cadence-core 1.0.0

Cadence: A neural model for next clinical event prediction from EHR sequences using the Narrative Velocity framework

Cadence

Single-model neural network for clinical next-event prediction on MIMIC-IV (50 classes), with the goal of beating XGBoost on both top-1 accuracy and MAE simultaneously. No ensembles, no knowledge distillation from competitors, no competitor models.

Active subproject

• clinical-record-prediction/ — the active project. Read clinical-record-prediction/STATE.md for current status, best result, and next steps.

On hold

• ecg-cardiac-trajectory/ — MIMIC-IV-ECG project, paused.

Project rules and agent workflow

See .claude/CLAUDE.md for the prime directive, orchestrator protocol, canonical folder structure, and training safety rules.

Resumption protocol for a fresh Claude session

1. .claude/CLAUDE.md — rules
2. clinical-record-prediction/STATE.md — current state, best result, next steps
3. clinical-record-prediction/dev/LEADERBOARD.md — ranked clean experiments

Layout

cadence/
├── .claude/                       agent rules, skills, docs
├── clinical-record-prediction/    active project
│   ├── STATE.md                   ⭐ session resumption file
│   ├── src/                       promoted production code
│   ├── dev/                       experiments + LEADERBOARD.md
│   ├── results/                   promoted model artifacts
│   ├── paper/                     manuscript
│   └── data/                      → /mnt/data/clinical_data
├── ecg-cardiac-trajectory/        on-hold subproject
├── files/                         shared raw data
└── archive/                       orphan/historical

Development

Uses uv for Python dependency management.