Medical-grade wearable

A medical-grade wearable is a small body-worn device that measures physiological signals — heart rate, blood oxygenation (SpO2), single-lead or multi-lead ECG, respiratory rate, body temperature, motion — with sufficient accuracy to inform clinical decisions, while being unobtrusive enough that a patient will wear it. It looks like a consumer fitness tracker but is held to a different standard: signal quality has to be clinical, the regulatory burden is substantial, and the firmware lifecycle is governed by IEC 62304. The Cortex-M inside is the same family as in a smartwatch; the engineering rigour around it is not.

This page covers the analog front-end choices for PPG and ECG, the sampling and DSP that turns raw signals into clinical-quality measurements, the BLE-based connectivity to a paired phone or gateway, the regulatory framework that constrains everything, and the battery-life arithmetic that makes the device wearable for days or weeks rather than hours.

PPG — photoplethysmography

PPG is the technology in every modern smartwatch heart-rate sensor: an LED illuminates the skin; the photodiode measures the reflected light. As blood pulses through the capillaries with each heartbeat, the absorption changes slightly — the AC component of the photodiode signal is the heart-beat-derived "plethysmogram". By measuring this at two wavelengths (red ~660 nm and infrared ~940 nm), the ratio of pulsation between them yields SpO2 (the Lambert-Beer pulse oximetry equation).

An AFE like the MAX86150 integrates:

• Three LED drivers (red, IR, green; the green is mainly for HR, the red/IR for SpO2).
• Photodiode preamp with selectable gain.
• 18-bit ADC sampling each LED's contribution separately (time-multiplexed: pulse LED1, sample; pulse LED2, sample; ambient sample for dark cancellation).
• I²C interface to the MCU.

Signal challenges:

• Motion artefact dominates. A person walking, running, or even gesturing produces optical-modulation an order of magnitude larger than the cardiac signal. Adaptive noise cancellation (using a 3-axis accelerometer as reference) is mandatory.
• Skin tone variation. Darker skin absorbs more; LED currents must auto-adjust to keep the photodiode in its linear range.
• Perfusion — fingertip PPG has 5x the signal-to-noise of wrist PPG. Wrist-worn designs need every bit of SNR they can extract.
• Ambient light. Sun, room lights add DC and 50/60 Hz components. The AFE does ambient-only samples for subtraction.

Signal processing pipeline:

Raw 100-400 Hz PPG → Ambient subtraction → DC removal (HPF 0.5 Hz)
  → Bandpass 0.5-4 Hz (cardiac band) → Motion artefact rejection (accelerometer-correlated removal)
  → Peak detection / FFT for HR → Cardiac quality assessment

A typical HR algorithm uses Welch's method or autocorrelation in a sliding window, returns "60-200 bpm with confidence X". The confidence is what distinguishes a clinical from a consumer device — clinical devices flag low-confidence intervals rather than report a number you might trust when you shouldn't.

SpO2 uses the red/IR ratio:

R = (AC_red / DC_red) / (AC_IR / DC_IR)
SpO2 = lookup_table(R)   // calibrated against arterial blood gas measurement

The calibration table comes from clinical trials — humans with measured arterial SpO2 wearing the device. FDA / CE clearance for SpO2 requires this calibration to be documented and validated against a standard.

ECG — single-lead and multi-lead

ECG measures the electrical activity of the heart directly via skin electrodes. Single-lead (two electrodes) gives a single voltage trace — enough for rhythm classification (sinus, atrial fibrillation, etc.). Multi-lead (typically 3, 5, 6, or the clinical 12) gives spatial information about the heart's electrical axes and is required for ischemia / MI detection.

A wearable patch ECG typically has 2-3 electrodes: one positive, one negative, one reference (driven-right-leg amplifier). The signal at the electrodes is ~1 mV peak with a bandwidth of 0.05-150 Hz for diagnostic quality (and the QRS complex is ~5-30 ms wide, demanding sample rates of at least 200 Hz, ideally 500-1000 Hz).

The AFE (MAX30003, AD8232) provides:

• Instrumentation amplifier with very high CMRR (>100 dB) to reject the 50/60 Hz interference from mains.
• Front-end filter (HPF ~0.05 Hz to remove DC drift, LPF ~150 Hz to anti-alias).
• 24-bit delta-sigma ADC oversampling at 8-32 kHz, decimated to 200-1000 Hz output.
• Lead-off detection — checks if the electrodes are making good skin contact.
• Pace-pulse detection — fast circuitry to identify pacemaker spikes (which would otherwise look like QRS to a software detector).

Software processing:

Raw 24-bit ECG @ 250 Hz → Notch 50/60 Hz → Bandpass 0.5-40 Hz (diagnostic) or 5-15 Hz (monitor)
  → Pan-Tompkins QRS detector → Beat classification (normal, supraventricular, ventricular)
  → Rhythm classifier (sinus, AFib, AFlutter, brady, tachy)

The QRS detector is classic Pan-Tompkins (1985) for monitoring grade or a neural-network-based classifier for higher accuracy. AFib detection from a wearable is FDA-cleared on multiple consumer products now (Apple Watch, Fitbit Sense, KardiaMobile) — the model has been validated against cardiologist-read 12-lead ECG ground truth in clinical trials.

MCU choice

The MCU does the DSP, BLE, storage, and supervises the AFEs. Constraints:

• Sample rate × DSP load: a 500 Hz ECG with motion-correlated PPG + AFib classifier consumes 5-15% of a Cortex-M4F at 64 MHz.
• BLE concurrency: streaming ECG to a phone over BLE in real time needs ~50 kbps throughput, well within BLE 5.0 PHY capability.
• Battery life: target average current < 3 mA for a 200 mAh battery to give 3 days; <1 mA for a week. Most of this is the AFE and the LED drivers; the MCU is a fraction.
• Memory: store 24 hours of raw ECG at 500 Hz × 16 bits = 86 MB. Internal flash won't fit; either compress aggressively (ECG compresses 5-10:1 with lossless wavelet transforms) and store events of interest only, or include external NOR flash, or stream to phone continuously.
• Security: patient data is regulated. On-device encryption (AES-256), secure boot, TLS for cloud upload — all needed.

Popular choices:

Maxim and Ambiq have specifically targeted wearable medical with feature sets like integrated PMIC + BLE + AFE companion chips that minimise external components.

Regulatory framework

Going from "it works" to "it can be sold" requires:

IEC 60601-1 (general electrical safety). Covers leakage currents, dielectric strength, mechanical integrity, biocompatibility of contact materials. The wearable's housing must not leak more than a few microamps under fault conditions; the battery must not exceed a safe surface temperature.

IEC 60601-1-2 (EMC). Emissions and immunity. The device must not interfere with other medical equipment in a clinical environment, and must continue to operate (or fail safely) under typical hospital electromagnetic disturbance.

ISO 14971 (risk management). Required process for identifying, evaluating, and mitigating risks. Every hazard ("device falsely reports normal SpO2 during hypoxic event") gets a risk score and a mitigation strategy. Software bugs are first-class risks.

IEC 62304 (medical device software lifecycle). The software process standard. Categorises software by safety class (A: no injury risk; B: non-life-threatening injury; C: death or serious injury possible) and prescribes development rigour proportional to class. A Class C wearable (e.g. critical-care monitor) requires unit tests, integration tests, code reviews, traceability from requirements to implementation to test, and formal verification or extensive testing of safety-critical algorithms.

ISO 13485 (quality management). The company's quality management system. Covers design controls, change control, supplier management, post-market surveillance. ISO 13485 certification is a prerequisite to selling.

FDA 21 CFR Part 820 (Quality System Regulation). US-specific QMS. Equivalent to ISO 13485 but with some additional requirements.

Cybersecurity (FDA 2023+ premarket guidance). Wearables that transmit data must demonstrate threat modelling, secure development practices, and a vulnerability disclosure / patching plan. SBOM (Software Bill of Materials) for every third-party component.

The class affects the entire engineering process. A Class II SaMD (Software as a Medical Device) wearable is not a place to move fast and break things.

Power and form-factor budgeting

A wrist wearable target — 7 days on a single charge with a 250 mAh battery — needs an average current of 250 mAh / (7 × 24 h) ≈ 1.5 mA. Component contributions:

This budget barely fits. The first thing to drop is "continuous HR" — switching to 5-minute sampling with adaptive on-demand brings PPG average down to ~100 µA. Trade is loss of continuous data; in a clinical use case (Holter-style cardiac monitoring), continuous is the requirement and the battery sets the wear-time.

Battery technology improvements (silicon-anode Li-Po, hybrid Li-Po + thin-film) push capacity-per-volume up; the design budget improves slowly.

BLE and data flow

The device pairs with a phone (iOS or Android) via BLE GATT. Standard services for some signals:

• Heart Rate Service (0x180D) — HR + RR intervals.
• Pulse Oximeter Service (0x1822) — SpO2.
• Body Composition Service (0x181B) — body fat, lean mass.
• Custom services for ECG, motion, raw signals.

The phone app stores data, syncs to a backend, applies higher-level analytics, and presents to the user or clinician. Modern apps also support FHIR (Fast Healthcare Interoperability Resources) for integration with electronic health records.

For continuous medical-grade transmission, BLE alone is insufficient — it requires the phone to be in range. Premium products add LTE-M directly to the wearable for cellular fallback, at the cost of significant battery drain (cellular TX is 10-100x BLE).

Cross-references

• Nordic nRF53 — BLE option.
• STM32 low-power — U5 with TrustZone for secure medical.
• Smartwatch system — consumer-grade cousin.
• IoT sensor node — same low-power patterns.
• IEC 60601-1 (4th edition) and 60601-1-2.
• IEC 62304:2006 + AMD1:2015.
• FDA "Software as a Medical Device (SaMD)" guidance and cybersecurity premarket guidance.
• Pan & Tompkins, "A real-time QRS detection algorithm", IEEE Trans Biomed Eng 1985.