Bipolar disorder (BD) is a severe, chronic, and recurrent psychiatric disorder, affecting 1-3% of the global population. It is associated with relevant morbidity and economic burden (1): indeed, WHO’s World Mental Health Surveys ranked BD as the second most disabling psychiatric illness in terms of lost productivity (2). Nonetheless, BD diagnosis and management remain challenging, primarily due to its heterogeneous presentation, the high rate of frequent comorbidities, and the absence of reliable biomarkers of disease progression.

According to the (51), BD prevalence has been stable worldwide and is comparable to that of schizophrenia, underlining its high heritability, with genetic heritability estimated around 70–90% (4). Although genome-wide association studies (GWAS) have identified multiple risk loci, each of them exerts only minor effects, supporting the notion of a multifactorial model in which gene-environment interactions contribute to disease vulnerability (5). In fact, the pathophysiology of BD involves a complex interplay of genetic, neurobiological, and environmental factors that lead to epigenetic, endocrine, and inflammatory dysregulation. These mechanisms drive neuronal alterations that contribute to illness onset and progression (5–7). Within this framework, the concept of neuroprogression has been proposed, positing that BD may follow a progressive trajectory marked by cumulative molecular, neuroanatomical, and functional changes that contribute to clinical worsening (8). Evidence supporting this hypothesis comes from neuroimging studies reporting progressive cortical thinning, hippocampal atrophy, and disruptions in white matter integrity, particularly in patients with multiple episodes and prolonged illness duration (9). On a molecular level, BD progression has been associated with altered levels of Brain-Derived Neurotrophic Factor (BDNF), inflammatory cytokines, and oxidative stress markers, which fluctuate depending on illness stage and symptomatic state (3, 10, 11, 13). From a clinical perspective, this process manifests as increased susceptibility to recurrent affective episodes, reduced likelihood of sustained remission, as well as progressive cognitive and functional impairment (5, 14, 15). However, the notion of neuroprogression in BD remains debated (16, 17). Recent prospective longitudinal studies have questioned the consistency of progressive decline. For example, first-treatment BD cohorts followed for ~10 years show long-term cognitive stability or improvement with trajectories paralleling healthy controls, and no evidence of kindling or cumulative deterioration (18, 19). Taken together, these findings underscore the heterogeneity of evidence and highlight the need for further multimodal longitudinal studies capable to clarify illness-related progression.

In light of these considerations, clinical staging models have been developed to better characterize BD progression and guide stage-specific interventions. The natural course of illness typically includes an asymptomatic at-risk phase, followed by a prodromal stage with non-specific symptoms, a first syndromic mood episode, and subsequent recurrences without full inter-episode recovery (20, 21). Despite theoretical advancements, the clinical applicability of BD staging models remains limited, with only a few studies assessing their real-world feasibility. In 2019, Van der Markt and colleagues applied the Kupka & Hillegers’ model to a sample of BD outpatients, confirming progressive illness trajectories and identifying biphasic mood episodes at onset and male gender as risk factors for faster progression (22). A follow-up study by Van der Markt and colleagues in 1396 BD I patients strengthened the construct validity of staging frameworks but highlighted inconsistencies between classification systems, suggesting the need for refined stratification approaches (23). In the same year, a Spanish group applied a k-means clustering approach integrating clinical, cognitive, and functional measures in a sample of 224 BD patients, 129 of whom followed for three years, showing functional deterioration and increasing treatment complexity with stage progression (24). More recently, Macellaro and colleagues from our research group conducted a ten-year retrospective study of multiple BD staging models. The findings further supported the hypothesis of significant stage progression and highlighted the influence of age at first elevated episode and duration of untreated illness on stage advancement (25). Expanding upon this, Cremaschi and colleagues focused on the Kupka & Hillegers staging model and applied a multistate Markov model to estimate transition probabilities across different BD stages. Notably, a high hazard of transition from stage 2 to 3 was observed, with the probability of remaining in stage 2 decreasing to 14% after three years (26).

Beyond clinical symptomatology, pharmacological treatments has emerged as a critical modulator of disease progression in BD. Current international guidelines, including those from the Canadian Network for Mood and Anxiety Treatments (CANMAT) and the International Society for Bipolar Disorders (ISBD), recommend first-line pharmacological strategies for both acute phases (mania and bipolar depression) and long-term maintenance, typically starting with lithium or valproate, either as monotherapy or in combination with atypical antipsychotics (27, 28). As extensively reviewed by Berk et al. (29), lithium has been shown to reduce inflammation, counteract oxidative stress, and promote neurotrophic activity, especially through effects on BDNF expression and methylation (29, 30). Early intervention through timely diagnosis, comprehensive biopsychosocial treatment, recognition of neuropsychiatric and physical comorbidities, and personalized management is likely to be more effective and to improve long-term outcomes (28, 31). In this regard, there is also consistent evidence that early pharmacological intervention may attenuate illness progression by reducing relapse rates and cognitive decline (32–34). Furthermore, treatment adherence has consistently been identified as a major determinant of illness trajectory, with poor adherence linked to higher relapse rates, increased suicidality, and poorer functional outcomes, particularly in the early stages of BD (35). These findings support the importance of accounting for pharmacological variables and adherence to treatment when interpreting biomarker-based staging outcomes.

Despite ongoing efforts, BD staging may be daunting, primarily due to the lack of reliable biomarkers for illness vulnerability and progression (29, 36, 37). In this perspective, recent advances in biological markers, neuroimaging, and computational models now offer an opportunity to enhance the precision and applicability of clinical staging models, enabling more personalized patient care and tailored treatment.

In this context, the BOARDING PASS study aims to refine clinical staging in BD by integrating traditional frameworks with advanced biological and neuroimaging data. Specifically, this multi-site, longitudinal, observational study adopts Kupka & Hillegers’ staging model (38), enriched with biomarkers such as BDNF, inflammatory, and epigenetic markers, as well as structural and functional neuroimaging data. Moreover, supervised machine learning (ML) algorithms will be trained on the resulting multimodal dataset to support biologically-informed, personalized risk stratification, early detection of the disorder, and the development of precision treatment strategies.

BOARDING PASS study has the following objectives:

A total of 126 subjects have already been enrolled across three Italian recruiting sites: Unit 1 (ASST Fatebenefratelli-Sacco, Milan; n=44), Unit 2 (ASST Papa Giovanni XXIII, Bergamo; n=41), and Unit 3 (ASL 2 ABRUZZO Lanciano-Vasto-Chieti; n=41). Participants are currently undergoing an 18-month follow-up, with 3 time points: T1 (6 months after baseline), T2 (12 months after baseline), T3 (18 months after baseline). At each time point, the clinical stage will be assigned according to the Kupka & Hillegers’ staging model, and changes in a subset of baseline clinical variables will be assessed to register potential progression in clinical staging.

Additionally, a set of psychometric measures will be administered. To investigate gene transcription regulation, pro-inflammatory cytokines (e.g., TNF-α, IL-6, IL-19), and BDNF profiles, biological samples will be collected at T0 and subsequently at T2 and T3. Biological samples from each recruitment site will be centralized at Unit 4 (University of Teramo) for standardized processing and analysis.

Neuroimaging assessments, including structural (sMRI) and resting-state functional MRI (rs-fMRI), will be conducted at each site at T0, T2, and T3, to evaluate structural connectomes based on the gyrification indices and resting-state functional connectivity. MRI datasets from all sites will be centralized at UO3 for standardized pre-processing and analysis, conforming to ENIGMA Network standards (39).

Finally, ML algorithms will be developed using MATLAB’s Statistics and Machine Learning Toolbox. A range of supervised ML models will be explored to identify the most accurate configuration. Support Vector Machines and Bayesian approaches will be tested to train the models on multimodal datasets comprising clinical, biological and neuroimaging markers.

The primary outcome of the BOARDING PASS study is the longitudinal assessment of clinical stage progression in BD at 6, 12, and 18 months, according to the Kupka & Hillegers’ staging model. Recent findings from our research group (25, 26) have supported the feasibility and sensitivity of the Kupka & Hillegers’ model in reporting significant stage transitions over time. Furthermore, the potential association between stage progression and socio-demographic and/or clinical characteristics will be investigated.

Secondary endpoints include the evaluation of biological and neuroimaging characteristics throughout the course of illness, along with their predictive power on illness progression. Specifically:

In relation to the ML application, the endpoints include: the development and optimization of a supervised ML algorithm to predict clinical stage transitions based on clinical, biological, and neuroimaging data; the evaluation of ML predictive performance through cross-validation techniques (e.g., leave-one-out validation); and the external validation of ML-based predictions against independent clinical stage assignments made by expert psychiatrists.

For stage progression analysis, data from repeated assessments of clinical stages across follow-up visits will be analyzed using generalized estimating equations implemented via the PROC GENMOD procedure in SAS^®, applying a multinomial distribution with a logit link function. This approach allows evaluating associations and interactions between stage transitions and relevant sociodemographic and clinical variables over time. Additionally, stage transition probabilities will be modeled using a multi-state Markov approach provided by the ‘mstate’ package in R. Statistical significance was set at p≤.05. Biological and neuroimaging data analyses will be performed by linear mixed-effects models to evaluate the associations between biomarkers and neuroimaging outcomes across BD stages at different time points. Omnibus one-way ANOVA will be used for molecular measures, complemented by post-hoc analyses for multiple comparisons when significant effects are identified. ML analyses will be conducted using MATLAB’s Statistics and Machine Learning Toolbox, supported by NeuroMiner software (74).

Sample size estimation was guided by multiple methodological considerations. Given the multidimensional nature of the study, several independent measures - clinical, biological, and neuroimaging - will be collected longitudinally and subsequently integrated into a unified ML predictive model. Therefore, the sample size was determined to ensure sufficient statistical power for traditional analyses, while also supporting the training and validation of robust ML algorithms. Specifically, the required sample size has been determined based on two complementary criteria. First, among the multiple domains investigated, a power analysis conducted using G*Power 3.1.9.7 (75) indicated that a total of N = 126 participants was sufficient to detect significant differences across five BD stages (0, 1, 2, 3, 4), assuming a statistical power of 0.80, an alpha=0.05, and a medium effect size (f=0.32). Second, from a computational perspective, this sample size was considered adequate for training supervised ML models with acceptable performance. Previous ML-based studies in BD have shown that models trained on ~50 subjects reached an accuracy of ~64%, while those trained on samples approaching 90 participants achieved accuracy up to 99% (76). Based on these observations and accounting for an estimated dropout rate of 15%, a total sample size of 126 participants was deemed both feasible and methodologically robust, ensuring sufficient accuracy of the ML models.

As of May 2025, a total of 126 subjects had been enrolled in the study. Among them, 54 completed T1 and 25 completed T2. At present, T1 and T2 assessments are ongoing and T3 is scheduled to begin next July. The current dropout rate is below 10%, consistent with our initial expectations. Preliminary internal tests on partial datasets suggest promising accuracy levels (>70%) in stage prediction using supervised ML models. These results, although not formally included in the current manuscript, support the feasibility of our analytical pipeline and will be detailed in future dedicated publications.

Although several staging models for BD have been conceptualized to date, current classifications still rely predominantly on clinical symptomatology, with limited integration of biological and neuroimaging correlates of illness progression. Incorporating these dimensions may significantly enhance the patient characterization by identifying specific dysfunctional trajectories and overcoming the traditional, diagnosis-centered approaches, ultimately leading to more targeted and effective treatment strategies. Moreover, the application of ML techniques may implement the accuracy in stage classification, enabling the development of data-driven and personalized staging systems. BOARDING PASS study aligns with the growing need for a standardized and transdiagnostic staging framework that integrates immuno-inflammatory, epigenetic, and neuroimaging biotypes. Such an approach is crucial to enhance the validity of staging in BD and to optimize its long-term clinical management. Preliminary results, after having successfully completed the initial planned sample recruitment, are expected for late 2026. Future directions may consider longer time of assessment and the inclusion of a matched control group in order to strengthen the study design and scientific validity of the findings.