PTSD Stratification and Predictive Modeling Workshop Summary

PTSD Stratification and Predictive Modeling Workshop Summary

By Heather Lasseter and Lee Lancashire

The PTSD Stratification and Predictive Modeling Workshop was conducted by Cohen Veterans Bioscience (CVB) on October 25th in Bethesda, MD as part of the Research Alliance for PTSD/TBI Innovation and Discovery Diagnostics (RAPID-Dx) Initiative. This meeting included researchers from academia, government, industry, and the non-profit sector, thereby bringing together experts in data analytics, predictive modeling, and clinical research in PTSD and related fields. The main objectives of the meeting were to discuss the latest advances in statistical and computational modeling for biomarker identification and patient stratification using case studies from industry and academia.

Dr. Magali Haas (CEO & President, Cohen Veterans Bioscience) established the RAPID-Dx program to advance biomarkers from discovery to validation. This involves many steps such as establishing and advancing a roadmap for discovery, replication and qualification of biomarkers by integrating large, high quality biomarker and imaging legacy and ongoing study datasets into a centralized cloud-based data platform for interrogation by a multi-disciplinary group of experts.

The meeting began with Dr. Lee Lancashire (CIO, Cohen Veterans Bioscience) outlining the urgent need to find effective pharmacologic treatments for PTSD given the enormous clinical, social, and financial burden associated with trauma-related brain disorders.  Dr. Nikolaos Daskalakis (Director of Data Science and Translational Medicine, Cohen Veterans Bioscience) spoke about the cohorts available as part of the RAPID-Dx program that will be used as part of the discovery and validation efforts. These subjects have undergone neurocognitive assessments and have provided biofluid samples either in a cross-sectional or longitudinal fashion resulting in diverse cohorts with data that requires harmonization across many types of measures including clinical assessments, neurocognitive testing, neuroimaging, and psychophysiological tests. 

Dr. Lancashire then framed the diagnostic challenge by describing the critical need in understanding diverse PTSD trajectories. Given these challenges, a paradigm shift will move away from the concept of PTSD as a single disease and instead adopt an endophenotype definition of PTSD in which symptoms can be deconstructed. Machine learning approaches such as forms of unsupervised clustering (NMF, t-SNE etc.) and deep learning approaches may enhance our ability to discover PTSD endophenotypes and biomarkers that facilitate our understanding of the underlying mechanisms. 

Dr. Guillermo Cecchi (IBM) gave examples of how speech and image based analytics of psychiatric disorders are being used to quantify psychiatric conditions. Dr. Cecchi’s group is also using multi-modal imaging approaches (e.g., structural MRI, fMRI, ROI volumes, and DTI) together with genetics and cognitive/motor tasks in Huntington’s disease to build markers for use as clinical endpoints or to help stratify patient populations based on risk of disease progression.

Dr. Hayete introduced the GNS platform for discovery of causal relationships from data. Two recent applications in neurology were described where GNS used clinical, genetic, imaging data and baseline molecular and clinical variables to develop novel predictors for motor and cognitive outcomes in Parkinson’s Disease (PD).  These studies indicate that causal modeling can be used to help identify novel predictors of PD motor progression.

Dr. Brian Kidd (Mt. Sinai) spoke about how systems biology medicine allows for the incorporation of complex data (e.g., clinical data, genomics, transcriptomics, proteomics, EMR). Examples were presented highlighting “The Promise of Big Data in Medicine,” such as the identification of subgroups in type 2 diabetes through topological analysis of patient similarity, and used EMRs and genotype data to predict the future of patients by applying deep learning methods.

The meeting culminated in an active discussion regarding how to address current challenges in PTSD translational neuroscience. The main conclusions were:

  • We need to consider what types of data and cohorts are needed to frame the problem.
  • There is a need to find a meaningful endpoint for treatment.
  • We need to revisit how patients are characterized, determine which symptoms “really matter,” and focus on those for phenotyping.
  • There are three critical biomarker domains; predictive, stratification, and monitoring.
  • The accurate clustering of patients accurately has been difficult. This should include other trauma-induced pathologies such as anxiety and major depressive disorder.
  • While a prior clustering might not always be informative, clustering conditional relationships from a model may be much more informative.
  • A challenge with large datasets is dimensionality reduction. Multiple types of data are available and need to be reduced to smaller sets.
  • Working Groups must be formed to tackle individual problems in which experts in computational modeling can be paired with clinical partners to fully address these problems.