MIRA
A multi-stage clinical-triage research prototype, designed for multilingual, low-resource settings.
Mohammed Omar Farhat · Syeda Fahada Zia · L. Sunil Kumar
Research paper (DOI) · Zenodo record · Code
Overview
MIRA is a fixed six-stage pipeline of five LLM-backed agents (Intake → Language → Diagnostic → Triage → Referral, with Language run again to translate the reply). A deterministic layer of eight red-flag rules at the Triage stage overrides the language model's tier when that stage is reached.
Evaluation — the main finding
In a local run on 30 synthetic cases written by the first author, 17 of 30 cases (56.7%) stopped at the Intake stage and never reached the Triage red-flag layer, and none of those received a safety or referral message. This is a system-level finding, not an accuracy estimate.
| Measure | Value |
|---|---|
| Stopped at Intake | 17/30 (11 asked for sex, 6 for duration/severity) |
| Agreement with Rater 1, explicit written tiers | 1/13 (95% CI 1.4–33.3%) |
| Agreement, incl. 10 inferred from "agree" | 8/23 (95% CI 18.8–55.1%) |
| "14/29" | Reconstructed historical v1.2 convention — not a performance result |
Methodology
30 synthetic messages; author-assigned provisional tiers; local harness on 2026-09-11 with hosted open-weight models; retrospective selection of infrastructure-complete records across 17 attempts; partial, non-blind clinical review.
Limitations
Synthetic and author-written cases; small samples; incomplete, non-blind review (Rater 1 is the first author's mother and a co-author); placeholder retrieval corpus; no real-patient data; WhatsApp path not exercised; exact reproduction limited because prompts, model revisions and sampling settings were not preserved.
Version history
- TABIB — June 2026 hackathon prototype (GitHub).
- v1.0–v1.2-interim — superseded; contained errors corrected in v1.3.
- v1.3-corrected — 2026-10-04, 10.5281/zenodo.23136983. The report's appendix lists every correction.
What I did
Designed and built the system, wrote the 30-case set, ran the evaluation, and wrote the report. Dr. Zia provided clinical review (Rater 1). AI coding assistants were used and are disclosed in the report.
Cite
Farhat, M. O., Zia, S. F., & Kumar, L. S. (2026). MIRA: A Multi-Stage Triage Pipeline with a Triage-Stage Red-Flag Layer, Designed for WhatsApp-Based Use in Low-Resource, Multilingual Settings — An Interim Evaluation on a 30-Case Synthetic Set (v1.3-corrected) [Preprint]. Zenodo. https://doi.org/10.5281/zenodo.23136983