What this chapter locks down

A sleep monitoring headband succeeds or fails at the system boundary—not at any single IC. This chapter defines (1) what the headband measures, (2) where decisions are made, and (3) what “done right” means using measurable acceptance metrics.

Scope note: This page focuses on hardware evidence chains (noise, artifacts, timing, power, data integrity). It does not claim medical diagnosis or cover cloud/app architecture.

Target signals and what each contributes (engineering view)

• EEG (microvolt domain): primary channel. The engineering challenge is not “band selection,” but input-referred noise, common-mode rejection, and electrode impedance drift that can bury low-frequency content. EEG is a “contact + AFE + system coupling” problem.
• Acceleration (optional): used mainly as a discriminator. If EEG variance spikes exactly when acceleration indicates a roll-over, treat it as artifact until proven otherwise.
• PPG (optional): a secondary channel for trend features (e.g., HR/variability cues) and quality gating. Its LED switching introduces a high-risk coupling path into the EEG chain, so timing and power partitioning must be designed.

Pipeline overview (where smart-wake actually lives)

The pipeline must be split into two loops: (A) an on-device closed loop that can trigger wake reliably without depending on a phone, and (B) a data delivery loop that preserves a clean timeline for review and analysis.

• Closed loop (real-time): electrodes → EEG AFE → ADC → MCU/DSP features → smart-wake decision → vib/buzzer driver. This loop is judged by latency and false trigger control.
• Delivery loop (integrity): sample timestamp → packetization (seq + CRC + gap marker) → BLE transport → phone (optional sync) + local flash ring buffer (for reconnection recovery).

“Done right” success criteria (measurable, testable)

Define acceptance in metrics that can be measured on the bench and in overnight pilots:

First measurements to de-risk the whole project

• Two-channel capture: EEG AFE output + a chosen rail ripple point (AFE rail or PMIC node).
• Contact quality trace: impedance estimate (or proxy) over time alongside EEG baseline.
• Timeline proof: packet sequence counter + timestamp drift across a full overnight session.
• Current profile: average current + peak current during BLE bursts and haptics events.

These measurements will directly support H2-2 (reality check), H2-8 (power), H2-9 (BLE), and H2-11 (validation).

Why EEG is “system-hard” (not component-hard)

EEG lives in the microvolt domain and is dominated by low-frequency noise and common-mode interference. In practice, the limiting factors are often electrode contact, CMRR under real impedance imbalance, and coupling from power/RF switching—not the advertised ADC resolution.

Bandwidth planning (engineering-first, minimal theory)

• Passband intent: set a low-frequency corner that preserves slow dynamics while avoiding baseline runaway. Keep room for artifact discrimination rather than chasing aggressive filtering.
• Anti-alias reality: the ADC sampling rate and analog front-end bandwidth must ensure that out-of-band interference (switching edges, RF burst harmonics) does not fold into the EEG band through non-idealities.
• Notch strategy: 50/60 Hz mitigation is a symptom response. The real goal is to maximize CMRR + contact balance so notch filtering is not doing all the work.

Noise budget (who dominates, and when)

Organize noise into three buckets that can be measured independently:

• AFE intrinsic noise (incl. 1/f): dominates when contact is stable and layout is quiet. Signature: smooth low-frequency “grain” with no event correlation.
• Electrode/skin interface noise: dominates when impedance drifts or becomes imbalanced. Signature: baseline wander, slow steps, sporadic saturation; often correlates with posture changes or sweat.
• System coupling noise (power/RF/switching): dominates when rails bounce or RF bursts inject common-mode. Signature: periodic spikes or envelope modulation aligned with BLE events, charging, or LED pulses.

Interference map (source → coupling path → signature → proof)

Each interference source must be paired with a proof measurement. The goal is to avoid ambiguous blame (e.g., “the algorithm is bad”) when the root cause is a measurable coupling path.

Decision goal: pick a topology that survives real-world impedance imbalance

EEG is a microvolt signal chain that fails most often in the real-world CMRR regime: electrode impedance is not balanced, contact changes over the night, and the system injects common-mode energy (mains, charging, RF bursts). This chapter selects the AFE architecture by the specs that actually map to field symptoms and measurable evidence.

Front-end options: integrated EEG AFE vs discrete INA + ADC

Both approaches can work, but they fail differently. Choose based on what you can control and validate.

• Integrated EEG AFE: typically combines input biasing/lead-off, high input impedance, low-noise PGA/ADC, and sometimes DRL/RLD support. Strength is faster convergence to a stable baseline and fewer “board-level” traps. Risk is less freedom to tune protection, reference, and gain staging.
• Discrete INA + ADC: allows selecting a specific INA, anti-alias network, reference, and ADC bandwidth/ENOB. Strength is controllability; risk is that layout/return paths, reference noise, and protection leakage can dominate unless the team can measure and iterate.

Specs that matter (and what they actually protect you from)

The most expensive mistake is optimizing a “headline spec” while ignoring the spec that dominates the field failure mode. Use this list as the minimum checklist.

• Input impedance (and balance): high input impedance reduces loading, but the critical issue is impedance imbalance across electrodes. Even a great datasheet CMRR can collapse when imbalance grows.
• Input bias current / leakage paths: bias current times electrode impedance becomes a slow-varying offset. In a headband, this shows up as baseline drift, slow recovery after motion, or “saturation that takes minutes to settle.”
• CMRR (real-world): the true target is not the lab CMRR number, but CMRR under intentionally imbalanced electrode impedances. This is the strongest predictor of 50/60 Hz hum complaints.
• Input range / headroom: EEG is small, but artifacts are not. Headroom must survive motion-induced steps, half-cell potentials, and recovery from transients without flatlining.
• Programmable gain (PGA): gain must be set by worst-case artifact headroom, not by “max sensitivity.” Over-gain yields clipping, then downstream filters create false structure.
• ADC ENOB in-band (not nominal bits): evaluate ENOB or noise density in the EEG band. “High resolution” without low in-band noise does not improve readability.

Reference and ground strategy (the hidden AFE input)

In practice, the reference node and return paths behave like an additional input to the AFE. The goal is to keep power/RF switching energy out of the AFE reference and to control the return current geometry.

• Analog ground island: keep the EEG AFE + reference network inside a quiet return region. Join to digital ground at a controlled point near the ADC/AFE boundary.
• Reference buffering + decoupling: the ADC/AFE reference noise directly appears as measurement noise. Place short decoupling loops; avoid routing switching nodes near the reference network.
• ADC clock hygiene: avoid routing clock edges through sensitive input regions; clock-coupled interference can look like periodic texture or elevated noise floor.

Spec → symptom mapping (convert complaints into measurable causes)

Use this table as a diagnostic bridge between datasheet thinking and field behavior.

This mapping is designed to be reused in H2-11 (validation) and H2-12 (FAQ) without scope creep.

Why electrodes are a first-class design variable

The electrode-skin interface sets the headband’s real-world noise floor and real-world CMRR. Over an overnight session, sweat, pressure relaxation, and micro-slip can change impedance by orders of magnitude, which converts common-mode energy into in-band noise and causes saturation and slow recovery.

Electrode options (dry / semi-dry / fabric) and what they tend to break

• Dry electrodes: simplest and cleanest mechanically, but impedance is highly skin-dependent. Failure mode is impedance steps and imbalance during motion, which reduces real-world CMRR.
• Semi-dry / gel-assisted: improves impedance stability and reduces hum sensitivity, but introduces risks: leakage paths, contamination, and time-varying wetting that can change bias conditions.
• Fabric electrodes: comfortable and wearable-friendly, but performance depends on pressure distribution, sweat paths, and connector strain relief. Failure mode is slow drift plus sporadic motion artifacts.

Measuring contact quality (impedance trend + imbalance + artifact counters)

“Good contact” must be measurable. A practical headband uses a layered approach: impedance estimation, imbalance detection, and signal-quality counters that feed a single Contact Score.

• Impedance estimate (trend): use a low-impact check (test tone or bias observation) to track impedance over time. The trend is more valuable than an absolute number.
• Imbalance indicator: when one side drifts more than the other, the system’s real-world CMRR collapses. Track a simple imbalance metric and flag “hum-risk.”
• Artifact counters: count clipping segments, saturation recovery time, and sudden baseline steps. These counters separate “contact failure” from “algorithm failure.”
• Contact Score (0–100): combine impedance stability, imbalance, hum peak proxy, and clipping rate into one score. Use it to gate smart-wake decisions and to annotate logs.

Mechanical layout: tension zones, strain relief, and sweat paths

Mechanical design is signal conditioning. The goal is stable pressure at electrodes without discomfort, and isolation of motion and sweat from the electrode interface.

• Tension zones: define a stable electrode zone (controlled pressure), a comfort buffer zone, and a routing zone that avoids pulling on electrode nodes.
• Strain relief: every connector or flex transition must protect the electrode node from tugging, otherwise micro-slip appears as baseline steps.
• Sweat paths: route sweat away from the reference/electrode junctions; avoid creating electrolyte bridges that change leakage and bias conditions.

Impedance-to-symptom evidence (what to log overnight)

These logs feed H2-10 smart-wake quality gating and H2-11 validation pass/fail criteria.

Objective: prove “artifact vs real EEG change” before filtering

Motion artifacts are not just “noise”—they have repeatable signatures that can be distinguished from physiological changes using a minimal evidence chain. This chapter builds a discriminator that combines EEG waveform cues with accelerometer correlation and contact/impedance indicators, then applies common-mode control (DRL/RLD) and minimal filtering without erasing sleep-relevant features.

Artifact signatures: a practical waveform dictionary

Treat each artifact class as a diagnosable failure mode. The goal is not to “make the plot look smooth,” but to isolate the cause so the fix survives overnight variability.

• Saturation / clipping: flat-topped segments or rails-hit behavior, often followed by slow recovery. Typical causes include insufficient headroom, sudden contact steps, or bias/leakage shifts. Evidence: clipping counter, % time clipped, recovery time constant.
• Baseline wander: low-frequency drift that rides on the signal like a slow wave. Typical causes include impedance drift, sweat-driven leakage, or electrode imbalance that collapses real-world CMRR. Evidence: drift trend correlates with impedance trend or imbalance indicator.
• Transient spikes / bursts: narrow impulses or spike trains that appear during micro-slip, cable strain, ESD recovery, or coupling from switching/RF events. Evidence: time alignment with accel spikes or event markers; spike rate increases during motion windows.

Discriminators: how to separate artifact from real EEG change

A single signal channel is rarely enough. Use a minimal multi-sensor discriminator so decisions remain stable across users and overnight conditions.

Common-mode control: DRL/RLD loop (engineering-level)

DRL/RLD is a common-mode control loop: it senses common-mode behavior at the AFE input and drives the reference electrode to pull the system into a more controlled operating region. The goal is improved effective CMRR under real-world impedance imbalance.

• What it improves: reduces sensitivity to mains/common-mode injection and impedance imbalance.
• What can go wrong: unstable or over-aggressive loop gain can introduce low-frequency oscillation-like behavior, extra artifacts, or “texture” that looks like signal.
• Stability mindset: validate worst-case electrode impedance conditions (high + imbalanced) and confirm the loop remains well-behaved under motion and sweat-driven drift.

Filtering strategy: minimal notch + controlled high-pass (do not erase features)

Filtering should be a controlled, evidence-driven step—not a replacement for contact and common-mode control. Over-filtering can suppress sleep-relevant dynamics and leave the system blind to quality issues.

• Notch (50/60 Hz): use as an assist, not as the primary fix. If notch is required constantly, the system is likely limited by real-world CMRR or contact imbalance.
• High-pass (baseline control): apply conservatively to reduce drift, then verify that discriminators still separate artifact windows from stable windows.
• Quality flags: when artifacts are detected, mark the window rather than hiding it—so later logic can avoid false decisions.

Scope: PPG as an optional helper signal (coexistence-first)

If PPG is included in a sleep headband, it should be treated as an optional helper channel for trend metrics and quality gating—not as a separate product category. The engineering focus is the optical chain and the coexistence problem: LED switching and return currents can contaminate the EEG front-end unless power/ground partitioning and scheduling are designed in from the start.

When PPG is worth adding (and when it is not)

• Worth adding: when HR/HRV trends help smart-wake gating, and when PPG can provide a robust “LED on/off subtraction” self-check to validate optical integrity.
• Risky to add: when the system cannot schedule LED pulses or cannot isolate LED current return paths, because EEG cleanliness will degrade and debugging becomes ambiguous.

Optical chain (engineering view): LED → tissue → PD → TIA → ADC

A headband PPG chain is a time-structured measurement. LED pulses are a strong interference source, while the photodiode/TIA stage is highly sensitive to ground/reference integrity.

• LED driver: controlled current pulses; edge rates and peak current define EMI and rail droop risk.
• Photodiode + TIA: converts small current to voltage; vulnerable to ground noise and coupling.
• Ambient cancellation timing: measure LED-on and LED-off (or multi-phase) to subtract ambient and isolate the optical component.

EEG–PPG coexistence: coupling paths and scheduling windows

The primary hazard is that LED switching introduces noise into the EEG chain through shared rails, shared return paths, and near-field coupling. The fix is to treat coexistence as a system requirement.

• Power coupling: LED current pulses cause rail droop and ripple; if EEG AFE shares that rail or poor decoupling exists, in-band noise rises.
• Return-path coupling: LED current return flowing through AFE reference/AGND regions injects common-mode energy.
• EM coupling: LED driver switching nodes routed near electrodes/reference or AFE inputs create spikes and texture.
• Scheduling windows: schedule LED pulses in bounded windows, generate event markers, and flag EEG windows affected by LED activity.

What to measure (minimal) to prove coexistence is healthy

Objective: prevent “quiet failure” from drift and gaps

Sleep analytics can fail silently when timestamps drift, packets reorder, or sampling gaps are masked by retry logic. A reliable system treats time as an engineered asset: the analysis timeline should be anchored to a sample counter, while RTC provides coarse session alignment. Transport events (e.g., BLE connection timing) are not a substitute for a trustworthy sampling timebase.

Clock + timestamp policy (counter-first, RTC for alignment)

The sampling timeline should be defined by the device, not inferred from the phone. Use a monotonic sample counter (or frame counter) as the primary axis for analysis, then attach coarse wall time using RTC to segment a night and support session boundaries.

• Sampling clock choices: ADC/AFE clock, MCU timer-derived clock, or external crystal-driven clock. The key is predictable rate and bounded drift over hours.
• Drift mindset: ppm-level error accumulates across an entire night; treat drift per night as a tracked metric, not as an assumption.
• Timestamp layering: (1) sample counter for per-sample spacing, (2) RTC for coarse absolute alignment, (3) transport timing only as a delivery marker.

Packetization strategy: seq + CRC + explicit gap markers

Packetization should make integrity observable. The system should detect missing, duplicate, and out-of-order payloads, and it should explicitly encode sampling discontinuities so the analytics stack does not “smooth over” real gaps.

• Sequence counters: detect drop, reorder, and duplicate delivery at the packet level. Use wrap-safe arithmetic.
• CRC: detect payload corruption in transport or storage. Track CRC-fail counts as a nightly metric.
• Gap markers: insert an explicit gap event when sampling pauses (brownout, buffer overflow, sensor restart), including the counter range or duration.

Local buffering vs streaming (what to log when the link drops)

Streaming alone can hide data loss behind reconnects and retries. Local buffering preserves continuity and allows integrity auditing even when the phone link is unstable.

• Ring buffer concept: store frames + integrity fields into flash/FRAM with write pointer wrap. Track high-water marks and overwrite events.
• Link drop behavior: continue sampling into the buffer; when the link recovers, transmit either buffered data or a summarized integrity report, depending on power/bandwidth constraints.
• Minimum metadata to keep: max continuous gap, gap count, buffer overflow count, reset/brownout reason, and a timebase drift estimate per night.

Nightly integrity metrics (minimum required)

Convert integrity into measurable pass/fail indicators. These numbers should be computed every night and retained alongside the session summary.

• Missing sample rate: missing samples / expected samples (from counter continuity)
• Max continuous gap: worst discontinuity duration (counter delta to time)
• Drift per night: relative drift estimate vs RTC/session boundary references
• Reorder/duplicate counts: from sequence counter anomalies
• CRC fail count: from payload verification

Objective: runtime is an architecture problem, not a parts list

Night-long operation depends on state architecture, rail partitioning, sequencing, and burst handling. A “low-power part” can still drain the battery if the system spends too much time in high-current states, or if burst events cause droop, brownouts, and data gaps.

Power states: define and mark them (firmware markers)

Build the power tree around a state machine. Each state should be observable in logs and current traces using firmware markers, enabling fast correlation between energy events and data integrity issues.

• Deep sleep: RTC alive, wake sources armed, minimal rails enabled.
• Sensing-only: EEG AFE + ADC running, MCU/DSP steady, transport minimized.
• BLE burst: packet transmit / retry, RF rail peak current, memory reads, scheduler activity.
• Vib/buzzer event: short high-current pulse, potential ground/rail injection risk.

Rail partition: keep the AFE quiet while RF and haptics burst

Partition rails so that pulsed loads do not share sensitive reference paths with the EEG front-end. The goal is not maximum separation everywhere, but controlled return paths and predictable behavior under bursts.

• AFE rail: prioritize low noise and stable reference; validate ripple and transient response.
• Digital rail: efficiency-driven, but with controlled switching ripple and return routing.
• RF rail: supports peak current; local decoupling and short loops matter.
• Haptics rail: isolate pulsed return paths; avoid contaminating AFE ground island.

PMIC strategy: buck + LDO where it matters (noise vs efficiency)

Use DC-DC conversion for efficiency where ripple tolerance exists, and use LDO (or a quiet supply strategy) where the EEG AFE requires isolation. Sequencing should ensure the AFE enters a stable operating region before bursty domains begin aggressive switching.

• Buck domains: digital + RF rails with validated peak current support.
• LDO domains: AFE rail, references, and any ultra-sensitive analog nodes.
• Sequencing: avoid “RF starts first” scenarios; ensure AFE bias/reference stability before high-activity windows.
• Brownout robustness: track UVLO/brownout counters and reset reasons; treat them as nightly health metrics.

Current profiling plan: shunt + coulomb counter + state markers

Current profiling should yield repeatable waveform archetypes for each state, plus rail noise checkpoints. Combine fast waveform capture (shunt + scope) with long-term energy accounting (coulomb counter/fuel gauge) and firmware markers so spikes can be attributed, not guessed.

• Shunt + scope: capture burst peaks and inrush; verify the droop margin on RF and haptics.
• Coulomb counter: estimate nightly energy per state; compare against expected duty cycles.
• Markers: log transitions (deep sleep → sensing → BLE burst → vib) to overlay on current traces.
• Noise points: measure ripple on AFE rail and reference nodes during worst-case bursts.

Objective: make BLE auditable, recoverable, and front-end friendly

Overnight sessions need predictable delivery and fast recovery, not peak throughput. Robustness means: the link exposes quality counters, reconnection behavior is deterministic, and RF bursts do not inject noise into sensitive analog domains. Transport success is measured by integrity metrics, not by RSSI alone.

Connection policy: stable rhythm beats raw bandwidth

Treat BLE as a scheduled delivery channel. Choose connection settings that keep energy and burst activity bounded, while maintaining a consistent reporting cadence that aligns with local buffering and integrity checks.

• Connection interval: shorter intervals increase airtime and burst frequency; overnight logging often prefers a calmer cadence with periodic bursts.
• Slave latency: skipping events can save power, but must not hide prolonged delivery stalls; use quality counters to detect “connected but not delivering.”
• PHY choice: select for robustness vs airtime as a system tradeoff; verify using PER/retry counters rather than assumptions.
• Application integrity: rely on seq/CRC/gap markers for truth; link-layer retries must not mask gaps in the analysis timeline.

Reconnection playbook: advertise in windows, keep sampling locally

Disconnections should be treated as routine. The device continues sampling into local storage and uses windowed advertising to regain the link without draining the battery.

• Detect: missing connection events and rising retry/PER counters trigger a “quality degraded” state.
• Recover: use a fast reconnection window first, then shift to low-duty advertising windows to protect battery life.
• Resume: after reconnect, transmit either buffered data or an integrity summary first (gap count/max gap/seq continuity), then backfill.
• Never guess continuity: reconnection success does not imply data continuity; gaps must be explicit (H2-7).

RF burst coexistence: prevent ground bounce from contaminating EEG

BLE transmissions can create pulsed current demand on the RF rail and return paths. If analog reference nodes share these paths, microvolt-level EEG measurements can be contaminated. Coexistence must be engineered across placement, decoupling, and scheduling.

• Placement: antenna keep-out near sensitive analog nodes; avoid routing RF return loops near AFE reference/inputs.
• Decoupling: dedicate RF rail decoupling close to the radio; avoid sharing burst return paths with AFE ground islands.
• Scheduling: when feasible, align transmit bursts with less sensitive analysis windows, while keeping sampling continuous.

Privacy (high-level): identity rotation without breaking integrity

Keep privacy mechanisms high-level and conservative. Identity rotation can reduce tracking risk, but it must not compromise session integrity, reconnection reliability, or the ability to audit nightly data continuity.

• Rotate identifiers: limit persistent identity exposure while maintaining stable session accounting on-device.
• Integrity first: seq/CRC/gap markers and quality counters remain the primary truth source.

Objective: a quality-controlled decision pipeline, not a black box

Smart wake should be explained as an engineering pipeline: signals are converted into simple features, quality gates prevent decisions on noisy data, and a scheduler enforces a wake window with a hard deadline. The system must always fail safe to a normal alarm.

On-device boundary: what runs locally vs what is post-processed

The headband should make wake decisions locally, using features that are compute- and power-feasible. Post-processing can refine reporting, but the wake action must not depend on unpredictable phone availability.

• Device-side: sampling → feature extraction → quality gating → lightweight classifier → wake scheduling → haptics.
• Session logging: store reason codes and quality counters for explainability and validation.
• Constraints: bounded CPU load, fixed memory, predictable latency, and minimal burst activity.

Engineering features: simple, stable, and testable

Features should be selected for robustness under real contact variability and limited compute. Keep the feature set small and interpretable so failures can be attributed to signal quality, not to unknown model behavior.

• EEG: bandpower ratios, band energy stability, and coarse trend indicators over fixed windows.
• Motion: movement index and event density (micro-movements vs large motion).
• Optional PPG: HR trend and coarse variability trend, used only when optical quality flags are good.

Confidence gating: do not wake on questionable data

A quality gate should block smart wake decisions when signal integrity is compromised. Use quality inputs that are already engineered elsewhere: contact score, noise flags, saturation events, and gap markers.

• Inputs to the gate: contact score, AFE saturation events, noise rise flags, gap markers (H2-7), and optional optical quality.
• Outputs: allow smart-wake decision, block and wait, or force fallback to normal alarm scheduling.
• Explainability: record the gate decision and the top blocking reason as a reason code.

Scheduler + fail-safe: windowed wake with a hard deadline

The wake scheduler enforces safety and predictability. It allows smart wake only inside a defined window, applies cooldown to avoid repeated triggers, and guarantees a hard deadline fallback to the normal alarm.

• Wake window: only allow early wake decisions within a bounded time range.
• Hard deadline: at the scheduled alarm time, wake must occur regardless of smart-wake confidence.
• Cooldown: prevent repeated wake attempts from transient spikes or unstable contact.
• Metrics: false-wake count, blocked-by-quality count, and reason-code distribution per night.

Intent: produce proof, not opinions

Validation is treated as a closed loop: fixtures create controlled stress, the device logs markers/counters, metrics are computed from raw streams, and pass/fail gates generate an actionable result. Bench tests establish hardware limits; pilot sessions quantify real wear variables; field evidence packages make failures reproducible.

Bench (1) EEG noise floor & band integrity

Lock the baseline performance using consistent input conditions (impedance emulation + stable references). Measure in the analysis band with a long enough capture to expose 1/f and slow drift behavior.

• Capture: long record with markers (start/stop, mode transitions, radio bursts, LED pulses if present).
• Metrics: band-limited noise (0.5–40 Hz), mains peak magnitude (50/60 Hz), saturation event rate, recovery time after saturation.
• Outputs: a single “noise report” per firmware build, directly comparable across revisions.

Bench (2) CMRR injection & common-mode sensitivity

Prove (or rule out) “mains hum” by injecting controlled common-mode disturbance into the electrode-equivalent network. Effective CMRR should be derived from injected amplitude versus observed output response.

• Injection: sweep frequency around 50/60 Hz and a few harmonics; vary amplitude in safe steps.
• Readouts: output spectral peak, noise delta in-band, saturation counts, and any quality-flag changes.
• Interpretation: if output hum scales with injection, prioritize CMRR/return-path/reference strategy; if not, suspect contact instability or rail coupling.

Bench (3) Impedance emulation: contact drift, imbalance, and saturation

Convert “skin contact variability” into a reproducible bench variable. Use a resistance decade box or a programmable network to emulate electrode impedance magnitude, imbalance, and time-varying drift.

• Variables: impedance level (kΩ to MΩ range), electrode mismatch, step changes, slow ramps (sweat/loosening proxy).
• Checks: contact score sensitivity, bias-current induced drift, baseline wander, saturation probability under high impedance.
• Outputs: contact-score vs ground-truth curve; saturation rate vs impedance curve.

Bench (4) Rail-noise injection & RF burst correlation

Quantify how supply ripple and burst current events translate into EEG noise and quality-flag behavior. The goal is a numeric transfer metric (rail disturbance → in-band noise delta), plus correlation evidence between RF markers and analog noise rise.

• Rail injection: inject ripple/steps onto AFE rail and RF rail independently (same amplitude, separate paths).
• Correlation: log firmware markers (TX start/stop) and compute EEG noise delta during RF windows.
• Outputs: “noise delta during burst” and “rail-to-EEG sensitivity” metrics, tracked over time.

Bench (5) ESD spot checks: “pass in lab, fail in field” prevention

ESD checks at key external touch points (electrode area, buttons, charge contacts, enclosure seams) should validate both survivability and graceful degradation (no latch-up, no silent data corruption, no stuck states).

• Targets: reset reasons, brownout counters, AFE saturation behavior after discharge, BLE reconnection behavior after event.
• Evidence: event timestamp + reset reason + integrity report (seq continuity, gap markers, quality counters).

Pilot users: overnight completeness under real wear variables

Pilot validation converts real-world variability into annotated, analyzable data. Each overnight session should produce a standardized “session package” with synchronized markers, counters, and summary metrics.

• Completeness: dropout %, max continuous gap, reconnection p50/p95/p99, buffer high-water mark.
• Motion scenarios: side-sleep, roll-over, sit-up; compare artifact flags vs accelerometer correlation.
• Sweat/temperature drift: correlate contact score drift and noise rise with temperature/skin coupling changes.
• Smart-wake reliability: blocked-by-quality count, false-wake count, reason-code distribution.

Metrics dashboard: field names that stay stable across revisions

Define a single dashboard schema and never rename metrics casually. Stable field names enable trend tracking across AFE/layout/firmware iterations.

Test matrix + acceptance thresholds (spec-style gates)

Use a matrix so every “fix” maps to a test case. Gates should be separated into hard pass/fail thresholds, soft optimization targets, and alarm triggers that increase logging depth.