toxic exposure extraction¶

Preview — under active development. This module is still being refined. Patterns, mappings, and category boundaries should be validated against your own corpus before being relied on for analysis or publication.

exposure_extraction.py walks the dashboard bundle and surfaces environmental, occupational, and toxic exposures relevant to ALS risk factor literature. It combines two extraction layers — structured ICD-10-CM codes and regex patterns over clinical narrative — and tags each finding with a representative ECTO (Environmental Conditions, Treatments, and Exposures Ontology) term where one has been verified.

Module source (exposure_extraction.py)

What it captures¶

The patterns and codes are drawn from the published ALS environmental risk factor literature. Each category is grounded in ECTO; the specificity of the ontology grounding varies (see ECTO grounding below).

Structured codes (ICD-10-CM)¶

Family	Codes	Captures
Nicotine	`F17.x`, `Z87.891`	Current and former nicotine dependence, history of smoking
Occupational	`Z57.x` (0–9)	Noise, radiation, dust, agricultural toxins, industrial chemicals, temperature, vibration
Substance contact	`Z77.0xx`	Lead, uranium, arsenic, aromatic compounds, asbestos, other hazardous chemicals
Pollution	`Z77.110–.118`	Air, water, soil pollution
Physical hazards	`Z77.121–.128`	Noise, mold, harmful algae / cyanotoxins
Body fluids	`Z77.21`, `Z77.22`, `Z77.29`, `Z77.9`	Hazardous body fluids, environmental tobacco smoke, other
Military	`Z91.82`, `Z56.82`	Personal history of military deployment, military deployment status
Trauma	`Z87.820`, `S06.0X0A`, `S06.9X0A`	Personal history of TBI, concussion, unspecified intracranial injury

Regex patterns (clinical narrative)¶

The pattern set covers 14 categories across the major ALS risk factor domains:

Category	Example patterns
Smoking	current / former smoker, pack-years, chewing tobacco
Military service	general service, Gulf War / OIF / OEF / OND, Vietnam + Agent Orange, Camp Lejeune
Pesticides	general pesticide, organophosphate, paraquat, glyphosate / Roundup, DDT, agricultural occupations
Heavy metals	lead, mercury, manganese / welding fumes, arsenic, cadmium
Solvents	TCE / PCE (context-required), benzene, formaldehyde, general organic solvents, welding
Asbestos	asbestos exposure, asbestos-containing material, mesothelioma risk
Head trauma	TBI history, multiple concussions, football career, contact sports, military blast injury
EMF / electrical	electrical worker, lineman, electrician, EMF / electromagnetic field exposure
Cyanotoxins	BMAA, cyanobacteria, blue-green algae, harmful algal blooms
Mold	mold exposure, black / toxic mold, stachybotrys, mycotoxin
Air pollution	PM2.5 / PM10, particulate matter, diesel exhaust, traffic-related air pollution
Occupational dust	silica, coal, wood, grain, cotton dust; silicosis; coal workers pneumoconiosis

Patterns that historically caused false positives are documented in comments inside the module:

NFL alone is not a contact-sport match — in an ALS cohort, NFL overwhelmingly means neurofilament light (a serum biomarker for motor-neuron-disease activity). NFL only matches when paired with explicit football vocabulary (played NFL, NFL player, NFL career, in the NFL, etc.).
Bare PERC is not a perchloroethylene match — in clinical notes, PERC is the Pulmonary Embolism Rule-out Criteria. PCE / PERC matches require explicit exposure context.
Bare TCE requires exposure / contamination / occupational context for the same reason — TCE clinical trial does not match.

ECTO grounding¶

Each exposure category is mapped to a representative term in ECTO. ECTO is an OBO Foundry ontology with ~2,700 exposure terms; it is the recommended terminology for exposure annotation in GA4GH Phenopackets and the GA4GH Human Exposome Data Standards Study Group, which is extending Phenopackets with formal schemas for exposure data.

The mapping table inside the module distinguishes three statuses:

verified — the ECTO term ID has been confirmed against the published SSSOM mapping file. These appear in the dashboard as clickable links to the OLS browser.
pending — no verified term ID yet. The dashboard displays "curation pending" rather than a guessed identifier. Adopters should search the OLS ECTO browser for the closest term and replace the None placeholder.
catch_all / out_of_scope — buckets that don't correspond to real ECTO concepts, or that ECTO doesn't cover (e.g. psychosocial exposures).

Currently verified IDs:

Category / agent	ECTO term	Label
smoking (all sub-patterns)	`ECTO:9000250`	exposure to nicotine
lead	`ECTO:9000945`	exposure to lead
mercury	`ECTO:0001571`	exposure to mercury
arsenic	`ECTO:9000032`	exposure to arsenic
asbestos	`ECTO:9000033`	exposure to asbestos
air pollution (PM inhalation)	`ECTO:0000977`	exposure to ultrafine respirable suspended particulate matter via inhalation

The remaining patterns and categories are flagged as pending and appear in the dashboard's ECTO Mapping tab as a curation worklist.

Outputs¶

File	Description
`exposure_codes.csv`	Structured ICD-10-CM matches
`exposure_extractions.csv`	Regex matches against narrative text

Both feed into exposure_dashboard.py for visual review, and into the downstream Phenopackets generator for formal encoding.

Usage¶

import exposure_extraction

result = exposure_extraction.main(
    bundle_path = './dashboard_data.json',
    out_root    = './',
)

print(result)
# {'exposure_codes_csv': './exposure_codes.csv',
#  'exposure_extractions_csv': './exposure_extractions.csv',
#  'code_rows': N,
#  'extraction_rows': M}

Important caveat¶

Most environmental and occupational exposure information lives in social-history narrative sections of clinical notes rather than as structured ICD-10 codes. The Z57.x and Z77.x families are systematically under-coded in clinical practice; the bulk of signal in any ALS-cohort run of this module typically comes from the regex side scanning narratives. Adopters should not over-interpret the absence of structured codes as absence of exposure, and detection sensitivity will vary substantially across EHR vendors and clinical-documentation conventions.

Exposure dashboard — visualization layer for the outputs of this module.
Note extraction — companion module that surfaces structured ALS findings (ALSFRS-R scores, FVC% predicted, El Escorial category, etc.) from the same narrative sources.
GA4GH Phenopackets — downstream consumer of these exposure annotations.