Neuroscience · Machine Learning
Josh Threlkeld
Building at the intersection of neuroscience and machine learning. B.S. Neuroscience, USC. Interested in clinical AI, BCI, and how biosignals become decisions.
Research
Anxiety Classification from EEG Signals Using Machine Learning
Investigates the temporal limits of EEG-based anxiety classification using the OpenNeuro ds007609 dataset. Demonstrates a systematic degradation in classification accuracy as target constructs become more temporally distal — eyes-open/closed states outperform state anxiety, which outperforms trait anxiety.
Music Genre Classification Using Machine Learning on the GTZAN Dataset
Applies machine learning to the GTZAN benchmark for automatic music genre classification. Identifies classical and metal as top performers due to spectral distinctiveness. Explains the blues anomaly through harmonic structure and random forest nonlinear feature interactions.
Projects
Clinical NLP RAG Assistant
Conversational Q&A system over seven peer-reviewed and preprint papers on ambient AI documentation in clinical settings. Sources span Abridge, Suki, and UCI Health deployments. Enables natural language access to the literature without manual search. Containerized with Docker for reproducible deployment.
EEG RAG Assistant
Question-answering system grounded in a published EEG anxiety classification paper. Ingests the PDF, embeds chunks with a local sentence transformer, stores them in ChromaDB, and retrieves context for Claude to generate answers. Built on top of original research to demonstrate RAG applied to domain-specific neuroscience literature.
CMS Readmission Analysis
Which Medicare patients get readmitted within 30 days, and what does it cost? Built a patient-level longitudinal pipeline over 66,773 inpatient claims in BigQuery, joining annual beneficiary snapshots for demographics and chronic condition burden. Found a 29.6% readmission rate concentrated in a handful of DRGs, and a $492 per-claim payment gap that held up under hypothesis testing. Also caught an apparent decline across 2008-2010 that turned out to be an artifact of the synthetic data, not a real trend. Run on CMS SynPUF, so the methodology is the deliverable, not the numbers.
Talkative Listener
Real-time voice agent built as a Python streaming pipeline handling speech-to-text, response generation, speech synthesis, and interruption mid-utterance. The harder problem is knowing whether it holds up, so the project ends in a two-bot adversarial simulator where one agent plays a difficult caller against a clinic scheduling agent, surfacing the failure modes that a scripted test would never trigger.
Articles
Write-up of “From Prompt to Product: A Human-Centered Benchmark of Agentic App Generation Systems”
Quome Inc. benchmarked Replit, Bolt, and Firebase Studio across 288 apps and 1,071 pairwise comparisons. Isolated ratings are nearly identical — but side-by-side comparison reveals a clear hierarchy, and a gap between the signals investors rely on and the ones that actually surface user preference.
OpenAI's GPT-5.3-Codex: Sandbagging and the Shifting Safety Calculus
Apollo Research scored GPT-5.3-Codex 0.88/1.00 on a sabotage benchmark — outperforming human ML experts. A breakdown of what that means for the "it can't, not it won't" safety argument.
About
I'm a neuroscience graduate from USC (B.S., Cum Laude, 2025) working at the intersection of brain science and applied machine learning. My research background spans wet lab molecular work at USC's Zilkha Neurogenetic Institute, where I contributed to a PNAS publication on fragile X syndrome, and computational projects applying ML to biosignals and audio.
I'm currently building toward roles in clinical AI — particularly companies using ML to close the loop between clinical data and care decisions. Longer term, I'm interested in health and life sciences venture.