inferred from GitHub
parkscreen
Detect cognitive decline using visual and voice dynamics
Core trick
Fuses per-channel classifier probabilities from voice and face-video recordings via AUC-weighted late fusion, then Claude writes the report.
Evidence
README matched ""Cerebral Valley" "Opus 4.7"": README.md
README excerpt
# ParkScreen [](https://kiwihu0923.github.io/parkscreen/) **ParkScreen** is a multimodal Parkinson's disease screening decision aid. It fuses three task-matched short recordings — sustained-vowel phonation, rapid /pa-ta-ka/ diadochokinesis, and a smile-task face video — through per-channel classifiers trained on the NeuroVoz (speech) and UFNet (face) corpora, combined via AUC-excess-weighted late fusion. The fused probability, per-channel scores, and channel-agreement flags become structured context for Claude Opus 4.7, which generates a clinical-style Markdown report with mandatory disclaimers and hedged language. Built for the *Built with Claude: Life Sciences* hackathon (Anthropic × Gladstone × Cerebral Valley), July 7–13 2026. **Development Track.** > ⚠️ ParkScreen is a **screening decision aid**, not a clinical diagnostic tool. Any elevated score should be interpreted only as a prompt to seek qualified clinical evaluation. See [`docs/LIMITATIONS.md`](docs/LIMITATIONS.md) before interpreting any output. <p align="center"> <img src="demo/screenshots/pd_demo_report-1.png" alt="ParkScreen example screening report — PD demo case, Grade C, fused 0.70" width="640"> </p> *Example output — end-to-end run on the self-recorded PD demo case (Grade C, fused 0.70). Full PDFs + PNG snapshots for all three demo cases (HC Grade A, PD Grade C, NeuroVoz held-out Grade C with facial channel N/A) in [`demo/screenshots/`