All Research Contributions
  • Oral Presentation
  • Featured Technology

ToxLumen — Featured During the MedSpeech Presentation at NACCT 2026

A NACCT 2026 platform session on the MedSpeech documentation pilot, which then introduced ToxLumen as the broader clinical platform. I am responsible for ToxLumen’s end-to-end architecture and software development.

My role · ToxLumenLead Software Engineer & System Architect
Conference
NACCT 2026 — North American Congress of Clinical Toxicology
Format
Platform Session · Oral Presentation
Date
October 25, 2026 · 11:30–11:45 AM PT
Location
San Diego, California, USA
Application screenshotThe production ToxLumen application — organization home, linking the Clinical and Simulation products.

Overview

The oral presentation focused on MedSpeech, an AI-enabled voice-to-text platform for poison-center documentation, and reported its randomized documentation pilot.

After the pilot results, the talk introduced ToxLumen as the broader clinical platform extending beyond the original documentation workflow: one case record that carries a call from intake, through extracted facts, dose math, safety signals and guidance, to a SOAP draft the specialist reviews and signs. MedSpeech is no longer a standalone product — its capabilities now live inside ToxLumen, the actively developed platform.

My Engineering Contribution

  • System architecture — one Django 5 + DRF backend of sixteen domains under a strict routes → views → serializers → services/selectors layering, with Organization → Workspace → Case ownership.

  • Backend and frontend development — the unified backend, the customer React + TypeScript SPA, and a separate back-office.

  • AI workflow integration — OpenAI, Groq and Deepgram behind a provider layer, with background stages on Huey and credits priced from measured provider usage.

  • Multimodal clinical case intake — typed narrative, recorded or uploaded audio, and live calls into one case record.

  • Speech transcription and structured fact extraction — including read-only live-call supervision over Django Channels, with patient audio streamed browser → provider and never proxied or stored.

  • Clinical analysis and documentation workflows — the Case → Analysis → Guidance → Note spine, audit trail and finalized-case lock.

  • Service integration and production infrastructure — external identity and billing services applied idempotently, Docker, Nginx and production deployment.

Attribution

  • The MedSpeech pilot — its design, data collection, analysis and findings — belongs to the clinical research team. I am not an author of the presentation or of the research.
  • Clinical content, guidance entries and the medical science behind ToxLumen come from the physicians’ clinical expertise. I did not develop or clinically validate the underlying medicine.
  • My contribution is engineering: the architecture, implementation and infrastructure of ToxLumen, documented in full in the ToxLumen case study.

ToxLumen as Presented

The slides present ToxLumen as one case record, from call to signed note, in which each step is reviewable and the specialist can correct any step before sign-off. The architecture behind that workflow — the unified backend, ownership model, AI provider layer and real-time transcription surface — is described in the ToxLumen engineering case study.

  • Input — typed, dictated, uploaded or live.

  • Facts — extracted, with missing items flagged.

  • Dose math — computed in code.

  • Safety signals — graded by severity.

  • Guidance — traced to curated entries.

  • SOAP draft — for specialist review.

Conference Presentation

Presented in a NACCT 2026 platform session on October 25, 2026, 11:30–11:45 AM Pacific Time, in San Diego. The disclosure slide identifies MedicalToxic.com as the developer of MedSpeech and ToxLumen.

The slides label every ToxLumen function beyond documentation as subsequent development not evaluated in the pilot, and present multi-center validation on ToxLumen as the next step — with its added functions to be validated in their own right, separately from the pilot.

Platform Session · Oral Presentation

“MedSpeech: An AI-Enabled Voice-to-Text Platform for Poison-Center Documentation”

Speakers

  • Omid MehrpourMD, FACMT · Florida/USVI Poison Information Center – Jacksonville · University of Florida
  • Farshad M. ShiraziMD, FACMT · Arizona Poison and Drug Information Center · University of Arizona

Speakers as listed on the title slide. I am not a speaker or an author of this presentation. The talk introduced ToxLumen — the platform I architected and built — after presenting the MedSpeech pilot; there was no separate presentation titled “ToxLumen”.

MedSpeech Pilot Findings

These results belong to the MedSpeech documentation pilot only: 60 incoming calls randomly assigned to MedSpeech-assisted or usual documentation, 58 analyzed (29 per group), six volunteer specialists in one pilot setting.

Median documentation time, call end to signed note
3.2 vs 5.8 min
Shorter median documentation time
≈45%
Field-level information-extraction accuracy vs original audio
96%
Analyzed in a small randomized pilot
58 calls

Scope of these results

  • These are not validation results for the broader ToxLumen platform. ToxLumen’s added functions, including decision support, were described as subsequent development requiring separate clinical evaluation.
  • Extraction accuracy is not clinical decision-making accuracy.
  • Patient outcomes, clinical-error rates and clinician workload were not measured; with 58 calls in one setting, the estimates are imprecise and may not generalize.

Evidence Gallery

  • Application screenshotThe production ToxLumen application — organization home, linking the Clinical and Simulation products.

Official References