Enhancing Early Diagnosis of Bipolar Disorder in Adolescents 
through Multimodal Neuroimaging

doi:10.21203/rs.3.rs-3915859/v1

Download PDF

Article

Enhancing Early Diagnosis of Bipolar Disorder in Adolescents through Multimodal Neuroimaging

https://doi.org/10.21203/rs.3.rs-3915859/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Bipolar Disorder (BD), a severe neuropsychiatric condition, often manifests during adolescence. Traditional diagnostic methods, relying predominantly on clinical interviews and symptom assessments, may fall short in accuracy, especially when based solely on single-modal MRI techniques.

Objective

This study aims to bridge the diagnostic gap in adolescent BD by integrating behavioral assessments with a multimodal MRI approach. We hypothesize that this combination will enhance the accuracy of BD diagnosis in adolescents at risk.

Methods

A retrospective cohort of 309 subjects, including BD patients, offspring of BD patients (with and without subthreshold symptoms), non-BD offspring with subthreshold symptoms, and healthy controls, was analysed. Behavioral attributes encompassing psychiatric familial history and assessments were integrated with MRI morphological and network features derived from T1, fMRI, and DTI. Three diagnostic models were developed using GLMNET multinomial regression: a clinical diagnosis model based on behavioral attributes, an MRI-based model, and a comprehensive model integrating both datasets.

Results

The comprehensive model outperformed the clinical and MRI-based models in diagnostic accuracy, achieving a prediction accuracy of 0.83 (CI: [0.72, 0.92]), significantly higher than the clinical diagnosis approach (accuracy of 0.75) and the MRI-based approach (accuracy of 0.65). These findings were further validated with an external cohort, demonstrating a high accuracy of 0.89 (AUC = 0.95). Notably, structural equation modelling revealed that factors like Clinical Diagnosis, Parental BD History, and Global Function significantly impacted Brain Health, with Psychiatric Symptoms having a marginal influence.

Conclusion

This study underscores the substantial value of integrating multimodal MRI with behavioral assessments for early BD diagnosis in at-risk adolescents. The fusion of phenomenology with neuroimaging promises more accurate patient subgroup distinctions, enabling timely interventions and potentially improving overall health outcomes. Our findings suggest a paradigm shift in the diagnostic approach for BD, highlighting the necessity of incorporating advanced imaging techniques in routine clinical evaluations.

Health sciences/Biomarkers/Diagnostic markers

Biological sciences/Neuroscience

Bipolar Disorder

Adolescent Psychiatric Assessment

Behavioral and Neuroimaging Integration

Multimodal MRI

GLMNET Multinomial Regression

Bipolar Disorder (BD), a neuropsychiatric condition with a high heritability rate of approximately 70%, predominantly affects younger populations (Gordovez & McMahon, 2020; McIntyre et al., 2020). This disorder is characterized by its progressive nature, often presenting milder initial symptoms before evolving into more classic manifestations. The need for early intervention is underscored by findings that up to 70% of individuals exhibit mood symptoms by the age of 21 (Vieta et al., 2018). However, the challenge in clinical diagnostics is particularly pronounced in adolescents at risk of BD, where subthreshold symptoms may lack the clarity necessary for a definitive diagnosis (McIntyre et al., 2020).

The clinical diagnosis of BD necessitates a nuanced evaluation, encompassing an individual's symptoms, family history, and the longitudinal course of the condition (Carvalho et al., 2020). This approach is critically important for those at risk, such as offspring of BD patients, who may exhibit subthreshold symptoms that do not fully align with diagnostic criteria, yet indicate a potential progression toward mood disorders (Axelson et al., 2015; Birmaher et al., 2021; Duffy et al., 2019). The challenge in clinical practice is to accurately characterize and monitor these individuals over time, facilitating early and appropriate intervention (Vieta et al., 2018).

Historically, research on BD, particularly in patients and offspring of BD parents, has predominantly utilized single-modality MRI approaches (Bora et al., 2021; Cattarinussi et al., 2022; De Zwarte et al., 2019; Luna et al., 2022; Roberts et al., 2018; Roberts, Lenroot, et al., 2022; Roberts, Perry, et al., 2022). However, the inherent complexity of familial risk factors and psychiatric symptoms presents challenges in their detection using a unimodal MRI approach. This limitation often leads to complications in interpretation and potential overfitting when categorization relies solely on structural or functional attributes. As a result, a comprehensive understanding of the structural and functional alterations in genetically predisposed individuals and other at-risk adolescents has remained largely elusive (Lin et al., 2015).

In this study, we aim to address these methodological limitations by evaluating the impact of various diagnostic approaches on diverse neuropsychiatric groups. These groups include offspring of BD patients with and without subthreshold symptoms, non-BD offspring exhibiting subthreshold symptoms, BD patients, and healthy controls. We have developed three distinct GLMNET multinomial classification models: (1) a comprehensive model integrating behavioral factors, encompassing psychiatric familial history and psychiatric assessments, with anatomical and functional features derived from multimodal MRI; (2) a data-driven model focusing solely on MRI features; and (3) a clinical diagnosis model rooted in behavioral variables, highlighting the importance of traditional clinical metrics in the diagnosis process.

Study design and participant selection

This study utilizes data sourced from the Recognition and Early Intervention of Prodromal Bipolar Disorders (REI-PBD) initiative (ClinicalTrials.gov Identifier: NCT01863628) (Lin et al., 2015; Lin, Shao, Geng, et al., 2018; Lin, Shao, Lu, et al., 2018; Lu et al., 2024). The Institutional Review Board of the Affiliated Brain Hospital of Guangzhou Medical University gave their ethical approval for the study. Written informed consent was acquired from all participants or their respective guardians. Recruitment included advertisements, referrals from psychiatric professionals, and word-of-mouth promotion. To balance the age and subjects number of five groups, 309 subjects selected from 392 subjects, 83 BD patients whose age were above 20 years old were excluded from further analysis. The remaining dataset included offspring of BD patients with subthreshold symptoms but without a formal BD diagnosis (OBDs, n = 48), offspring of BD patients without a BD diagnosis yet with subthreshold symptoms (OBDns, n = 63), non-BD offspring with subthreshold symptoms but without a formal BD diagnosis (nOBDs, n = 62), as well as matching BD patients (BD, n = 50) and healthy controls (HC, n = 86). The demographic and clinical characteristics of the enrolled subjects can be found in Table 1.

Table 1

Demographic and clinical characteristics of participants.
	BD (n = 50)	HC (n = 86)	nOBDs (n = 62)	OBDns (n = 63)	OBDs (n = 48)	Test Statistic (F/χ²)	P value
	Mean (SD) /n (%)					Test Statistic (F/χ²)	P value
Age(years)	16.7 (1.46)	16.0 (3.23)	16.0 (3.94)	17.5(5.55)	17.5(5.96)	1.865	0.116
Education (years)	8.7 (4.2)	9.5 (3.4)	10.1 (2.3)	9.6 (2.9)		1.379	0.241
Sex (female)	26 (53.1%)	30 (63.8%)	29 (56.9%)	40 (55.6%)		1.236	0.872
HAMD	4.4 (5.3)	0.4 (0.9)	6.9 (7.6)	0.6(1.0)	7.7(8.6)	26.42	< 0.001
HAMA	3.1 (4.8)	0.4(1.1)	4.9(5.5)	0.4(0.8)	5.8(7.5)	20.36	< 0.001
YMRS	1.3(2.2)	0.2(1.6)	1.8(2.3)	0.4(1.3)	2.2(2.7)	14.7	< 0.001
BPRS	20.3(2.6)	18.3(0.8)	21.6(5.1)	17.7(3.3)	23.0(5.5)	21.85	< 0.001
GAS	79.8(10.1)	94.2(3.5)	93.6(13.4)	77.2(4.2)	75.1(11.8)	68.03	< 0.001
PBD (with/without)	8/42	0/86	0/62	63/0	48/0
PDD	3/47	1/85	2/60	3/45	4/59
PSP	0/50	0/86	0/62	1/47	0/63
PSUB	1/49	1/85	6/56	8/40	0/63
FAMDD	2/48	1/85	5/57	6/42	4/59
FAMSP	1/49	0/86	0/62	1/47	0/63
FAMSUB	0/50	1/85	0/62	2/46	0/63
Note: BD, Bipolar Disorder patients; nOBDs, non-BD offspring with subthreshold symptoms; OBDs, BD offspring with subthreshold symptoms; OBDns, BD offspring without subthreshold symptoms; HAMD, Hamilton Depression Rating Scale; HAMA, Hamilton Anxiety Rating Scale; YMRS, Young Manic Rating Scale; BPRS, Brief Psychiatric Rating Scale; GAS, Global Assessment Scale; PBD, parent with bipolar disorder history; PDD, parent with depressive disorder history; PSUB, parent with substance use disorder history; PSP, parent with schizophrenia history; FAMDD, first-degree family history of depressive disorder; FAMSP, first-degree family history of schizophrenia; FAMSUB, first-degree family history of substance use disorder. The psychiatric history is indicated by with history/without history.

Participant evaluation

We undertook systematic clinical evaluations for all participants to ascertain their symptomatology. Scales such as the Hamilton Anxiety Rating Scale (HAMA) (Hamilton, 1959), the Hamilton Depression Rating Scale (HAMD) (Hamilton, 1986), the Young Mania Rating Scale (YMRS) (Youngstrom et al., 2003), and the Brief Psychiatric Rating Scale (BPRS) (Overall & Gorham, n.d.) were implemented to evaluate the severity of anxiety, depression, mania, and psychotic symptoms. We employed semi-structured tools like the Schedule for Affective Disorders and Schizophrenia for School-Aged Children: Present and Lifetime Version (K-SADS-PL DSM-IV) (for participants under 18 years) or the Structured Clinical Interview for DSM-IV Axis I Disorders, Patient Edition (SCID-I/P) (for those 18 years and older) to validate psychiatric disorder diagnoses (Kaufman et al., 1997). The Family Interview for Genetic Studies (FIGS) was employed to validate the familial history of psychiatric conditions, entailing a broad spectrum of specific diagnoses. This encompassed PBD, indicating a parental history of bipolar disorder; PDD, representing a parental history of depressive disorder; PSUI, symbolizing parental history of substance use disorder; and PSP, reflecting a parental history of schizophrenia. In addition, first-degree family history markers were considered, namely FAMDD for depressive disorder, FAMSUI for substance use disorder, FAMSP for schizophrenia, and FAMSUB for substance use disorder. Global functionality was assessed using the Global Assessment of Functioning Scale (GAS) or the GAS for children (Endicott et al., 1976).

Criteria for participant inclusion and exclusion

In the OBDs (n = 48) and nOBDs (n = 63) groups, subthreshold symptoms were characterized as major depression or hypomania that did not fulfil the DSM-IV criteria for a major depressive episode (MDE) or hypomania due to insufficient symptom count (2–4 for MDE lasting at least seven days and 2–3 for hypomania persisting at least four days) or inadequate duration (less than two weeks for MDE or four days for hypomania). Those with subthreshold symptoms were excluded if they had any DSM-identified psychiatric disorders and/or were taking psychotropic medication. A group of OBDns (n = 62) with parents’ bipolar disorder history and without subthreshold symptoms was also recruited.

Furthermore, a cohort of BD patients (n = 50), selected from a larger sample of 133 patients aged under 20, was incorporated into the study. These patients had maintained clinical stability for a minimum of three months prior to recruitment. In addition, the HC group (n = 86) comprised individuals with no current or prior psychiatric disorders, no first- or second-degree family history of psychiatric disorders, and did not meet the criteria for the defined subthreshold syndrome. The total sample size was 309. The detailed information of demographics were collected in Table 1.

Multimodal MRI acquisition parameters

The MRI images were obtained using a 3.0 T Phillips scanner outfitted with an 8-channel SENSE head coil. Three modalities of MRI images were acquired for all participants, namely T1-weighted image, Diffusion Tensor Imaging (DTI), and resting-state functional magnetic resonance imaging (rsfMRI). The parameters for T1-weighted high-resolution anatomical images included: TR = 8.2 ms, TE = 3.7 ms, FOV = 256 × 256 mm², voxel sizes = 1 × 1 × 1 mm³, matrix size = 256 × 256, slices = 188, slice thickness = 1 mm, slice gaps = 0 mm, and flip angle = 7 °. Whole brain DTI data was captured through the SE-EPI weighted sequence, and the established parameters included: TR = 6000 ms, TE = 70 ms, FOV = 256 × 256 mm², Voxel size = 2 × 2 × 2 mm³, Matrix = 128 × 128, Diffusion direction = 32, slices = 75, slice gap = 0 mm, b value = 1000 s/mm², and flip angle = 90°. The resting-state fMRI data were collected using a gradient echo planar imaging sequence, and the operational parameters were collected: TR = 2000 ms, TE = 30 ms, FOV = 220 × 220 mm², voxel sizes = 3.44 × 3.44 mm², matrix size = 64 × 64, slices = 33, slice thickness = 4 mm, slice gap = 0.6 mm, flip angle = 90°, and volumes = 240.

Quality assurance was maintained through rigorous protocol implementation. We excluded images that demonstrated artifacts, insufficient resolution or contrast, motion distortions, missing modalities, or inadequate registration.

Multimodal MRI data analysis

Morphometric measures calculation

The morphometric measures, obtained during the construction of Morphometric Similarity Networks (MSN), characterize the physical attributes of the brain's structures (Seidlitz et al., 2018, 2020). These measures, including volume, thickness, surface area, Gaussian curvature, mean curvature, Fractional Anisotropy (FA), and Mean Diffusivity (MD), were extracted from the structural and diffusion MRI data.

The processing of anatomical MRI data involved intricate preprocessing procedures and stringent quality assurance protocols to enable accurate morphometric analysis. T1-weighted images underwent initial correction for low-frequency bias fields and denoising (Manjón et al., 2010; Tustison et al., 2010), subsequently being processed through the Freesurfer recon-all pipeline (version 7.3) (Fischl, 2012). This protocol enables the construction of cortical surfaces, generating five morphometric features at each vertex: cortical thickness, gray matter volume, surface area, gaussian curvature, and mean curvature. After this step, native surface images were aligned with the standard Conte69-32K template (Van Essen et al., 2012) using the Workbench (https://github.com/Washington-University/workbench).

In parallel, the DTI data were subjected to a comprehensive preprocessing pipeline using the MRtrix3 (Tournier et al., 2019), which included denoising, Gibbs ringing artifacts removal, correction for distortions and head movements, and bias field correction. Following these steps, Fractional Anisotropy (FA) and Mean Diffusivity (MD) maps were derived and co-registered with the T1-weighted images as well as the Montreal Neurological Institute (MNI) template.

The cortical parcellation process was accomplished using the Schaefer 400 parcellation atlas (Schaefer et al., 2018). This atlas comprises 400 parcels, enabling effective integration of structural and functional analysis (Eickhoff et al., 2018). To capture regional morphometric features, the parcellated atlas in the template space (either Conte69 or MNI) was transformed into each participant's native surface or volumetric space, yielding 7 measures for each of the 400 parcels per subject.

Functional connectivity (FC) and structural connectivity (SC) Calculation

Calculations of FC and SC were achieved through meticulous preprocessing procedures and detailed computations (Cruces et al., 2022). The FC estimation involved the preprocessing of BOLD signals, commencing with slice timing correction to address temporal incongruities in slice acquisition times. This phase was followed by motion correction to rectify spatial distortions resulting from participant head movements. The application of ICA-FIX denoising was undertaken to minimize non-neural noise (Griffanti et al., 2014). A subsequent nuisance regression controlled for potential confounding factors such as white matter, cerebrospinal fluid, and 24 motion parameters. Band-pass filtering (0.01–0.1 Hz) was applied to focus on frequencies associated with resting-state networks. The data were then projected onto the Conte69 surface for surface-based analysis and smoothed using a 10mm kernel to improve the signal-to-noise ratio and mitigate potential effects of inter-individual anatomical variability.

SC estimation involved a comprehensive preprocessing and computation protocol for T1-weighted and DTI data (Tournier et al., 2019). T1-weighted images were processed to remove non-brain tissue and for gray matter parcellation. DTI data were subjected to denoising, Gibbs ringing artifacts removal, correction for distortions and head movements, and bias field correction. These collective preprocessing steps facilitated the extraction of whole brain masks, and the estimation of diffusion tensors at each voxel, allowing for the derivation of fractional anisotropy (FA) and axial diffusivity (AD) maps, among other measures. Anatomically constrained tractography (ACT) was employed, requiring five-tissue-type (5TT) segmented images generating from T1-weighted image for each subject. Fiber orientation distributions (FOD) were estimated from DTIs, and probabilistic tractography was performed to yield 10 million streamlines per subject. The fiber-tracking data were subsequently filtered to identify an appropriate cross-section multiplier for each streamline.

The Schaefer 400 parcellation atlas was utilized to construct FC and SC matrices. FC matrices were built from the mean time series for each region, excluding negative correlations to minimize potential spurious anticorrelations. SC matrices were derived based on the number of streamlines connecting the designated regions.

Graph measures of FC and SC

The computational exploration of complex networks involves the application of various graph-theoretic measures to encapsulate the unique properties of network nodes and their interrelations. Leveraging the BCT package (Bullmore & Sporns, 2012; Rubinov & Sporns, 2010), we computed these measures on connectivity matrices, focusing on both regional and global aspects. The collective deployment of these measures provides a robust toolkit to probe the intricate structural and functional properties of diverse networks (Hansen et al., 2022). Furthermore, they capture critical aspects of network topology, mirroring the delicate balance between segregated and integrated information processing within the brain, thereby furnishing a comprehensive depiction of functional and structural brain networks.

Regional measures predominantly concentrate on individual nodes and their immediate vicinity, encapsulating the local features of the network. They include degree centrality, which quantifies the number of direct connections a node possesses, thereby denoting its immediate influence within the network. The clustering coefficient is another such measure, assessing the extent to which a node's neighbors are interconnected, hinting at the potential for formation of closely-knit communities. Local efficiency, a complementary measure, evaluates the proficiency of information exchange within a node's immediate vicinity. Further, betweenness centrality acts as an indicator of a node's significance with respect to the shortest paths passing through it, thus underscoring its control over information flow within the network.

In contrast, global measures offer a holistic perspective of the network's characteristics. The characteristic path length, a foundational measure, calculates the average shortest path length between all pairs of nodes, signifying the network's overall navigability. Global efficiency, inversely, assesses the smoothness and swiftness of information transfer throughout the network. Modularity indicates the degree to which the network can be partitioned into distinct communities or modules. In our study, the regional and global measures were computed interhemispheric for FC and separately for each hemisphere of SC due to potential inaccuracies in estimating interhemispheric connections.

GLMNET multinomial classification

GLMNET, situated within the Generalized Linear Model (GLM) framework, excels in both regression and classification tasks (Tay et al., 2023). Its strength derives from the combined use of Lasso (L1) and Ridge (L2) regularizations, which adeptly manage the challenges of high-dimensionality and multicollinearity. This ensures that complications from interrelated predictor variables are mitigated. Particularly noteworthy is GLMNET's proficiency in multinomial classification, which models relationships between a response variable and multiple predictors, especially when outcomes exceed two. This capability facilitates the categorization of intricate, multivariate datasets. Moreover, the L1 regularization promotes robust feature selection, emphasizing only pivotal predictors. This not only streamlines interpretation but also controls model complexity, preventing overfitting.

In our study, we classified 309 subjects into five groups, subsequently all the subjects were divided into the training and testing sets at an ratio of 80:20. We harnessed grid searching and 10-fold cross-validation via Caret to optimally tune the alpha and lambda parameters. Through L1 regularization, GLMNET excels in managing a high-dimensional feature space, differentiating key predictors from unnecessary ones, enhancing both the predictive strength and model interpretability.

Understanding the efficacy of integrating MRI features with traditional diagnostic methods can revolutionize early intervention strategies for at-risk adolescents. By enhancing the diagnostic precision, this study holds potential to pave the way for timely and targeted treatments, ultimately improving the prognosis and quality of life for those afflicted with BD.

Comprehensive GLMNET model

The comprehensive model encompassed all the imaging and behavioral features. The feature selection via GLMNET aimed to identify the most predictive variables from a myriad of morphometric measures, graph measures, and behavioral attributes. These include morphometric Measures: Area, thickness, volume, mean curvature, Gaussian curvature, fractional anisotropy, and mean diffusivity. Graph Measures: Degree centrality, clustering coefficient, local efficiency, betweenness centrality, characteristic path length, global efficiency, and modularity derived from FC and SC. Behavioral Characteristics: HAMD, HAMA, YMRS, BPRS, GAS, and familial history variables (PBD, PDD, PSP, PSUB, FAMDD, FAMSP, FAMSUB), as well as age, gender, brain volume, and education years.

Clinical diagnosis GLMNET model

In clinical settings, mental disorders are often diagnosed based on clinical interviews and psychological assessments. Drawing from this, the first model was mainly focused entirely on the behavioral attributes, excluding morphometric and graph measures for the brain. This model represented the traditional diagnostic approach, enabling us to gauge how clinicians' interview-based assessments might compare with models incorporating MRI data.

MRI-based GLMNET model

The MRI-based model was exclusively focused on morphometric and graph measures from multimodal MRI data, omitting behavioral attributes and family history. These include morphometric measures: area, thickness, volume, mean curvature, Gaussian curvature, fractional anisotropy, and mean diffusivity; Graph Measures: degree centrality, clustering coefficient, local efficiency, betweenness centrality, characteristic path length, global efficiency, and modularity derived from FC and SC. This model could provide insights into the predictive power of anatomical and network features in isolation, elucidating the added value of behavioural data.

By contrasting the clinical diagnosis, MRI-based, and comprehensive models, we tried to gain a richer understanding of the relative importance of MRI and behavioral data. This deepened the interpretability and relevance of the models we craft.

Cross validation of the GLMNET models

To examine the robustness and reliability of GLMNET models, a cross-validation strategy was employed in the current study. For these three models, we used the same one-time split of subjects into an 80:20 training and testing set ratio. Recognizing the limited insights this provided, we enhanced our method by randomly splitting the subjects 100 times to perform systematic cross-validation. In each iteration, the model was trained on the selected 80% of data, and its predictive accuracy evaluated on the remaining 20%. This process was conducted for all three models, bolstering the credibility of their performance, reducing potential overfitting, and improving the reliability of our findings.

Structural equation modelling

To probe the interplay between observed variables and latent constructs, and the dynamics among the latent constructs themselves, we employed a structural equation modelling (SEM) approach using the lavaan package in R. Two distinct latent constructs were delineated: Psychiatric Symptoms and Brain Health. The manifest indicators constituting the Psychiatric Symptoms latent construct encompassed the BPRS, HAMA, HAMD, and YMRS. Meanwhile, the Brain Health latent construct was defined by the top 20 MRI features, which had been ranked according to L2-norms in the GLMNET coefficients. The regression schema was articulated as follows: the variable Clinical Diagnosis was predicated upon the Parental BD History and Psychiatric Symptoms. The Brain Health latent construct was regressed on the Psychiatric Symptoms latent construct, Clinical Diagnosis, parental BD history, and GAS. The SEM was subsequently fitted to the dataset in order to derive path coefficient estimates and to gauge the adequacy of the model fit, with the Maximum Likelihood estimation method being employed for this purpose.

Independent validation using external data

To enhance the robustness of our comprehensive GLMNET model, we performed an independent validation using an external dataset from the Affiliated Nanjing Brain Hospital of Nanjing Medical University. This external dataset was comprised exclusively of BD patients (n = 40, age = 17.6 ± 2.01) and HC (n = 34, age = 18.21 ± 1.70), reflecting a common limitation in psychiatric research where data from other conditions are less accessible. The detailed information of demographics and methodology can be found in Supplementary Materials.

The validation dataset encompassed all the measures of morphometric measures and graph measures derived from FC and SC, with the exact same processing procedures like the exploring dataset. However, education years, GAS scores and familial psychiatric history were not included in this dataset due to the limitation of the data. The validation process involved logistic regression analysis, utilizing the coefficients from the comprehensive GLMNET model to maintain consistency. This approach was critical in evaluating the model's predictive validity and generalizability to an external cohort, thereby ensuring that its performance was not an artifact of overfitting to the original dataset.

Demographics

Table 1 presents a comparative analysis of different demographic and clinical measures across five groups: BD, nOBDs, OBDns, OBDs, and HC. The groups exhibited similar ages [F_{(5, 304)} = 1.865, p = 0.116], suggesting that age was not a significant distinguishing factor. Education levels also did not significantly vary between the groups [F_(5,304) = 1.379, p = 0.241]. The gender distribution across the groups was found to be balanced, with no significant difference in the proportion of females [χ²_(5,304) = 1.236, df = 4, p = 0.872]. When it comes to the clinical measures, the groups demonstrated marked differences. HAMD scores were significantly different across groups [F_(5,304) = 26.42, p < 0.001], as were the HAMA [F_(5,304) = 20.36, p < 0.001], YMRS [F_(5,304) = 14.7, p < 0.001], BPRS [F_(5,304) = 21.85, p < 0.001], and GAS scores [F_(5,304) = 68.03, p < 0.001]. Further, the groups varied significantly in terms of parental psychiatric history. The prevalence of parental history of BD was concentrated in OBDs and OBDns groups, as well as in some participants in BD group. Parental history of depressive disorder, substance use disorder, and schizophrenia was commonly low in all groups. Family histories of depressive disorders, substance use disorders, and schizophrenia in the first-degree relatives also showed small variation across groups. The characteristics of demographic and clinical measures in external validation dataset can be found in Supplementary Table 1.

Classification analysis with GLMNET

Data encompassing behavioral attributes like PBD, GAS, YMRS, in addition to multimodal MRI images, resulted in a comprehensive feature set of 6006 metrics and 16 behavioral variables for each participant. The averaged MRI measures of all participants can be found in Fig. 1. Then, we tried to investigate the multinomial classification of five discrete categories: BD, HC, nOBDs, OBDns, and OBDs, as delineated in Fig. 2A. 309 subjects were assigned to training and testing sets in a random split of 80:20. Optimal tuning parameters (alpha and lambda) were determined via grid search combined with ten-fold cross-validation, using the Caret package integrated with the GLMNET model for Elastic-Net regularization. The whole dataset was employed in three GLMNET models for multinomial classification and essential feature identification.

Comprehensive model

In the main analysis, the morphometric and graph measures and behavioral attributes were all included (6022 in total), the training process yielded 2500 unique tuning iterations and optimal parameters (alpha: 0.669, lambda: 0.049), achieving peak model accuracy of 0.827 (Fig. 2B). Figure 2B also illustrates the GLMNET coefficients plot of the comprehensive model, demonstrating an inverse correlation between coefficients and lambda, indicative of increasing model simplicity with enhanced regularization.

The optimized model classified the test set, resulting in classification accuracies for BD (60%), HC (94.1%), nOBDs (66.7%), OBDns (100%), and OBDs (88.9%), as illustrated in the confusion matrix (Fig. 2C). The model's overall accuracy was approximately 0.833, with a kappa statistic of 0.787, signifying significant agreement beyond chance. The accuracy confidence interval ranged from 0.715 to 0.917, and a near-zero p-value (1.2e-18) confirmed model superiority over a random classifier.

Further evaluation of the model's discriminatory capability was conducted via Receiver Operating Characteristic (ROC) curves (Fig. 2D). The resulting Area Under the Curve (AUC) values for BD (0.9), HC (0.98), nOBDs (0.93), OBDns (1), and OBDs (0.98), signified superior discriminative performance across all groups. Classification performance by class exhibited sensitivity from 0.600 (BD) to 1.000 (OBDns), and specificity from 0.883 (HC) to 1.000 (BD).

Clinical diagnosis model

This clinical diagnosis model was exclusively grounded on behavioral variables (16 features), explicitly omitting the morphometric and graph measures. By focusing solely on behavioral variables such as HAMD, HAMA, YMRS, BPRS, GAS, and family history variables, it offers a clear insight into the diagnostic power rooted in behavioral observations.

The classification model achieved a training accuracy of 0.78, and a comprehensive assessment of the model’s performance yielded an overall prediction accuracy of 0.75 (Fig. 3A), backed by a Kappa statistic of 0.68. The range for the model’s predictive accuracy was defined between 0.62 (Accuracy Lower) and 0.85 (Accuracy Upper), emphasizing confidence in the model’s capabilities. The p-value of 9.3e-14 denoted a statistically significant deviation from a random classifier. The Receiver Operating Characteristic (ROC) curve for the data could illustrate the model's discriminative ability between the enrolled groups, here, the ROC curves demonstrated high AUC values for different groups (Fig. 3B): BD (0.76), HC (0.97), nOBDs (0.91), OBDns (1), and OBDs (0.98), which demonstrated the fair discriminative ability of the clinical diagnosis model. The coefficients plot of the model can be found in Supplementary Fig. 2.

MRI-based Model

The MRI-based model focused on morphometric measures and graph measures (6006 features in total), purposely excluding the behavioral variables. This enables the examination of the unique predictive power of the anatomical and network features.

Employing the same parameter tuning strategy and training/testing sets from the clinical diagnosis model, the MRI-based model achieved a training accuracy of 0.63 (Fig. 3C), demonstrating its potential to differentiate between classes using only anatomical and network features. Model performance evaluation revealed an overall predicting accuracy of 0.65 (Fig. 3C), with a Kappa statistic of 0.55. The Accuracy Lower and Accuracy Upper statistics were 0.52 and 0.77 respectively, indicating confidence in the model's predictive accuracy. Null Accuracy was 0.283, suggesting the model's superior performance over a naive, majority-class prediction approach. The p-value for the accuracy was 4.0e-9, indicating statistical significance against the random classifier. Further performance quantification via Receiver Operating Characteristic (ROC) curves (Fig. 3D) showed Area Under the Curve (AUC) values for BD (0.93), HC (0.9), nOBDs (0.8), OBDns (0.99), and OBDs (0.8). The coefficients plot of the model can be found in Supplementary Figs. 3 & 4.

Reliability of the three models

Systematic cross-validation provided evidence of the reliability of three GLMNET models (Fig. 3E). The comprehensive model exhibited a mean accuracy of 0.804 across 100 iterations. The MRI-based model demonstrated a lower mean accuracy of 0.637. A t-test comparing the comprehensive model to the MRI-based model indicated a significant difference (t = 23.371, df = 198, p < 2.2e-16). Lastly, the clinical diagnosis model has a mean accuracy of 0.726. A t-test between the comprehensive model and the clinical diagnosis model also showed a significant difference in their accuracies (t = 12.087, df = 198, p < 2.2e-16).

Discriminative behavioural and MRI features in comprehensive model

For the comprehensive model, the detailed GLMNET coefficient analysis for each group revealed the most influential features, including both MRI and behavioral variables, suggesting a complex relationship between behavioral and neuroimaging factors. In the features identification conducted using GLMNET, the top 20 L2-norm coefficients of all features were determined (Fig. 2E). Notably, parental BD history (PBD) and global assessment of functioning (GAS) emerged as the most discriminative variables.

The PBD exhibited the highest coefficient, valued at approximately 1.801, followed by GAS with a coefficient of 0.735. Other features, such as the volume of left frontal operculum insula region 1, volume of right precuneus posterior cingulate cortex region 4, and volume of right dorsal prefrontal cortex medial prefrontal cortex region 4, had coefficients of 0.256, 0.253, and 0.244 respectively. YMRS, MD of right posterior region 3and volume of right frontal operculum insula region 7 presented coefficients in the range of 0.2, specifically 0.223, 0.219, and 0.187. The features, namely the mean curvature of left visual region 20, volume of right posterior region 17, and gaussian curvature of right orbital frontal cortex region 6, ranged from 0.177 to 0.159. Moreover, gaussian curvature of left prefrontal cortex region 12, gaussian curvature of left lateral prefrontal cortex regions 3 and 5 had coefficients between 0.158 and 0.146. The remaining features, such as the thickness of left precuneus posterior cingulate cortex region 9, BPRS, thickness of right somatomotor region 12, local efficiency of structural connectivity of right frontal eye fields region 1, volume of left lateral prefrontal cortex region 8, and volume of left somatomotor region 29, demonstrated coefficients descending from 0.145 to 0.124.

Furthermore, Fig. 4 also illustrated the significant role of behavioral variables in the classification of five groups, with GLMNET coefficients indicating specific variable associations. Variables such as age, volume, gender, PDD, PSUI, FAMMD, FAMSUI, FAMSP, and FAMSUB had negligible impact across all groups due to zero coefficients. Here, only the top 10 features were displayed. In the BD group (Fig. 4A), MRI features such as the volume of right precuneus posterior cingulate cortex (0.253), volume of right dorsal medial prefrontal cortex (0.243), and MD of right posterior cortex (0.219) were dominant. For the HC group (Fig. 4B), PBD (0.643), global function GAS (0.556), and mania symptoms YMRS (-0.171) were the principal contributors. The nOBDs group (Fig. 4C) featured both MRI and behavioral variables, including PBD (0.477), GAS (-0.348), mania symptoms YMRS (0.097), mean curvature of left visual cortex (0.177), and gaussian curvature of left lateral prefrontal cortex (-0.145). The OBDns (Fig. 4D) group was influenced by PBD (-1.198), GAS (0.327), psychotic symptoms BPRS (-0.074), and MRI features such as volume of left frontal operculum insula (0.255). Lastly, in the OBDs group (Fig. 4E), PBD (-1.080), BPRS (0.113), YMRS (0.104), and certain MRI features were prominent.

Behavioral variables (such as PBD, BPRS, GAS and YMRS) and certain MRI features appeared across groups, indicating their significance in neuropsychiatric contexts. Surprisingly, anatomical attributes but not functional attributes, such as volume and curvature, were the most distinctive MRI features. As per Fig. 4F, discriminative MRI features, characterized by their unique distributions across the cerebral landscape, were mapped onto the brain's surface, further accentuating each group's distinct brain mapping patterns.

Relationships among Clinical Diagnosis, Parental BD History, Psychiatric Symptoms, Brain Health and Global Function

The structural equation model was constructed to understand the relationships between Psychiatric Symptoms, Brain Health, Clinical Diagnosis, and Parental BD History. The model was estimated using 309 observations and 60 parameters. The chi-square test yielded a significant result for the user model [χ²₍₃₇₂₎ = 929.683, p < 0.001] and the baseline model [χ²₍₄₀₅₎ = 2542.440, p < 0.001]. The Comparative Fit Index (CFI) was 0.740 and the Tucker-Lewis Index (TLI) was 0.717. The Root Mean Square Error of Approximation (RMSEA) was 0.069 with a 90% confidence interval ranging from 0.064 to 0.075. The Standardized Root Mean Square Residual (SRMR) was 0.101. In summary, the model presents a moderate to acceptable fit to the data based on the fit indices. While some indices like RMSEA suggest an acceptable fit, others like CFI, TLI, and SRMR indicate a moderate fit.

In terms of parameter estimates (Fig. 5A), Psychiatric Symptoms were significantly influenced by BPRS, HAMA, HAMD, and YMRS. Brain Health was significantly associated with several brain volume and curvature measures. Clinical Diagnosis had a significant positive effect on Parental BD History (β = 0.769, p < 0.0001) and Psychiatric Symptoms (β = 0.143, p < 0.0001). Brain Health was significantly influenced by Clinical Diagnosis (β = 0.487, p < 0.0001), Parental BD History (β = -0.380, p < 0.0001), and Global Function (β = 0.578, p < 0.0001), with a marginal influence from Psychiatric Symptoms (β = -0.112, p = 0.056).

External validation results

Finally, the comprehensive GLMNET model was independently validated using external dataset from Nanjing Medical University, distinguishing BD patients from HC. The model demonstrated high accuracy (89.19%) and a significant improvement over random classification, as indicated by a p-Value of 8.073e-11 (Fig. 5B). The Kappa statistic of 0.7814 suggests a substantial agreement between predicted and actual classifications.

Sensitivity (92.50%) and specificity (85.29%) metrics were high, indicating the model's effectiveness in correctly identifying both BD and HC cases. Positive predictive value (88.10%) and negative predictive value (90.62%) reinforce the model's reliability (Fig. 5B). Balanced Accuracy stood at 88.90%, confirming the model's consistent performance across both conditions. The ROC curve for the independent validation data could illustrate the model's discriminative ability between BD and HC groups, and the AUC value is 0.95, indicating the excellent performance for the comprehensive model (Fig. 5C). These results support the robustness and transferability of the GLMNET model to external datasets for BD detection.

Enhanced diagnosis through MRI metrics integration in a real-world setting

The accurate diagnosis of neuropsychiatric conditions, particularly bipolar disorder among at-risk adolescents, remains a substantial challenge in clinical settings. Early and accurate classification is crucial for effective intervention and management (Vieta et al., 2018). Our study leverages an innovative approach by integrating multimodal MRI data with behavioral attributes to improve the precision of bipolar disorder diagnoses. This integrative approach has advantages over conventional methods, reflecting a multidimensional understanding of psychiatric conditions.

The research methodology employed a sophisticated, data-driven approach involving three GLMNET multinomial classification models: one focusing on behavioral data, another on MRI data, and a comprehensive model integrating both. This approach was designed to identify and leverage patterns within the data that might be indicative of bipolar disorder, thus aiding in more accurate classification of at-risk adolescents. Most notably, the integration of MRI metrics not only mirrored the intricacies of real-world diagnoses but also significantly amplified diagnostic precision. This breakthrough underscores the indispensable role of MRI metrics in refining and elevating neuropsychiatric assessments.

Unravelling the predictive potential of behavioral attributes

The comprehensive model, which integrated both behavioral and MRI data, achieved a prediction accuracy rate significantly higher than those models which relied on a single type of data. The study's analytical techniques allowed for the identification of specific discriminative features critical in classifying bipolar disorder among at-risk adolescents. Within the behavioral data, parental bipolar disorder history (PBD) and Global Assessment Scale (GAS) scores emerged as significant markers. These elements likely reflect the genetic predispositions and the overall functional impact of the disorder on individuals.

Results showed that the influence of PBD was remarkably evident in the HC and nOBDs groups, indicating a positive association. Conversely, a negative influence was observed in the OBDns and OBDs groups. This suggests that a strong positive association signifies the characterization of HC and nOBDs by the absence of parental BD history, while an even stronger negative association signifies the characterization of OBDns and OBDs by the presence of parental BD history.

The findings of this study are remarkably noteworthy, as the model successfully identified the most discriminative feature, namely PBD, among the pool of 6022 features. The selection of PBD as the primary marker for group classification is critical. However, it is worth noting that the discriminative power of PBD is not evident within the BD group. The BD group contains both the patients with and without PBD, we speculate that the consideration of genetic factors or behavioral assessments no longer solely characterizes the condition once it has progressed into a diseased state. In other words, once BD has developed, genetic factors cease to hold meaningful or significant implications for characterizing the affected population. Instead, the primary indicators of the BD group manifest predominantly in the features observed through MRI. This finding attests to the GLMNET model's efficacy in discerning the most influential factors from an extensive feature set to differentiate participants based on their familial history.

GAS is a clinical tool used in mental health assessments to measure and track the overall functioning and severity of symptoms in individuals. It provides a standardized method for evaluating an individual's level of functioning across various domains of life, such as work, social relationships, self-care, and psychological well-being. Higher scores denote better functioning, while lower scores indicate poorer functioning and more severe symptoms. GAS positively impacted the HC and OBDns groups while negatively influencing the nOBDs group, affirming its utility to distinguish the groups with or without subthreshold symptoms.

Furthermore, certain variables, such as BPRS and YMRS though not as widely recognized as PBD and GAS, remain relevant in various groups. The BPRS, a measure of psychiatric symptoms, exhibited a slight negative influence on OBDns, but positively impacted OBDs. In the case of YMRS, a scale for assessing manic symptoms, it had a negative impact on HC, while positively affecting both nOBDs and OBDs. As for the anxiety symptom scale HAMA, it had a minimal impact, specifically on OBDs. Lastly, the depression symptom scale HAMD did not demonstrate significant effects on any of the groups mentioned. It is reasonable to observe these impacts since these groups are characterized by symptoms of BD.

Discriminative capability of MRI metrics

The examination of the data-driven model which considered only MRI measures for prediction, highlighted the inherent discriminative potential of anatomical and network features. However, it also showcased the substantial role that behavioral variables play in enriching the classification models for neuropsychiatric conditions.

In MRI-based and comprehensive models, distinct MRI features were observed to exert significant influence within each neuropsychiatric group. Notably, the frontal operculum insula, lateral prefrontal cortex, prefrontal cortex, precuneus, posterior cingulate cortex, orbital frontal cortex, temporal pole, and dorsal medial prefrontal cortex emerged as consistently discriminative brain regions across the five groups. As partially confirmed by previous research on the adolescents at high genetic risk for BD (Bora et al., 2021; Roberts et al., 2016, 2018; Roberts, Lenroot, et al., 2022; Roberts, Perry, et al., 2022), compared to the HC, these adolescents showed greater cortical thickness in orbital frontal cortex, prefrontal cortex, structural dysconnectivity in left and right prefrontal regions, fronto-temporal regions.

Studies have reported structural and functional abnormalities in these regions in BD (Chen et al., 2022). A large-scale study found that, cortical gray matter was thinner in frontal, temporal and parietal regions of both brain hemispheres in BD (Hibar et al., 2018). This includes the frontal operculum insula (Pastrnak et al., 2021), a region critical to emotional processing and awareness, where abnormalities may contribute to mood dysregulation observed in bipolar disorder. The lateral prefrontal cortex and the prefrontal cortex (Carlén, 2017; Whittaker et al., 2018), both implicated in cognitive control, emotion regulation, decision-making, and social behaviour, are thought to be affected possibly leading to cognitive deficits and emotional dysregulation; further, the precuneus and the dorsal medial prefrontal cortex, key to self-awareness, episodic memory, and social cognition, may harbour changes affecting self-perception and memory. The posterior cingulate cortex and orbitofrontal cortex (Rolls et al., 2022), involved in internally directed thought, emotional processing, decision-making, and impulsivity control, may demonstrate altered activity associated with mood symptoms and poor decision-making during manic episodes.

Finally, dysfunction in the temporal pole (Chen et al., 2022), critical for social cognition and emotional processing, may contribute to social and emotional dysregulation. It should be emphasized that these brain-region involvements are intricate, interrelated, and form a neural network underlying bipolar disorder. This understanding, however, continues to evolve, and ongoing research aims to elucidate these relationships further.

In the MRI-based model, each group was identified by distinctive MRI feature profiles, albeit certain brain regions were recurrently implicated. It is notable that no intersecting features were observed among the five groups under study. These individualized feature profiles underline the complex, multifaceted nature of neuropsychiatric disorders. The subtle interplay between the anatomical and network features in each group suggests that the underpinnings of these conditions extend beyond generic markers and require a more personalized approach for identification and treatment. Interestingly, anatomical attributes such as volume and curvature emerged as crucial markers, taking precedence over features derived from FC and SC graph measures. This observation implies that the anatomical changes observed in neuropsychiatric conditions might have more discriminative capacity than the functional alterations.

In comparison of the two models, the comprehensive model revealed PBD and GAS as the most discriminative features. Notably, the BD group was primarily influenced by MRI features, whereas HC, nOBDs, OBDns, and OBDs groups were influenced by both MRI and behavioral variables. Intriguingly, the MRI-based model that survived more MRI features did not necessarily produce superior results. Without adding the behavioral attributes into the model, the predicting accuracy dropped significantly. The comprehensive model with fewer, but more selective, features demonstrated higher discriminative power. This reinforces the hypothesis that the integration of behavioral variables with MRI features produces a synergistic effect, thereby enhancing the model's ability to discern between neuropsychiatric conditions.

Interplay of Psychiatric Symptoms, Brain Health, Parental BD History, Clinical Diagnosis and Global Function

The results from SEM elucidate the relationships between Psychiatric Symptoms, Brain Health, Clinical Diagnosis, Parental BD History and Global Function, we identified key interactions between the two primary latent constructs: Psychiatric Symptoms and Brain Health.Notably, BPRS, HAMA, HAMD, and YMRS were significant indicators for Psychiatric Symptoms, emphasizing their clinical relevance. The strong influence of the top 20 MRI features on Brain Health underscores the importance of neuroimaging in assessing brain health and its potential role in psychiatric disorders.

The significant association between Clinical Diagnosis and Parental BD History highlights a potential hereditary component in psychiatric disorders. While Brain Health was significantly influenced by Clinical Diagnosis and Parental BD History, its relationship with Psychiatric Symptoms was marginally significant, suggesting a more subtle interplay.

Generalizability of the comprehensive model

A critical aspect of any diagnostic model is its generalizability and applicability to diverse populations and settings. The study's approach to generalizability involves testing the classification models using external validation dataset. This step is crucial in assessing the robustness and reliability of the model beyond the initial study population. The use of external datasets helps ascertain whether the predictive power of the models holds up across different demographics, clinical settings, and even variations in MRI equipment and protocols. It is a stringent test of the model's utility and a necessary step towards clinical application. The study's findings, when validated externally, provide stronger evidence for the generalizability of the multimodal approach, supporting its adoption in broader clinical practice.

Limitations

There are several limitations that should be acknowledged. Firstly, the retrospective nature of the study may introduce selection biases, potentially impacting the generalizability of the findings. The cohort, derived from a specific population, may not fully represent the diversity seen in broader populations, including variations in ethnicity, socio-economic backgrounds, and environmental factors. Another limitation is the reliance on existing clinical assessments and MRI data. The accuracy of these inputs is crucial, and any inconsistencies or errors in initial evaluations could influence the study's outcomes. Additionally, the complexity and variability inherent in adolescent development and psychiatric symptomatology pose challenges in ensuring that the diagnostic models developed are universally applicable across different adolescent subgroups.

This study marks a significant advancement in diagnosing Bipolar Disorder in adolescents, demonstrating the efficacy of integrating multimodal MRI with behavioral assessments for greater diagnostic precision. This integrated approach promises to enhance early detection and intervention strategies, potentially reshaping the diagnosis and management of BD in clinical practice. The findings encourage further exploration into advanced imaging in psychiatry, with implications for improving patient outcomes.

Code and data availability: The project's code is accessible on the GitHub repository at https://github.com/JinFeng-Wu/BipolarCohorts_GLMNET. Data can be accessed upon request for legitimate academic or research purposes.

Funding: This project was supported by grants from the National Natural Science Foundation of China (No. 82171531, 82102031, 81671347) and the Science and Technology Program of Guangzhou, China (202007030012).

Competing interests: Dr. Roger S. McIntyre has received research grant support from CIHR/GACD/National Natural Science Foundationof China (NSFC) and the Milken Institute; speaker/consultation fees from Lundbeck, Janssen, Alkermes, Neumora Therapeutics, Boehringer Ingelheim, Sage, Biogen, Mitsubishi Tanabe, Purdue, Pfizer, Otsuka, Takeda, Neurocrine, Neurawell, Sunovion, Bausch Health, Axsome, Novo Nordisk, Kris, Sanofi, Eisai, Intra-Cellular, NewBridge Pharmaceuticals, Viatris, Abbvie and Atai Life Sciences. Dr. S. Roger McIntyre is a CEO of Braxia Scientific Corp.

Axelson, D., Goldstein, B., Goldstein, T., Monk, K., Yu, H., Hickey, M. B., Sakolsky, D., Diler, R., Hafeman, D., Merranko, J., Iyengar, S., Brent, D., Kupfer, D., & Birmaher, B. (2015). Diagnostic Precursors to Bipolar Disorder in Offspring of Parents With Bipolar Disorder: A Longitudinal Study. American Journal of Psychiatry, 172(7), 638–646. https://doi.org/10.1176/appi.ajp.2014.14010035
Birmaher, B., Merranko, J., Hafeman, D., Goldstein, B. I., Diler, R., Levenson, J. C., Monk, K., Iyengar, S., Hickey, M. B., Sakolsky, D., Axelson, D., & Goldstein, T. (2021). A Longitudinal Study of Psychiatric Disorders in Offspring of Parents With Bipolar Disorder From Preschool to Adolescence. Journal of the American Academy of Child & Adolescent Psychiatry, 60(11), 1419–1429. https://doi.org/10.1016/j.jaac.2021.02.023
Bora, E., Can, G., Zorlu, N., Ulas, G., Inal, N., & Ozerdem, A. (2021). Structural dysconnectivity in offspring of individuals with bipolar disorder: The effect of co-existing clinical-high-risk for bipolar disorder. Journal of Affective Disorders, 281, 109–116. https://doi.org/10.1016/j.jad.2020.11.122
Bullmore, E., & Sporns, O. (2012). The economy of brain network organization. Nature Reviews Neuroscience, 13(5), 336–349. https://doi.org/10.1038/nrn3214
Carlén, M. (2017). What constitutes the prefrontal cortex? Science, 358(6362), 478–482. https://doi.org/10.1126/science.aan8868
Carvalho, A. F., Firth, J., & Vieta, E. (2020). Bipolar Disorder. New England Journal of Medicine, 383(1), 58–66. https://doi.org/10.1056/NEJMra1906193
Cattarinussi, G., Kubera, K. M., Hirjak, D., Wolf, R. C., & Sambataro, F. (2022). Neural Correlates of the Risk for Schizophrenia and Bipolar Disorder: A Meta-analysis of Structural and Functional Neuroimaging Studies. Biological Psychiatry, 92(5), 375–384. https://doi.org/10.1016/j.biopsych.2022.02.960
Chen, G., Wang, J., Gong, J., Qi, Z., Fu, S., Tang, G., Chen, P., Huang, L., & Wang, Y. (2022). Functional and structural brain differences in bipolar disorder: A multimodal meta-analysis of neuroimaging studies. Psychological Medicine, 52(14), 2861–2873. https://doi.org/10.1017/S0033291722002392
Cruces, R. R., Royer, J., Herholz, P., Larivière, S., Vos De Wael, R., Paquola, C., Benkarim, O., Park, B., Degré-Pelletier, J., Nelson, M. C., DeKraker, J., Leppert, I. R., Tardif, C., Poline, J.-B., Concha, L., & Bernhardt, B. C. (2022). Micapipe: A pipeline for multimodal neuroimaging and connectome analysis. NeuroImage, 263, 119612. https://doi.org/10.1016/j.neuroimage.2022.119612
De Zwarte, S. M. C., Brouwer, R. M., Agartz, I., Alda, M., Aleman, A., Alpert, K. I., Bearden, C. E., Bertolino, A., Bois, C., Bonvino, A., Bramon, E., Buimer, E. E. L., Cahn, W., Cannon, D. M., Cannon, T. D., Caseras, X., Castro-Fornieles, J., Chen, Q., Chung, Y., … Van Haren, N. E. M. (2019). The Association Between Familial Risk and Brain Abnormalities Is Disease Specific: An ENIGMA-Relatives Study of Schizophrenia and Bipolar Disorder. Biological Psychiatry, 86(7), 545–556. https://doi.org/10.1016/j.biopsych.2019.03.985
Duffy, A., Goodday, S., Keown-Stoneman, C., & Grof, P. (2019). The Emergent Course of Bipolar Disorder: Observations Over Two Decades From the Canadian High-Risk Offspring Cohort. American Journal of Psychiatry, 176(9), 720–729. https://doi.org/10.1176/appi.ajp.2018.18040461
Eickhoff, S. B., Yeo, B. T. T., & Genon, S. (2018). Imaging-based parcellations of the human brain. Nature Reviews Neuroscience, 19(11), 672–686. https://doi.org/10.1038/s41583-018-0071-7
Endicott, J., Spitzer, R. L., Fleiss, J. L., & Cohen, J. (1976). The Global Assessment Scale: A procedure for measuring overall severity of psychiatric disturbance. Archives of General Psychiatry, 33(6), 766–771.
Fischl, B. (2012). FreeSurfer. NeuroImage, 62(2), 774–781. https://doi.org/10.1016/j.neuroimage.2012.01.021
Gordovez, F. J. A., & McMahon, F. J. (2020). The genetics of bipolar disorder. Molecular Psychiatry, 25(3), 544–559. https://doi.org/10.1038/s41380-019-0634-7
Griffanti, L., Salimi-Khorshidi, G., Beckmann, C. F., Auerbach, E. J., Douaud, G., Sexton, C. E., Zsoldos, E., Ebmeier, K. P., Filippini, N., Mackay, C. E., Moeller, S., Xu, J., Yacoub, E., Baselli, G., Ugurbil, K., Miller, K. L., & Smith, S. M. (2014). ICA-based artefact removal and accelerated fMRI acquisition for improved resting state network imaging. NeuroImage, 95, 232–247. https://doi.org/10.1016/j.neuroimage.2014.03.034
Hamilton, M. (1959). THE ASSESSMENT OF ANXIETY STATES BY RATING. British Journal of Medical Psychology, 32(1), 50–55. https://doi.org/10.1111/j.2044-8341.1959.tb00467.x
Hamilton, M. (1986). The Hamilton Rating Scale for Depression. In N. Sartorius & T. A. Ban (Eds.), Assessment of Depression (pp. 143–152). Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-70486-4_14
Hansen, J. Y., Shafiei, G., Vogel, J. W., Smart, K., Bearden, C. E., Hoogman, M., Franke, B., Van Rooij, D., Buitelaar, J., McDonald, C. R., Sisodiya, S. M., Schmaal, L., Veltman, D. J., Van Den Heuvel, O. A., Stein, D. J., Van Erp, T. G. M., Ching, C. R. K., Andreassen, O. A., Hajek, T., … Misic, B. (2022). Local molecular and global connectomic contributions to cross-disorder cortical abnormalities. Nature Communications, 13(1), 4682. https://doi.org/10.1038/s41467-022-32420-y
Hibar, D. P., Westlye, L. T., Doan, N. T., Jahanshad, N., Cheung, J. W., Ching, C. R. K., Versace, A., Bilderbeck, A. C., Uhlmann, A., Mwangi, B., Krämer, B., Overs, B., Hartberg, C. B., Abé, C., Dima, D., Grotegerd, D., Sprooten, E., Bøen, E., Jimenez, E., … Andreassen, O. A. (2018). Cortical abnormalities in bipolar disorder: An MRI analysis of 6503 individuals from the ENIGMA Bipolar Disorder Working Group. Molecular Psychiatry, 23(4), 932–942. https://doi.org/10.1038/mp.2017.73
Kaufman, J., Birmaher, B., Brent, D., Rao, U., Flynn, C., Moreci, P., Williamson, D., & Ryan, N. (1997). Schedule for Affective Disorders and Schizophrenia for School-Age Children-Present and Lifetime Version (K-SADS-PL): Initial Reliability and Validity Data. Journal of the American Academy of Child & Adolescent Psychiatry, 36(7), 980–988. https://doi.org/10.1097/00004583-199707000-00021
Lin, K., Shao, R., Geng, X., Chen, K., Lu, R., Gao, Y., Bi, Y., Lu, W., Guan, L., Kong, J., Xu, G., & So, K.-F. (2018). Illness, at-risk and resilience neural markers of early-stage bipolar disorder. Journal of Affective Disorders, 238, 16–23. https://doi.org/10.1016/j.jad.2018.05.017
Lin, K., Shao, R., Lu, R., Chen, K., Lu, W., Li, T., Kong, J., So, K.-F., & Xu, G. (2018). Resting-state fMRI signals in offspring of parents with bipolar disorder at the high-risk and ultra-high-risk stages and their relations with cognitive function. Journal of Psychiatric Research, 98, 99–106. https://doi.org/10.1016/j.jpsychires.2018.01.001
Lin, K., Xu, G., Wong, N. M. L., Wu, H., Li, T., Lu, W., Chen, K., Chen, X., Lai, B., Zhong, L., So, K., & Lee, T. M. C. (2015). A Multi-Dimensional and Integrative Approach to Examining the High-Risk and Ultra-High-Risk Stages of Bipolar Disorder. EBioMedicine, 2(8), 919–928. https://doi.org/10.1016/j.ebiom.2015.06.027
Lu, W., Wu, J., Shao, R., Zou, W., Zhang, R., Li, X., Kong, J., Zheng, D., Tian, X., Gao, Y., Yau, S., Goldstein, B., Xu, G., So, K.-F., Wang, J., & Lin, K. (2024). Genetic and symptomatic risks associated with longitudinal brain morphometry in bipolar disorder. Nature Mental Health. https://doi.org/10.1038/s44220-023-00194-x
Luna, L. P., Radua, J., Fortea, L., Sugranyes, G., Fortea, A., Fusar-Poli, P., Smith, L., Firth, J., Shin, J. I., Brunoni, A. R., Husain, M. I., Husian, M. O., Sair, H. I., Mendes, W. O., Uchoa, L. R. A., Berk, M., Maes, M., Daskalakis, Z. J., Frangou, S., … Carvalho, A. F. (2022). A systematic review and meta-analysis of structural and functional brain alterations in individuals with genetic and clinical high-risk for psychosis and bipolar disorder. Progress in Neuro-Psychopharmacology and Biological Psychiatry, 117, 110540. https://doi.org/10.1016/j.pnpbp.2022.110540
Manjón, J. V., Coupé, P., Martí-Bonmatí, L., Collins, D. L., & Robles, M. (2010). Adaptive non-local means denoising of MR images with spatially varying noise levels: Spatially Adaptive Nonlocal Denoising. Journal of Magnetic Resonance Imaging, 31(1), 192–203. https://doi.org/10.1002/jmri.22003
McIntyre, R. S., Berk, M., Brietzke, E., Goldstein, B. I., López-Jaramillo, C., Kessing, L. V., Malhi, G. S., Nierenberg, A. A., Rosenblat, J. D., Majeed, A., Vieta, E., Vinberg, M., Young, A. H., & Mansur, R. B. (2020). Bipolar disorders. The Lancet, 396(10265), 1841–1856. https://doi.org/10.1016/S0140-6736(20)31544-0
Overall, J. E., & Gorham, D. R. (n.d.). THE BRIEF PSYCHIATRIC RATING SCALE.
Pastrnak, M., Simkova, E., & Novak, T. (2021). Insula activity in resting-state differentiates bipolar from unipolar depression: A systematic review and meta-analysis. Scientific Reports, 11(1), 16930. https://doi.org/10.1038/s41598-021-96319-2
Roberts, G., Lenroot, R., Frankland, A., Yeung, P. K., Gale, N., Wright, A., Lau, P., Levy, F., Wen, W., & Mitchell, P. B. (2016). Abnormalities in left inferior frontal gyral thickness and parahippocampal gyral volume in young people at high genetic risk for bipolar disorder. Psychological Medicine, 46(10), 2083–2096. https://doi.org/10.1017/S0033291716000507
Roberts, G., Lenroot, R., Overs, B., Fullerton, J., Leung, V., Ridgway, K., Stuart, A., Frankland, A., Levy, F., Hadzi-Pavlovic, D., Breakspear, M., & Mitchell, P. B. (2022). Accelerated cortical thinning and volume reduction over time in young people at high genetic risk for bipolar disorder. Psychological Medicine, 52(7), 1344–1355. https://doi.org/10.1017/S0033291720003153
Roberts, G., Perry, A., Lord, A., Frankland, A., Leung, V., Holmes-Preston, E., Levy, F., Lenroot, R. K., Mitchell, P. B., & Breakspear, M. (2018). Structural dysconnectivity of key cognitive and emotional hubs in young people at high genetic risk for bipolar disorder. Molecular Psychiatry, 23(2), 413–421. https://doi.org/10.1038/mp.2016.216
Roberts, G., Perry, A., Ridgway, K., Leung, V., Campbell, M., Lenroot, R., Mitchell, P. B., & Breakspear, M. (2022). Longitudinal Changes in Structural Connectivity in Young People at High Genetic Risk for Bipolar Disorder. American Journal of Psychiatry, 179(5), 350–361. https://doi.org/10.1176/appi.ajp.21010047
Rolls, E. T., Deco, G., Huang, C.-C., & Feng, J. (2022). The human orbitofrontal cortex, vmPFC, and anterior cingulate cortex effective connectome: Emotion, memory, and action. Cerebral Cortex, 33(2), 330–356. https://doi.org/10.1093/cercor/bhac070
Rubinov, M., & Sporns, O. (2010). Complex network measures of brain connectivity: Uses and interpretations. NeuroImage, 52(3), 1059–1069. https://doi.org/10.1016/j.neuroimage.2009.10.003
Schaefer, A., Kong, R., Gordon, E. M., Laumann, T. O., Zuo, X.-N., Holmes, A. J., Eickhoff, S. B., & Yeo, B. T. T. (2018). Local-Global Parcellation of the Human Cerebral Cortex from Intrinsic Functional Connectivity MRI. Cerebral Cortex, 28(9), 3095–3114. https://doi.org/10.1093/cercor/bhx179
Seidlitz, J., Nadig, A., Liu, S., Bethlehem, R. A. I., Vértes, P. E., Morgan, S. E., Váša, F., Romero-Garcia, R., Lalonde, F. M., Clasen, L. S., Blumenthal, J. D., Paquola, C., Bernhardt, B., Wagstyl, K., Polioudakis, D., De La Torre-Ubieta, L., Geschwind, D. H., Han, J. C., Lee, N. R., … Raznahan, A. (2020). Transcriptomic and cellular decoding of regional brain vulnerability to neurogenetic disorders. Nature Communications, 11(1), 3358. https://doi.org/10.1038/s41467-020-17051-5
Seidlitz, J., Váša, F., Shinn, M., Romero-Garcia, R., Whitaker, K. J., Vértes, P. E., Wagstyl, K., Kirkpatrick Reardon, P., Clasen, L., Liu, S., Messinger, A., Leopold, D. A., Fonagy, P., Dolan, R. J., Jones, P. B., Goodyer, I. M., Raznahan, A., & Bullmore, E. T. (2018). Morphometric Similarity Networks Detect Microscale Cortical Organization and Predict Inter-Individual Cognitive Variation. Neuron, 97(1), 231-247.e7. https://doi.org/10.1016/j.neuron.2017.11.039
Tay, J. K., Narasimhan, B., & Hastie, T. (2023). Elastic Net Regularization Paths for All Generalized Linear Models. Journal of Statistical Software, 106(1). https://doi.org/10.18637/jss.v106.i01
Tournier, J.-D., Smith, R., Raffelt, D., Tabbara, R., Dhollander, T., Pietsch, M., Christiaens, D., Jeurissen, B., Yeh, C.-H., & Connelly, A. (2019). MRtrix3: A fast, flexible and open software framework for medical image processing and visualisation. NeuroImage, 202, 116137. https://doi.org/10.1016/j.neuroimage.2019.116137
Tustison, N. J., Avants, B. B., Cook, P. A., Yuanjie Zheng, Egan, A., Yushkevich, P. A., & Gee, J. C. (2010). N4ITK: Improved N3 Bias Correction. IEEE Transactions on Medical Imaging, 29(6), 1310–1320. https://doi.org/10.1109/TMI.2010.2046908
Van Essen, D. C., Glasser, M. F., Dierker, D. L., Harwell, J., & Coalson, T. (2012). Parcellations and Hemispheric Asymmetries of Human Cerebral Cortex Analyzed on Surface-Based Atlases. Cerebral Cortex, 22(10), 2241–2262. https://doi.org/10.1093/cercor/bhr291
Vieta, E., Salagre, E., Grande, I., Carvalho, A. F., Fernandes, B. S., Berk, M., Birmaher, B., Tohen, M., & Suppes, T. (2018). Early Intervention in Bipolar Disorder. American Journal of Psychiatry, 175(5), 411–426. https://doi.org/10.1176/appi.ajp.2017.17090972
Whittaker, J. R., Foley, S. F., Ackling, E., Murphy, K., & Caseras, X. (2018). The Functional Connectivity Between the Nucleus Accumbens and the Ventromedial Prefrontal Cortex as an Endophenotype for Bipolar Disorder. Biological Psychiatry, 84(11), 803–809. https://doi.org/10.1016/j.biopsych.2018.07.023
Youngstrom, E. A., Gracious, B. L., Danielson, C. K., Findling, R. L., & Calabrese, J. (2003). Toward an integration of parent and clinician report on the Young Mania Rating Scale. Journal of Affective Disorders, 77(2), 179–190. https://doi.org/10.1016/S0165-0327(02)00108-8

Dr. Roger S. McIntyre has received research grant support from CIHR/GACD/National Natural Science Foundationof China (NSFC) and the Milken Institute; speaker/consultation fees from Lundbeck, Janssen, Alkermes, Neumora Therapeutics, Boehringer Ingelheim, Sage, Biogen, Mitsubishi Tanabe, Purdue, Pfizer, Otsuka, Takeda, Neurocrine, Neurawell, Sunovion, Bausch Health, Axsome, Novo Nordisk, Kris, Sanofi, Eisai, Intra-Cellular, NewBridge Pharmaceuticals, Viatris, Abbvie and Atai Life Sciences. Dr. S. Roger McIntyre is a CEO of Braxia Scientific Corp.

SupplementaryMaterials.pdf
supplemental material

Download PDF

Version 1

posted

You are reading this latest preprint version

Enhancing Early Diagnosis of Bipolar Disorder in Adolescents through Multimodal Neuroimaging

Status:

Version 1

Abstract

Background

Objective

Methods

Results

Conclusion

Figures

Introduction

Methods

Study design and participant selection

Participant evaluation

Criteria for participant inclusion and exclusion

Multimodal MRI acquisition parameters

Multimodal MRI data analysis

Morphometric measures calculation

Functional connectivity (FC) and structural connectivity (SC) Calculation

Graph measures of FC and SC

GLMNET multinomial classification

Comprehensive GLMNET model

Clinical diagnosis GLMNET model

MRI-based GLMNET model

Cross validation of the GLMNET models

Structural equation modelling

Independent validation using external data

Results

Demographics

Classification analysis with GLMNET

Comprehensive model

Clinical diagnosis model

MRI-based Model

Reliability of the three models

Discriminative behavioural and MRI features in comprehensive model

Relationships among Clinical Diagnosis, Parental BD History, Psychiatric Symptoms, Brain Health and Global Function

External validation results

Discussion

Enhanced diagnosis through MRI metrics integration in a real-world setting

Unravelling the predictive potential of behavioral attributes

Discriminative capability of MRI metrics

Interplay of Psychiatric Symptoms, Brain Health, Parental BD History, Clinical Diagnosis and Global Function

Generalizability of the comprehensive model

Limitations

Conclusion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1