Development of a Software Tool, Using Artificial Intelligence, That Integrates Clinical, Biological, Genetic and Imaging Data to Predict Diagnosis and Outcome of Depressed Patients in Order to Enhance Prognosis and Limiting Healthcare Costs.

Overview

Based on robust evidence from literature, the investigators hypothesize the presence of disease-specific neurobiological underpinnings for bipolar and unipolar disorder, which may serve as biomarkers for differential diagnosis. However, the group comparison approaches adopted in psychiatric research fail to translate the emerging knowledge to the diagnostic routine. How can physicians predict differential diagnosis and treatment response by using cutting-edge knowledge obtained in the last decade? How can such extensive knowledge be useful and applicable in clinical practice? With this project, the investigators propose a solution to these challenges by developing a software tool that integrates the available clinical, biological, genetic and imaging data to predict diagnosis and outcome of new individual patients. The decision support platform will employ artificial intelligence, specifically machine learning techniques, which will be "trained" through data in order to predict the category to which a new observation belongs to. By doing this, existing and newly acquired multimodal datasets of bipolar and unipolar patients will be translated into predictors for personalized patient diagnosis and prognosis. The project can have a great impact on psychiatric community and healthcare system. Identifying predictive biomarkers for UD and BD will provide an essential tool in the early stages of the disease, ensuring accurate diagnosis, enhancing prognosis and limiting health care costs. The investigators will recruit 80 bipolar patients, 80 unipolar patients and 80 healthy controls for the MRI study. Clinical, genetic and inflammation data will be acquired from all subjects. The following data will be obtained: age, gender, number of episodes, recurrence, age of illness onset, lifetime psychosis, BD or UD familiarity, tempted suicide, medication, scores at HDRS, Beck Depression Inventory and BACS battery. MRI will be performed on 3.0 Tesla scanners. MRI acquisitions will include SE EPI DTI, T1-weighted 3D MPRAGE and fMRI sequences during resting state and a face matching paradigm, which previously allowed defining the connectivity in mood disorder. Blood samples samples will be collected and plasma will be extracted and stored at -80. Pro- and anti-inflammatory cytokines will be measured using the Bioplex human cytokines 27-plex. Genetic variants associated considered for differential diagnosis will be evaluated using the Infinium PsychArray-24 BeadChip. This cost-effective, high-density microarray was developed in collaboration with the Psychiatric Genomics Consortium for large-scale genetic studies focused on psychiatric predisposition and risk. The relevance of the single clinical, genetic, molecular and image-based features as bipolar and unipolar disorder signatures will be evaluated by considered the cutting-edge literature and estimated on a independent already existing dataset (30 subjects per group). General Linear Model analyses followed by two sided t-tests will be used to identify whether each parameter significantly differs among groups, while removing the contribution of age, gender, length of illness and other confounding factors. A multiple kernel learning (MKL) algorithm will project the multisource features to a higher-dimensional space where the three subject groups will be maximally separated. The selected features will be used both separately and in combination. The nuisance effects of age, gender, length of illness and MRI system will be corrected during the training phase of the algorithm. The MKL classifier will be tested using a k-fold nested cross-validation strategy with hyperparameter tuning. The training dataset is already made available and includes about 550 subjects. The software architecture will be designed in Matlab environment by integrating quantitative imaging methods, machine learning algorithm and statistical analyses as separate modules in a user-friendly interface, which will facilitate the sharing of computational resources in the clinical community.