Machine Learning in Psychiatry
Project Summary
Despite substantial efforts, the causes of most psychiatric disorders remain unclear; even categorizing such disorders precisely has been difficult. The diagnostic systems in psychiatry have mostly relied on descriptive phenomenology that does not fully consider the heterogeneous symptoms or their biological mechanisms and etiology. Recent approaches to psychiatric classification such as Research Domain Criteria (RDoC) have moved toward characterization of biomarkers that cut across symptom-based diagnoses, but map on to translational domains from cellular to circuitry and behavioral levels. An increasing amount of neuroimaging data has been established in recent years to understand the complex brain functions in both healthy and pathologic mental conditions.
To quantify the complex brain signal data, an approach that integrates mathematics, physics, and computational neuroscience is required. Nonlinear dynamical approaches to quantify the complexity of brain signal data may have the potential to develop useful markers to extract fundamental features from spatial-temporal neuroimaging data at multiple levels. Recently, we have applied the multiscale entropy analysis, a method for quantifying the complexity of physiologic signals, to analyze the temporal dynamics of resting-state functional magnetic resonance imaging (fMRI) data in both health and pathological conditions, such as psychosis.
We found that the alteration of complexity of brain activity may be a defining feature of psychosis, which could serve as a marker to phenotype the psychotic spectrum disorders (e.g., psychotic bipolar, schizoaffective disorder, and schizophrenia). Therefore, this proposal aims to extend our current findings by investigating comprehensively the complexity of resting-state fMRI data, and develop the RDoC strategy with machine learning-based computational algorithms to phenotype psychotic spectrum disorders. We anticipate that these approaches will provide better characterization of the heterogeneity of psychosis and will highlight a subset of dynamical brain markers that will lead to the translational research utilizing these computational models in the clinic.
To quantify the complex brain signal data, an approach that integrates mathematics, physics, and computational neuroscience is required. Nonlinear dynamical approaches to quantify the complexity of brain signal data may have the potential to develop useful markers to extract fundamental features from spatial-temporal neuroimaging data at multiple levels. Recently, we have applied the multiscale entropy analysis, a method for quantifying the complexity of physiologic signals, to analyze the temporal dynamics of resting-state functional magnetic resonance imaging (fMRI) data in both health and pathological conditions, such as psychosis.
We found that the alteration of complexity of brain activity may be a defining feature of psychosis, which could serve as a marker to phenotype the psychotic spectrum disorders (e.g., psychotic bipolar, schizoaffective disorder, and schizophrenia). Therefore, this proposal aims to extend our current findings by investigating comprehensively the complexity of resting-state fMRI data, and develop the RDoC strategy with machine learning-based computational algorithms to phenotype psychotic spectrum disorders. We anticipate that these approaches will provide better characterization of the heterogeneity of psychosis and will highlight a subset of dynamical brain markers that will lead to the translational research utilizing these computational models in the clinic.
Hypothesis
Phenotyping psychosis has been a critical research topic in recent years. Approaches to phenotype psychosis have been largely based on behavioral pathology, such as cognitive deficit. However, little progress has been made to phenotype psychosis based on structural or functional brain abnormality such as those derived from magnetic resonance imaging. Increasing amount of neuroimaging data has been established in recent years to understand the psychosis with fruitful results. We hypothesize that the subtyping of psychosis is more related to brain pathology but not behavioral pathology, which the former is a fundamentally top-down approach to understand psychosis.
Aim
To phenotype psychotic spectrum disorders with structural and functional brain markers
The aim of this project is to use computational framework based on the machine learning to identify biological classifications of psychotic spectrum disorders by incorporating developed structural and functional brain markers.
The aim of this project is to use computational framework based on the machine learning to identify biological classifications of psychotic spectrum disorders by incorporating developed structural and functional brain markers.
Methods
We have recently established a large-scale schizophrenia neuroimaging database (more than 300 patients) at a single site in Taiwan. Additionally, I have access to the data from Bipolar and Schizophrenia Network for Intermediate Phenotypes. We plan to use these data as the basis to investigate the brain phenotype of psychosis.
Briefly, we will develop computational algorithms to phenotype psychosis using both supervised and unsupervised learning strategies, ranging from statistical classifiers to deep learning networks. In each strategy, the input is the parametric brain image mapping from structural and functional brain analysis; the output is the DSM diagnoses in supervised learning, or biological classifications in unsupervised pattern recognition. In general, the data will be divided into three parts: training (60%), validation (20%) and testing (20%). Here we list these strategies as follow:
1. To develop supervised learning classifiers of DSM diagnoses using multinomial logistic regression and support vector machine.
2. To develop supervised learning classifiers of DSM diagnoses using convolutional neural network.
3. To develop unsupervised learning classifiers of psychosis using hierarchical clustering methods and decision trees.
4. To develop unsupervised learning classifiers of psychosis using the autoencoder neural network.
5. To identify the clinical correlates of unsupervised classification results as assessed and validated by symptom profiles, outcome measures, and cognitive assessments.
Briefly, we will develop computational algorithms to phenotype psychosis using both supervised and unsupervised learning strategies, ranging from statistical classifiers to deep learning networks. In each strategy, the input is the parametric brain image mapping from structural and functional brain analysis; the output is the DSM diagnoses in supervised learning, or biological classifications in unsupervised pattern recognition. In general, the data will be divided into three parts: training (60%), validation (20%) and testing (20%). Here we list these strategies as follow:
1. To develop supervised learning classifiers of DSM diagnoses using multinomial logistic regression and support vector machine.
2. To develop supervised learning classifiers of DSM diagnoses using convolutional neural network.
3. To develop unsupervised learning classifiers of psychosis using hierarchical clustering methods and decision trees.
4. To develop unsupervised learning classifiers of psychosis using the autoencoder neural network.
5. To identify the clinical correlates of unsupervised classification results as assessed and validated by symptom profiles, outcome measures, and cognitive assessments.
|
|
|
Your comments and suggestions are welcome.
Please send your comments by email to Dr. Albert C. Yang (Laboratory of Precision Psychiatry).
Location: No. 155, Section 2, Linong Street, Beitou District, Taipei, Taiwan
Please send your comments by email to Dr. Albert C. Yang (Laboratory of Precision Psychiatry).
Location: No. 155, Section 2, Linong Street, Beitou District, Taipei, Taiwan