KithMed

Inspiration

In humanitarian disaster zones, conflict regions, and remote rural clinics, front-line Community Health Workers (CHWs) are routinely forced to make life-or-death triage decisions under intense pressure. Most modern AI clinical models fail in these austere environments because they demand high-bandwidth cloud APIs, hallucinate confident answers on ambiguous vitals, and provide uninterpretable black boxes that clinicians cannot verify.

We built KithMed to bridge this gap: an open-source, edge-native clinical decision engine that couples gold-standard physiological scoring frameworks (NEWS2, qSOFA, Shock Index) with Conformal Uncertainty Quantification to deliver transparent, deterministic, and fail-safe triage intelligence with zero internet connectivity.

What it does

KithMed provides both an ultra-responsive Tactical Sunlight Web Console and a lightweight Field CLI:

1. Multi-Modal Clinical Deterioration Detection:
   • NHS NEWS2 (0-20): Continuously evaluates 6 physiological telemetry markers (HR, SBP, RR, SpO2, Temperature, Glasgow Coma Scale) to detect physiological collapse.
   • Sepsis-3 quick SOFA (0-3): Rapidly identifies patients with suspected infection who are at high risk of in-hospital mortality.
   • Shock Index (SI = HR / SBP): Detects occult hemorrhagic and hypovolemic shock before systolic hypotension manifests.
2. Conformal Uncertainty Bounding:
   • Rather than forcing ambiguous single-point predictions, KithMed computes rigorous conformal epistemic bounds based on vital parameter proximity to volatile decision thresholds.
   • If uncertainty exceeds the 35% safety boundary, KithMed alerts the operator and mandates physician tele-consultation or urgent transport dispatch.
3. Tactical Emergency Console & Offline Passes:
   • High-contrast sunlight-readable interface engineered for extreme outdoor visibility.
   • Integrated Web Audio acoustic alarms for emergent and critical classifications.
   • Cryptographic offline emergency pass generator to enable seamless handoffs between transport units and regional hospitals.

How we built it

• Deterministic Inference Engine: Written in pure, dependency-free Python 3.10+ and portable client-side JavaScript for instantaneous local execution on smartphones, rugged tablets, or battery-powered field laptops.
• Conformal Calibration Layer: Evaluates aleatoric and epistemic uncertainty distributions across borderline vital thresholds and vulnerable cohorts (pediatric/maternal).
• Tactical Web Console: Built with zero external CDN dependencies, leveraging native Web Audio API, responsive CSS grid layouts, and sunlight-optimized typography.
• Production Deployment: Edge-deployed via Vercel for instant worldwide offline caching.

Challenges we ran into

• Calibrating real-time physiological decision boundaries to prevent false negatives in occult shock without overloading rural field workers with alarm fatigue.
• Ensuring strict sub-millisecond execution times on low-power edge devices with 0% network reliance.

Accomplishments that we're proud of

• Achieving 100% deterministic test coverage across 8 comprehensive clinical shock, sepsis, hypoxia, and pediatric boundary scenarios in 0.000s.
• Building a fail-safe medical UI that prioritizes clinician interpretability and clear actionable intervention bundles over opaque black-box outputs.

What we learned

• How crucial uncertainty quantification is in safety-critical medical machine learning—knowing when the model does not know is more valuable than false confidence.

What's next for KithMed

• Integration with LoRa mesh transceivers for ad-hoc distributed mass-casualty disaster response networks.
• Field testing partnerships with community health worker networks and disaster relief NGOs.

Built With

• conformal-prediction
• css3
• html5
• javascript
• nhs-news2
• python
• sepsis-3-qsofa
• shock-index
• typescript
• vercel
• webassembly