AI + Human-in-the-Loop Explorer — Occupational Asthma

1) Basic parameters

Prevalence in the screened workforce

How common occupational asthma is in the group you’re screening (not in the general population).

AI sensitivity (true positive rate)

Chance the AI flags a worker who truly has occupational asthma.

AI specificity (true negative rate)

Chance the AI does not flag a worker who does not have occupational asthma.

Referral threshold (AI risk score cut-off)

Lower threshold → more referrals (more sensitive) but more false alarms. Higher threshold → fewer referrals but more misses.

Uncertainty in sensitivity/specificity

Real-world performance varies by site, device, protocol, and population. This adds a distribution (not a single number).

Human expert in the loop

Adds a second review step that can catch some false positives/negatives, especially around the threshold.

Expert catch rate for AI false negatives

If AI misses a true case, how often does a clinician/occupational physician still catch it (history, spirometry, PEF diary, exposure clues)?

Expert dismiss rate for AI false positives

If AI flags a non-case, how often does expert review prevent unnecessary escalation (alternative diagnosis, poor data quality, transient symptoms)?

Borderline share (hard cases near the threshold)

Real occupational asthma workups are messy; these are the cases where human judgement most changes the path.

Positive predictive value (PPV)

—

If the AI flags someone, probability they truly have occupational asthma.

Negative predictive value (NPV)

—

If the AI does not flag someone, probability they truly do not have occupational asthma.

Expected referrals per 1,000 workers

—

AI triage burden on clinics and on-site occupational health.

Expected missed cases per 1,000

—

The safety-critical metric (under-diagnosis risk).

Notes: This is a simplified educational model. Occupational asthma diagnosis typically involves expert clinical assessment plus objective testing (e.g., serial peak expiratory flow, spirometry with reversibility, bronchial provocation, exposure linkage). AI outputs should be treated as decision support — not definitive diagnosis.

2) Uncertainty visualisations

Posterior P(disease | AI flag) Posterior P(disease | AI not flag) With human-in-the-loop (if enabled)

False negatives False positives True positives True negatives

3) What humans add (qualitatively)

Human expertise	Why it matters for occupational asthma
Exposure reasoning	Links symptoms to tasks/agents (isocyanates, flour, latex, cleaning agents), timing, controls, PPE, and co-worker clusters.
Data quality checks	Interprets incomplete questionnaires, noisy spirometry, poor adherence to peak flow diaries, and confounders (smoking, infections).
Alternative diagnoses	Distinguishes asthma from COPD, vocal cord dysfunction, rhinitis, anxiety/hyperventilation, cardiac disease.
Ethical & legal judgement	Decides safe actions when evidence is uncertain (workplace adjustments, temporary removal, further tests, reporting).

Tip: Try dropping prevalence to 1% and watch PPV collapse even with “good” sensitivity/specificity — a classic reason AI needs expert oversight in low-prevalence screening.