Project status
MedDelta has a working browser prototype using synthetic examples. Independent evaluation and clinical validation are planned; no clinical performance or patient outcome claims have been established. Christian Lopez Blanco is developing the project as a solo fourth-year Neuroengineering PhD student at the University of Alabama at Birmingham.
Inspiration
Discharge documents can give conflicting instructions for the same medication. MedDelta asks a focused question: do two documents tell the same patient to do incompatible things with the same medicine over overlapping or unspecified effective dates? The goal is to make these disagreements easier for a clinician to inspect and resolve.
What the prototype does
The demo compares two editable synthetic documents and groups findings by medication. It detects supported status, dose, and frequency disagreements, recognizes explicitly non-overlapping instruction periods as potentially intentional changes, and requests clarification when evidence is insufficient. Each finding links to exact source passages and line numbers.
Seven example scenarios cover stop/continue conflicts, dose and frequency differences, explicit timing, consistent aliases, ambiguous wording, and formulation mismatches. Users can edit the text, compare again, inspect evidence, record a review outcome and reason, and download a JSON review record. Review decisions remain in the current browser session; editing documents requires a new comparison before recording a review.
How it is built
The current prototype uses TypeScript, React, and Vinext, hosted on Cloudflare Workers through Sites. Comparison runs in the browser. A narrow deterministic parser extracts medication facts into a fixed schema, normalizes supported doses and a small local set of medication aliases, and applies explicit comparison rules. It currently supports eight medications and one instruction per line. It does not use a language model, a live RxNorm connection, Synthea-generated cases, PDF ingestion, or an EHR integration.
A future extraction layer could use a language model and RxNorm to expand coverage while preserving exact source evidence. These integrations remain planned.
Challenges and learning
The main challenge is separating disagreement from uncertainty. Different dates do not automatically make one document authoritative; a missing fact does not prove agreement. The prototype therefore surfaces unsupported lines and uncertain instructions instead of treating them as safe. Building the demo also exposed a usability issue: Compare needed visible completion feedback and direct navigation to the findings, which are now implemented.
Differentiation
Medication reconciliation products and discharge-auditing research already exist. MedDelta focuses on cross-document instruction conflicts, explicit timing and uncertainty, and a review record tied to source evidence. This is a proposed distinction to evaluate, not a claim that medication reconciliation is new. Planned baselines are normalized medication-list matching and direct language-model review.
Evaluation plan
The code has 36 automated rule tests, which are development checks rather than an independent performance benchmark. We plan 60–100 synthetic document bundles with separate development and evaluation templates, including clean controls, contradictions, intentional changes, and ambiguous wording. Independent pharmacist review of a subset is planned and has not yet been secured.
Planned metrics include conflict precision and recall, false alerts per bundle, evidence-span correctness, abstention rate, and reviewer time. Engineering targets are at least 90% precision and recall on predefined synthetic categories, source evidence for every flag, and a 30% reduction in median review time in a paired exercise. These are targets, not achieved results or evidence of clinical effectiveness.
Boundaries and next steps
The demo is for synthetic text only. It does not verify patient identity, determine the correct treatment, prescribe, or change patient instructions. “Consistent” applies only to the supported fields checked and does not establish regimen safety or completeness. A public synthetic demonstration is available through the Try it out link.
Next steps are independent synthetic evaluation, clinician usability feedback, broader extraction coverage, and continued improvements to the demonstration. Any use in patient care would require further validation and appropriate institutional review.
Selected references
- Dawaiwala et al. (2025), Impact of Clinical Pharmacist Interventions on Medication Reconciliation at Discharge in an Indian Tertiary Care Hospital: https://www.ijopp.org/article/18/2/219
- Dasula et al. (2026), CareTransition-Audit: https://arxiv.org/abs/2604.05435
- Synthea synthetic patient generator: https://synthetichealth.github.io/synthea/
- National Library of Medicine, RxNorm: https://www.nlm.nih.gov/research/umls/rxnorm/overview.html
AI tools assisted concept research and drafting. Planned prototype resources include the open-source libraries, Synthea, and RxNorm identified above. Third-party resources will be acknowledged in the implementation.
Log in or sign up for Devpost to join the conversation.