AI refers to the simulation of human intelligence processes by machines, especially computer systems [9]. These processes include learning (the acquisition of information and rules for using the information), reasoning (using rules to reach approximate or definite conclusions), and self-correction.

Machine Learning (ML) is a subset of AI that focuses on the development of algorithms and statistical models that enable computers to perform tasks without explicit instructions, relying on patterns and inference instead [10]. It involves the construction of systems that can learn from and make predictions or decisions based on data. Among the various techniques under the broad umbrella of ML, ANNs are computational models inspired by the structure and function of biological neural networks in the brain [11,12]. In general, an ANN consists of interconnected nodes, called neurons, organized into layers. Each neuron receives input signals, processes them through an activation function, and passes the output to the next layer. Through a process of training, ANNs adjust the connection strengths (weights) between neurons to learn the patterns and relationships within the input data. ANNs are fundamental components of Deep Learning algorithms [12] and are used for pattern recognition, classification, regression, and representation learning. Deep models contain multiple layers that can extract increasingly abstract features from images, language, speech, and other complex inputs. For the purposes of this review, ANNs are treated as a coherent family of models rather than as inherently superior to alternative machine-learning or statistical approaches. ANNs are a subset of ML and are widely used for pattern recognition, classification, regression, and clustering, all of which are relevant to assessment and diagnosis. Their historical development reflects biologically inspired, distributed approaches to learning [13], although contemporary ANNs should not be interpreted as literal models of the human brain. Their practical value lies in learning complex and potentially nonlinear relations and generalizing them to previously unseen inputs, provided that training data and validation procedures are adequate. In the next section we report some essential information to introduce ANNs; for a more detailed exposition and discussion, it is possible to refer to Urban and Gates [14].

At a topological level, the configuration of connections defines the network architecture. Information flows through the network from the input layer, where data are received externally, through one or more hidden layers—which do not have direct contact with the external environment—to the output layer, where the final outcome is produced [15].

As introduced previously, artificial neurons process incoming information using an activation function [16]. Specifically, this activation is calculated by summing all the input values multiplied by their respective weights [17]. Various types of functions can then be applied to this sum to determine the neuron’s output. The most commonly used are the step function (where the output is 0 or 1 if the input exceeds a set excitability threshold), the continuous linear function (a linear transformation of the input), and the sigmoid or logistic function (which ranges from 0 to 1 and can be likened to probability distributions) [18].

Ultimately, what an ANN “learns” resides in this pattern of connections, in relation to its architecture and the information flow. For example, the flow can go in one direction only, from input to output, in a feed-forward NN, or there can be recursive feedback connections, as it happens in a recurrent NN [19,20].

The behavior of a neural network, including its response to a given input, is determined by its architecture and the configuration of its weights. Accordingly, a network’s learned representation is encoded in its weight configuration [20]. Finding an appropriate configuration for a task constitutes the learning process. Supervised learning uses labeled target outputs, whereas unsupervised learning seeks structure in data without labeled targets [21,22].

In supervised learning, an external teaching signal specifies the desired output for each training example [21]. Through repeated updates, the network adjusts its weights to reduce the discrepancy between predicted and target outputs. The delta (Widrow–Hoff/LMS) rule applies to single-layer adaptive networks, whereas backpropagation extends gradient-based learning to multilayer networks by propagating error derivatives through hidden layers [20,23].

In unsupervised learning, no target output is specified, and the network learns structure from the input data [22]. Kohonen’s self-organizing maps are a canonical example of unsupervised learning [24]. Hopfield networks, by contrast, are recurrent associative-memory models designed to store and retrieve patterns, including from partial inputs [25]. In addition to these two primary paradigms, semi-supervised learning represents a hybrid approach [10]. It leverages a small amount of labeled data alongside a much larger pool of unlabeled data during the training phase. This method is particularly valuable in psychological and clinical research, where acquiring expertly labeled diagnostic data is often expensive and time-consuming, while raw, unannotated data is more readily available [26].

The architecture is a key element of ANNs and determines their behaviors and what learning algorithm can be applied to them. Several neural network architectures have proven particularly relevant to psychological research and assessment [14]. Feedforward Neural Networks (FNN) are the simplest type of neural network architecture, where information flows in one direction—from input nodes through intermediate nodes (hidden layers) to output nodes—without cycles or loops. They are commonly used for tasks like pattern recognition, classification, and regression; in psychometrics, they are frequently employed for predicting clinical outcomes from questionnaire data [27].

Convolutional Neural Networks (CNN) are designed specifically for processing grid-like data, such as images or videos. They consist of convolutional layers, pooling layers, and fully connected layers, leveraging convolution operations to detect patterns and spatial hierarchies in the input data. This makes them highly effective for image classification and object detection; in clinical psychology, CNNs are increasingly applied to analyze complex spatial representations, such as brain imaging scans or digitized clinical drawing tests [28]. To handle sequential data, Recurrent Neural Networks (RNN) introduce loops within the network architecture, allowing information to persist over time. A more advanced variant, Long Short-Term Memory (LSTM) networks, addresses the vanishing gradient problem by introducing gated cells. These cells allow the network to retain long-term dependencies and remember information over extended time intervals [20]. While widely used for natural language processing and sequence generation, in psychological research they are highly effective for dealing with time-series data, such as longitudinal behavioral tracking or analyzing patient speech patterns [29].

Finally, Autoencoders are a type of neural network architecture used for unsupervised learning and dimensionality reduction. They consist of an encoder network that compresses the input data into a lower-dimensional latent space representation and a decoder network that reconstructs the original input from the encoded representation. While employed in general tasks like anomaly detection and feature learning, in psychometrics they are exceptionally useful for compressing complex behavioral data or aiding in the item-reduction process of psychological scales by extracting latent features from raw inputs [30].

These architectures are tailored to different data modalities and tasks. We selected ANNs as the organizing focus because they constitute a coherent and influential family of models across psychometric and clinical applications; this choice does not imply that they are uniformly preferable to regression, support vector machines, random forests, or boosting methods. Model selection should depend on the research question, data modality, sample size, class balance, interpretability requirements, and intended clinical use. Large Language Models were excluded because their scale, training objectives, and predominantly generative use cases require a separate review framework and would reduce comparability across the task-specific models examined here.

In the present study, we propose a scoping review of the use of ANNs in the field of psychological measures, assessment, and diagnosis. Rather than approaching ANNs only as abstract computational models, we examine their role in tasks that are directly relevant to applied mental health practice, including psychological evaluation, diagnostic classification, symptom monitoring, risk detection, and prediction of clinically meaningful outcomes. In this sense, ANNs are considered not simply as psychometric tools, but as methods that may help translate complex patient data into information that can support clinical interpretation and decision-making.

Moving beyond the theoretical level, we examine how these models may function as decision-support tools for clinicians and psychiatric nursing staff. ANNs may organize complex information, identify cases requiring closer assessment, and support continuous monitoring or early-warning workflows. They do not independently establish diagnoses or determine treatment, and translation into routine practice requires data standardization, algorithmic transparency, robust external validation, workflow integration, and accountable human oversight. We did not include studies whose primary objective was to connect neuroimaging or physiological patterns to clinical conditions without an explicit psychological outcome, in order to maintain a focused review of behavioral and psychometric assessment. The potential integration of validated biological, behavioral, and psychometric information is considered as a future research direction. Specifically, this review aims to summarize ANN applications and validation practices, distinguish diagnostic classification from prognosis, symptom monitoring, risk detection, and personalized intervention, and identify the conditions under which ANN outputs could responsibly support clinical assessment.

Studies were eligible when they: (1) were empirical articles published in English and in final publication form between 2018 and 26 January 2024; (2) were retrieved within the psychology subject area and through the psychology keyword filter applied in the databases; (3) used ANNs as a method of data analysis; and (4) investigated cognitive, emotional, or behavioral outcomes relevant to psychological measurement, psychometric development or scoring, diagnostic classification, symptom monitoring, prognosis, risk detection, or treatment-related applications. Studies were excluded if they did not use ANNs; addressed non-psychological fields, such as economics or law; analyzed physiological variables without an explicit psychological outcome; fell outside the clinical, health, or psychometric scope of the review; or were theoretical rather than empirical. The included studies were grouped according to their principal application and function in psychological assessment.

Scopus and Web of Science were searched. The final search in both databases was conducted on 26 January 2024. The search was restricted to the peer-reviewed literature indexed in these databases; grey literature and other complementary sources were not searched.

Titles, abstracts, and keywords were searched in both Scopus and Web of Science using the same Boolean search string: “artificial” PRE/0 “neural” PRE/0 “networks” OR “artificial” PRE/0 “neural” PRE/0 “network” OR “deep” PRE/0 “learning” OR “machine” PRE/0 “learning” OR “artificial” PRE/0 “intelligence”. Results were limited to English-language articles in final publication stage, published from 2018 to the date of the final search, within the psychology subject area and using the keyword “psychology.” The identification, manual duplicate-removal, screening, and eligibility process is summarized in the PRISMA 2020 flow diagram (Fig. 1).

Duplicate records were identified and removed manually before screening. Three reviewers independently screened each remaining record on the basis of its abstract. At this stage, records were excluded when the abstract did not mention a machine-learning method of data analysis or a psychological variable. Full texts of potentially eligible reports were then assessed independently against the eligibility criteria. Disagreements were resolved through discussion and consensus among the three reviewers. No automation tools were used in duplicate removal or study selection.

Three reviewers independently extracted data from each included report. A shared charting framework was used to collect the topic and area of application, authors, year, country or region, sample size and data characteristics, psychological outcome, ANN architecture and methodological approach, comparator models, training and evaluation strategy, performance metrics, and principal findings. Funding sources of the included studies were not charted.

The charting framework was refined through discussion as the reviewers became familiar with the heterogeneous evidence base, while transparent communication and consensus were used to maintain consistency in data extraction and interpretation. Performance metrics and study findings were charted as reported by the original authors. No contact with study authors was required, and no automation tools were used in data charting.

The charted data were summarized descriptively and synthesized thematically according to the principal functions served by ANN models in psychological assessment, including psychometric applications, diagnostic classification, symptom monitoring, prognosis, risk detection, and treatment-related applications. The studies were tabulated according to their general and methodological characteristics. To support thematic mapping, a co-occurrence network analysis of two-word groups extracted from the included abstracts was performed using the bibliometrix R package [32]. The resulting 25-node network represented term co-occurrences, node size reflected centrality, and communities were identified using the Walktrap algorithm. No formal critical appraisal, reporting-bias assessment, or certainty assessment was conducted because the purpose of the review was to map the range and characteristics of the evidence rather than to estimate intervention effects.

ANNs are increasingly used to enhance psychological assessment and diagnosis across two complementary levels: (a) the methodological (psychometric) level, where ANNs support the construction, optimization, and scoring of instruments; and (b) the clinical-applicative and health psychology level, where ANNs improve diagnosis, differential diagnosis, outcome prediction, and the personalization of intervention strategies. In this section, we integrate both levels: first the methodological dimension (psychometrics) and then the applicative domains of diagnosis, differential diagnosis, monitoring and intervention.

ANNs belong to a predictive modelling tradition that complements, rather than replaces, explanatory psychological research [59,60]. A model may accurately estimate an outcome without identifying its causes, and variables that optimize prediction are not automatically appropriate targets for intervention. Prediction should therefore be separated from clinical decision-making: diagnosis and treatment require integration of model output with history, interview findings, comorbidity, contextual factors, patient preferences, and professional standards.

The clinical value of these models cannot be inferred from discrimination metrics alone. Patient-centered outcomes include timeliness and appropriateness of referral, symptom improvement, prevention of adverse events, patient understanding and acceptability, clinician workload, access to care, and the distribution of benefits and harms across populations. ANNs may support these goals, but evidence of accurate prediction is not itself evidence of improved care.

The reviewed studies indicate that architecture should follow the data and task. Feed-forward networks were common for structured psychometric variables, CNNs for spatial or image-derived representations, and recurrent or LSTM models for longitudinal and sequential data. This task-specific flexibility is an advantage, but it does not make ANNs universally preferable. Simpler models may provide comparable performance, better calibration, and clearer explanations, particularly in small structured datasets.

The narrative synthesis distinguishes five functions that should not be conflated: psychometric construction and scoring; diagnostic classification and differential diagnosis; symptom monitoring and prognosis; risk detection; and treatment personalization or adherence. Each function entails different validation requirements, error costs, clinical pathways, and patient-centered outcomes.

The promising applications must be weighed against pronounced dataset heterogeneity. Samples ranged from 22 participants to tens of thousands of individuals and included psychometric responses, text, images, speech, behavioral traces, and sensor streams. Small samples combined with high-dimensional inputs create substantial risks of overfitting, data leakage, class imbalance, unstable feature importance, and optimistic performance estimates. Exceptionally high accuracy should therefore be treated as a signal for independent replication rather than evidence of clinical certainty.

Validation procedures were also inconsistent. Many studies relied on a single training/testing split, while others used different forms of cross-validation; external validation in an independent clinical population was uncommon. Internal resampling can estimate performance within the source dataset, but it cannot establish transportability across hospitals, countries, languages, devices, demographic groups, or changes in clinical practice. Multicenter external validation, prospective temporal validation, calibration assessment, and subgroup-specific error analysis are necessary before deployment.

These requirements distinguish research-oriented applications from clinically deployable systems. A promising retrospective classifier is a proof of concept; a deployable system must additionally demonstrate reproducibility, workflow compatibility, prospective benefit, safety under distribution shift, maintenance procedures, and accountable ownership throughout its lifecycle. The predominance of clinical classification studies over other psychological domains may reflect differences in data availability as well as uneven methodological maturity.

Intervention-delivery systems such as virtual therapists, chatbots, and AI-supported psychotherapy were outside the assessment-focused scope of this review [61]. Their clinical effectiveness and safety require separate evaluation rather than inference from the performance of assessment models.

The substantial variation in outcomes, data modalities, model architectures, sample sizes, and validation strategies limits direct comparison of model performance across studies.

Architectures, activation functions, optimizers, preprocessing pipelines, class-balancing procedures, and hyperparameter choices varied considerably, limiting direct comparison across studies.

Training, validation, and testing procedures were also inconsistent. Studies used different performance metrics and data partitions, and many did not report calibration, uncertainty, threshold selection, or subgroup-specific errors.

Many studies did not provide accessible code, sufficiently detailed analytic workflows, or appropriately governed data resources, which further limits replication and independent verification.

These sources of heterogeneity should be treated as central determinants of the reported results and of the comparability of model performance across studies.

Although studies based solely on neuroimaging or physiological signals were excluded, multimodal psychiatric assessment is a logical future direction. Imaging, EEG, wearable physiology, speech, behavior, ecological self-report, and psychometric instruments may provide complementary information at different temporal and functional levels [62,63,64,65,66,67,68,69]. Objective biomarkers should complement, not displace, psychometric and clinical assessment, and their incremental validity, fairness, cost, burden, missingness, and interpretability must be evaluated against simpler pathways.

The black-box challenge is particularly consequential in psychological assessment because model outputs may influence labels, treatment, education, employment, or access to services. Post-hoc explanations can indicate which features contributed to a prediction, but they do not necessarily reveal a stable causal mechanism or prove that the model relied on clinically legitimate information [70].

Transparent reporting is essential for reproducibility. Architectures, preprocessing, feature selection, class balancing, hyperparameter tuning, threshold selection, missing-data handling, and every partition used during model development should be documented. Open code and appropriately governed data can facilitate replication, while reporting guidance such as TRIPOD+AI provides a common basis for evaluating clinical prediction-model studies [71].

Algorithmic bias may arise when training data underrepresent demographic, cultural, linguistic, or socioeconomic groups; when measures function differently across populations; or when historical service inequalities are encoded in labels and outcomes [3]. Overall performance can conceal clinically important subgroup failures. Studies should report sample composition, assess measurement invariance where relevant, compare calibration and error rates across groups, and test whether deployment could widen existing disparities.

Ethical implementation also requires proportionate and transparent data use, privacy and cybersecurity safeguards, meaningful human oversight, mechanisms for contesting or correcting outputs, and clear communication about uncertainty. The WHO guidance on AI for health emphasizes that ethics and human rights should be embedded in design, deployment, and governance rather than added after technical development [6].

Regulatory and medico-legal responsibility cannot be delegated to a model. In the European Union, high-risk AI systems may be subject to requirements concerning risk management, data governance, transparency, human oversight, accuracy, cybersecurity, and post-market monitoring [7]. Services must define who verifies model suitability, who acts on an alert, how disagreement between clinician and system is documented, and how patients are protected when an error occurs.

Routine implementation requires more than installing software. Psychiatric services need secure data pipelines, interoperability with clinical records, acceptable response times, staff training, technical support, and procedures for missing or low-quality data. Decision thresholds should reflect the clinical pathway and the relative consequences of false-positive and false-negative results rather than being selected only to maximize a summary metric. ANNs should be positioned as decision-support tools that can prioritize information, prompt further assessment, or identify changes requiring review. Clinicians remain responsible for interpreting outputs in context and discussing them with patients. Prospective evaluations should examine how a system changes decisions and outcomes, including automation bias, alert fatigue, unnecessary referrals, and reduced attention to information not represented in the model. Deployment also requires continuous monitoring for calibration drift, changes in population characteristics, software updates, and emerging inequities. Early-stage evaluation frameworks such as DECIDE-AI emphasize human factors, workflow integration, and real-world safety before large-scale implementation [72]. Table 3 summarizes the principal barriers and corresponding priorities.

The historical relationship between psychology and AI remains relevant [73], but current clinical adoption must be governed by evidence rather than technological novelty. Professionals who use ANN outputs need sufficient understanding of model limitations, uncertainty, and failure modes to interpret them responsibly.

This scoping review has several limitations. The search was restricted to Scopus and Web of Science and to literature available by 26 January 2024, which may have excluded relevant grey literature and later developments. Study selection and thematic extraction involved qualitative judgment, and no formal risk-of-bias instrument was applied. Excluding Large Language Models, purely neuroimaging studies, and physiological models without explicit psychological outcomes improved conceptual coherence but limits coverage of the wider AI ecosystem. Publication bias may also favor studies reporting strong performance, while heterogeneity prevented quantitative synthesis.

Future research should prioritize multicenter and cross-cultural external validation; prospective studies of clinical impact; calibration and error analysis across demographic and socioeconomic subgroups; standardized reporting of data, training, and validation procedures; and open, reproducible workflows where privacy permits. Studies should compare ANNs with simpler baselines, evaluate multimodal models only when each modality has a justified contribution, incorporate patient and clinician perspectives, and define patient-centered endpoints before model development.