AI-assisted summarization of intrapartum case sheets, so clinicians get an accurate, up-to-date picture of a mother's condition in seconds at shift handovers, consultant rounds, and emergency referrals without re-reading the full case sheet.
Part of Doto Health's open source initiative supported by UNICEF, targeting the intrapartum period in high-volume, resource-constrained hospital settings.
⚠️ Clinical disclaimer: This solution is a clinical documentation assistant, not a diagnostic or decision-making tool. It does not perform diagnosis, risk prediction, or treatment planning, and every summary is reviewed and signed off by a clinician before use. It is not a certified medical device.
Explore the architecture » · Project Charter · Developer Docs
Labour room clinicians manage a large amount of maternal clinical information: obstetric history, vitals, fetal monitoring, labour progression, medications, and clinical notes. During shift changes, consultant rounds, and emergency referrals, they need a quick, accurate read on the mother's current status, but reviewing the full case sheet takes time, and there is no standard way to produce a short summary.
This solution is building an AI assisted maternal case summarization framework. It converts detailed case sheets into concise, clinically relevant summaries, without altering or interpreting the original clinical documentation.
This repository covers Part 1 only: the AI summarization pipeline, from data intake through to a summary endpoint. Part 2, a Dashboard (portal) for clinicians to view summaries, is a separate, later phase that has not started.
The pipeline is a hybrid of deterministic clinical logic and AI-based language generation, split into six phases:
- Intake : Pull the latest patient record and the previous summary, and compute values like gestational age. parameters (e.g. gestational age).
- Deterministic Detection : Check the record against WHO Labour Care Guide thresholds to flag danger signs, no AI involved. This also classifies the patient as Normal or High-risk. involved. Classifies the patient as Normal or High-risk.
- AI Background Drafting : An LLM writes the mother's clinical background from free text notes. This runs once and is reused across summaries.
- Merger LLM : Assembles the final summary: current situation, what changed, and active alerts. The Recommendation is always left blank for the clinician to fill in. alerts as clear points. The Recommendation is always left blank for the clinician.
- Verify (governance) : a deterministic check confirms every number, alert, and change exactly matches the source data before anything reaches a clinician. Failing that check regenerates the summary (max 2 attempts), then falls back to a rules-only summary.
- Deliver : the clinician reviews, writes the Recommendation, and signs off.
See ARCHITECTURE.md for the full breakdown, including the design
This is the planned layout. Pipeline code has not been published yet — see Status & Roadmap.
├── src/ # pipeline code (planned, not yet published)
│ ├── intake/ # Phase 1
│ ├── rules_engine/ # Phase 2 — WHO Labour Care Guide checks
│ ├── llm_summary/ # Phases 3–4 — background + merger LLM stages
│ └── verification/ # Phase 5 — deterministic grounding checks
├── data/
│ └── synthetic/ # synthetic dataset (planned)
├── configs/ # model config — points to a local/HF model path, no weights committed
├── docs/ # developer documentation (GitHub Pages)
├── ARCHITECTURE.md
├── PROJECT_CHARTER.md
├── QA_PROCESS.md
├── CONTRIBUTING.md
├── CODE_OF_CONDUCT.md
├── LICENSE
└── README.md
This solution does not lock users into one hosting approach. A set of open models is being evaluated and will be published as recommended, tested defaults, but the pipeline is designed so a deployer can plug in either a locally hosted model or an API based model instead, whichever suits their setup.
No model weights are committed to this repository (see .gitignore). Recommended local models are downloaded from Hugging Face at setup time. Model selection is not finalized yet; see docs/model-setup.md for what is being evaluated and how model choice will be configured.
This solution is in active solution design and prototyping. It is not production ready and there is no public API yet.
- Problem definition, literature review (SBAR, I-PASS, SOAP), and solution architecture
- Six phase pipeline design finalized
- Evaluating candidate models, both local and API based
- Real clinical dataset collection and synthetic data generation underway
- Pipeline code (intake, rules engine, LLM stages, verification)
- Model selection finalized
- API and summary endpoints(software / system agnostic)
- Evaluation harness run against the finalized pipeline
- Part 2, Dashboard (portal) (not started, separate phase)
Planned for this repository next: the synthetic dataset, pipeline code, setup and run instructions, and instructions for connecting a chosen model, local or API, to the summary endpoints.
This solution is pre-alpha. The architecture is settled, implementation is in progress. Contributions, questions, and issues are welcome. See CONTRIBUTING.md for setup notes and the PR process, and QA_PROCESS.md for how summaries are evaluated for clinical grounding.
All contributors are expected to follow the Code of Conduct.
Licensed under the Apache License 2.0.
DOTO Software - software@dotohealth.com
Project Link: https://github.com/DOTO-Health/ai-clinical-case-summarisation.git
DOTO and the DOTO logo are trademarks of DOTO Health. Licensed under Apache 2.0 — trademark use is not covered by the code license. See LICENSE.md.