This systematic review was conducted in accordance with the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses; Checklist 1) 2020 guidelines.

Studies were eligible if they: (1) reported original research applying classical AI, ML, DL, or hybrid models; (2) focused on treatment of individuals with BD; (3) focused on adult populations (aged ≥18 years) or included mixed samples with both adults and adolescents; (4) were written in English; and (5) were published after 2015.

Studies were excluded if they met at least one of the following criteria: (1) were reviews, meta-analyses, or commentaries; (2) were clinical trial protocols; (3) were conference abstracts, editorials, letters, lectures, or book chapters; (4) did not include original data; and (5) exclusively enrolled populations aged <18 years.

A systematic literature search was conducted across 4 electronic databases, including PubMed, Web of Science, Scopus, and Embase. The searches were performed on November 17, 2025. Queries are detailed in Multimedia Appendix 1. All retrieved records were imported into the “Rayyan” (Rayyan Systems, Inc) application [28] and duplicates were removed.

Following deduplication, records were screened in 2 sequential stages. In the first stage, titles and abstracts were independently assessed by 3 reviewers (SDF, DA, and AR) to identify studies potentially eligible based on the predefined inclusion and exclusion criteria. In the second stage, full texts of the selected articles were retrieved and independently evaluated for eligibility by the same reviewers. Any uncertainties arising during either stage were resolved through discussion, and when necessary, a consensus meeting was convened to reach a final decision regarding study inclusion or exclusion.

In addition, the annual trend of publications was examined to assess changes in research activity over time.

From each study, the following information was extracted:

Studies were grouped into five categories based on their primary clinical aim: (1) acute symptomatic response, referring to the prediction of short-term (4/12 weeks) treatment success based on symptom severity reduction; (2) long-term (more than 6 months) maintenance response, referring to the ability to remain episode-free over time; (3) relapse and readmission risk, including the prediction of manic and depressive relapse as well as hospital readmission; (4) safety and dose optimization, focused on treatment personalization through dose adjustment and adverse-event prevention; and (5) brain aging and phenotyping aimed at identifying biologically or clinically meaningful patient subgroups.

This framework was designed to capture the main clinical areas in BD in which AI provides practical benefit.

Studies were further categorized according to the type of biomarkers included: only clinical data (clinical); genomic and clinical data (genomic); imaging and clinical data (imaging); wearable devices (including smartphone sensors); and clinical data (wearable). They were also sorted according to the modeling approach used (ie, logistic regression, Naïve Bayes, decision trees, random forest, support vector machines (SVM), gradient boosting, and neural networks).

All studies were classified into 3 tiers of validation strategy. Tier 1 included studies where validation on an independent data cohort was performed, Tier 2 were those studies where only internal validation strategies (eg, cross-validation or leave-one-out approaches) were performed, and Tier 3 were exploratory studies.

The performance metrics considered included accuracy, area under the curve (AUC), sensitivity, specificity, and their CIs computed on the validation or test sets (when available or derivable).

To explore potential publication bias, we constructed a funnel plot. The plot displays the best AUC for each study on the X-axis and the inverse square root of the sample size on the Y-axis. This approach allowed visualization of the relationship between model performance and study size to identify potentially biased studies.

The assessment of methodological quality, risk of bias, and applicability of the included predictive models was performed using the PROBAST+AI method [27]. The evaluation process was structured into 2 main components, model development, which analyzed the quality and generalizability of the algorithm’s construction, and model evaluation, which assessed the credibility and clinical applicability of performance estimates that may be affected by systematic bias. The analysis was structured into 4 fundamental domains, including participants and data sources, predictors, outcomes, and analysis. An additional overall judgment was provided following the PROBAST+AI guidelines.

To assess the average performance of the models stratified by the defined categories (ie, acute symptomatic response, long-term maintenance response, relapse and readmission risk, safety and dose optimization, and brain aging and phenotyping), a pooled AUC was calculated from study-specific AUCs and their 95% CIs (when available) using a random-effects meta-analysis based on the DerSimonian-Laird method [29]. The weights were calculated based on the 95% CIs for each study’s AUC. Heterogeneity was summarized using the inconsistency index (I²) and the between-study variance (τ²) measure.

To further evaluate potential publication bias, the Egger test was performed on logit-transformed AUC values. This statistical test assesses asymmetry in the funnel plot, indicating whether smaller studies tend to report systematically higher or lower performance estimates compared to larger studies. The logit transformation was applied to stabilize variance across studies.

All analyses were performed using Python (version 3.8.8; Python Software Foundation).

The PubMed query returned 127 articles; the SCOPUS query returned 297 articles; the Web of Science query returned 172 articles; the Embase query returned 119 articles. After removing duplicates, a sample of 327 studies was screened; 35 studies met all the eligibility criteria and were included in the final synthesis (Figure 1).

Over the past decade, scientific interest in applying AI to the treatment and long-term management of BD has grown substantially. The publication trend of the 327 records screened in this review, normalized on the total number of publications per year in PubMed, follows an exponential trajectory (R²=0.979); while contributions in the early 2010s were relatively limited, the volume of research has steadily increased, with particularly pronounced growth in the past 6 years (Figure 2).

The final set comprised 35 [22,30-63] studies published between 2015 and 2025. Most studies were conducted in the United States, China, the European Union (EU), Canada, and the United Kingdom, with sample sizes ranging from about 10 to approximately 530 thousand participants. A detailed summary of study characteristics is provided in Table 1. The aims of the studies are reported in Table S1 in the Multimedia Appendix 1. Data sources varied widely and included different clinical assessment tools and rating scales (eg, Young Mania Rating Scale [YMRS] and Hamilton Depression Scale [HAMD]), electronic health records (EHR), neuroimaging data, digital phenotyping (smartphone sensors and wearables), and multimodal datasets. AI methodologies included classical ML models (eg, random forests [RF], SVMs, DL architectures (convolutional neural networks [CNNs] and recurrent neural networks [RNNs]), NLP approaches, and hybrid computational frameworks. Based on the validation strategy, we classified most studies [32-34,36,37,46-49,51-54,56,58-63] (n=20, 57%) as Tier 2, a significant proportion [22,30,31,35,38,41-45,50,55,57] (n=13, 37%) as Tier 3, while only 2 studies [39,40] were categorized as Tier 1.

The target outcomes for AI models are reported below for the 5 main categories.

Figure 6 provides a comparative overview of AI model performance across different marker domains (ie, clinical, imaging, genomics, and digital-wearable). For studies assessing multiple models, only the main models are shown, either the best-performing models or the ones developed specifically for the study. Clinical-based models were represented across the full range of models, achieving better accuracies, although the highest average accuracies were observed in imaging- and genomics-based models, which were often derived from comparatively smaller BD sample sizes (smaller bubbles).

The potential publication bias showed that studies with small sample sizes tended to report exceptionally high AUC values, often close to perfect, while as the sample size increased, the AUC values converged toward a more moderate and realistic range (Figure 7). The Egger test did not reveal evidence of funnel plot asymmetry (intercept=−4.63; P=.56).

The methodological evaluation using PROBAST+AI found that the great majority of the included studies [22,30-63] were at high risk of bias (Figure 8). In the model development phase, most studies were judged at high risk, primarily due to limitations in the analysis domain (Figure S1 in Multimedia Appendix 1), including inadequate handling of overfitting (ie, a model that performs better on training data than on independent data), insufficient reporting of model performance measures, and lack of external validation. While the participants’ and predictors’ domains were frequently rated as low risk, concerns regarding predictors’ applicability and overall methodological quality contributed to unfavorable overall judgments.

Similarly, in the model evaluation phase, the overall risk of bias was predominantly rated as high. This was mainly due to shortcomings in the analysis domain, such as limited external validation, inappropriate performance assessment, or insufficient sample sizes. Although predictors and outcomes were often judged as low or unclear risk, these did not offset the high risk associated with analytical issues. Overall applicability in the evaluation phase was judged as low in a small subset of studies but remained unclear or high in most cases, reflecting heterogeneity in validation settings with limited generalizability and fairness.

The present systematic review synthesized evidence from 35 studies investigating the application of AI methodologies to treatment optimization in BD, restricting inclusion to studies published from 2015 onward, in line with the marked growth of research on classical AI in psychiatry during this period [71]. Across the literature, AI approaches were primarily used for five clinical purposes: (1) acute symptomatic response, (2) long-term maintenance response, (3) relapse and readmission risk, (4) safety and dose optimization, and (5) brain aging and phenotyping. While the rapid growth of studies reflects increasing interest in precision psychiatry from 2020 onward, our critical analysis revealed substantial methodological weaknesses that limit the clinical utility of most models. Although the pooled performance metrics across the 5 categories appear encouraging at first glance, the high heterogeneity across studies and the predominance of high risk of bias identified by the PROBAST+AI assessment suggest that these results must be handled with caution. In most cases, reported accuracies represent optimistic estimates derived from small or internally validated samples rather than robust, generalizable clinical tools.

The forest plot for this group showed a heterogeneous pattern of results, with a modest pooled AUC indicating only moderate discriminative performance. This suggests that acute symptomatic response remains difficult to predict reliably, likely because these are short-term treatment outcomes influenced by multiple interacting clinical and biological factors. Due to the dynamic and multifactorial nature of acute treatment response, the high heterogeneity (I²=93.6%) reflects substantial differences that are unlikely to be explained by sampling error alone. Therefore, the pooled AUC estimate should be interpreted prudently, as it summarizes studies that differ markedly in treatment modality (eg, first or second generation of antidepressants), predictor type (clinical vs neuroimaging, with this one representing the most informative one), and methodological approaches.

Despite this, the estimated between-study variance was low in absolute terms (τ²=0.0074). However, this does not eliminate concerns about heterogeneity. Rather, it suggests that the dispersion of effects on the AUC was limited in absolute magnitude, but variability within studies is large. This combination is consistent with a modest AUC that was not consistently reproduced across the studies, particularly in neuroimaging-based models.

Differences in validation strategies, with more rigorous cross-site validation yielding lower but more realistic performance, further contribute to the observed variability. Overall, these findings suggest that AI models may detect a biological signal associated with acute response, but their current predictive utility is limited and generalizability remains uncertain.

This group showed clinically meaningful predictive performance, indicating fair discriminative ability of AI models in identifying pharmacological (lithium) responders. Between-study heterogeneity was moderate (I²=56.3%), suggesting that some of the variability across study results reflected genuine differences. Moreover, the estimated between-study variance was low in absolute terms (τ²=0.0022), suggesting limited dispersion of the underlying effects on the AUC. These findings highlight that, although methodological and clinical differences across studies represent a constraint, they do not preclude a consistent overall predictive performance comparison and reinforce the idea that drug-treatment response (lithium) is a predictable target. However, the observed heterogeneity suggests that improvements in predictive accuracy may depend on the multimodal integration of sociodemographic, clinical, and biological data, in line with the current trend in precision psychiatry, with genetic predictors appearing the most informative. Additionally, assessment using PROBAST+AI identified a high risk of bias in each domain across almost all studies in this group; therefore, the findings must be interpreted cautiously. This category included 4 studies characterized by significant publication bias due to small sample size associated with very high AUC [43,53,57] or performance outside the confidence interval [56], which must be minimized in future AI development efforts.

In this group, AI models achieved moderate predictive performance for clinically critical outcomes, indicating fair-to-good discriminative ability for identifying patients at risk of relapse, hospitalization, or mortality. This performance is supported by both high-accuracy models in specific contexts and the stability of large registry-based studies. The most informative predictors were wearable and clinical ones. Interpretation of this pooled estimate requires caution because between-study heterogeneity was high (I²=93.0%), reflecting differences in outcome definitions (eg, manic relapse, suicidal behavior, and all-cause mortality), data modalities, and sample size, which ranged from large-scale EHR-based studies to smaller multimodal approaches integrating neuroimaging. Despite this variability, the low between-study variance (τ²=0.0039) suggests that the pooled AUC may still capture a broad average predictive signal consistent across studies, indicating a shared set of clinically relevant risk factors. Indeed, converging evidence highlights the importance of variables such as prior self-harm, hospitalization history, and medical comorbidities. Moreover, despite the differences in sample size, the weights assigned to each model in the forest plot are similar, indicating that no single study disproportionately influences the pooled estimate of predictive performance. From a clinical perspective, these models showed potential to support proactive risk stratification and personalized monitoring strategies, with external validations across different health care systems further supporting their generalizability. Overall, although methodological fragmentation persists, the findings indicate that AI can reliably capture the trajectory of illness severity in BD. Stronger conclusions about generalizability will require more homogeneous outcome definitions and more consistent external validation.

The group of studies focusing on safety and dose optimization highlights the translational potential of AI in moving beyond generalized clinical guidelines toward individualized treatment strategies based on patient-specific risk and pharmacokinetic profiles. AI models showed strong performance in predicting blood concentrations of key mood stabilizers, with approaches such as artificial neural networks and gradient boosting algorithms identifying clinically relevant predictors (ie, age, renal function, co-medications, and bilirubin levels) to support personalized dosing. In parallel, advanced causal inference methods have been applied to better quantify the cardiometabolic risks associated with antipsychotic treatments, enabling more precise risk stratification and informing safer treatment switches using informative predictors, such as age, sex, baseline health status, psychiatric comorbidities, previous exposure to medications with cardiometabolic effects, and other clinical data. AI-driven pharmacovigilance has also shown promise in identifying predictors of adverse effects, including endocrine complications such as antipsychotic-induced hyperprolactinemia. Furthermore, the integration of NLP applied to EHR has revealed important discrepancies between real-world prescribing patterns and controlled trial evidence, underscoring the need for AI-based clinical decision support systems. These findings suggest that AI can play a key role in optimizing both the efficacy and safety of pharmacological treatments in BD, facilitating a more precise and data-driven approach to clinical management.

Studies focusing on phenotyping and brain aging highlight the potential of AI to move beyond traditional diagnostic categories by identifying biologically and clinically meaningful dimensions of disease progression. Unsupervised ML approaches indicate that clusters based on real-world treatment patterns explain longitudinal functional outcomes better than the conventional BD-I/BD-II distinction, advising that illness trajectories are closely linked to specific pharmacological response phenotypes. In parallel, AI-derived metrics such as BrainAGE and the QRI have revealed, through informative imaging predictors, an accelerated brain aging process in BD, particularly affecting white matter and subcortical structures, driven independently by illness severity and cardiometabolic comorbidities, with hypertension emerging as a key risk factor. Notably, converging evidence supports a neuroprotective role for lithium, which is associated with a younger brain age comparable to that of healthy controls. Overall, these findings suggest that AI can provide objective biomarkers of “brain frailty,” paving the way for precision psychiatry approaches aimed at monitoring disease progression and personalizing interventions to mitigate long-term cognitive decline. However, the limited number of studies in this group precludes definitive conclusions.

The relationship between AI models and classification accuracy across different application domains in BD research highlights substantial heterogeneity in both performance and data typology. Gradient boosting models were most frequently adopted, particularly in the clinical domain, where they achieve a wide range of accuracies (from random to good) and were often trained on larger sample sizes due to the relative ease of collecting such data. DL models tended to achieve higher accuracies (up to approximately 0.95), although typically with more restricted sample sizes, raising potential concerns regarding their downstream generalizability. Genomics-based studies, despite smaller cohorts, frequently reported high accuracies even with different models, suggesting strong informative signal-to-noise ratios but with the highest risks of overfitting. Wearable-based applications remain limited in number, although they demonstrate promise. This emphasizes the need for balanced methodological choices during model development to ensure robust, unbiased, and clinically applicable AI systems for patients with BD.

Although the Egger test did not reveal evidence of funnel plot asymmetry, its interpretation is limited by the small number of studies reporting AUC. Visual inspection of the funnel plot suggests that most small-sample studies are distributed on the right side of the central dashed line (AUC≈0.8), indicating estimates of high performance. Conversely, there is a clear gap in the lower left portion, where small studies with poor performance should be located. This distribution suggests that scientific journals tend to publish small studies only when the results are very good, leading to a systematic overestimation of the true effectiveness of AI. Studies with modest performance but small samples probably remain unpublished or confidential. As the sample size increases, AUC values tend to converge toward a more realistic range. Attention should be moved from “near-perfect” AUC in small samples to more robust, externally validated models that demonstrate stable and reliable performance in heterogeneous populations.

Across the 5 categories, PROBAST+AI analysis revealed that most studies were at high risk of bias, particularly in the analysis domain. Indeed, only 2 [37,40] studies received a low-risk judgment across all domains of the PROBAST+AI assessment, due to rigorous data handling, robust validation (including external validation), adherence to standards, and comprehensive evaluation of model performance, including calibration and fairness. In the other studies, applicability concerns were common, reflecting nontransparent AI system modeling or limited generalizability to real-world clinical settings. Only a minority of studies demonstrated consistently low risk across domains, while several judgments remain unclear due to insufficient methodological reporting. Furthermore, only 7 [36,39,40,43,45,48,50] out of the 35 included studies provided access to their models, underscoring a lack of transparency and availability of predictive tools, which limits reproducibility and the possibility of external validation. In light of the limitations identified through the PROBAST+AI assessment, future model development in the context of BD spectrum treatment should focus on enhancing dataset diversity to reduce bias, improving external validation across different clinical settings, and integrating clinically interpretable features to support decision-making. Additionally, incorporating longitudinal patient data and multimodal inputs may strengthen predictive performance and generalizability, ensuring that AI-driven recommendations are both robust and clinically applicable.

Data-related challenges represented a major constraint. Multicenter datasets were often affected by differences in clinical protocols and data acquisition procedures, whereas high-dimensional modalities such as omics introduced technical artifacts, including batch effects, that can distort predictive signals. This variability undermines AI models’ stability and reproducibility across independent cohorts. A further concern was the imbalance between the number of predictors and the available sample sizes. Many studies [43,57] relied on complex omics or neuroimaging features derived from very small cohorts, creating conditions prone to overfitting. Although some studies [37,41,43] used regularization or feature selection, these strategies were not externally validated. Calibration and real-world integration were rarely assessed. Most models were evaluated primarily using discrimination metrics, leaving their practical impact in clinical scenarios uncertain.

From a regulatory perspective, all psychiatric AI systems are likely to be classified as high-risk under the EU AI-Act [72], requiring even stronger evidence of robustness, technical documentation, transparency, fairness, and ethically grounded development.

Several regulatory and methodological frameworks have been proposed to guide the responsible development and application of both classical and generative AI in the life sciences and health care sectors. The European Medicines Agency (EMA) has issued a discussion document on the use of AI throughout the medicinal product lifecycle, outlining considerations for the safe and effective integration of AI/ML at all stages from drug development to postauthorization, while emphasizing the need to align its use with EU regulatory requirements and data governance principles [73]. In addition, EMA guidance on AI highlights the importance of ensuring transparency, robustness, and appropriate validation of AI-based tools to support regulatory decision-making. Internationally, initiatives such as the Food and Drug Administration (FDA) draft guidance on AI-enabled medical products [74] and the joint Good Machine Learning Practice (GMLP) guiding principles developed by the FDA, Health Canada, and the Medicines and Healthcare Products Regulatory Agency (MHRA) [75] further emphasize life cycle–wide considerations, including data quality, model performance monitoring, and documentation to ensure safety and effectiveness in real-world applications. Furthermore, consensus frameworks such as Fair Universal Traceable Usable Robust Explainable–AI (FUTURE-AI) [76] propose structured best practices covering the entire AI lifecycle in health care, from design and development to deployment and monitoring, with an emphasis on trustworthiness, explainability, and clinical validity. Collectively, these methodological recommendations emphasize the importance of a risk-based, transparent, and continuously evaluated approach to AI implementation in the life sciences.

The medico-legal status of AI in the life sciences remains complex and continues to evolve. Although AI-based tools are increasingly used to support research, drug development, and clinical decision-making, their legal framing depends on their intended use and degree of autonomy. In the EU, AI systems fall under the regulatory framework for medical devices when used for diagnostic or therapeutic purposes, requiring compliance with the Medical Device Regulation (EU 2017/745 MDR) [77] and, where applicable, the In Vitro Diagnostic Regulation (EU 2017/746 IVDR) [78]. As such, the safe integration of AI into life sciences requires not only technical validation but also regulatory compliance and robust governance frameworks.

Several systematic reviews have highlighted the increased use of AI models in psychiatric disorders beyond BD, particularly in major depressive disorder, schizophrenia, anxiety disorders, and suicide risk prediction. In major depressive disorder, ML approaches applied to clinical, neuroimaging, and digital phenotyping data have shown moderate to high performance in identifying diagnoses and predicting treatment response, although generalizability remains limited by small sample sizes and heterogeneous study designs [79,80]. In schizophrenia, AI models have been used to support diagnosis and symptom classification using neuroimaging and speech-based features, demonstrating promising accuracy but requiring further external validation [14]. Similarly, in anxiety disorders, wearable and smartphone-derived data have been explored to detect symptom fluctuations and stress-related patterns, although methodological variability and lack of standardized endpoints limit clinical translation [81]. A substantial body of literature has also focused on suicide risk prediction, where NLP and multimodal ML models have achieved encouraging predictive performance using EHR and social media data, but raise important concerns regarding interpretability, ethical use, and false-positive rates in clinical practice [82,83].

The current evidence suggests that, consistently with other psychiatric condition models, AI in BD treatment is still largely in its exploratory or proof-of-concept phase. While the conceptual shift toward precision psychiatry is evident, most existing models are not yet suitable for routine clinical decision-making.

The translational gap does not lie in algorithmic sophistication but rather in: (1) data quality and representativeness; (2) validation strategies; (3) transparent and interpretable modeling from human experts (ie, human in the loop); and (4) prospective clinical evaluation. Until these issues are addressed, AI tools remain prone to the risk of producing optimistic results at the benchside that are not at the bedside and in real-world settings.

Overall, although AI applications in psychiatry show significant potential across multiple conditions, the current evidence remains largely exploratory, and robust prospective validation studies are needed before real-world clinical implementation.

Several limitations should be considered when interpreting the findings of this review. The included studies showed heterogeneity in sample characteristics, outcome definitions, follow-up duration, relapse types, and treatment strategies, which was only partially mitigated using broad search strategies. This variability reduced comparability and limited the interpretability of pooled performance estimates. Furthermore, many investigations were based on small cohorts or exploratory designs, restricting the generalizability of their findings.

The conclusions of this review should be interpreted in light of several methodological limitations inherent in this literature. First, there is a lack of standardized diagnostic criteria for BD that use heterogeneous definitions across ICD-9/10 (International Classification of Diseases, 9/10th Revision), DSM-IV (Diagnostic and Statistical Manual of Mental Disorders [Fourth Edition]), and DSM-5 (Diagnostic and Statistical Manual of Mental Disorders [Fifth Edition]), complicating direct comparisons and limiting generalizability. A further limitation is that we did not account for time since diagnosis or the presence of complex medical and psychiatric comorbidities, often relying instead on static factors that fail to capture the longitudinal progression of the disease. Moreover, our analysis was restricted to peer-reviewed publications, excluding proprietary or industrial models protected by patents or trade secrecy, which may perform differently but are not accessible for independent scientific validation. Another key limitation of this review is that none of the included studies addressed nonpharmacological interventions. This restricts the generalizability of our findings and prevents a comprehensive assessment of AI tools across the full spectrum of treatment strategies. Although this review focuses on AI applications, most models rely on ML techniques, with other classical AI approaches being underrepresented. This limited diversity of methods should be considered a potential limitation of the current evidence. Finally, this systematic review was not prospectively registered. While preregistration is considered best practice to enhance transparency and minimize the risk of selective outcome reporting, the review was conducted in accordance with PRISMA guidelines, and all methodological steps, including eligibility criteria and outcomes, were defined a priori and consistently applied. Nonetheless, the absence of registration remains a limitation.

Future research should prioritize methodological maturity and translationality over incremental performance gains. This will require larger, harmonized multicenter benchmarking datasets on which to perform systematic external validation and to favor AI-model calibration, real decision impact, and workflow integration. The development of open, shared repositories (eg, ENIGMA [Enhancing NeuroImaging Genetics through Meta-Analysis; ENIGMA Consortium] and PsychENCODE [Psychiatric Encyclopedia of DNA Elements; PsychENCODE Consortium]) and public computational e-infrastructures (eg, OpenNeuro [Stanford Center for Reproducible Neuroscience], Zenodo [CERN], NewPsy4U [Laboratory of Neuroinformatics, IRCCS Istituto Centro San Giovanni di Dio Fatebenefratelli]) will be propaedeutic to improve reproducibility and support the transition from experimental models to clinically reliable AI systems in BD.

Classical AI approaches show potential for improving treatment response prediction, relapse risk estimation, and patient stratification in BD, supporting the transition toward precision psychiatry. However, the current evidence base remains methodologically fragile. Most studies showed a high risk of bias, limited external validation, small and heterogeneous samples, and inconsistent outcome definitions when assessed with the PROBAST+AI method. Consequently, reported performance metrics are likely to be overly optimistic and, at present, not generalizable to routine clinical practice.

Progress will depend on rigorous study design, standardized methods, including benchmarking datasets, transparent reporting, and prospective validation to bridge the gap between proof-of-concept studies and deployable clinical decision support systems.