithm itself, which is subsequently utilized to formulate responses to queries. In contrast, non-deep learning machine learning algorithms are constrained to the utilization of structured data with a substantial sample size, such as electroencephalography (EEG) or resting-state functional MRI data.

The primary aim of this review was to provide a comprehensive and systematic examination of how AI algorithms have been applied over the years in studies involving individuals diagnosed with schizophrenia spectrum disorders (SSDs). In particular, the review sought to map key characteristics of these studies, including the year of publication, the type of AI algorithm employed, the populations under investigation, the reported outcomes, and the instrumental techniques utilized. Focusing on outcomes and methodological approaches was intended to create a structured framework that could guide and inform future research in the field. As a secondary objective, the review also aimed to explore whether any of these study characteristics were associated with journal Impact Factor (IF), thereby providing insight into the topics currently attracting the greatest scientific attention.

Peer-reviewed articles were identified through searches of three electronic databases: PubMed, Scopus, and PsycINFO. Searches were conducted on titles, abstracts, and keywords using a strategy adapted for each database to optimize retrieval, as follows:

• Pubmed: (machine learning OR artificial intelligence OR deep learning OR natural language processing OR neural network OR unsupervised learning OR supervised learning OR data mining) AND (schizophrenia OR psychosis OR psychotic).

• Scopus: ((machine AND learning) OR (artificial AND intelligence*) OR (deep AND learning) OR (natural AND language AND processing) OR (neural AND network*) OR (unsupervised AND learning) OR (supervised AND learning) OR (data AND mining)) AND ((schizophrenia) OR (psychosis) OR (psychotic)).

• PsycINFO: ((machine AND learning) OR (artificial AND intelligence*) OR (deep AND learning) OR (natural AND language AND processing) OR (neural AND network*) OR (unsupervised AND learning) OR (supervised AND learning) OR (data AND mining)) AND ((schizophrenia) OR (psychosis) OR (psychotic)).

The present systematic review included studies published in peer-reviewed journals up to December 2, 2024. Only publications written in English were eligible for inclusion.

According to PICOS Reporting System [31] we defined the following:

- Population: studies with subjects diagnosed with a SSD were included. Papers on early or first-episode psychosis (FEP) were included, reflecting the growing research focus on this population. Those focusing on brief psychotic disorder, substance/medication-induced psychosis, or psychotic disorders secondary to other medical conditions were excluded due to their heterogeneous and typically transient nature, which may limit comparability with other SSDs.

- Intervention: studies employing AI algorithms belonging to the three categories described above (NLP, ML, and DL) were retained.

- Comparison: no pre-specified comparison groups were defined for this review. The only analytical comparison performed involved examining the distribution of journal IF across the study variables, in order to explore potential associations between IF and study characteristics.

- Outcomes: any results obtained, divided into the categories described later.

- Study: was not limited to a specific design; controlled and uncontrolled clinical studies, cohort studies, prospective case-control studies, and cross-sectional studies were considered for inclusion.

In line with the objectives of this review, a quantitative analysis was conducted on specific study parameters, including publication year, type of AI employed, study population, primary outcomes, and the instrumental techniques utilized.

A timeline analysis was conducted to describe the temporal evolution of the included publications. This analysis was performed systematically using the extracted data, aggregating studies by year of publication and reporting variations in the volume of scientific output over the examined period.

AI approaches were categorized as ML, DL, or NLP. When studies integrated multiple AI paradigms, the classification was based on the most specific algorithmic category; for instance, studies combining ML and DL methods were assigned to the DL category, as DL represents a specialized subset of ML techniques.

Study populations were divided into five categories: FEP, schizophrenia, SSDs including schizophrenia, SSDs excluding schizophrenia, and schizotypal personality disorder.

The outcomes of individual studies were organized into four overarching domains:

$\bullet$ Neurobiological correlates, encompassing the identification of brain regions and circuits potentially pathognomonic for the condition, estimation of brain age, and characterization of nervous system aging trajectories;

$\bullet$ Clinical characterization, including correlations of clinical, biological, behavioral, and functional parameters, such as psychopathological features, physical comorbidities, self- and hetero-injurious behaviors, objective functioning (e.g., service utilization), and subjective well-being;

$\bullet$ Diagnosis, covering both differentiation from healthy controls and the identification of differential diagnoses relative to other pathologies;

$\bullet$ Prognosis, addressing disease evolution, treatment outcomes (pharmacological and psychotherapeutic), treatment adherence, and prediction of disease progression or risk of transition.

Instrumental techniques employed for data collection were grouped as follows:

$\bullet$ Neurobiological measurements, including structural and functional neuroimaging (MRI, fMRI, PET, magnetoencephalography) and neurophysiological assessments (EEG);

$\bullet$ Audio/video recordings, captured during interviews or standardized tests;

$\bullet$ Biological samples, such as blood or saliva, analyzed for genetic, inflammatory, renal, hepatic, or metabolic markers;

$\bullet$ Other, comprising somatic measurements (e.g., height, weight, BMI) and movement tracking (e.g., eyes, head, limbs) using mobile applications equipped with gyroscopes and GPS sensors;

$\bullet$ No instrumental methods, for studies that relied exclusively on socio-demographic or psychometric data obtained through standardized tests and scales.

Studies employing multimodal datasets (e.g., combining MRI, EEG, or genetic data) were included by disaggregating the modalities and coding each instrumental technique separately.

Finally, journal IF was analyzed in relation to publication year. IF values were obtained from Journal Citation Reports (Clarivate Analytics). When the IF for the year of publication was unavailable, the value from the closest preceding year was used. This allowed us to assess potential associations between study characteristics and journal impact over time.

Following the removal of duplicates, the screening process—based on titles and abstracts—was independently conducted by two reviewers, working in pairs randomly assigned from a group of five (VB, CD, MD, MGre, MGhi). Any discrepancies or uncertainties were resolved by a sixth researcher (AZ), who served as adjudicator. Data extraction was subsequently carried out independently by four reviewers (VB, CD, MGre, MGhi), with the supervising author (AZ) performing a final verification to ensure consistency and accuracy across all extracted data. In addition, the reference lists of all included studies were manually reviewed to identify further relevant articles not captured by the electronic search. This multi-step, independent review process was designed to strengthen methodological rigor and minimize potential bias in study selection and data extraction.

All statistical analyses were performed using R (A language and environment for statistical computing. Version 4.4.2 (2024-10-31 ucrt). R Foundation for Statistical Computing, Vienna, Austria. https://www.r-project.org/), RStudio (Integrated Development Environment for R. Version 2025.05.0. Posit Software, PBC, Boston, MA, USA. https://posit.co/). A p-value of 0.05 or less was considered statistically significant. The primary objective was to analyze the entire database by describing the absolute and relative frequencies of each variable. As a secondary objective, a bibliometric analysis was conducted to assess the distribution of the IF across selected variables. For each study, the IF of the journal in which it was published during the year of publication was assigned in order to objectively assess the scientific impact of each paper [32, 33, 34]. Because the IF distribution was non-normal (as indicated by the Kolmogorov–Smirnov and Shapiro–Wilk tests, both significant at p 0.001), non-parametric tests were applied. Spearman’s rank correlation coefficient (R) was used to assess correlations involving IF over the years. For the remaining variables, the Kruskal–Wallis test was employed, followed, when significant, by pairwise comparisons using the Mann–Whitney (MW) test with Bonferroni correction. Studies reporting multiple outcomes or instrumental techniques were analyzed by treating each outcome or technique as a separate observation to provide a more detailed examination of methodological patterns and research trends. Although this approach may introduce some degree of non-independence among data points, the proportion of such cases was low, and its impact on the overall validity of the findings was considered minimal.

The temporal distribution of publications reveals a markedly accelerated growth of the field. As shown in Fig. 2, the earliest studies appeared in 1997 [35], followed by over a decade of sparse production, with annual outputs generally below ten articles. A substantive increase became evident after 2014, with a progressive escalation culminating in a sharp expansion from 2017 onwards. Annual publications rose from 35 in 2017 to 111 in 2020, ultimately reaching their peak in 2024 with 135 articles (15.83% of the overall sample). As detailed in Table 1, the cumulative distribution indicates that approximately half of all included studies had been published by the end of 2021 (52.05%), while the remaining 48% were produced in the subsequent three-year period, underscoring the rapid recent growth of research activity in this domain.

Fig. 2.

Number of articles per year and cumulative percentage.

As demonstrated in Table 2, the analysis focused predominantly on ML algorithms (592; 69.4%), with a subsequent emphasis on DL (215; 25.2%) and NLP (46; 5.4%).

Most of the studies analysed focused on subjects with Schizophrenia (616; 72.2%). Patients with SSDs including Schizophrenia, accounted for 144 of the studies (16.9%), while FEP accounted for 87 (10.2%). However, schizotypal personality disorder and SSDs excluding schizophrenia are poorly studied in AI (0.5% and 0.2%, respectively). The data are shown in Table 3.

In our systematic review, 101 studies (11.84%) evaluated two objectives, resulting in a total of 954 outcomes. The number of objectives increases to 960 if we analyze them more specifically, as described in the preceding paragraph (107 articles evaluated two objectives (12.54%)). The most common objective is the diagnosis of pathology in relation to healthy controls and other disorders (401; 42%), followed by the identification of neurobiological correlates (229; 24%), clinical characterisation (200; 21%) and prognosis (122; 12.8%). A rigorous analysis of the specific aims indicates that the two most common objectives, following diagnosis and differential diagnosis, are the identification of pathognomonic brain areas and circuits (216; 22.5%) and psychopathological aspects (146; 15.3%). For a more detailed analysis, please refer to Table 4a,4b.

Out of the 853 studies analysed, 143 (16.76%) did not use instrumental techniques and only collected socio-demographic variables and/or psychopathological aspects. Only one study (0.12%) used three instrumental techniques, 35 studies (4.10%) used two techniques, and the remaining 674 studies (79.01%) used only one technique. The categories were then divided into more specific techniques. The results showed that only 3 studies (0.35%) used three techniques, 49 studies (5.74%) used two techniques, and 658 studies (77.14%) used only one technique.

Neurobiological measurements are the most used instrumental techniques (515; 57.9%), primarily through neuroimaging (383; 42.3%) and to a lesser extent EEG (141; 15.6%). The second most used technique involves biological samples (120; 13.5%), with genetic component evaluation being the most prevalent (84; 9.3%). Audio and video recordings account for 7.6% of the total techniques used. Finally, “other” techniques make up 4.8% of the total in our sample.

For a detailed exposition of the instrumental techniques employed, please refer to Table 5a,5b.

The analysis revealed a very weak positive, yet statistically non-significant correlation between publication year and journal IF (Spearman’s $\rho$ = 0.038, 95% CI [–0.026, 0.108], p = 0.263), suggesting no meaningful temporal trend in the IF of published studies.

The non-parametric Kruskal-Wallis test was utilised to evaluate the relationship between the IF and the variables. In instances of statistical significance, subsequent comparisons between the variables were conducted employing the Mann-Whitney (MW) test, which was corrected for Bonferroni.

The variable “instrumental techniques” did not yield statistically significant results (p = 0.193), and thus no post-hoc comparisons were performed.

The overall Kruskal-Wallis test for “type of AI” reached statistical significance (p 0.05; $\eta$² = 0.006, 95% CI [–0.002, 0.019]), although the effect size was negligible. No pairwise comparison remained significant after Bonferroni correction ($\alpha$ = 0.016). The nominal ML–NLP difference (ML: median 4.3, IQR 3.3–5.3; NLP: median 3.9, IQR 2.9–4.5; uncorrected p = 0.019) did not withstand multiple-comparison adjustment and therefore should not be interpreted as substantively meaningful (detailed results are provided in Supplementary Table 1, included in the Supplementary Material B).

The statistical analysis performed using the Kruskal-Wallis test demonstrated a significant relationship between study outcomes and IF (p 0.001; $\eta$² = 0.02, 95% CI [0.003, 0.038]). Pairwise comparisons using the Mann-Whitney test revealed statistical significance for the comparison between the “diagnosis” (IF: median 4.0, IQR 3.1–5.1) and “neurobiological correlates” (IF: median 4.5, IQR 3.6–5.9) categories (p 0.001 after Bonferroni correction). Complete results are provided in Supplementary Table 2 and Supplementary Table 3 in the Supplementary Material B.

The Kruskal-Wallis test indicated a statistically significant association between recruited population and IF (p 0.05; $\eta$² = 0.002, 95% CI [–0.004, 0.01]), although the effect size was negligible. Subsequent pairwise comparisons using the Mann-Whitney test revealed no statistically significant differences. For further details, see Supplementary Table 4 and Supplementary Table 5 in the Supplementary Material B.

This review has several important limitations that should be acknowledged. First, restricting the included studies to those published in English may introduce a degree of publication bias; however, prior evidence suggests that this effect is generally minor [42, 43].

The heterogeneity of reported outcomes across studies further limited standardized comparisons and emphasized the descriptive nature of the review. Metrics such as sensitivity, specificity, and clinical utility were not consistently reported, precluding the adoption of a uniform framework for comparative analyses. Consequently, a descriptive rather than analytical approach was employed, which provides a broad overview of methodological trends but limits the ability to directly compare AI model performance and reduces the granularity of conclusions regarding clinical applicability.

Another limitation lies in the rigid categorization of AI algorithms into ML, DL, and NLP. While this classification enhances clarity and consistency, it may oversimplify hybrid or emerging approaches that span multiple categories, potentially underrepresenting more complex or multimodal AI methodologies.

The largely descriptive and bibliometric approach of the review also constrains inferential conclusions. Although some statistical tests were applied to compare variables, they do not support robust inferential claims, and readers should interpret these findings with caution. Additionally, some studies reported multiple outcomes or instrumental techniques, which were treated as separate observations. This may introduce a degree of non-independence among data points, potentially violating assumptions of non-parametric analyses; however, given the relatively low proportion of such cases, the overall impact is likely minimal. The use of journal IF as a proxy for study relevance represents a further limitation.

While IF is a widely recognized bibliometric indicator and provides a standardized measure of journal visibility and influence, it does not fully capture the quality, scientific rigor, or specific impact of individual studies. Therefore, IF should be interpreted cautiously, and complementary metrics may offer a more comprehensive assessment of research significance [44, 45].

Finally, most included studies lacked external validation using independent datasets, and prospective evaluations were generally absent. This limits the real-world generalizability and clinical applicability of the AI models [46, 47]. Moreover, the majority of studies focused exclusively on schizophrenia, leaving the applicability to other SSD subtypes uncertain. Given that algorithmic performance depends heavily on the characteristics of the training dataset, transferability across different SSD populations cannot be assumed.

It is also notable that the subcategory “ARMS” was not originally included, as it is not formally classified within SSDs. Terms such as “ARMS”, “CHR”, and “UHR” were absent from the search strategy, highlighting the need for future reviews to incorporate these groups to achieve a more comprehensive assessment.