Research Projects
Research Projects
Embedded Files
BEST: Biomarker Evaluation Supporting Clinical Translation in schizophrenia
BEST: Biomarker Evaluation Supporting Clinical Translation in schizophrenia
Role: Partner
Goal: To mechanistically refine existing descriptive biomarkers for individualized outcome prediction in patients with clinically defined high-risk states for psychosis, and patients with established psychosis spectrum disorders.
Motivation: Psychotic disorders are highly heterogeneous, posing significant challenges for clinicians in accurate diagnosis, prognosis, and treatment selection for individual patients. To enhance clinical and functional outcomes while mitigating the substantial healthcare burden, biomarker signatures that capture the cross-sectional and longitudinal complexity of these disorders are essential.
Partners: Natural and Medical Sciences Institute at the University of Tübingen, Central Institute of Mental Health Mannheim, University hospital for Psychiatry and Psychotherapy Tübingen
Duration: 11/2021 - 12/2024
Funding: German Federal Ministry of Education and Research with 2 Million Euros.
Lab contact: Lisa Hahn
CARE
CARE
Role: Partner/Co-Leadership
Goal: The project aims at the implementation of a computer-assisted, AI-based algorithms for an efficient treatment optimization through the use of diagnostic risk profiles and risk-adapted, risk-stratified therapy for patients in a clinical high-risk (CHR) state for psychosis. This new form of care targets the prevention of psychosis manifestation and the improvement of social and professional functional level.
Motivation: Psychosis is one of the most cost-intense and impairing psychiatric disease. Usually, the slowly progressing development is in the clinical high-risk state is overlooked and not treated appropriately.
Partners: Heinrich-Heine-Universität Düsseldorf (Coordinator), Universitätsklinikum Würzburg Zentrum für psychische Gesundheit, Bergische Universität Wuppertal, Uniklinikum Leipzig, Universitätsklinikum Augsburg, Vivantes Klinikum Am Urban Berlin, Universitätsklinikum Carl Gustav Carus an der Technischen Universität Dresden, Universitätsklinikum Hamburg-Eppendorf (UKE), Zentrum für Psychiatrie Reichenau, Universitätsklinik für Psychiatrie und Psychotherapie Tübingen, Rheinhessen-Fachklinik Alzey, Zentralinstitut für Seelische Gesundheit (ZI) Mannheim, LWL-Universitätsklinikum der Ruhr-Universität Bochum, Uniklinik RWTH Aachen, Universitätsklinikum Magdeburg A.ö.R., Charité Universitätsmedizin Berlin, Zentrum für Integrative Psychiatrie (ZIP) Lübeck, Zentrum für Integrative Psychiatrie Kiel, Universitätsklinikum Münster, LVR-Klinik für Kinder- und Jugendpsychiatrie Bonn, Universitätsklinikum Bonn, Uniklinik Köln, LMU Klinikum
Duration: 2022 - Ongoing
Funding: The project is funded by the Innovationsfond of the German Gesundheitsausschuss with 9.5 Million Euros.
Website: CARE – Computer-assistierte Risiko-Evaluation (care-network.eu)
Lab contact: Alessa Grund
COMMITMENT
COMMITMENT
Role: Partner
Goal: The goal of COMMITMENT is to identify shared neurobiological signatures that link psychotic disorders with common somatic comorbidities. By uncovering the biological pathways that contribute both to psychiatric illnesses and to medical conditions like metabolic, cardiovascular, and neurodegenerative diseases, COMMITMENT aims to:
Stratify patients based on shared biological risk profiles that cut across psychiatric and somatic domains.
Identify predictive biomarkers that inform about disease trajectory, treatment response, and risk of comorbidities early in the course of illness.
Develop biological tools for personalized treatment strategies that address both psychiatric symptoms and associated somatic risks.
Translate precision medicine approaches from fields like oncology to psychiatry, moving toward a biology-informed, individualized care model.
Motivation: Psychotic illnesses such as schizophrenia and bipolar disorder are among the most severe and complex mental health disorders. They pose a substantial burden on patients, families, and healthcare systems. Today, diagnosis and treatment are largely based on clinical symptoms, with little consideration of the biological mechanisms underlying these disorders. Despite significant differences in disease progression, treatment responses, and the frequent presence of serious comorbidities like type 2 diabetes, cardiovascular diseases, and neurodegenerative disorders, treatment remains largely uniform and generalized. There is an urgent need for biological tools that can help identify subgroups of patients, predict individual disease courses, and assess risks for comorbidities. Such tools would allow for personalized interventions, improve clinical outcomes, and reduce long-term health risks.
Partners: CMIH Mannheim, Bonn, Fraunhofer Institute for Algorithms and Scientific Computing, Heidelberg, Oslo
Duration: 2019 - Ongoing
Funding: German Federal Ministry of Education and Research (BMBF)
Lab contact: Lisa Hahn
DZPG Munich-Augsburg site, Data and Knowledge Management Infrastructure (DKM-INF)
DZPG Munich-Augsburg site, Data and Knowledge Management Infrastructure (DKM-INF)
Role: Lead
Goal: Establish an efficient and cost-effective data management system for the German Center for Mental Health (DZPG), integrating centralized and federated approaches to securely store and analyze multimodal research and clinical data in compliance with privacy regulations.
Motivation: In the DZPG, data is to be collected both within studies (research data) and in clinical care (real clinical data). These two data sources should not be considered separately in terms of content, as research data is usually generated in a circumscribed hypothesis-driven context and therefore cannot fully represent the heterogeneity/complexity of psychiatric disorders without integration with real clinical data. In addition, the accessibility of real data is essential for the clinical validation of findings from research projects. However, the management of these two types of data requires different technical solutions, as research data should primarily be managed centrally, while real clinical data must be stored via a federated infrastructure due to the legal framework. DKM-INF follows the principle that centralized analyses benefit from larger, aggregated datasets, while federated analyses allow for collaborative research respecting variable data-sharing permissions in the DZPG.
Partners: German Center for Neurodegenerative Diseases (DZNE) and German Cancer Research Center (DKFZ/DKTK).
Duration: 04/2023 - Ongoing
Funding: German Federal Ministry of Education and Research (BMBF)
Website: https://www.dzpg.org/
Lab contact: Renata De Souza Falguera
DZPG Munich-Augsburg site, MUC3 Retrospective Subproject
DZPG Munich-Augsburg site, MUC3 Retrospective Subproject
Role: Lead
Goal: To better understand the evolution of brain signatures during aging and their relationship to the transition to psychosis in childhood and adolescence stages.
Motivation: In this ongoing investigation, we apply Machine and Deep Learning models to predict psychosis transition in patients with Clinical High-Risk (CHR) syndromes using structural MRI (sMRI) data. Models are compared, trained, and evaluated in a cross-corpus strategy. We analyze the effects of age, particularly stages of childhood and early adolescence (12- to 20-year-olds), and the severity of the Schizophrenia Proneness Instrument Items (SPI-A), and their relation to the predictions.
Duration: 11/2023 - Ongoing
Funding: German Federal Ministry of Education and Research (BMBF)
Website: https://www.dzpg.org/
Lab contact: Esther Rituerto-González
Novel AI-based biomarkers for retinal gene therapy outcomes by integrating functional neuroimaging and retinal imaging.
Novel AI-based biomarkers for retinal gene therapy outcomes by integrating functional neuroimaging and retinal imaging.
Subproject 8 of the FOR5621 Research group OCU-GT: Addressing the Unmet Needs in Ocular Gene Therapy
Subproject 8 of the FOR5621 Research group OCU-GT: Addressing the Unmet Needs in Ocular Gene Therapy
Role: Subproject Lead
Goal: Create multimodal outcome measures to objectively assess visual function. Form prediction models for treatment success in patients with inherited retinal diseases. Improve knowledge about brainplasticity and functional changes in patients with progressing inherited retinal diseases with a focus on visual and auditory domains and attentional resources.
Motivation: Inherited retinal diseases (IRDs) affect about 1:3000 patients in europe, leading to progressive loss of visual function - until recently without any options for treatment. IRDs are highly heterogeneous with about 300 genes being known to be linked to IRDs. Recent progress in gene therapies leading to the development of Voretigene Neparvovec (Luxturna) to treat RPE65-related retinis pigmentosa has given hope to develop therapies to treat IRDs based on different genes as well. However, clinical trials are faced with the problem that outcome measures for treatment success are insufficient, leading to failures to meet endpoints in stage 2 and stage 3 clinical trials. This reveals an unadressed need for better understanding and objective measurement of visual function not only on an opthalmologic but also neural level.
Research Questions: This project attempts to address this need for objective outcome measures for visual function by combining established measures of Opthalmology with psychophysical and neuroscientific measures to create multimodal AI-driven tools to measure visual function and to predict the progression of change in visual function to identify biological markers that predict treatment outcome. It furthermore strives to use the data generated in the project to further knowledge about neuronal plasticity by investigating changes in networks of visual and auditory perception and attention in relation to changes in visual function due to the progress and treatment of IRDs.
Partners: Dr. Benedikt Schworm and Dr. Tobias Stückler from the Department of Opthalmology plan and conduct the opthalmologic measurements. Prof. Thomas Geyer at the chair of general psychology 1 is cosupervisor for Daniel Weinert for the psychophysical investigations. Dr. Daniel Keeser at the NICUM supports the MRI measurements and analyses. Prof. Christian Windischberger from the University of Vienna assists in the setup for population receptive field mapping as a collaboration. Cooperations with other subprojects of the FOR5621 group are in planning.
Duration: 10/2024 - 10/2028
Funding: Funded through a DFG research group
Lab contact: Daniel Weinert
ProNET
ProNET
Role: Partner / Centre
Goal: The project aims to better understand the causes and course of the earliest stages of psychosis, focusing on the clinical high risk syndrome that sometimes progresses to first episode psychosis.
Motivation: Early identification and treatment of psychosis produces better health outcomes, including reduced suicidal risk and symptom severity, improved long-term quality of life and lower economic burden. By studying a variety of biomarkers, we aim to predict the health outcomes of individuals at clinical high risk, particularly to identify individuals at high risk for transition to psychosis and individuals with symptom improvement.
Partners: Under the leadership of Yale University the ProNET network consists of 26 international sites.
Duration: 09/2020 - Ongoing
Funding: The project is funded by NIMH with 10 Million US Dollars.
Website: https://www.ampscz.org/about/
Lab contact: Christopher Eberle
Pronia
Pronia
Role: Project Lead
Goal: Develop an innovative prognostic tool based on self-learning algorithm to identify patients with psychological crisis who are at risk to develop psychosis.
Motivation: Psychoses typically commence in the most productive and critical period of life – late adolescence and early adulthood. Around 75-90% of people, who develop the full-blown illness, show early symptoms. Thus, there is an immediate need to detect these people at an early stage to be able to target them appropriately.
Partners: University of Basel, University of Cologne, University of Birmingham, University of Turku, University of Udine, University of Melbourne, Dynamic Evolution, GABO:mi, GE Global Research, GE Healthcare, University of Milan, ARTTIC, Westfaelische Wilhelms-Universitaet Muenster ,Universita degli Studi di Bari Aldo Moro, Heinrich-Heine-Universitaet Duesseldorf, Universitätsklinikum Bonn, Max-Planck-Gesellschaft zur Förderung der Wissenschaften e.V.
Duration: 10/2013 - 12/2019
Funding: The European Commission has awarded PRONIA with 6 Million Euros within the 7th Framework Programme.
Website: http://proniapredictors.eu/pronia/index.html
Lab contact: Maria Sacha
TYPIA
TYPIA
Role: Partner / Co-leadership
Goal: TYPIA aimed to uncover why some individuals with pronounced positive and negative schizotypal personality traits develop psychosis while others remain resilient. By using a population-based approach, the study sought to identify protective and risk factors for psychosis using a multimodal approach including, neurocognitive, phenotypical, structural and functional MRI, eye tracking, among other modalities that could contribute to a wider understanding of the causes of schizophrenia and other psychotic disorders.
Motivation: Psychotic disorders exist on a spectrum, with schizotypy representing a personality trait found in the general population without necessarily having clinical implications. However, while some individuals with high schizotypy remain resilient, others do develop a psychotic illness. TYPIA aimed to uncover the factors that determine this divergence, bridging the gap between personality traits and psychosis.
Partners: University of Bonn, University Hospital of LMU Munich
Duration: 2016 - 2020
Funding: Funded by the German Research Foundation (DFG).
Publications:
Lab contact: Maria F. Urguijo
VBT: VIRTUAL BRAIN TWIN FOR PERSONALISED TREATMENT OF PSYCHIATRIC DISORDERS
VBT: VIRTUAL BRAIN TWIN FOR PERSONALISED TREATMENT OF PSYCHIATRIC DISORDERS
Role: Partner
Goal: To develop a dynamic modular personalized virtual brain twin model, based upon multiscale and molecular signaling pathways related to treating schizophrenia symptoms.
Motivation: Psychotic disorders, including schizophrenia, exhibit substantial heterogeneity across clinical, neurocognitive, electrophysiological, neuroimaging, and genomic levels, complicating the development of personalized biomarkers. Given that distinct disease subtypes are associated with varying courses and treatment responses, future models must disentangle this complexity into clinically actionable dimensions. A key challenge is to advance beyond fluid biomarkers and develop approaches to model antipsychotic drug activity in individual patients, optimizing personalized treatment strategies.
Partners: EBRAINS AISBL, Universite d'Aix Marseille, PROTISVALOR MEDITERRANEE, CHARITE - Universitätsmedizin Berlin, Kungliga Tekniska Hoegskolan, Universiteit van Amsterdam, Universita Degli Studi Di Pavia, Centre National de la Recherche Scientifique, Forschungszentrum Jülich GMBH, Universidad Politecnica De Madrid, CODEMART SRL, Universidad Rey Juan Carlos, Universität Bonn, Assistance Publique - Hôpitaux De Marseille, ATHENA, European Federation of Families of People with Mental Illness, Global Alliance of Mental Illness Advocacy Networks-Europe, European Psychiatric Association.
Duration: 01/2025 - Ongoing
Funding: Horizon Europe with 10 Million Euros
Website: https://www.virtualbraintwin.eu/
Lab contact: Lisa Hahn
Individual Research Projects
Individual Research Projects
ProtAGE: Normative proteomic model for detecting abnormal aging and its multivariate links to brain patterns and clinical factors
ProtAGE: Normative proteomic model for detecting abnormal aging and its multivariate links to brain patterns and clinical factors
Aging is accompanied by complex molecular changes that can be captured through blood-based proteomic profiling. This project aims to develop and validate machine learning models that use peripheral protein signatures to predict chronological age and identify abnormal aging patterns in individuals with early-stage psychiatric conditions.
Key Highlights:
Objective: Test the hypothesis that peripheral proteomic profiles can accurately predict chronological age in healthy individuals and detect abnormal aging trajectories in psychiatric patients.
Data Acquisition: Proteomic data for N=587 participants from the PRONIA cohort were acquired using Orbitrap Liquid Chromatography Mass Spectrometry (LC-MS).
Machine Learning Models: Support vector regression models were developed using 277 serum proteins to predict age. Models were optimized within a nested leave-one-site-out cross-validation framework.
Clinical Relevance: We are investigating whether deviations from normative proteomic aging patterns are present in patients with recent onset psychosis (ROP), recent onset depression (ROD), and clinical high-risk states (CHR), to explore potential signs of early biological aging in psychiatric populations. Functional enrichment and network analyses are underway to identify key pathways and protein interactions, with the goal of linking any observed molecular deviations to clinical, cognitive, and neuroimaging outcomes.
This study establishes a robust proteomic signature for chronological age prediction and reveals early signs of biological aging in psychiatric conditions. Integrating proteomic, clinical, and neuroimaging data will advance efforts toward precision psychiatry by elucidating molecular mechanisms underpinning abnormal aging trajectories.
Interpretable Machine Learning for Personalized Psychiatry
Interpretable Machine Learning for Personalized Psychiatry
Machine learning offers powerful tools for identifying disease risk, predicting outcomes, and uncovering new subtypes in complex psychiatric disorders. With the increasing availability of large-scale clinical, neuroimaging, and biological data, predictive models are becoming more accurate - but also more complex and harder to interpret. This project line focuses on improving the clinical utility and transparency of machine learning models by developing interpretable approaches tailored to psychiatric research and practice.
Key Highlights:
Objective: Improve the transparency, clinical validity, and personalization of machine learning models in psychiatry by integrating methods to explain individual predictions and guide model complexity based on patient needs.
Interpretability Methods: We applied Shapley value analysis to a previously validated clinical risk calculator for predicting transition to psychosis (Koutsouleris et al., 2021, https://doi.org/10.1001/jamapsychiatry.2020.3604). Shapley values quantify the contribution of each feature to an individual prediction, enabling patient-level interpretability.
Subgroup Discovery: Clustering based on individual Shapley profiles revealed subgroups of patients characterized by distinct predictive patterns, suggesting potential new subtypes with differing underlying risk factors and treatment responses.
Model Selection: We trained a secondary model to identify patients who may benefit from more complex, multimodal prediction models (e.g., including neuroimaging or genetic data), thereby guiding model selection and resource allocation.
Clinical Relevance: This approach promotes efficient, personalized assessment by tailoring the complexity of predictive models to individual patients, while enhancing interpretability and supporting clinical decision-making.
This research underscores the importance of making machine learning models more transparent and clinically meaningful in psychiatric contexts. By leveraging individualized explanations and subgroup analyses, we aim to advance precision psychiatry - improving prediction accuracy, resource efficiency, and treatment personalization.
EEF-based Signatures of Schizophrenia, Depression, and Aberrant Aging
EEF-based Signatures of Schizophrenia, Depression, and Aberrant Aging
In this project, electroencephalography (EEG) and machine learning (ML) were used to investigate brain activity patterns in individuals with schizophrenia (SCZ) and major depressive disorder (MDD). The study acquired resting-state EEG recordings from a large cohort of participants, including healthy controls (HC) and patients with SCZ or MDD.
Key Highlights
Data Acquisition: Resting-state 19-channel EEG recordings were collected from a cohort of 735 participants, which included healthy controls, individuals diagnosed with SCZ and individuals with MDD.
Machine Learning Models: Support vector machine (SVM) models were developed to classify patients with SCZ or MDD from healthy controls and to predict age in healthy individuals. The age prediction model was then used to calculate the Electrophysiological Age Gap Estimation (EphysAGE) in patient groups.
Classification of Disorders: The ML models were able to distinguish between SCZ and HC with a balanced accuracy (BAC) of 72.7%, MDD from HC with 67.0% accuracy, and SCZ from MDD with 63.2% accuracy.
Key Predictive Features: A decrease in central alpha (8-11 Hz) power was the most consistent predictive feature for both SCZ and MDD.
Electrophysiological Age Gap Estimation (EphysAGE): The study introduced EphysAGE, which is the difference between an individual’s predicted brain age based on EEG data and their chronological age. Higher EphysAGE was associated with an increased likelihood of being misclassified as SCZ in both HC and MDD individuals.
Impact of Age: The study found that the classification performance of the models was affected by age, with the best performance seen in younger age groups.
The research suggests that machine learning models can extract electrophysiological signatures of MDD and SCZ from EEG recordings, which could have potential clinical applications, particularly in early recognition settings. The findings highlight the complex interplay between aging processes and the neural signatures of SCZ and MDD, demonstrating the potential of EEG-based measures for understanding severe mental illnesses.
Fig. 1. Topographical plots of overall mean of CVR in classifcation models—SCZ Model. The topographical representations with channel locations and names of CV ratio overall mean values in the frequency power domain for three classifcation models are illustrated. Frequency ranges: delta: 1–3 Hz, theta: 3–7 Hz, alpha: 8–11 Hz, beta: 12–19 Hz, low-gamma: 20–49 Hz, high-gamma: 50–70 Hz. SCZ Model: decreased alpha frequency power was predictive of SCZ likeness.
E. Sarisik et al., “EEG-based signatures of schizophrenia, depression, and aberrant aging: a supervised machine learning investigation,” Schizophrenia Bulletin, Sep. 2024, doi: 10.1093/schbul/sbae150.
Large Language Models in Psychotherapy and Mental Health Counseling: Scoping Review
Large Language Models in Psychotherapy and Mental Health Counseling: Scoping Review
This scoping review takes a systematic look at the state of research on the application of large language models (LLMs) within psychotherapy and mental health counseling. The aim is to synthesize existing research, identify current trends, and highlight gaps in how these tools are used and evaluated in clinical and non-clinical settings. Given the rapidly evolving nature of this field, this review provides a snapshot using a comprehensive and methodologically rigorous analysis.
Key Objectives and Scope:
Mapping the types of LLM-based tool applications in psychotherapy and non-formal counseling currently being researched
Summarizing the performance of these tools and their evaluation methods
Assessing how frequently they are evaluated with human clients versus solely on pre-existing datasets
Understanding client attitudes and acceptance toward these tools
Outlining research gaps and suggesting directions for future research to ensure clinical relevance, efficacy, and safety
Who will benefit:
Mental health professionals seeking to understand the potential and limitations of LLMs in psychotherapy
AI researchers and developers aiming to create clinically relevant and ethically sound LLM tools
Policymakers and regulators seeking insights to inform guidelines for responsibly integrating LLMs in therapeutic practice
This review is conducted by an interdisciplinary team with backgrounds in medicine, psychology, computer science, and systematic review methodology.
The ECNP-NNADR: A Collaborative Neuroimaging Data Repository for Advancing Psychiatric Research
The ECNP-NNADR: A Collaborative Neuroimaging Data Repository for Advancing Psychiatric Research
The study introduces the European College of Neuropsychopharmacology Neuroimaging Network Accessible Data Repository (ECNP-NNADR), a large-scale, multi-site initiative designed to address key challenges in psychiatric research. By integrating multi-modal neuroimaging and clinical data from diverse diagnostic groups, ECNP-NNADR facilitates the development and validation of machine learning (ML) models for diagnostic and prognostic applications.
Key Highlights
Objective: The repository was established to enable large-scale, collaborative psychiatric research by harmonizing multi-site and multi-diagnostic neuroimaging data. It aims to support the identification and validation of complex disease patterns across different imaging modalities.
Data Repository: ECNP-NNADR includes neuroimaging and clinical data from 4,829 participants across 21 cohorts, covering 11 psychiatric diagnoses, including schizophrenia (SZ), major depressive disorder (MDD), bipolar disorder I (BDI), bipolar disorder II (BDII), borderline personality disorder (BPD), obsessive-compulsive disorder (OCD), mild cognitive impairment (MCI), hoarding disorder (HD), generalized anxiety disorder (GAD), social phobia (SP), clinical high-risk for psychosis (CHR), and first-episode psychosis (FEP), along with large healthy control (HC) samples.
Data Harmonization & Sharing:
Multi-site data collection and harmonization protocols were established to ensure consistency across cohorts.
Clinical and MRI data were standardized using modality-specific data dictionaries, defining variable types, scales, and validity ranges.
MRI processing followed a harmonized morphometric analysis pipeline (CAT12) across sites.
Data sharing was implemented using the Virtual Pooling and Analysis of Research Data (ViPAR) platform, ensuring compliance with data privacy regulations.
Proof-of-Concept Analyses
Multivariate Classification: A machine learning model was trained on grey matter volume (GMV) region-of-interest (ROI) data to classify SZ patients versus HC individuals. The model achieved a balanced accuracy (BAC) of up to 71.13% across sites and atlases.
Normative Age Prediction: A regression model was developed using GMV data from HC individuals to estimate brain age and was applied to SZ patients. The model showed a mean absolute error (MAE) of 6.95 years and revealed an accelerated brain aging effect in SZ, with a brain age gap (BrainAGE) of 4.49 years.
Key Findings
Multi-atlas GMV data successfully differentiated SZ patients from HC individuals, achieving a BAC of up to 72.31% in discovery and 63.22% in validation.
SZ patients exhibited accelerated brain aging compared to HC individuals.
ECNP-NNADR provides a structured framework for evaluating novel methodologies in batch effect correction and symptom- or stage-oriented clinical phenotyping.
Clinical Relevance
ECNP-NNADR offers a robust, privacy-compliant platform for advancing neuroimaging research in psychiatry. By fostering collaboration and ensuring data harmonization, it enables the development of generalizable and interpretable ML-based diagnostic and prognostic models. The repository’s emphasis on widely available imaging sequences and software enhances the clinical feasibility of neuroimaging-based tools for psychiatric applications.
Conclusion
ECNP-NNADR represents a critical step toward large-scale, collaborative psychiatric neuroimaging research. By leveraging machine learning and standardized data sharing, it supports the development of individualized diagnostic and treatment strategies, ultimately contributing to precision psychiatry.
Fig 3. Reliable features for predicting the persistence of high FThD symptomatology from baseline to follow-up relative to other symptom courses based on A. slow-5 fALFF data, B. slow-4 fALFF data, C. slow-3 fALFF data, D. gray matter volume data, and E. white matter volume data. The reliability of the features is displayed using a grand mean cross-validation ratio, thresholded based on FDR-corrected sign-based consistency maps at α=.05 (detailed in Text S3). Warm colors represent voxels with increased activity/volume for individuals with persistently high FThD from baseline to follow-up, while cold colors indicate decreases for this subgroup.
Using Neuroimaging and Machine Learning to Predict Severity and Persistence of Formal Thought Disorder (FThD)
Using Neuroimaging and Machine Learning to Predict Severity and Persistence of Formal Thought Disorder (FThD)
This research explores the use of structural and functional neuroimaging data combined with machine learning to identify brain patterns that predict the severity and persistence of formal thought disorder (FThD) in individuals with recent-onset psychosis (ROP).
Key Highlights:
Data Acquisition: Whole-brain grey matter volume (GMV) and white matter volume (WMV) data were extracted from structural MRI and multiband fractional amplitude of low-frequency fluctuations (fALFF) were calculated from resting-state fMRI . 233 individuals with ROP were analyzed, divided into high and low FThD severity subgroups based on prior clustering analysis. Follow-up data was available for 153 participants for persistence analysis.
Machine Learning Models: Support vector machine (SVM) classifiers were developed to distinguish high vs. low FThD severity using baseline neuroimaging data and to predict persistence of high FThD severity at 1-year follow-up based on baseline neuroimaging. Multimodal stacked models were trained based on single-modality models to enhance prediction accuracy.
Cross-Sectional Results: GMV patterns in salience, dorsal attention, visual, and ventral attention networks classified FThD severity with 60.8% accuracy. High FThD severity was associated with higher GMV in cingulate cortex regions and lower GMV in the visual network.
Longitudinal Results (Persistence): Persistent high FThD severity was predicted by fALFF (BAC: 68.0%–73.2%), GMV (BAC: 62.7%) and WMV (BAC: 73.1%). Multimodal models combining these features achieved a BAC of 77%.
Clinical Relevance: The findings suggest that neuroimaging-based biomarkers could identify individuals at risk of persistent FThD, enabling stratified treatment approaches. FThD persistence is associated with poor clinical outcomes, emphasizing the need for early interventions.
The research provides evidence that structural and functional brain patterns can predict FThD severity and persistence in early psychosis. This work highlights the potential of using neuroimaging and machine learning to stratify patients and develop personalised treatments for those with severe FThD.
M.-O. Buciuman et al., “Structural and functional brain patterns predict formal thought disorder’s severity and its persistence in Recent-Onset psychosis: results from the PRONIA study,” Biological Psychiatry Cognitive Neuroscience and Neuroimaging, vol. 8, no. 12, pp. 1207–1217, Jun. 2023, doi: 10.1016/j.bpsc.2023.06.001.
Page updated
Google Sites
Report abuse