Breath-Sensor Project – Team Meeting Notes

Meeting Overview & Context

• Informal, multi-chapter discussion among a small R&D group (core speakers: project lead, Ananda, Vijay, Abhishek, Dheeraj, and unnamed team-mates).
• Primary theme: designing a breath-analysis device that accurately measures/infers ethanol, acetone, and other VOCs (volatile organic compounds) while rejecting confounding signals (e.g.
  smoking-related gases).
• Deliverable sought: a concrete “plan of action” to be finalized in tomorrow’s follow-up meeting.

Core Objectives & Immediate Tasks

• Determine realistic post-drinking ethanol concentrations in exhaled breath, specifically 3 h after first drink.
• Translate that concentration into the equivalent response on the acetone sensor already in use.
• Decide whether two well-chosen sensors + AI selectivity are sufficient, or if multiple specialty sensors are still required.
• Survey, analyse, and shortlist commercial acetone / total-VOC sensors (including one found in China and others shared at 4 a.m.).
• Read two peer-reviewed papers (sent via chat) that remove ethanol interference using lightweight ML (PCA + KNN) suitable for TinyML deployment.
• Summarise findings and text/call the lead by 10 a.m. tomorrow as a reminder.

Ethanol Concentration in Breath (Post-Drink)

• Normal baseline (sober): essentially $0\,\text{PPM}$.
• Observed post-drink, 3 h mark:
  • Common field values reported $50-60\,\text{PPM}$.
  • Upper extreme witnessed one evening: $\approx 200\,\text{PPM}$.
• Device must therefore reliably quantify/flag ethanol anywhere within $0-200\,\text{PPM}$ (or higher for safety margin).

Current Acetone Sensor Behaviour

• Baseline output (no detectable VOC): $1.5\,\text{V}$.
• Example calibration point supplied during meeting:
  • $100\,\text{ppb}$ acetone ⇒ $2.0\,\text{V}$.
  • Net change from baseline: $\Delta V = 2.0 - 1.5 = 0.5\,\text{V}$.
• All future discussion framed as “voltage change from baseline” rather than absolute volts.
• Need to derive $\Delta V_{\text{ethanol}}$ for the 0–$200\,\text{PPM}$ ethanol window so AI can subtract it out.

Acetone-Equivalent Mapping for Ethanol

• Goal: produce a function $f(\text{ethanol PPM}) \rightarrow \Delta V_{\text{sensor}}$.
• If $\Delta V<em>{\text{ethanol}} \ll \Delta V</em>{\text{acetone}}$ at same molar concentration, AI selectivity should succeed easily.
• Mapping unknown; requires lab data or spec-sheet interpolation.

AI / Machine-Learning Strategy

• Literature shows ethanol artefact removal using off-the-shelf multivariate algorithms:
  • PCA (Principal Component Analysis) for dimensionality reduction.
  • KNN (k-Nearest Neighbours) for classification/regression.
• Advantages:
  • Lightweight enough for TinyML (microcontroller-scale deployment).
  • Avoids cloud dependency; lower latency & privacy risk.
• Two key papers shared in chat (still uploading during call); team must read before next meeting.

Candidate Sensors & Performance Constraints

• Chinese acetone sensor highlighted by Ananda:
  • Promising slope (lower ethanol cross-sensitivity) but detection range starts at $1\,\text{PPM}$—may be marginal for sub-PPM needs.
• Previously used Cigaro sensors at IIT:
  • Successfully pushed to $600–700\,\text{ppb}$ by electronic tweaking.
  • Limit of detection (LOD) historically $1.5–2.0\,\text{PPM}$.
  • Acceptable if electronic noise $\pm500\,\text{PPM}$ remains manageable.
• “Other links” acetone sensor (shared 4 a.m.) also on watchlist; awaiting vendor response.
• Team discussing possible discount with “Iron Science” supplier.

Experimental Range & Electronic Considerations

• Acceptable electronic margin: handling $0.5\,\text{PPM}$ to $1\,\text{PPM}$ gradations is feasible.
• Baseline drift, noise, and calibration constants (ethanol historically used for MOx sensors) still major pain-points.
• Emphasis on characterising $\pm$ range rather than single points for robust ML training.

Smoking, Cannabis & Total-VOC (TVOC) Complications

• Additional use-case: detecting cannabis breath signature while excluding normal smoking artefacts.
• Smoking changes many <10 eV VOCs: benzene, toluene, nitric oxide, etc.
  • These minor (~1 %) contributors can cumulatively jump $\approx100–200\,\text{ppb}$.
• Idea: maintain median TVOC profile (PPM / PPB) for typical smoker vs non-smoker vs cannabis user.
• If TVOC for cannabis ≈ smoking TVOC and acetone fixed, uniqueness disappears—making separation impossible.
• Therefore plan to catalogue every gas <10 eV that rises in cigarette & hookah smoke to build exclusion model.

Multiple-Sensor vs Two-Sensor Debate

• Original thought: deploy an array of diverse MOx/NDIR sensors to tease apart signatures.
• Current thinking: may be “over-complicating.”
  • Two carefully chosen sensors + strong AI selectivity might suffice.
  • Reduces BOM cost, power, firmware complexity.
• Decision postponed until tomorrow’s ‘final plan of action’ meeting.

Logistical & Administrative Notes

• Bills need to be handed over to VIT for reimbursement.
• Team members’ availability:
  • Major follow-up call/meeting on Tuesday; conflicts 8–10 a.m. & 5 p.m.
  • Reminder text to be sent at 10 a.m. tomorrow.
• Some colleagues not joining night shift (Abhishek, Dheeraj).
• Voice notes to be recorded & shared for asynchronous coordination.

Ethical, Practical & Real-World Relevance

• Breath diagnostics intersect with public safety (alcohol detection), metabolic health (diabetic ketoacidosis via acetone), and substance monitoring (cannabis).
• On-device ML preserves user privacy; avoids transmitting sensitive biometric data to cloud servers.
• Accurate discrimination prevents false-positive DUI or medical misclassification.
• Over-engineering (too many sensors) increases cost and may limit manufacturability; under-engineering risks mis-detection.

Numerical & Formula Reference Sheet

• Normal sober ethanol: $\approx0\,\text{PPM}$.
• Observed post-drink range: $50-200\,\text{PPM} \; (3\text{h})$.
• Calibration example: $100\,\text{ppb acetone} \rightarrow 2.0\,\text{V}$ (baseline $1.5\,\text{V}$, so $\Delta V = 0.5\,\text{V}$).
• LOD for Cigaro sensor: $1.5-2.0\,\text{PPM}$.
• Noise tolerance mentioned: $\pm500\,\text{PPM}$ (context: electronic gambling comment).
• Target TVOC anomalies from smoking/cannabis: $\approx100-200\,\text{ppb}$ aggregated.

Immediate Next Steps (Action Items)

• Look up ethanol ↔ sensor response curve; compute $\Delta V_{\text{ethanol}}$ for $0-200\,\text{PPM}$.
• Read & summarise two ML papers; prepare talking points for TinyML implementation.
• Receive vendor quotations / slopes for short-listed acetone sensors (Chinese + 4 a.m. link + “other links”).
• Decide tomorrow: two-sensor ML architecture vs multi-sensor array.
• Assemble all bills; deliver to VIT admin for reimbursement.
• Share voice note detailing order numbers and pending sensor purchases.