Course Overview
The full stack of clinical AI deployment — from regulation to hardware to data
AI in Wearables & Healthcare · Fall 2025 · UNT · Prof. Dr. Mahdi Pedram · 15 activities
No other course covered all of: FDA regulatory frameworks, embedded microcontroller programming, data collection from real sensors, edge model architectures, hyperparameter search, mobile app deployment, and real clinical dataset analysis (MIMIC-IV). The breadth is the point — deploying AI in healthcare requires understanding every layer of this stack simultaneously.
FDA's Software as a Medical Device (SaMD) framework and mHealth regulation. Multi-level design analysis of the Dexcom CGM: mechanical (waterproof transmitter, applicator), software (colour-coded glucose display, trend arrows), and system (Bluetooth + cloud sync) levels. Analysis of activity monitor accuracy degradation in obesity: hip-worn monitors show altered gait effects, wrist-worn affected by placement at higher BMI. Proposed solution: combined multi-site sensing with subject-specific calibration models — the same approach used in multi-site MRI harmonisation.
Hands-on data collection from the Seeed Studio XIAO nRF52840 Sense (BLE microcontroller with IMU). Three activity types at 50Hz: Sitting, Walking, Stairs. Gesture recognition: Flex vs Punch wrist movements showing distinct temporal acceleration profiles. PPG pulse sensor comparison: Finger provides stronger, more stable signal at rest (BPM ~75); both show motion artefacts during walking (BPM increases to 94–100). First-hand understanding of the gap between lab-grade sensors and wearable-grade sensors — directly informative for WECARE's ECG + IMU design decisions.
186 smartphone photos collected (Activities 7–8), YOLOv8 trained (Activity 14). Overall: Precision 0.915 · Recall 0.874 · mAP@50 = 0.913 · mAP@50-95 = 0.849. Burger class: Recall 100%, mAP@50 = 0.995. Systematic grid search over 230 hyperparameter combinations (embedding × threshold) rather than manual tuning. MIT App Inventor deployment on iPhone. First complete end-to-end experience: data collection → annotation → training → mobile deployment — in a single project.
Analysis of MIMIC-IV — de-identified ICU admissions across unit types and patient demographics. MICU (Medical ICU) admits the highest number of patients aged >89 years (50,000+). Emergency admission type: 42,071 cases — the most common. Working with a real clinical dataset (understanding schema, admission type distribution, patient age distribution) provided the bridge between building models on clean benchmarks and deploying them where data is messy, incomplete, and clinically nuanced.