AS7263 NIR Spectral Sensor — Bring-up & Cold Stimulus Test

Overview

The AS7263 (ams-OSRAM) is a 6-channel near-infrared spectral sensor covering 610–860 nm. Each channel has a 20 nm FWHM narrowband filter, making it essentially a miniature spectrometer on a chip. It communicates via I2C at address 0x49 using a virtual register protocol.

This document covers sensor bring-up on the i.MX8MP EVK and an initial cold stimulus experiment for skin tissue monitoring.

Hardware Setup

Wiring on J21 (I2C3):

AS7263 Pin	J21 Pin	Signal
3.3V	Pin 1	VEXT_3V3
GND	Pin 9	GND
SDA	Pin 3	I2C3_SDA_3V3
SCL	Pin 5	I2C3_SCL_3V3

Pull-up resistors are built into the SparkFun Qwiic breakout board.

Bring-up

I2C Detection

Sensor detected immediately on first scan:

$ i2cdetect -y 2
     0  1  2  3  4  5  6  7  8  9  a  b  c  d  e  f
40: -- -- -- -- -- -- -- -- -- 49 -- -- -- -- -- --

Chip identification via the virtual register protocol:

HW version: 0x40
FW version: 0x20
Sensor type: 0x3F (AS7263 NIR confirmed)

Platform Adaptation

The Yocto image does not include smbus2, and the EVK has no internet access for pip. Rewrote the I2C interface using raw /dev/i2c-2 with fcntl.ioctl — zero external dependencies:

fd = os.open("/dev/i2c-2", os.O_RDWR)
fcntl.ioctl(fd, 0x0703, 0x49)   # I2C_SLAVE
os.write(fd, bytes([reg]))       # write register
data = os.read(fd, 1)[0]         # read byte

This also served as a good exercise in understanding how Linux exposes I2C hardware through character devices.

Note on BME280

Before AS7263 testing, two GY-BME280 modules were debugged extensively on the same J21 pins — neither responded. The AS7263 working on the exact same wiring confirmed that the I2C bus, 3.3V power rail, and level shifters are all functional. The BME280 modules were defective (same batch, both dead).

Lesson learned: when an I2C device isn't detected, swap in a known-good device before diving into software debugging.

Key Channels for Tissue Monitoring

Channel	Wavelength	Role
S	680 nm	Deoxyhemoglobin (Hb) absorption peak — rises during ischemia
U	760 nm	NIR reference, no special physiological significance
W	860 nm	Oxyhemoglobin (HbO2) absorption peak — high when perfusion is good

Tissue Oxygenation Index:

TOI = (W_860 - S_680) / (W_860 + S_680)

Cold exposure → vasoconstriction → less oxygenated blood → W drops, S rises → TOI decreases.

Raw TOI values are negative because the on-board LED has uneven spectral output (shorter wavelengths are brighter), so S is always larger than W in absolute terms. White-reference calibration would correct this. However, the trend (change over time) is still valid without calibration.

Measurement Results

Test Conditions

Gain: 64x (maximum, needed for skin reflectance measurement)
Integration time: 140 ms (INT_T = 50 × 2.8 ms)
LED: 100 mA (required for skin — ambient light is blocked when sensor is pressed against skin)
Mode: One-shot (mode 3), triggered every 2 seconds

Finding: LED Must Be Enabled

With LED off, pressing the sensor against skin blocks all ambient light — the 860 nm channel drops to zero. Active illumination is mandatory for contact-mode skin measurements.

Warm Skin Baseline (LED 100mA)

Channel	Wavelength	Value (µW/cm²)
R	610 nm	9000
S	680 nm	2780
T	730 nm	582
U	760 nm	374
V	810 nm	490
W	860 nm	405

Chip temperature: 38–39°C. TOI = −0.745. Signal stability: ±2% across 15 consecutive samples.

Cold Stimulus — Short Exposure (2–3 min ice pack)

Applied ice pack to forearm, then immediately placed sensor on cooled skin. 15 samples collected.

Metric	Warm	Cold	Change
Temperature	38°C	28→31°C	−10°C
S_680	2780	2500	−10%
W_860	405	422	+4%
TOI	−0.745	−0.710	+0.035

Short cold exposure reduced blood flow (S dropped 10%) but did not cause desaturation — consistent with mild, reversible vasoconstriction.

Cold Stimulus — Extended Exposure (5 min ice pack)

Applied ice pack for 5 minutes, then monitored recovery for 1 minute (30 samples at 2s intervals).

Time	Temp	S_680	W_860	V_810	TOI
0s	28°C	2941	518	644	−0.700
15s	29°C	2900	500	617	−0.706
30s	30°C	2889	476	578	−0.718
45s	31°C	2882	460	556	−0.725
60s	33°C	2811	432	514	−0.734

Observations:

Temperature recovered from 28°C to 33°C during sampling, confirming blood flow restoration
W_860 dropped 17% (518→432) during rewarming — counterintuitive at first glance
TOI shifted from −0.700 to −0.734 (more negative during recovery)
This pattern is consistent with reactive hyperemia: after cold stimulus is removed, blood vessels reopen and a surge of blood enters the tissue, temporarily increasing local oxygen consumption
30 consecutive samples with zero communication errors — sensor is reliable

Virtual Register Protocol

The AS7263 does not support direct I2C register access. All reads/writes go through three physical registers:

0x00 — Status register (TX_VALID / RX_VALID flags)
0x01 — Write register (send virtual address or data)
0x02 — Read register (retrieve response)

Each virtual register access requires polling the status register for readiness. If communication is interrupted mid-transaction (e.g., loose wire), the state machine desyncs and returns garbage data. The only recovery is a power cycle.

Files

scripts/as7263_monitor.py — Data acquisition script (raw I2C, no dependencies)
docs/as7263_cold_test.png — Cold stimulus test charts

White Reference Calibration

Before skin measurements, a white paper reference was taken to characterize the LED spectral profile. White paper has roughly uniform reflectance across 610–860 nm, so white-paper readings primarily reflect the LED's own spectral shape rather than sample properties. Dividing skin readings by this reference yields normalized tissue reflectance, removing the LED spectral bias.

Method: sensor face-down on standard A4 white paper, LED 100mA, 10 samples collected, last 6 averaged for stability.

Channel	Wavelength	White Ref (µW/cm²)
R	610 nm	3449
S	680 nm	938
T	730 nm	231
U	760 nm	165
V	810 nm	249
W	860 nm	193

The R:S:W ratio of 18:5:1 confirms severe LED spectral non-uniformity — this is why raw TOI is always negative.

Calibrated TOI example (warm skin baseline):

S_norm = 2780 / 938 = 2.96
W_norm = 405 / 193 = 2.10
TOI_cal = (2.10 - 2.96) / (2.10 + 2.96) = -0.170

After calibration, TOI shifts from −0.745 to −0.170, much closer to physiologically meaningful values.

Known Issues

Signal jitter in TOI curve: The raw TOI plot shows noticeable sample-to-sample fluctuation (~±0.005). Two main causes:

Contact pressure variation — hand-held measurement means small movements change the optical path and reflection angle, causing readout fluctuation
ADC quantization noise — the AS7263 outputs calibrated IEEE754 floats, but the underlying ADC is 16-bit. On weaker channels like W_860, quantization steps become visible in the ratio calculation

Mitigation: apply a 3–5 point moving average to smooth the curve, or use a mechanical fixture to stabilize sensor-skin contact.

Calibrated Skin Baseline

With the white reference established, skin readings can be normalized to remove LED spectral bias:

normalized_reflectance = skin_reading / white_reference

Forearm skin at room temperature, LED 100mA, 15 samples averaged:

Metric	Raw	Calibrated	Note
S_680 normalized	2268 µW/cm²	2.42	skin / white_ref(938)
W_860 normalized	302 µW/cm²	1.56	skin / white_ref(193)
TOI (raw)	−0.766	—	dominated by LED spectral bias
TOI (calibrated)	—	−0.215	physiologically meaningful
Temperature	31–32°C

Observations:

Calibration shifted TOI by +0.55 (from −0.77 to −0.21), confirming that LED non-uniformity was the dominant factor in raw readings
TOI_cal stability across 15 samples: −0.19 to −0.22 (±0.015), good repeatability
Compared to the previous warm-hand session (38°C, TOI_cal = −0.17), today's baseline is slightly lower (32°C, TOI_cal = −0.21) — cooler skin means less blood flow, consistent with physiology
This establishes the room-temperature skin baseline at TOI_cal ≈ −0.21 for future cold stimulus comparison

The calibration step is standard practice in diffuse reflectance spectroscopy — without it, the raw TOI is an artifact of the light source, not the tissue. The fact that calibrated values track skin temperature across sessions validates the measurement approach.

Calibrated Cold Stimulus Test

Applied ice pack until skin was thoroughly cold (~5 min), then immediately placed sensor on cooled forearm. 30 samples with calibrated TOI:

Time	Temp	TOI_raw	TOI_cal	vs baseline (−0.21)
0s	25°C	−0.721	−0.118	+0.09
15s	26°C	−0.726	−0.128	+0.08
30s	27°C	−0.734	−0.145	+0.07
45s	27°C	−0.744	−0.167	+0.04
60s	29°C	−0.747	−0.173	+0.04

Observations:

Cold skin TOI_cal (−0.12) is higher than warm baseline (−0.21) — seemingly counterintuitive, but explained by physiology: vasoconstriction reduces total blood volume in the tissue. Both Hb and HbO2 decrease, but S_680 (Hb) drops proportionally more, shifting the normalized ratio upward
TOI_cal drifts from −0.12 → −0.17 during rewarming — reactive hyperemia again. Blood vessels reopen, oxygen-rich blood surges in but is consumed rapidly, temporarily lowering the oxygenation ratio
Cross-session reproducibility confirmed: warm baseline TOI_cal = −0.17 (previous session at 38°C) matches the rewarming endpoint here (−0.17 at 29°C), validating the calibration approach
Temperature rose from 25°C to 29°C during the 60s measurement window, consistent with active reperfusion

Methodological note: In this test, there was a delay between removing the ice pack and the start of data acquisition (sensor placement + script startup). The initial 25°C reading likely underestimates the peak cold because the skin had already begun rewarming.

Countdown-Triggered Cold Stimulus Test

To address the measurement latency issue, a GTK3 fullscreen countdown script was developed. The workflow: 10-second countdown on HDMI → user places cold hand during countdown → LED and data acquisition start simultaneously at t=0.

First few samples (#1–#4) were discarded (hand not yet stable), and the last sample (#30) was an outlier (hand lifted). Valid data: samples #5–#29.

Phase	Samples	Temp	TOI_cal	Note
Active cooling	#5–#10	27°C	−0.12	skin still cooling, vasoconstriction
Peak cold	#11–#16	27°C	−0.14 → −0.17	TOI drops as cold penetrates deeper
Rewarming	#17–#24	27→28°C	−0.17 → −0.19	reactive hyperemia begins
Near baseline	#25–#29	29°C	−0.19 → −0.20	converging toward baseline −0.21

Key findings:

Captured a complete cooling → rewarming transition — temperature dropped from 30°C to 27°C then recovered to 29°C, with a clear TOI response throughout
Monotonic TOI descent from −0.12 to −0.20 — much cleaner than earlier tests where rewarming had already started before the first sample
Three independent sessions now converge to TOI_cal ≈ −0.17 to −0.21 at rewarming endpoint — strong evidence for measurement reproducibility
The countdown approach successfully reduced placement delay, capturing colder initial readings than the manual method

Mild vs Strong Cold Stimulus Comparison

After fixing the countdown dual-instance bug, a second test was run with milder cold exposure (starting at 30°C vs 25°C in the previous test). The result was strikingly different:

Metric	Strong cold (test 3)	Mild cold (test 4)
Start temp	25°C	30°C
TOI_cal range	−0.12 → −0.20	−0.14 → −0.17
TOI swing	0.08	0.03
Signal jitter	noticeable	minimal
Curve shape	steep descent	gentle, smooth

Three factors explain the smoother, narrower curve:

Cold stimulus intensity drives response amplitude — stronger vasoconstriction (25°C) produces larger TOI shifts than mild cooling (30°C). This is a dose-response relationship: the sensor can distinguish different levels of cold exposure, which is exactly what a frostbite warning system needs
Signal strength vs noise floor — milder cold means more blood in the tissue, stronger reflected signal, better SNR. At 25°C with severe vasoconstriction, S_680 dropped to ~1400 µW/cm² where ADC quantization noise becomes more visible in the TOI ratio
Operator experience — hand placement stability improved with practice, reducing contact pressure variation. This is a real consideration for a clinical device (sensor mounting matters)

Five-Session Reproducibility Summary

Session	Condition	Endpoint TOI_cal
Baseline	Warm skin, 38°C	−0.21
Cold test 1	Manual, strong cold	−0.17
Cold test 2	Calibrated, strong cold	−0.20
Cold test 3	Countdown, strong cold	−0.17
Cold test 4	Countdown fixed, mild cold	−0.17

All rewarming endpoints converge to −0.17 to −0.21 — consistent across different cold intensities, measurement methods, and sessions. This level of reproducibility validates the white-reference calibration approach and confirms the sensor can serve as a reliable tissue perfusion monitor.

DPF Correction — From Negative TOI to Physiological StO2

The white-reference calibration removes LED spectral bias but not tissue scattering. Shorter wavelengths scatter more in tissue (higher path length), so more 680nm light returns to the sensor regardless of hemoglobin content. This is why even calibrated TOI remains negative.

Clinical NIRS instruments solve this using the modified Beer-Lambert law with Differential Pathlength Factors (DPF) — wavelength-dependent coefficients that normalize for scattering path length differences.

Measurement site: fingertip (middle finger), not forearm. Fingertip tissue is much thinner than forearm, with denser capillary beds and shorter scattering path. DPF values are approximately 40–50% of forearm values (Scholkmann & Wolf, 2013).

Wavelength	Forearm DPF	Fingertip DPF	Ratio
680 nm	6.51	3.0	46%
860 nm	5.86	2.5	43%
DPF ratio (680/860)	1.11	1.20	—

The higher DPF ratio for fingertip (1.20 vs 1.11) means the scattering correction is amplified — a direct consequence of thinner tissue where wavelength-dependent scattering differences are more pronounced.

DPF-corrected calculation (fingertip):

A_corrected(λ) = |ln(I_reference / I_skin)| / DPF(λ)

Warm baseline (fingertip):
A_680 = |ln(938 / 2268)| / 3.0 = 0.882 / 3.0 = 0.2940
A_860 = |ln(193 / 302)|  / 2.5 = 0.448 / 2.5 = 0.1792

StO2 = A_860 / (A_860 + A_680) = 0.1792 / (0.1792 + 0.2940) = 0.38

Three-Method Comparison on Mild Cold Stimulus Data

Phase	Samples	TOI_raw	TOI_cal (white ref)	StO2 (fingertip DPF)
Cooling	#4–#8	−0.733	−0.14	0.43
Slow descent	#9–#18	−0.740	−0.16	0.40
Rewarming	#19–#30	−0.746	−0.17	0.24
Warm baseline	prev. session	−0.766	−0.21	0.38

Summary of the three correction levels:

Raw TOI: −0.73 to −0.77. Always negative, no physiological meaning, but trend is valid
White-reference TOI: −0.14 to −0.21. Removes LED bias, still negative (scattering not corrected)
DPF-corrected StO2 (fingertip): 0.24 to 0.43. Positive values with much wider dynamic range than forearm DPF (which gave a compressed 0.36–0.39)

Key observations with site-corrected DPF:

Cold fingertip StO2 = 0.43, warm baseline = 0.38 — same trend direction as white-ref TOI
Rewarming phase drops to StO2 = 0.24 — reactive hyperemia with rapid oxygen consumption is now clearly visible. Forearm DPF compressed this to 0.36, nearly indistinguishable from baseline
The fingertip DPF ratio (1.20) amplifies the correction compared to forearm (1.11), revealing physiological dynamics that were hidden with incorrect site parameters
Using the wrong measurement site's DPF masks real physiological changes — this is why clinical NIRS protocols always specify the measurement site

Why Measurement Site Matters for DPF

The DPF absolute values differ between body sites, but what actually determines measurement sensitivity is the DPF ratio (680nm / 860nm):

Site	DPF_680	DPF_860	DPF ratio	StO2 range	Effect
Forearm	6.51	5.86	1.11	0.36–0.39	Compressed — cold vs warm nearly indistinguishable
Fingertip	3.0	2.5	1.20	0.24–0.43	Spread out — cold/warm clearly separated

A higher DPF ratio amplifies the scattering correction, improving the resolution of physiological changes. With the wrong site's DPF, the reactive hyperemia signal (StO2 = 0.24) gets compressed to 0.36, nearly indistinguishable from the 0.38 baseline — the real physiological dynamics are masked.

This is why clinical NIRS protocols always specify the measurement site. The same person at the same moment can show 10–20 percentage points difference in StO2 between fingertip and forearm, purely due to tissue geometry and scattering properties.

Ref: Duncan A. et al. (1995), Phys Med Biol 40(2):295-304; Scholkmann F. & Wolf M. (2013), "General equation for the DPF of the adult head, neonatal head, adult forearm, and adult leg", J Biomed Opt 18(10):105004.

Frostbite Warning Threshold

Threshold Definition

Based on 5 independent cold stimulus sessions, the strongest cold exposure (ice pack until skin was barely tolerable, starting temperature 25°C) is used as the "near-frostbite" reference point. At this intensity, the subject reported the sensation as approaching the pain/tissue-damage boundary.

Key data from the near-frostbite session (calibrated cold stimulus test, 25°C start):

Metric	Value	Meaning
Skin temperature	25°C	Severe cold, approaching tissue damage
TOI_cal (peak cold)	−0.118	Maximum vasoconstriction observed
StO2 (fingertip DPF)	0.43	Tissue-level oxygenation during severe cold

Important: the direction is counterintuitive. In this reflectance-mode system, vasoconstriction reduces total blood volume, and Hb (680nm) drops proportionally more than HbO2 (860nm), so TOI_cal rises (becomes less negative) during cold exposure. The danger signal is TOI_cal moving toward zero, not away from it.

Three-Level Warning System

Level	TOI_cal Range	StO2 (fingertip)	Skin Temp	Action
SAFE (green)	< −0.15	< 0.40	> 28°C	Normal perfusion, continue cooling
WARNING (yellow)	−0.15 to −0.12	0.40 to 0.43	25–28°C	Significant vasoconstriction, monitor closely
DANGER (red)	> −0.12	> 0.43	< 25°C	Near-frostbite, stop cooling immediately

Threshold Derivation

The thresholds are derived from the five-session dataset:

Session	Start Temp	Peak TOI_cal	Endpoint TOI_cal	Severity
Warm baseline	32°C	−0.21	−0.21	None
Cold test 1 (manual)	28°C	~−0.15	−0.17	Mild
Cold test 2 (calibrated)	25°C	−0.118	−0.173	Severe
Cold test 3 (countdown)	27°C	−0.12	−0.20	Moderate
Cold test 4 (mild)	30°C	−0.14	−0.17	Mild

The WARNING threshold (−0.15) is set at the point where mild cold starts — all mild exposures peaked near this value. The DANGER threshold (−0.12) is set at the peak of the severe exposure where tissue damage risk begins. Values beyond −0.12 have not been tested (would require prolonged exposure beyond ethical limits).

Reactive Hyperemia as a Secondary Indicator

After removing the cold stimulus, healthy tissue exhibits reactive hyperemia — a surge of blood reperfusing the tissue, causing a temporary StO2 dip to ~0.24 (TOI_cal ~−0.20). This is actually a good sign: it means the microcirculation is intact.

If reactive hyperemia is absent after cold removal (StO2 stays flat instead of dipping), it would suggest microvascular damage — a more serious frostbite indicator. This has not been observed in our experiments (all sessions showed clear reactive hyperemia), but it would be the critical warning in a clinical deployment.

DPF Reference Table — Human Body Sites

Differential Pathlength Factor (DPF) values vary significantly across body sites due to differences in tissue thickness, scattering properties, and blood vessel density. Using the wrong site's DPF can compress or mask real physiological changes, making it critical to match the DPF to the actual measurement location.

DPF Values by Body Site and Wavelength

Body Site	DPF @ ~690nm	DPF @ ~830–860nm	DPF Ratio (short/long)	StO2 Sensitivity	Source
Adult Forehead	6.51	5.38–5.42	1.20–1.21	High	Duncan 1995; Scholkmann & Wolf 2013
Adult Head (temporal)	6.26	5.13	1.22	High	Duncan 1995
Neonatal Head	5.38	4.53	1.19	High	Duncan 1995
Adult Calf	5.51	4.74	1.16	Medium	Duncan 1995; van der Zee 1992
Adult Thigh	~5.0–5.5	~4.3–4.8	~1.16	Medium	Extrapolated from calf
Adult Forearm	4.16	3.59	1.16	Medium	Duncan 1995; van der Zee 1992
Fingertip	~3.0	~2.5	1.20	High	Essenpreis 1993 (estimated, thin tissue)
Thenar (palm)	~3.5–4.0	~3.0–3.5	~1.15	Medium	Limited published data

Why the DPF Ratio Matters More Than Absolute DPF

The absolute DPF values affect the magnitude of computed absorbance, but the DPF ratio (short wavelength / long wavelength) determines how much the scattering correction amplifies or compresses the physiological signal:

Higher DPF ratio → stronger scattering correction → wider StO2 dynamic range → better sensitivity

Our experimental verification:

Site	DPF Ratio	StO2 Range (cold→warm)	Can Distinguish Cold/Warm?
Forearm (DPF 6.51/5.86)	1.11	0.36–0.39	No — compressed, nearly flat
Fingertip (DPF 3.0/2.5)	1.20	0.24–0.43	Yes — 19 pp dynamic range

This is why clinical NIRS devices always specify the measurement site in their operating protocol. The same person at the same moment can show 10–20 percentage points difference in StO2 between fingertip and forearm, purely due to tissue geometry.

Scholkmann & Wolf 2013 General Equation (Forehead Only)

For the adult forehead, DPF can be computed as a function of wavelength (λ, nm) and age (A, years):

DPF(λ, A) = 223.3 + 0.05624 × A^0.8493 − 5.723×10⁻⁷ × λ³ + 0.001245 × λ² − 0.9025 × λ

Valid for λ = 690–832 nm, age 1–50 years. Not applicable to other body sites.

Caveats

Fingertip and thenar data are sparse — most rigorous DPF measurements used phase-resolved or time-resolved instruments on head, forearm, and calf
DPF depends on source-detector separation — literature values assume 3–4 cm spacing; compact sensors like AS7263 (< 5 mm) sample more superficial tissue, potentially lowering effective DPF
DPF varies with age, adipose thickness, and skin pigmentation — the table values are population averages from predominantly Caucasian cohorts
For compact reflectance sensors: many pulse oximetry papers fold DPF into an empirical calibration constant rather than using literature values directly

Ref: Duncan A. et al. (1995), Phys Med Biol 40(2):295-304; Scholkmann F. & Wolf M. (2013), J Biomed Opt 18(10):105004; van der Zee P. et al. (1992), Adv Exp Med Biol 316:143-153; Essenpreis M. et al. (1993), Applied Optics 32(4):418-425.

Kernel Customization for BLE HID

Motivation

Wanted to use Logitech MX Master 3S (mouse) and MX Keys (keyboard) via Bluetooth with the EVK's Weston desktop, enabling direct interaction with the GTK3 dashboard.

Kernel Config Changes

Analyzed the existing .config — most BLE infrastructure was already present. Only 3 options were missing:

CONFIG_UHID=y                  # BLE HID-over-GATT (HOGP)
CONFIG_HID_LOGITECH_HIDPP=m    # Logitech HID++ protocol for MX series
CONFIG_INPUT_MOUSEDEV=y        # /dev/input/mice device node

Applied via Yocto do_configure:append in a custom layer (meta-custom), since NXP's linux-imx recipe doesn't merge .cfg fragments through the standard kernel-yocto mechanism:

# meta-custom/recipes-kernel/linux/linux-imx_%.bbappend
do_configure:append() {
    ${S}/scripts/config --file ${B}/.config --enable UHID
    ${S}/scripts/config --file ${B}/.config --module HID_LOGITECH_HIDPP
    ${S}/scripts/config --file ${B}/.config --enable INPUT_MOUSEDEV
    oe_runmake -C ${S} O=${B} olddefconfig
}

Kernel Image replaced on the existing SD card — no full reflash needed, all WiFi config, DTB patches, scripts, and wallpaper preserved.

USB Input — Working

Logitech Bolt receiver (046d:c548) was plug-and-play after kernel update:

Bus 001 Device 002: ID 046d:c548 Logitech, Inc. Logi Bolt Receiver
Input: Logitech USB Receiver Mouse → /dev/input/event3 (mouse0)

Key learning: Bolt receiver stores pairing internally — pair on any PC via Logi Options+, then move the receiver to the target device.

BLE Pairing — NXP Firmware Limitation

Both MX Master 3S and MX Keys were detected via hcitool lescan, and L2CAP connections established. However, SMP pairing consistently failed:

Bluetooth: hci0: security requested but not available

Root cause: the NXP 88W8997 Bluetooth firmware does not fully implement BLE Secure Connections. All kernel crypto is present (CRYPTO_ECDH=y, CRYPTO_CMAC=y, BT_LE=y), but the firmware rejects SMP security requests. Workaround: use USB Bolt/Unifying receivers.

Self-Screenshot via Cairo

Weston's weston-screenshooter requires special permissions that couldn't be configured reliably. The GTK3 dashboard was modified to capture its own rendering:

surface = cairo.ImageSurface(cairo.FORMAT_ARGB32, w, h)
cr = cairo.Context(surface)
self.on_draw(self.da, cr)
surface.write_to_png("/tmp/dashboard_screenshot.png")

Screenshots auto-saved every 5 seconds, pulled via SCP for remote analysis — a visual feedback loop without physical access to the HDMI display.

Methodological Reflections — Limitations Found in Practice

After the 5 cold-stimulus sessions reported above, a deeper review revealed several fundamental issues with the original framing of "real-time frostbite detection." These notes are recorded here so that anyone inheriting this work understands what worked, what didn't, and why the next iteration should look different.

AS7263 Was Never Designed for Human Tissue

The AS7263 is an agricultural / industrial NIR spectrometer. ams-OSRAM's official application notes target plant health monitoring (NDVI), food composition (moisture / sugar / fat), plastic recycling, and printed-color verification. SparkFun's hookup guide markets the breakout board as a "plant health detector" and uses leaf NDVI as its first example.

Two literature searches confirm this is more than marketing positioning:

PubMed search for AS7263 returns only agricultural / food papers (notably corn frost-damage classification — the sensor really does detect "frostbite" in crops).
PubMed search for frostbite + NIRS returns zero human clinical studies across decades of NIRS research.

This is not a coincidence. It reflects a real mismatch between what NIRS measures (hemoglobin in flowing blood) and what frostbite actually is (ice crystallization in tissue, which displaces blood from the affected microvasculature).

Four Reasons NIRS Cannot Detect Frostbite Onset

The signal disappears at the critical moment. When tissue actually freezes, blood is forced out of the affected microvasculature. NIRS depends on blood absorption — no blood, no signal. The sensor sees the precursor (vasoconstriction) but loses the actual freezing event.
Vasoconstriction is non-specific. During therapeutic cooling, vasoconstriction is the desired outcome. Approaching frostbite, vasoconstriction is the pathological warning. Same signal, opposite clinical meaning.
The reversibility threshold is invisible. Endothelial damage and microvascular thrombosis (the transition from reversible frostnip to irreversible damage) are not in NIRS's observable. The closest proxy — absent reactive hyperemia after cold removal — can only be assessed after cold withdrawal, not in real time.
Zero clinical literature. Frostbite + NIRS has zero human PubMed hits across decades. A negative result spanning that many research groups is itself informative.

Geometric Constraint of the SparkFun Breakout

The SparkFun breakout (25.4 × 17.8 mm, 1.6 mm FR4) acts as a thermal insulator when sandwiched between an ice pack and skin. Steady-state estimate: the skin patch directly under the sensor stays around 22°C while surrounding skin reaches 5–10°C. A >10 K thermal gradient exists at the sensor edge.

This means the 5 sessions documented above did not measure local cold-induced frostbite progression. They measured peri-cooling reflex perfusion — vasoconstriction propagating from adjacent cooled tissue into the sensor's region of view. The data is still valid, but the research question must be reframed accordingly:

❌ Real-time frostbite detection (not feasible with this geometry or this sensor)
✅ Peri-cooling tissue oxygenation response (the actual phenomenon captured)
✅ Cold therapy effectiveness monitoring (auxiliary to skin temperature)
✅ Reactive hyperemia magnitude as microcirculation health proxy (post-cooling assessment)

Where AS7263 Actually Fits in a Cryotherapy Device

The viable role for AS7263 (or a similar compact NIRS module) is auxiliary perfusion monitoring complementing skin temperature — not as the primary frostbite warning. Concrete capabilities:

Therapy effectiveness check — confirms vasoconstriction is actually occurring. A cold pad pressed loosely against skin can show low temperature without engaging perfusion response; only NIRS reveals this.
Soft warnings — detects (a) over-cooling (StO2 falling without reaching steady state) and (b) microcirculation damage after cold removal (absent reactive hyperemia → potential NFCI). Neither is visible in temperature signal.

What it cannot do:

Replace skin temperature as primary safety signal — specificity is too low to distinguish "therapeutic vasoconstriction" from "pathological vasoconstriction"
Provide an absolute calibrated StO2 — only relative trends within a session
Operate during actual ice formation — the signal disappears with the displaced blood

Hardware Path Forward: Custom Compact PCB

The SparkFun breakout served well as a prototyping platform — it confirmed AS7263 works on Yocto / i.MX 8M Plus and produced the 5-session calibration data. For productization, the next iteration should be:

A custom 8 × 8 mm PCB with only AS7263 (LGA-20) + a single white LED (SMD 0402)
I2C signal routed via flex cable; power and ground from main controller board
Board thickness ≤ 0.6 mm, embedded directly into the cold pad to bypass the FR4 thermal barrier
This places the sensor's optical view in the core cooled region rather than at the thermal edge

This is a transition from "off-the-shelf prototyping" to "embedded integration." With it, the geometry constraint above is resolved and the system can finally measure the same skin patch that the cold pad is treating.

Recommended Reframing for EMBC / Future Publication

The manuscript should not claim "NIRS for real-time frostbite detection." Defensible framings supported by literature:

NIRS as auxiliary perfusion marker complementing skin temperature in portable cryotherapy
Peri-cooling reflex perfusion response captured by compact NIR spectroscopy
Cold therapy effectiveness monitoring via tissue oxygenation index

The DEVICE_TEMP register should not be cited as evidence of skin temperature change — it reads die temperature, contaminated by LED self-heating.

FilesExpand file tree

10-nir-spectral-sensor.md

Latest commit

History