Search

Introduction About 60% of bipolar disorder (BD) cases are initially misdiagnosed as major depressive disorder (MDD), preventing BD patients from receiving appropriate treatment. An urgency exists to identify reliable biomarkers for improving differential diagnosis (DD). Machine learning methods may help translate current knowledge on biomarkers of mood disorders into clinical practice by providing individual-level classification. No study so far has combined biological data with clinical data to provide a multifactorial predictive model for DD. Objectives Define a predictive algorithm for BD and MDD by integrating structural neuroimaging and inflammatory data with neuropsychological measures (NM). Two different algorithms were compared: multiple kernel learning (MKL) and elastic net regularized logistic regression (EN). Methods In a sample of 141 subjects (70 MDD; 71 BD), two different models were implemented for each algorithm: 1) structural neuroimaging measures only (i.e. voxel-based morphometry (VBM), white matter fractional anisotropy (FA), and mean diffusivity (MD)); 2) VBM, FA, and MD combined with NM. In a subsample of 71 subjects (36 BD; 38 MDD), two similar models were implemented: 1) VBM, FA, and, MD combined with only NM; 2) VBM, FA, and MD combined with NM and peripheral inflammatory markers. Finally, the best model was selected for comparison with healthy controls (HC). Results Overall, the EN model based on all the modalities achieved the highest accuracy (AUC = 90.2%), outperforming MKL (AUC=85%). EN correctly classified BD and MDD with a diagnostic accuracy of 78.3%, sensitivity of 75%, and specificity of 81.6%. The most significant predictors of BD (variable inclusion probability (VIP) > 80%) were the parahippocampal cingulate, interleukin 9, chemokine CCL5, posterior thalamic radiation, and internal capsule, whereas MDD was best predicted by chemokine CCL23, the anterior cerebellum, and the sagittal stratum. In contrast, NM did not help to differentiate between MDD and BD. However, they help to distinguish patients from HC. Psychomotor coordination and speed of information processing discriminated between MDD and HC (VIP>90%), whereas fluency, working memory, and executive functions differentiated between BD and HC (VIP>80%). Conclusions In summary, BD was predicted by a strong proinflammatory profile, whereas MDD was identified by structural neuroimaging data. A multimodal approach offers additional instruments to improve personalized diagnosis in clinical practice and enhance the ability to make DD. Disclosure of Interest None Declared

Introduction Depression is the predominant mood alteration in bipolar disorder (BD), leading to overlapping symptomatology with major depressive disorder (MDD). Consequently, in clinical assessment, almost 60% of BD patients are misdiagnosed as affected by MDD. This calls for the creation of a framework for the differentiation of BD and MDD patients based on reliable biomarkers. Since machine learning (ML) enables to make predictions at the single-subject level, it appears to be particularly suitable for this task. Objectives We implemented a ML pipeline for the differentiation between depressed BD and MDD patients based on structural neuroimaging features. Methods Diffusion tensor imaging (DTI) and T1-weighted magnetic resonance imaging (MRI) data were acquired for 282 depressed BD (n=180) and MDD (n=102) patients. Axial (AD), radial (RD), mean (MD) diffusivity, and fractional anisotropy (FA) maps were extracted from DTI images, and voxel-based morphometry (VBM) measures were obtained from T1-weighted images. Each feature was entered separately into a 5-fold nested cross-validated ML pipeline differentiating between BD and MDD patients, comprising: confound regression for nuisance variables removal (i.e., age and sex), feature standardization, principal component analysis, and an elastic-net penalized regression. The models underwent 5000 random permutations as a test for significance, and the McNemar’s test was used to assess whether there was any significant difference between the models (significance threshold was set to p

Introduction Every year at least one million people die by suicide, with major depressive disorder (MDD) being one of the major causes of suicide deaths. Current suicide risk assessments rely on subjective information, are time consuming, low predictive, and poorly reliable. Thus, finding objective biomarkers of suicidality is crucial to move clinical practice towards a precision psychiatry framework, enhancing suicide risk detection and prevention for MDD. Objectives The present study aimed at applying machine learning (ML) algorithms on both grey matter and white-matter voxel-wise data to discriminate MDD suicide attempters (SA) from non-attempters (nSA). Methods 91 currently depressed MDD patients (24 SA, 67 nSA) underwent a structural MRI session. T1-weighted images and diffusion tensor imaging scans were respectively pre-processed using Computational Atlas Toolbox 12 (CAT12) and spatial tract-based statistics (TBSS) on FSL, to obtain both voxel-based morphometry (VBM) and fractional anisotropy (FA) measures. Three classification models were built, entering whole-brain VBM and FA maps alone into a Support Vector Machine (SVM) and combining both modalities into a Multiple Kernel Learning (MKL) algorithm. All models were trained through a 5-fold nested cross-validation with subsampling to calculate reliable estimates of balanced accuracy, specificity, sensitivity, and area under the receiver operator curve (AUC). Results Models’ performances are summarized in Table 1.Table 1. Models’ performances. Input features Algorithm Specificity Sensitivity Balanced accuracy AUC VBM SVM 55.00% 50.00% 52.50% 0.55 FA SVM 72.00% 54.00% 63.00% 0.62 VBM and FA MKL 68.00% 54.00% 61.00% 0.58 Abbreviations: AUC, area under the receiver operator curve; FA, fractional anisotropy; VBM, voxel-based morphometry. Conclusions Overall, although overcoming the random classification accuracy (i.e., 50%), performances of all models classifying SA and nSA MDD patients were moderate, possibly due to the imbalanced numerosity of classes, with SVM on FA reaching the highest accuracy. Thus, future studies may enlarge the sample and add different features (e.g., functional neuroimaging data) to develop an objective and reliable predictive model to assess and hence prevent suicide risk among MDD patients. Disclosure of Interest None Declared

Introduction One of the main obstacles in providing effective treatments for major depressive disorder (MDD) is clinical heterogeneity, whose neurobiological correlates are not clearly defined. A biologically meaningful stratification of depressed patients is needed to promote tailored diagnostic procedures. Objectives Using structural data, we performed an unsupervised clustering to define clinically meaningful clusters of depressed patients. Methods T1-weighted and diffusion tensor images were obtained from 102 MDD patients. In 64 patients, clinical symptoms, number of stressful life events, severity and exposure to adverse childhood experiences were evaluated using the Beck Depression Inventory (BDI), Schedule of Recent Experiences (SRE), Risky Family Questionnaire (RFQ), and Childhood Trauma Questionnaire (CTQ). Clustering analyses were performed with extracted tract-based fractional anisotropy (TBSS, FSL), cortical thickness, surface area, and regional measures of grey matter volumes (CAT12). Gaussian mixture model was implemented for clustering, considering Support Vector Machine (SVM) as classifier. A 10x2 repeated cross-validation with grid search was performed for hyperparameters tuning and clusters’ stability. The optimal number of clusters was determined by normalized stability, Akaike and Bayesian information criterion. Analyses were adjusted for total intracranial volume, age, and sex. The clinical relevance of the identified clusters was assessed through MANOVA, considering domains of clinical scales as dependent variables and clusters’ labels as fixed factors. Discriminant analysis was subsequently performed to assess the discriminative power of these variables. Results Cross-validated clustering approach identified 2 highly stable clusters (normalized stability=0.316, AIC=-80292.48, BIC=351329.16). MANOVA showed a significant between-clusters difference in clinical scales scores (p=0.038). Discriminant analysis distinguished the two clusters with an accuracy of 78.1%, with BDI behavioural and CTQ minimisation/denial domains showing the highest discriminant values (0.325 and 0.313). Conclusions Our results defined two biologically informed clusters of MDD patients associated with childhood trauma and specific clinical profiles, which may assist in targeting effective interventions and treatments. Disclosure of Interest None Declared

Introduction Bipolar patients (BP) frequently have cognitive deficits, that impact on prognosis and quality of life. Finding biomarkers for this condition is essential to improve patients’ healthcare. Given the association between cognitive dysfunctions and structural brain abnormalities, we used a machine learning approach to identify patients with cognitive deficits. Objectives The aim of this study was to assess if structural neuroimaging data could identify patients with cognitive impairments in several domains using a machine learning framework. Methods Diffusion tensor imaging and T1-weighted images of 150 BP were acquired and both grey matter voxel-based morphometry (VBM) and tract-based white matter fractional anisotropy (FA) measures were extracted. Support vector machine (SVM) models were trained through a 10-fold nested cross-validation with subsampling. VBM and FA maps were entered separately and in combination as input features to discriminate BP with and without deficits in six cognitive domains, assessed through the Brief Assessment of Cognition in Schizophrenia. Results The best classification performance for each cognitive domain is illustrated in Table 1. FA was the most relevant neuroimaging modality for the prediction of verbal memory, verbal fluency, and executive functions deficits, whereas VBM was more predictive for working memory and motor speed domains.Table 1. Performance of best classification models. Input feature Balance Accuracy (%) Specificity (%) Sensitivity (%) Verbal Memory FA 60.17 51.31 43 Verbal Fluency FA 57.67 62 53.33 Executive functions FA 60 63.33 56.67 Working Memory VBM 56.50 56 57 Motor speed VBM 53.50 47.67 59.33 Attention and processing speed VBM + FA 58.33 49.17 67.5 Conclusions Overall, the tested SVM models showed a good predictive performance. Although only partially, our results suggest that different structural neuroimaging data can predict cognitive deficits in BP with accuracy higher than chance level. Unexpectedly, only for the attention and processing speed domain the best model was obtained combining the structural features. Future research may promote data fusion methods to develop better predictive models. Disclosure of Interest None Declared