🧠 In-Ear Neurothermal Anomaly Detection via Embedded ML

An experimental embedded machine learning proof-of-concept that classifies temporal infrared thermal patterns from the ear canal using non-contact thermometry and a quantized neural network running entirely on a microcontroller.

⚠️ This is a research prototype. It is NOT a medical device. See Disclaimer.

---

🔬 Project Overview

This repository contains firmware and a compiled inference SDK that continuously sample tympanic-region temperature, extract spectral features from the time series, and classify each window as normal or stroke using an on-device neural network.

The tympanic region is anatomically adjacent to cerebral arterial pathways. This project investigates whether temporal patterns in ear-canal infrared readings—processed locally on a microcontroller—can distinguish baseline thermal profiles from anomalous ones.

Scope: This codebase covers the embedded inference stage only. Model training was performed externally using Edge Impulse. The training dataset and training scripts are not included.

---

🛠️ Hardware

Component	Part	Details
IR Sensor	MLX90614	Non-contact infrared thermometer. Factory-calibrated. I²C address 0x5A.
Interface	I²C / SMBus	Two-wire serial bus. Requires SDA, SCL, VCC, GND.
MCU	Arduino-compatible board	Any board supporting <Wire.h> and the Edge Impulse C++ SDK (e.g., ESP32, SAMD21, RP2040).

Wiring

MLX90614 Pin	MCU Pin	Notes
VCC	3.3 V or 5 V	Per sensor module board specification
GND	GND	—
SDA	SDA	4.7 kΩ pull-up to VCC if not on breakout board
SCL	SCL	4.7 kΩ pull-up to VCC if not on breakout board

[ Hardware photograph placeholder ]

---

💻 Software Stack

Layer	Component	Notes
Firmware	neuro.ino (Arduino C++)	Sensor polling, sample buffering, inference invocation, serial logging
Sensor Driver	Adafruit MLX90614 Library	I²C driver; provides readObjectTempC()
Inference SDK	Neuro_Frost_inferencing/	Auto-generated Edge Impulse C++ library (v1.0.2) containing DSP and NN code
NN Runtime	TensorFlow Lite for Microcontrollers	EON-compiled int8 model; no interpreter overhead
DSP	Edge Impulse Spectral Analysis	FFT-based feature extraction built into the SDK

---

🧠 Machine Learning Pipeline

1. Data Acquisition

The firmware reads mlx.readObjectTempC() in a synchronous loop with delay(EI_CLASSIFIER_INTERVAL_MS) between samples.

Parameter	Value	Source
Sampling interval	500 ms (2 Hz)	EI_CLASSIFIER_INTERVAL_MS · model_metadata.h:96
Samples per window	8	EI_CLASSIFIER_RAW_SAMPLE_COUNT · model_metadata.h:88
Axes per sample	1 (temperature)	EI_CLASSIFIER_RAW_SAMPLES_PER_FRAME · model_metadata.h:89
Window duration	4.0 s	8 × 500 ms
Windowing strategy	Tumbling (non-overlapping)	neuro.ino:26-38,70 — buffer fills, infers, resets index to 0

2. DSP — Spectral Feature Extraction

The 8-sample buffer is passed to the Edge Impulse spectral analysis block, which produces 13 features.

DSP Parameter	Value	Source
Analysis type	FFT	model_variables.h:63
FFT length	16	model_variables.h:64
Spectral peaks	3	model_variables.h:65
Power edge frequencies	0.1, 0.5, 1.0, 2.0, 5.0 Hz	model_variables.h:67
Log scaling	Enabled	model_variables.h:68
Output feature count	13	EI_CLASSIFIER_NN_INPUT_FRAME_SIZE · model_metadata.h:87

3. Neural Network

Layer	Type	Shape	Activation	Quantization
Input	—	(1, 13)	—	int8 affine
Dense 1	Fully Connected	(13) → (20)	ReLU	int8
Dense 2	Fully Connected	(20) → (10)	ReLU	int8
Output	Fully Connected + Softmax	(10) → (2)	Softmax	int8

Classification labels: normal, stroke — defined in model_variables.h:50.

Tensor arena size: 368 bytes (default) / 2944 bytes (largest arena) — tflite_learn_768102_9_compiled.cpp:99, model_metadata.h:112.

Registered ops: FULLY_CONNECTED, SOFTMAX — tflite_learn_768102_9_compiled.cpp:568-569.

Inference Flow

MLX90614 (I2C)                    Edge Impulse SDK              Serial Output
─────────────                     ────────────────              ─────────────
readObjectTempC()  ──►  features[0..7]  ──►  Spectral DSP (13 features)
                                             │
                                             ▼
                                        Dense NN (int8)
                                        13 → 20 → 10 → 2
                                             │
                                             ▼
                                        Softmax probabilities
                                             │
                                             ▼
                                        Serial.println()
                                        normal: 0.xxxx
                                        stroke: 0.xxxx

[ Architecture diagram placeholder ]

For a detailed technical breakdown, see docs/INFERENCE_PIPELINE.md.

---

📂 Repository Structure

├── neuro.ino                   Arduino sketch — sensor read loop, buffering, inference
├── Neuro_Frost_inferencing/    Edge Impulse C++ SDK (auto-generated, read-only)
│   ├── library.properties      Arduino library metadata
│   ├── examples/               Board-specific example sketches (from Edge Impulse)
│   └── src/
│       ├── model-parameters/   Model constants, DSP config, label definitions
│       ├── tflite-model/       EON-compiled int8 neural network weights and graph
│       └── edge-impulse-sdk/   TFLite Micro runtime, CMSIS-DSP, classifier engine
├── docs/
│   ├── DISCLAIMER.md           Research and medical disclaimer
│   └── INFERENCE_PIPELINE.md   Detailed sampling, DSP, and NN specification
├── LICENSE                     Apache License 2.0
└── .gitignore                  Build artifact exclusions

---

🚀 Build Instructions

Prerequisites

1. Arduino IDE v1.8.x or v2.x.
2. Adafruit MLX90614 Library — install via Sketch → Include Library → Manage Libraries.

Steps

1. Copy Neuro_Frost_inferencing/ into your Arduino libraries folder
   (e.g., ~/Documents/Arduino/libraries/Neuro_Frost_inferencing/).
2. Open neuro.ino in Arduino IDE.
3. Select your board and port under Tools.
4. Click Upload.
5. Open Serial Monitor at 115200 baud.

Expected Serial Output

MLX90614 + Edge Impulse inference started...
Sample 0: 36.47 C
Sample 1: 36.51 C
...
Sample 7: 36.49 C
Inference results:
normal: 0.9412
stroke: 0.0588
--------------------------

---

⚠️ Limitations

Limitation	Detail
Not a medical device	This is an engineering proof-of-concept. No clinical validation has been performed.
Blocking acquisition loop	neuro.ino uses delay(500) inside a while loop, blocking the MCU for ~4 s per inference cycle.
No sensor error handling	I²C read failures or NaN returns from the sensor are not checked before buffering.
Tumbling window only	Non-overlapping 8-sample blocks; no sliding window or overlap. Temporal discontinuities at block boundaries are not captured.
No ambient compensation	Readings may drift with ambient temperature, sensor repositioning, or body movement.
Training data not included	The dataset and Edge Impulse project configuration used to train this model are not in this repository. Model performance metrics cannot be independently verified from repository contents alone.
Boot lockup on sensor failure	If mlx.begin() fails at startup, the MCU enters while(1) and cannot recover without a hardware reset.

---

⚖️ Disclaimer

This project is an experimental prototype and research exploration. It is NOT a clinically validated medical diagnostic device.

Do not use this system to diagnose, treat, cure, or prevent any disease. See docs/DISCLAIMER.md for the full disclaimer.

---

📚 References & Acknowledgements

Resource	Role in This Project
Edge Impulse	Model training platform, DSP block generation, C++ SDK export
TensorFlow Lite for Microcontrollers	On-device neural network inference runtime
CMSIS-DSP	Arm DSP library for signal processing acceleration
Adafruit MLX90614 Library	Arduino I²C driver for the MLX90614 sensor

---

🏷️ Suggested GitHub Topics

tinyml · edge-impulse · tflite-micro · embedded-ml · mlx90614 · arduino · infrared-thermometer · proof-of-concept