Bipolar Disorder Integrative Staging: Incorporating Biomarkers Into Progression Across Stages

Bipolar Disorder Integrative Staging: Incorporating Biomarkers Into Progression Across Stages

Bipolar disorder (BD) is a lifelong, recurrent condition with growing evidence supporting a neuroprogressive course, entailing the need to adopt staging models to guide stage-speci c interventions. Although different approaches have been proposed, their application remains limited and largely based on clinical features. BOARDING-PASS is an Italian government-funded, multicenter, prospective, and observational study aimed at advancing current knowledge of BD progression through the integration of clinical, biological, neuroimaging data, alongside machine learning (ML) methodologies. The study will enroll 120 subjects (age 18-70 years), classified according to the Kupka & Hillegers' staging model, and recruited from three secondary-level psychiatric services in Italy. The primary outcome is the longitudinal assessment of clinical stage progression over an 18-month period, with evaluations conducted at baseline (T0), T1 (6 months), T2 (12 months), and T3 (18 months after baseline). At each time point, clinical variables will be collected, as well as clinical stages assigned. Additionally, at T0, T2, and T3, peripheral blood and unstimulated saliva samples will be collected to assess epigenetic regulation of gene expression - including DNA methylation, histone modi cations, and exosomal miRNAs - with a focus on key biomarkers such as C-reactive protein, proinflammatory cytokines, and BDNF, as well as microbial signatures of major oral bacterial phyla. Structural and resting state functional MRI scans will also be acquired at the same time points:structural data will be used to compute the structural connectome based on gyrification-based covariance networks, while resting-state data will be used to assess functional connectome alterations via graph theory metrics. Finally, all multimodal data will be integrated within a supervised ML algorithm based on Support Vector Machine, with the goal of developing a re ned, data-driven staging model for BD. BOARDING PASS project aligns with the growing need for a standardized, biologically informed staging framework that integrates clinical, inflammatory, epigenetic, and neuroimaging pro les to enhance prognostic accuracy and support tailored therapeutic interventions in BD.

Rationale Bipolar Disorder (BD) is a highly prevalent (approximately 1% in the general population) and disabling condition, characterized by a chronic and progressive course. The natural history of BD typically includes an initial asymptomatic phase, followed by a prodromal stage, and subsequently the onset of a first syndromic episode (depressive or manic), which is usually followed by recurrent episodes, often in the absence of full inter-episodic recovery. A growing body of evidence supports the notion that the longitudinal course of BD is associated with an active process of neuroprogression, characterized by progressive brain alterations and functional impairment. Several clinical factors may influence the trajectory of the disorder, including the number of affective episodes, the presence of psychiatric and medical comorbidities, exposure to stressful life events, and a family history of psychiatric disorders. Neuroimaging studies have consistently supported the neuroprogressive hypothesis in BD, demonstrating structural and functional brain alterations that progressively emerge as the illness advances. Abnormalities have also been described at the level of large-scale neural networks. In particular, the Default Mode Network appears to exhibit altered patterns of hyper- and hypo-connectivity with affective, fronto-parietal, and attentional systems, involving key hubs responsible for efficient large-scale brain communication. Moreover, structural and functional alterations have also been identified in unaffected relatives of patients with BD. Studies investigating additional biomarkers in BD have further provided evidence supporting a stage-related progression of the disorder, particularly with respect to Brain-Derived Neurotrophic Factor (BDNF) and inflammatory cytokines. Overall, recent advances in the field of clinical staging of BD, together with progress in the identification of biological markers (e.g., gene transcription modulation, neurotrophic signaling, immuno-inflammatory pathways, microbiota) and neuroimaging (structure and function of large-scale brain systems), provide a strong rationale for the development of a multidimensional staging model. Such a model would integrate clinical and neurobiological data, allowing for greater diagnostic, prognostic, and therapeutic precision.

Specific aims and experimental design

BOARDING PASS study has the following objectives:

1. to longitudinally assess BD clinical progression over an 18-month period using the Kupka & Hillegers' staging model;
2. to investigate the role of biological and neuroimaging markers in BD stage transitions (i.e. gene transcription regulation, inflammation, microbiotic, structural and functional neuroimaging measures);
3. to implement a predictive ML model based on the integration of clinical, biological, and neuroimaging data, in order to provide an individualized and data-driven prediction of BD stage transitions.

The study involves the consecutive recruitment of 120 subjects enrolled at three of the four participating centers: UO1 (ASST Fatebenefratelli-Sacco, Milan), UO2 (ASST Papa Giovanni XXIII, Bergamo), and UO3 (ASL 2 Abruzzo, Lanciano-Vasto-Chieti). Inclusion and exclusion criteria are detailed below.

Inclusion Criteria:

1. Subjects affected and unaffected by BD whose clinical stage falls within those defined by the Kupka and Hillegers model, namely:

   stage 0 (increased risk: having a first-degree relative with BD, in the absence of psychiatric symptoms); stage 1 (having a first-degree relative with BD, in the presence of non-specific psychiatric symptoms or depressive episode(s)); stage 2 (first hypo/manic episode allowing a diagnosis of BD type I or II according to DSM-5; APA, 2013); stage 3 (recurrent episode(s): depressive, hypo/manic, or mixed); stage 4 (persistent non-remitting disorder: chronic depressive, manic, or mixed episodes, including rapid cycling);

2. Individuals of both sexes;

3. Age ≥ 18 years and ≤ 70 years;

4. Ability to provide valid written informed consent.

Exclusion Criteria:

1. Inability to provide valid written informed consent;

2. Presence of intellectual disability;

3. Presence of a severe concomitant medical condition;

4. Presence of a current substance use disorder. After providing written informed consent to participate, enrolled subjects will undergo baseline assessment (T0) and will then enter an 18-month follow-up period consisting of three subsequent time points: T1 (6 months after T0), T2 (12 months after T0), and T3 (18 months after T0). At T0 and at each subsequent time point, participant assessment will include: (i) clinical and psychometric evaluation; exclusively at T0, T2, and T3, (ii) biological marker assessment; and (iii) brain magnetic resonance imaging for acquisition of structural and functional MRI data.

   Clinical and Psychometric Assessment

   At baseline (T0), for each study participant, the main sociodemographic and clinical variables will be collected and entered into an anonymized database. These include:

   (a) sociodemographic data: age, sex, ethnicity, educational level, marital status, and occupational status; (b) clinical characteristics: family history of psychiatric disorders, BD subtype, age at onset of BD and associated stressful life events, illness duration, duration of untreated illness, age at first depressive and hypo/manic episode, polarity of the first and most recent affective episode, predominant polarity, total lifetime number of affective episodes, presence of mixed or rapid-cycling features, current and previous psychopharmacological treatments, medical and psychiatric comorbidities, history of substance use disorder, lifetime number of hospitalizations, and suicide attempts.

   These variables will be used to assign the clinical stage at T0 according to the Kupka and Hillegers model (Kupka & Hillegers, 2012). At each subsequent time point, sociodemographic and clinical data will be updated in order to reassign the corresponding stage and to evaluate potential associations between clinical variables and the probability of progression to more advanced stages of illness.

   Clinical assessment will further include the administration of the following psychometric scales and questionnaires: the Test di Intelligenza Breve (TIB; Sartori, 1997; Colombo, 2002), administration time approximately 5 minutes; the Hamilton Depression Rating Scale (HDRS-21; Hamilton, 1960; Cassano et al., 1991), administration time approximately 20 minutes; the Hamilton Anxiety Rating Scale (HARS; Hamilton, 1959; Cassano et al., 1991), administration time approximately 15 minutes; the Montgomery-Åsberg Depression Rating Scale (MADRS; Montgomery & Åsberg, 1979; Palma et al., 1999), administration time approximately 15 minutes; the Young Mania Rating Scale (YMRS; Young et al., 1978; Palma et al., 1999), administration time approximately 15 minutes; and the Global Assessment of Functioning (GAF; Hall, 1995; APA, 1996), administration time approximately 2 minutes; Family Interview for Genetic Studies (FIGS); Drugs Abuse Screening Test (DAST-10), administration time has been estimated in nearly 5 minutes.TWEAK test with a completion time of roughly 2 minutes. Childhood Trauma Questionnaire (CTQ), it requires 5-10 minutes and will be administered at baseline. Paykel Scale for Recent Life Events, its administration takes approximately 15 minutes and will be conducted at baseline; Clinician Rating Scale (CRS), recorded in nearly 2 minutes at each time point and aimed at evaluating patients' adherence to pharmacological treatment.

   Biological marker assessment: Gene transcription regulation, inflammation, microbiome data

   Biological samples for gene expression, inflammation, and microbiome analyses will be collected at baseline, T2 and T3. Specifically:

   1. unstimulated saliva samples -i.e., whole saliva collected under resting conditions without gustatory, masticatory, or pharmacological stimulation- will be obtained using cotton buccal swabs (Salivette, Sarstedt, Nümbrecht, Germany) and stored at -20 °C until genomic DNA (gDNA) extraction. Exosomes wil be also isolated from saliva and miRNAs purified using an exosome RNA isolation kit.

   2. peripheral venous blood samples will be collected in two 5 ml vacutainer tubes containing sodium citrate. Serum and cellular components will be separated and total RNA as well as gDNA will be extracted from PBMCs.

      All biological samples collected at the three recruiting sites (Units 1, 2 and 3) will be transferred to the central laboratory (Unit 4) for standardized processing and analyses, including:

      - LIPIDOMICS to analyze short chain fatty acids (SCFAs) extracted from saliva derivatization for LC-MS/MS analysis will be carried out.

      Molecular biology studies:

      - gene expression analysis. Relative abundance of mRNA species in PBMCs will be assessed by real-time RT-PCR and Digital PCR.

      - DNA methylation in both blood and saliva cells a. general DNA methylation status will be analyzed using the Reduced representation bisulfite sequencing (RRBS) method (SBS sequencing of the SURFseq5000 platform); b. gene-specific DNA methylation study will be performed on amplified bisulfite (BS) treated DNA and methylation levels analyzed using PyroMark Q48 (64).

      - salivary exosomal miRNAs: miRNOme analysis and selected miRNAs after networking analysis by RealTime PCR and Digital PCR.

      • Transcriptional factors DNA-binding. ALPHAScreenTM assay technique to verify if identified recognition elements at genes promoter bind to different transcriptional factors and if this binding is directly modulated by the methylation degree of CpG motifs.
      • Salivary MICROBIOTA COMPOSITION by 16S rRNA Microbiome sequencing.

      Neuroimaging assessment

      MRI assessments will be performed using 3T scanners at T0, T2, and T3, and comprised:

      - Structural MRI (sMRI): 3D T1-weighted images will be acquired using a SPGR sequence (TE = minimum (full); flip angle, 6°; FOV, 250 mm; bandwidth, 31.25; matrix, 256 x 256) with 124 axial slices of 1.3 mm thickness. Following cortical surface reconstruction, local gyrification indices will be computed for 68 parcellated cortical regions based on the Desikan Atlas using FreeSurfer v7.1.0. A jackknife bias estimation procedure will then be applied to determine each individual's contribution to group-level covariance structure, generating a 68×68 individual-wise distance matrix. The topological organization of the resulting structural covariance networks will subsequently be analyzed using the Graph Analysis Toolbox.

      - Resting-state functional MRI: rs-fMRI images will be acquired using a gradient-echo EPI sequence with 36 axial slices (TE = 30 ms; TR = 2000 ms; voxel size: 3×3×4 mm3; matrix size: 64× 64; FOV: 192×192 mm2);, acquired in interleaved order. Each resting-state session will consist of 400 volumes. Pre-processing will be conducted using a combination of FMRIB's Software Library (FSL) and custom MATLAB scripts. The pipeline will include the following steps: (1) reorientation to standard space; (2) detection of outlier volumes, followed by spline-based interpolation of outlier timepoints; (3) spatial and temporal preprocessing, including motion correction (MCFLIRT), temporal high-pass filtering, and spatial smoothing (FWHM = 5 mm); (4) brain extraction of the structural image; (5) nonlinear registration to the MNI152 standard space using FSL-FNIRT. Static and dynamic functional connectomes will be estimated by calculating z-transformed Pearson correlation coefficients between all pairs of brain regions in the adopted parcellation scheme. Dynamic connectivity will be computed using a sliding-window approach with a window length of 30 TRs and a step size of 2 TRs. These steps will be implemented through in-house software developed in MATLAB. Graph-theoretical measures will be computed through the Brain Connectivity Toolbox (MATLAB).

      To minimize inter-site variability in neuroimaging data, both structural and functional MRI acquisitions will performed using harmonized protocols across the two imaging centers, each equipped with a 3 T scanner. ML algorithms A ML framework will be developed to predict clinical stage transitions in BD by integrating collected clinical, biological, and neuroimaging data. To manage the integration of multimodal data, we will adopt robust pre-processing pipelines including data normalization, outlier detection, and imputation methods for handling missing values (e.g., k-nearest neighbor or multiple imputation). ML analyses will start at month 6 of the study and will be conducted using MATLAB's Statistics and Machine Learning Toolbox, initially supported by NeuroMiner software (http://proniapredictors.eu/neurominer/index.html), a validated software platform designed to manage heterogeneous datasets. NeuroMiner provides a broad range of cross-validation frameworks, preprocessing strategies, supervised learning algorithms, feature selection tools, and external validation methods. The ML approach will be based on supervised classifiers, primarily Support Vector Machine (SVM) and Bayesian models. In the preliminary phase, classifiers will be trained and tested on preliminary datasets to compare alternative predictive models in a controlled setting and to identify the best-performing algorithmic configuration. Subsequently, feature selection procedures will be employed to identify the most discriminative variables from the pool of candidate features. This step is crucial to enhance model interpretability and to prevent overfitting, especially in small-sample, high-dimensional datasets typical of multimodal studies. In fact, by reducing the number of input variables, feature selection minimizes noise, lowers model complexity, and improves generalizability of the ML predictions. If needed (i.e. if the dimensionality of the feature space is still too high), to further reduce overfitting and the computational burden, Principal Component Analysis might be applied. SVM classifiers will be prioritized due to their robustness in handling high-dimensional, small-sample data. In addition, class weighting will be applied so that errors on minority stages are penalized more heavily, reducing the bias introduced by uneven group sizes. Model performance will be assessed using cross-validation techniques (e.g. leave-one-out or stratified k-fold, depending on data structure and class distribution) and evaluated in terms of sensitivity, specificity, F1 score and overall accuracy, in order to account for potential class imbalance. A predefined minimum target accuracy of 90% is required for model deployment on the full dataset. Study sites and organizational structure BOARDING-PASS study will be conducted across four operational units (UOs), each with defined roles to ensure high-quality data acquisition and standardized clinical procedures.

      - UO1 (ASST Fatebenefratelli-Sacco, Milan) is the coordinating center, responsible for patient recruitment and baseline diagnostic assessments, standardized collection of clinical and biological data. UO1 investigators will provide coordination and oversight within the multicenter research network, maintaining effective communication among clinicians, researchers, and technical staff.

      • UO2 (ASST Papa Giovanni XXIII, Bergamo) is responsible for participant recruitment, diagnostic assessment, collection of biological samples, structural and resting-state MRI data acquisition also on behalf of UO1, and structural neuroimaging analysis.
      • UO3 (ASL 2 Lanciano-Vasto-Chieti) will conduct participant recruitment and assessment, data collection and management, biological sample collection, as well as structural and functional neuroimaging data acquisition.
      • UO4 (University of Teramo) serves as the centralized facility responsible for biological analyses, specifically gene transcription regulation and microbiota analyses. UO4 is also responsible for the analysis of functional neuroimaging data and for the implementation of ML algorithms.

Inclusion Criteria:

1. A range of bipolar spectrum manifestations meeting the criteria of the Kupka & Hillegers' staging model
2. Both genders
3. Age ≥18 and ≤70 years
4. Written informed consent obtained

Exclusion Criteria:

1. Inability to provide informed consent
2. Diagnosis of intellectual disability
3. Presence of a severe medical condition (e.g., previously diagnosed neurological disorders, including chronic migraine, hematological diseases, renal disorders, history of stroke, or diabetes mellitus, even if compensated; controlled hypertension allowed)
4. Current substance use disorder or within six months prior to screening