AEGIS · PRAS Console

Pulmonologist reviewing a spirometry flow-volume loop on a diagnostic console

No spirometer needed for the screening call. Enter age, smoking history, and two cheap symptom/risk-factor scores — the console flags who should get a spirometry test next, it does not replace one.

1 · Patient — Screening Inputs

Symptom and risk-factor inputs only. No FEV1/FVC reading required at this stage.

Age (yrs)55

Pack-history (years)12

CAT score (0–40)5

Peak flow (L/min)430

Current or ex-smoker

Smoker

Non-smoker

Asthma comorbidity

Yes

Result — Screening Call + Severity Estimate

Logistic-regression screen decides the class; linear-regression estimate grades predicted FEV1% (both distilled from the trained SVM/SVR pair, see footer).

NORMAL

AT-RISK

OBSTRUCTED

Confidence in screening call—

Estimated FEV1 (% predicted)—

30 (severe)110 (normal)

ROUTINE — monitor

No obstructive signal detected.

screen

2 · Medical resource planning — COPD / pulmonary population

🌍 GLOBAL COVERAGE — planning anchored to the real health-system spend of 24 countries. Choose your market; COPD-cohort cost, bed-days and avoidable spend re-compute in your economy.

Screened pulmonary / COPD-risk patients in your service

Market (per-capita cost anchor)

Annual care cost (pulmonary cohort)—

Expected inpatient bed-days / yr—

Cost avoidable via early cessation / bronchodilator—

P1 · NASNurses for follow-up (@500/nurse)—

Spirometry referrals to pre-empt (shadow cohort)—

Ties the individual call to the system view: every At-risk patient flagged here becomes a spirometry referral and nurse-follow-up load (Project 1). Catching airflow limitation early — before COPD — is where the avoidable cost lives.

A pulmonologist studying spirometry waveforms, making the final decision

The system flags who needs a spirometry test. A clinician makes every call.PRAS surfaces its own uncertainty and hands the judgement to a human — augmenting the clinician, never replacing them.

How the numbers are derived — data vintage & method

Data sources & vintage

L1 · Airflow model	Author’s doctoral thesis — Vol.1 Ch.3 “Lung Ventilation Modeling & Assessment” (damped-oscillator R/C indices). Single illustrative case, so L1 substitutes real GOLD spirometry (disclosed).
L2 · Spirometry / severity	101 real GOLD-consistent COPD patients (FEV1/FVC, severity labels). Normal class = documented ATS/ERS reference (FEV1/FVC ≈ 0.80 ± 0.05) — a disclosed modelling assumption, not real controls.
L3 · Population risk	300 real asthma-cohort patients (smoking / comorbidity signal).
L4 · Hospitalisation	UCI Diabetes 130-US-Hospitals respiratory subset (ICD-9 490–496 / 466 / 480–488 / 786). Encounters 1999–2008 (15,108; readmit30 8.1%, mean LOS 4.2 d).
L5 · Economics	Umbrella 24-country per-capita anchors (WHO / World Bank, compiled mid-2026, SGD), wired into the market-planning panel above (patients × COPD severity multiplier 1.0 / 1.5 / 3.0 × per-capita anchor; avoidable = At-risk × per-capita × 45% early-cessation effect).

Method — illustrative screening scaffold, not billed figures:

Screening call	logistic-regression surrogate of the full SVM (6 features: age, smoking, pack-years, CAT, asthma, peak flow) — 99.1% agreement with the full SVM (ACC 96.3%, macro-F1 0.90).
Severity estimate	linear-regression surrogate of the SVR predicting FEV1% predicted — R²=0.85 vs the full SVR.
Cost method	wired into the market-planning panel above: patients × COPD severity multiplier (Normal 1.0 / At-risk 1.5 / COPD-Obstructed 3.0) × per-capita anchor; avoidable = At-risk × per-capita × 45% (early smoking-cessation / bronchodilator effect). Per-capita anchors: WHO / World Bank, compiled mid-2026, SGD.

Cohort is the synthetic PRAS-RP dataset (6,000 patients). Every row carries the SYNTHETIC label; a clinician makes every call.

What this is. A screening-only console: predicts a 3-class pulmonary risk band (Normal / At-risk / COPD-Obstructed) and an estimated FEV1% from age, smoking history, symptom burden (CAT score), asthma comorbidity, and peak flow — deliberately excluding the actual FEV1/FVC spirometry reading, so it can be used before a spirometer is involved, to decide who needs one.

Champion models (server-side, full precision). SVM classifier (screening-only 6-feature set) — 96.3% accuracy, macro-F1 0.90, ROC-AUC (OvR) 0.986. Per-class recall: Normal 99%, At-risk 97%, COPD-Obstructed 70% — the model reliably catches Normal and At-risk, but under-catches confirmed obstruction from symptoms alone. That is the honest, expected finding for a screening tool built without a spirometry reading: severe disease still needs the actual test to confirm, not a symptom-based guess. SVR severity regressor (same 6 features, predicting FEV1% predicted) — R²=0.61, MAE=7.98.

Browser distillation. This page runs a logistic-regression surrogate of the SVM (99.1% agreement with the full model on held-out data) and a linear-regression surrogate of the SVR (R²=0.85 vs the full model) — same distillation pattern as the RRAS/DRAS consoles, so the browser needs no server and no Python at runtime.

Data sources (5-layer evidence stack, PRAS-RP synthetic cohort, 6,000 patients). L1+L2 spirometry/severity: 101 real COPD patients (GOLD-consistent FEV1/FVC, COPD severity labels) for the At-risk/Obstructed classes; the Normal class uses a documented ATS/ERS population reference (FEV1/FVC ≈ 0.80 ± 0.05) since the real cohort has no healthy controls — this is a modelling assumption, not real patient data. L3 population: 300 real asthma-cohort patients for smoking/comorbidity signal. L4 utilisation: 15,108 real respiratory hospital encounters (130-hospital dataset, ICD-9 490–496/466/480–488/786 subset; readmit30=8.1%, mean LOS=4.2 days), resampled by class. L5 economics: Umbrella 24-country per-capita anchors (WHO / World Bank, compiled mid-2026, SGD), wired into the market-planning panel above — patients × COPD severity multiplier (1.0 / 1.5 / 3.0) × per-capita anchor, avoidable = At-risk × per-capita × 45% early-cessation effect. Illustrative planning figures, not billed costs.

Theoretical grounding. Thesis Vol.1 Ch.3 “Lung Ventilation Modeling and Assessment”, S3.12–3.13, develops a damped-oscillator model of lung volume in terms of airflow-resistance (R) and compliance (C), proposing R/C-derived indices for differential diagnosis of obstructive lung disease — the same theoretical role Vol.1 Ch.1 played for DRAS and Vol.2 App.G played for RRAS. Unlike those two, Ch.3 does not contain a multi-subject R/C table to resample from (only a single two-lung illustrative case) — so this console's L1 layer substitutes real GOLD-consistent spirometry for resampled thesis kinetics. That substitution is disclosed here, not hidden.

Status: SYNTHETIC cohort and distilled screen — a research scaffold, not a validated clinical tool. ⚠ Human-in-command: every flag is a prompt for a clinician-ordered test, not a diagnosis.