About me

I am a computational scientist working at the intersection of neuroscience, psychometrics, biostatistics, and AI for science. I did my postdoctoral work in the Cognitive Network Modelling Lab at the Krembil Centre for Neuroinformatics at the Centre for Addiction and Mental Health, and earned my PhD from the University of Edinburgh.

My work spans multiple scales, from developing computational models and gamified cognitive assessments for studying specific disorders to addressing systemic probelms in research practices. At the disorder level, I have worked with reinforcement learning, (hierarchical) Bayesian inference, and active inference models to study multiple conditions including psychosis, autism (Karvelis et al., 2018), and suicidality (Karvelis et al., 2022). The goal of these models is to provide novel computational markers for diagnosis, prognosis, and treatment – yet their promise depends on overcoming multiple methodological barriers.

In my most impactful work, I have shown that these novel computational assays often suffer from poor psychometric (reliability and validity) properties due to inadequate research designs, and provided a roadmap for achieving clinical utility (Karvelis et al., 2023). I carried out further empirical (Karvelis et al., 2024) and simulation-based (Karvelis & Diaconescu, 2025) work to highlight the extent of measurement reliability issues and to provide new methods for dealing with them. In addition to the psychometric issues, I have also detailed common methodological problems when applying machine learning for treatment response prediction, highlighting the benefits of developing generative models for feature extraction (Karvelis et al., 2022).

To address these methodological and translational bottlenecks, I have developed an interactive open-source web tool, www.e2p-simulator.com. E2P Simulator evaluates biomarkers and prediction models through the lens of real-world predictive utility by accounting for measurement reliability and outcome base rates—two crucial but often overlooked factors. Its interactive nature makes it accessible and educational, helping fight methodological inertia and steer researchers towards best practices. See this preprint for an example of its application to research in precision psychiatry and medicine.

Lately, I have been focused on agentic science: I am developing an AI co-scientist that can answer complex biomedical questions by synthesizing evidence across 50+ public data sources: AI Co-Scientist; I am also exploring how to build high-quality knowledge graphs (important infrastructure for agentic science) via automated literature retrieval, curation, and claim extraction: Psychedelics Knowledge Graph. I am also working on applying psychometric principles to build better evals and benchmarks for AI (coming soon…).