Source-linked AI summary
Auditable Emergency Triage for Maternal and Newborn Care in India
Shobhit Jagga, Aman Dalmia, Niharika Priyadarshini, Neelima Devadas, Amrita K Prasen, Nikhil Nalin, Santhosh SJ, Sreeram Nurani Ramasubramanian, Muhammed Afeer K, Anubhav Arora
TL;DR
Emergency triage at Noora Health was opaque, costly to audit, and disconnected from clinicians’ context-sensitive decision process. The paper introduces TRACE, which extracts canonical symptoms and patient context with an LLM before applying deterministic clinical rules. TRACE raised recall from 0.565 to 0.810 and F1 from 0.606 to 0.702, while enabling stage-level auditing and independent rule updates.
Problem
The prior end-to-end LLM triage system was opaque, used a flat danger-sign list instead of clinicians’ context-sensitive decision tree, and required costly full evaluations after prompt changes.
Method
TRACE uses an LLM to extract canonical symptoms and patient context from a clinician-authored vocabulary, then applies a deterministic rule engine owned by the clinical team.
Results
Recall increased from 0.565 to 0.810 and F1 from 0.606 to 0.702 over the replaced system.
Takeaways & Limitations
The decomposition makes errors traceable to specific stages and lets clinicians add rules independently without changing the LLM prompt.
Takeaways & Limitations
TRACE is 1.5 times costlier than LLM-Rules, and results may not generalize beyond the program’s internally labeled dataset and supported scenarios.
Abstract
from arXiv · showhide
At Noora Health, our nurses answer more than 50,000 medical queries per month on our WhatsApp-based service that provides caregivers with on-demand support. Their most time-critical task is emergency triage: deciding which queries need immediate in-person attention. To support them, we built a system that uses a large language model (LLM) to classify whether a message is an emergency and provide a rationale for interpretability. But the system was opaque: analyzing mistakes meant reading reasoning chains for each message, which is infeasible at our scale. Prompt changes meant re-running a full evaluation to prevent regressions, which was both costly and operationally challenging. Clinicians follow a decision tree to make this call, but it was never documented or passed to the model, which relied on a flat list of danger signs. To address these issues, we decomposed triage into two steps: an LLM extracts canonical symptoms and patient context from the query using a clinician-authored vocabulary, and a deterministic rule engine captures the scenarios that indicate an emergency. We show that the new system raised recall from 0.565 to 0.810 and F1 from 0.606 to 0.702, with structured rules driving most of the accuracy gains while the decomposition provides auditability: clinical experts can inspect each stage of the new system to see whether the query was mistranslated, symptoms were incorrectly extracted, patient context was wrongly inferred, or the necessary rules were missing. They can add new rules independently without causing regressions and avoid running costly evaluations. Since deployment, the new system has triaged 152,421 patient queries and flagged 28,535 (18.7%) as emergencies. The over-escalation rate has been 17.8%, without any increase in missed emergencies. Clinicians have also added 48 new rules since deployment, evidence of the faster correction loop we set out to build.
1 Introduction
Noora Health’s high-volume emergency-triage service faced opaque, costly, and clinically misaligned end-to-end LLM decisions. TRACE separates interpretation from clinical decision-making with structured extraction and deterministic rules, improving reported performance and auditability.
- More than 50,000 medical queries per month arrive through Noora Health’s WhatsApp-based maternal and newborn care service.
- The initial system used an LLM and a flat danger-sign list to classify emergencies and generate reasoning chains, but over-escalations became challenging at scale.The system was optimized for higher recall and deployed with nurse overrides for monitoring.
- The end-to-end model was opaque, did not match clinicians’ context-sensitive decision tree, and required full evaluations after prompt changes to prevent regressions.Mistake analysis required reading one reasoning chain at a time, while changes required coordination across teams.
- TRACE separates LLM-based canonical symptom and patient-context extraction from a deterministic, clinician-owned rule engine.The decomposition makes symptom interpretation the only subjective stage and leaves the subsequent rules fixed.
- Recall increased from 0.565 to 0.810 and F1 from 0.606 to 0.702, with structured rules driving most accuracy gains and each error traceable to a specific stage.Since deployment, clinicians added 48 rules without changing the LLM prompt or rerunning an expensive evaluation.
2 Related Work
Prior work establishes the difficulty of clinical triage from messy multilingual inputs and motivates systems whose decisions remain reviewable. This paper combines clinician-curated symptom mapping with a deterministic, auditable decision component.
- Clinical NLP systems can approach clinician-level acuity ratings, but prior results often assume clean inputs and keep verdicts inside the model.
- Maternal-triage messages are short, code-switched, frequently romanized, and expressed in lay terminology that challenges existing language models.The paper addresses this gap with a clinician-curated mapping from lay phrases to canonical symptom keywords.
- Interpretability research favors inherently reviewable high-stakes systems over post-hoc explanations, with deterministic components commonly used after an LLM language front-end.
- Every verdict in this system traces to a specific rule that clinicians can audit and modify directly.This operationalizes hard guardrails while allowing expert corrections to affect the system’s decision logic.
3 System Design
The system formalizes triage as a context-dependent classification problem rather than a message-only judgment. It translates multilingual queries, extracts symptoms and context, then applies rules using enrollment and profile information.
- The system maps each message, care category, and patient profile to an emergency or non-emergency output.Patient profiles may include gestational trimester, delivery date, or infant age.
- Triage cannot rely on the message alone because the urgency of the same symptom varies with patient context.
- Queries are translated into English, after which an LLM extracts canonical symptoms, available patient context, and a short rationale.
- A rule engine evaluates the extracted symptoms and context together with the care category and patient profile to assign the emergency label.
4 Symptom Vocabulary and Structured Rules
TRACE represents colloquial and multilingual symptom descriptions with clinician-curated canonical vocabulary, then applies context- and category-specific rules. The rule set is intentionally scoped to supported emergencies and can be revised as uncovered cases appear.
- Symptom Vocabulary: The symptom vocabulary maps multiple colloquial expressions to canonical clinical symptoms.
- Structured Rules: Rules are scoped to care categories, so the same symptom can require different accompanying symptoms or patient contexts to indicate an emergency.For example, leg pain in antenatal care differs from post-delivery leg swelling.
- Structured Rules: Standalone rules trigger on specified symptoms alone, whereas Combination rules require another symptom or matching context such as trimester or infant age.
- Structured Rules: The clinical team prepared 231 rules, including separate Standalone and Combination cases.
- Scope and Revision: The vocabulary and rule engine cover only emergencies supported by the program, and clinicians revise them when production queries expose uncovered cases.
5 Dataset
The dataset contains 769 real patient queries across seven supported languages, with clinician-adjudicated emergency labels and separate validation and test sets. Its size is constrained by the clinical experts’ review capacity.
- 769 real WhatsApp queries contained 251 emergencies across Hindi, English, Telugu, Kannada, Punjabi, Marathi, and Odia.The data were collected and annotated in two batches.
- 342 validation queries contained 104 emergencies, while 427 unseen test queries contained 147 emergencies.The first batch was stratified by care category; the second batch was added entirely to the test set.
- 83% raw agreement was reached by two in-house clinicians, with Cohen’s κ = 0.626 and senior-clinician adjudication for disagreements.Both annotators had access to the same patient context available to the triage system.
- Dataset size was limited by clinical experts’ review bandwidth, while government agreements prohibited commercial annotation vendors.Clinician adjudication added work on top of existing clinical duties.
6 Experiments
The experiments isolate structured rules, decomposition, and context inference using a stepwise ablation evaluated primarily by recall and secondarily by F1. Structured rules produced most of the gains, while TRACE offered auditability without a measurable quality advantage over LLM-Rules and incurred higher cost.
- 6.1 Setup: Structured rules alone raised recall from 0.565 to 0.837 and F1 from 0.606 to 0.689 over the LLM-KB baseline.LLM-Rules replaces the baseline’s flat danger-sign list with structured rules.
- 6.3 Auditability has no measurable quality loss: TRACE achieved recall 0.810 and F1 0.702, compared with LLM-Rules recall 0.837 and F1 0.689; overlapping intervals prevent declaring one architecture more accurate.TRACE also had lower median latency: 2.55 seconds per query versus 2.72 for LLM-Rules.
- 6.4 Context inference shows no measurable gain: TRACE-NoCtx and TRACE both reached recall 0.810, with overlapping F1 intervals, so context inference showed no measurable performance gain.TRACE-NoCtx was fastest at 1.96 seconds per query but context remained necessary for auditing context-dependent rules.
- 6.5 Auditability comes at a cost: TRACE was 1.5 times costlier than LLM-Rules, creating a genuine scaling obstacle despite auditability’s operational value.The authors prioritized reducing missed emergencies and over-escalations in production simultaneously.
- 6.7 Performance varies across care categories: Performance differences across languages and care categories were not statistically significant because the confidence intervals overlapped.TRACE led on precision and recall for high-risk pregnancy cases, while LLM-Rules had higher recall in routine ANC cases.
- 6.8 Alternatives we discarded: Fine-tuning was rejected because it would embed clinical policy in unreadable, uneditable model weights and require retraining for each rule revision.The design keeps clinical policy outside model weights so medical teams can revise it directly.
7 Auditability in deployment
TRACE makes triage errors traceable to specific components and assigns clear ownership for correction. In deployment, clinicians independently added rules while maintaining stable escalation outcomes.
- Evaluation: Table 5 evaluates architectures on a held-out test set of 427 queries containing 147 emergencies.It reports 95% stratified-bootstrap confidence intervals and excludes the shared translation step from latency and cost measurements.
- Error tracing: Every TRACE error is mapped to mistranslation, symptom extraction or vocabulary, patient context, or a missing or incorrect rule.This isolates actionable next steps for clinical and ML teams.
- Ownership: Clinicians can add rules independently because the rule engine operates separately from the LLM.This avoids requiring the ML team to run an evaluation for each rule change.
- Deployment: 152,421 queries were triaged after deployment, with 28,535 (18.7%) flagged as emergencies.Nurses marked 17.8% of flagged cases as over-escalations and reported no increase in missed emergencies.
8 Conclusion
The paper concludes that decomposing triage to mirror clinician workflows made mistakes traceable and enabled clinicians to resolve issues more independently. It argues that domain-specific systems should prioritize expert-aligned processes and ownership over model optimization alone.
- Conclusion: Decomposition traced each mistake to its source and enabled clinicians to resolve issues independently.Earlier changes required coordination between multiple teams and had high error-resolution times.
- Broader lesson: Domain-specific solutions should more closely mimic domain experts’ operations than optimize only the model powering them.The proposed process gives experts ownership of production quality and clear responsibility for auditing and resolving issues.
9 Future Work
Future work will expand data and clinician review while testing whether nurse overrides provide reliable ground-truth labels. The cited evaluation materials also emphasize care-category comparisons and component-level error analysis.
- Future work: The dataset will be expanded across danger signs, languages, and patient contexts with more clinicians involved in review.This is intended to increase sample size and coverage.
- Future work: Nurse overrides will be checked for clinical accuracy before being treated as a source of continuous ground-truth labels.The current goal is to assess whether overrides are more than a noisy production-quality signal.
- Evaluation: Table 7 compares LLM-Rules and TRACE by care category, omitting categories with fewer than 40 queries.It reports the number of emergencies for each included category.
- Evaluation: Table 8 analyzes TRACE mistakes by component so each error can receive an actionable next step and clear owner.The analysis covers mistakes on the validation set.
10 Limitations
The evaluation is bounded by internal labels and a dataset whose metrics may not generalize beyond the served population. The system also cannot probe for missing context, and long-term effects on clinical outcomes remain unvalidated.
- Scope: Results may not generalize to other settings because labels follow internal protocols for the families served.The current dataset’s metrics also have meaningful variance.
- System limitations: Emergencies are missed when symptoms are absent from the vocabulary or scenarios are not covered by the rules.Many rules also depend on patient context that may be unspecified, missing from the profile, or outdated.
- System limitations: The system currently does not probe patients for missing context, unlike trained nurses.This limits handling of rules whose applicability depends on information not supplied or reliably stored.
- Clinical outcomes: Long-term effects on patient outcomes have not yet been validated.An upcoming randomized controlled trial will evaluate the overall impact of the WhatsApp chatbot.
Ethical Statement
The system supports nurses rather than acting autonomously, with nurses and in-house doctors retaining final clinical authority. Its design recognizes that missed and over-escalations carry different human and health-system burdens.
- The system is a decision-support tool for nurses, not a patient-facing or autonomous system.It does not communicate directly with patients or provide medical advice.
- Nurses and in-house doctors review escalations and retain final clinical authority.
- Missed escalations can cause serious harm, while over-escalations increase patient anxiety and public-health-system burden.