Skip to content

open-source-roadmap.md

Latest commit

 

History

History
84 lines (64 loc) · 3.19 KB

open-source-roadmap.md

File metadata and controls

84 lines (64 loc) · 3.19 KB

Open Source Roadmap

CaseCrawler is an open-source synthetic healthcare data platform for research, training-pipeline prototyping, benchmark construction, and release-package experimentation. It is not a medical device, and generated outputs require validation before any downstream clinical or operational use.

Maturity Levels

Level 1: Offline Synthetic Records

  • Deterministic structured EHR records.
  • Clinical notes with controlled variation.
  • Schema, privacy, and clinical-consistency validation.
  • Basic JSONL, FHIR NDJSON, parquet, and fine-tuning exports.

Level 2: Evidence-Aware Dataset Generation

  • Grounding metadata from public references and local fixtures.
  • Public reference fixture import for benchmark comparisons.
  • Benchmark profile scoring.
  • Human review queues and review summaries.

Level 3: Training-Ready Release Packages

  • Train, validation, and test split packages.
  • Dataset cards, model cards, manifests, checksums, and provenance.
  • Objective coverage audits.
  • Strict placeholder rejection for release-gated multimodal packages.
  • Export transparency summaries that describe record origins and limitations.

Level 4: Clinician-Reviewed Benchmark Suites

  • Clinician-reviewed golden cases.
  • Specialty-specific coverage targets.
  • Condition-specific clinical content packs.
  • Contributor-maintained benchmark and validation fixtures.

Level 5: Multimodal Research Platform

  • Real image backend contracts and policy metadata.
  • Time-series model adapters.
  • Image/report alignment validators.
  • Optional external judge support.
  • Public evaluation suites for generated release packages.

Current Strengths

  • No-key deterministic generation path for local development.
  • Multimodal record model covering structured EHR, notes, labs, vitals, medications, allergies, orders, time series, and imaging metadata.
  • Fine-tuning export profiles for SFT, chat, tool use, note-fact extraction, clinical observations, medication reconciliation, multimodal tasks, time-series tasks, DPO/RL, FHIR NDJSON, and parquet.
  • Release-readiness gates with quality reports, audit artifacts, and human review workflow support.
  • Pull-request CI tiers for backend, UI, and optional backend coverage.

Near-Term Priorities

  1. Convert hard-coded clinical profiles into contributor-friendly clinical content packs.
  2. Expand clinical coverage validation for condition-specific required artifacts.
  3. Grow public benchmark-suite fixtures and golden regression cases.
  4. Improve UI visibility into release readiness, human review status, and benchmark failures.
  5. Add more examples for open-source contributors building new conditions, model adapters, and export profiles.

Out Of Scope

  • Real patient data ingestion for training datasets.
  • Clinical decision support claims.
  • Automated diagnosis, treatment recommendation, or patient-care workflows.
  • Shipping generated data without provenance, validation, and review metadata.

Contributor Entry Points