MPC Signal Detection Project

AI-Assisted NPDS Verbatim Surveillance for Emerging Substance Threats

Updated March 18, 2026 | NACCT 2026 Abstract Deadline: May 4
1,339
Exposures identified
across 6 substances
7
Deaths
(all polysubstance)
98
Major effects
32+
Additional substances
in ecosystem discovery
Retatrutide: 17x Growth in 8 Months
Monthly exposures (all) vs Google Trends search interest (US)
Product code creation lag: 6 months. First verbatim case March 7, 2025. Product code created September 19, 2025. During this window, 38 cases were coded as semaglutide, tirzepatide, compounded GLP-1, or "unknown drug." Google Trends and NPDS exposures move in near-lockstep (r = 0.97), both driven by expanding grey-market availability.
New Finding: Multi-Substance Cases Are Where People Die
7-OH: Death + Major Rate by Exposure Type
All 5 deaths are multi-substance. Zero single-substance 7-OH deaths in the dataset. Co-ingestants on death dates include alcohol (3/5), gabapentin (2/5), fentanyl (1/5), and unknown substances (4/5).
Medetomidine: Death + Major Rate by Exposure Type
27.1% vs 6.9% severity. Multi-substance = illicit fentanyl adulterant population. Single-substance = clinical Precedex and veterinary Sileo. The data quantitatively proves this split.
Google Trends vs NPDS: Four Distinct Patterns
7-OH: NPDS Leads Google Trends (gas station distribution)
NPDS reached 20/month by March 2025. Google Trends didn't match until July 2025. For substances in physical retail, poison centers catch the signal first.
Melanotan II: Google Trends Leads NPDS by 3-6 Months
For substances discovered and ordered online, search interest is a genuine leading indicator. NPDS lags 3-6 months as the discovery-to-order-to-adverse-event pipeline plays out.
Medetomidine: GT Measures Awareness, NPDS Measures Poisoning
GT reflects DEA reports and news coverage. NPDS reflects actual clinical events concentrated in the PA/DE/WV fentanyl corridor. They measure different things.
Tesofensine & SLU-PP-332: Zero Google Trends Signal
0
Google Trends search volume
5 tesofensine exposures + 1 SLU-PP-332 exposure
All with NULL product codes
These substances spread through closed biohacking communities invisible to web search. Only NPDS verbatim surveillance catches them. The two data streams are synergistic, not redundant.
Code-Scattering: Same Substance, Multiple Generic Codes
Substances without product codes scatter more. Melanotan II landed in 7 different generic codes including Hematopoietics, Copper, and "Brain natriuretic peptide." Once a product code is created, cases consolidate. The verbatim is the only field that reconnects scattered cases.
Six Focus Substances at a Glance
Retatrutide
Exposures: 280 (258 rows)
Severity: 0D / 2 Major / 33 Mod
Generic codes: 4
Misspellings: 22
PC lag: 6 months
Grey-market GLP-1 triple agonist
7-Hydroxymitragynine
Exposures: ~831 combined
Severity: 5D / 48 Major
Deaths: ALL polysubstance
Bidirectional: 61 PC-only cases
Gas station opioid alkaloid
Medetomidine
Exposures: 165 (all) / 58 (single)
Severity: 1D / 32 Major (all)
Multi-sub rate: 27.1% D+Maj
Single rate: 6.9% D+Maj
Fentanyl adulterant, no antidote
Melanotan II
Exposures: ~36
Severity: 0D / 0 Major / 9 Mod
Generic codes: 7 (max scatter)
Miscoded as: brain natriuretic peptide, copper
Grey-market tanning peptide
Tesofensine
Exposures: 5
Severity: 0D / 1 Major / 2 Mod
Product codes: ALL NULL
100% verbatim-dependent
Abandoned pharma, grey-market weight loss
SLU-PP-332
Exposures: 1
Severity: No effect
Miscoded under: Evista (raloxifene)
Human safety data: ZERO
Exercise mimetic research chemical
Two Proposed NACCT 2026 Abstracts
Abstract #1: Primary (v4)
Focus: 6 substances, AI verbatim method + single-vs-multi severity
Lead finding: Retatrutide 6-month PC lag, 38 misattributed cases
Novel contribution: All 7-OH deaths are polysubstance; bidirectional detection gap
Scope: 1,339 exposures, 7 deaths, 98 major
Status: v4 drafted, ~497 words, pending Dan's JMP validation
Abstract #2: Ecosystem Discovery
Focus: 32 substances, social media intelligence pipeline
Lead finding: Growth hormone peptides 50% major rate; vendor brands in verbatims
Novel contribution: Reddit slang "GLP-3" finds cases invisible to all name searches
Scope: +562 exposures, 2 deaths, 14+ major from 32 additional substances
Status: v1 drafted, ~498 words, needs validation
Recommendation: submit both. If both accepted, they tell a two-part story: "here's the method" then "here's what the method found when we let it run." If only one accepted, ecosystem data becomes poster material for the primary abstract.
Decisions Needed
For the team
1. Submit one or two abstracts?
2. Title selection (3 options in v4 doc)
3. Include single-vs-multi in abstract or hold for poster?
4. HTC positioning (Al's read on v4 framing)
Timeline
Now to April 1: Dan validates in JMP; team reviews v4
April 1-15: Incorporate feedback, finalize
April 15: Team approval target
May 4: Submission deadline