Project
| # | Title | Team Members | TA | Documents | Sponsor |
|---|---|---|---|---|---|
| 32 | Insight: Cardiovascular Screening Device |
Ethan Pereira Jay Nathan Rishab Iyer |
Weiman Yan | design_document1.pdf final_paper1.pdf photo1.jpeg presentation1.pptx proposal1.pdf |
|
| After discussing with both Professor Rakesh Kumar and Professor Arne Filthelt, we have addressed some limitations and concerns regarding our initial project proposal. Here is our revised proposal. # Insight: Cardiovascular Screening Device ## Team Members: Jay Nathan (jayrn2) Rishab Iyer (riyer20) Ethan Pereira (ethanrp2) # Problem Cardiovascular disease (CVD) is the leading cause of death worldwide, responsible for nearly 20 million deaths annually, about one in three deaths overall. A significant share of these fatalities occur without prior diagnosis: approximately 45% of sudden cardiac deaths happen in individuals with no previously recognized heart disease, while nearly 20% of adults with hypertension and up to 23% of those with atrial fibrillation remain undiagnosed. These silent conditions, such as hypertension, arrhythmias, and sinus bradycardia risk factors, drive the majority of preventable CVDs. Current solutions remain fragmented, while comprehensive screening still requires multiple expensive clinical visits, such as blood pressure measurement, lipid panels, ECGs, and rhythm monitoring, creating barriers for uninsured or underserved populations. The impact is most severe in rural communities, where mortality rates are 20% higher than in urban areas due to limited access to screening. Yet the challenge extends to cities as well, where preventive tests are often costly, not covered by insurance, and therefore underutilized. Consumer devices like blood pressure cuffs, smartwatches, and single-lead ECGs are disjointed, expensive, and difficult to interpret. Critically, there is no affordable, comprehensive, and user-friendly screening solution that can detect CVD risks. # Solution We propose a low-cost, cardiovascular screener to detect the leading drivers of CVD: hypertension, atrial fibrillation (AFib), and sinus bradycardia (SB). The device combines Electrocardiogram (ECG), Photoplethysmography (PPG), and accelerometer-based noise correction for accurate measurements. Mathematical regression-based models will then analyze the digital signals and generate comprehensible readings. Technically, the system consists of a single custom PCB with built-in ECG and PPG sensors plus an accelerometer to capture physiological signals simultaneously. Both the ECG and PPG sensors will be placed on the board to focus on immediate automated screening. A data acquisition board with an Atmega microcontroller synchronizes and packages the data for transmission via USB. A connected PC runs lightweight mathematical regression models for hypertension, AFib, and SB detection. Results are displayed through a simple web dashboard for easy interpretation. # Solution Components ## Subsystem 1: Data Acquisition System (DAS) Function: The system captures physiological data from both an ECG and a PPG sensor (both on board). An accelerometer is also connected to support motion-based denoising of the signal. The microcontroller processes the data we’re reading and sends it into the ML pipeline on the PC through a USB port. Components: - ECG Sensor: AD8232 Analog Front End for ECG - PPG Sensor: MAX30102 - Accelerometer: Adafruit LIS3DH - Microcontroller: ATmega328P MCU - Comm Port: USB ## Subsystem 2: Firmware & Communication Function: Firmware handles ADC sampling, calibration routines, and packaging data for transmission. Supports firmware for reading data from both sensors. USB-based serial communication happens between the ATmega on the DAS board and the host PC. The firmware will send ECG/PPG data as formatted CSV files to be processed by the models. Components: - Microcontroller: ATmega328P MCU - Sensor Buses: SPI/I²C - Comm Port: USBC or micro-USB connector on the DAS PCB - Host Software: an application on the PC to receive and display data on the web app ## Subsystem 3: Backend Processing and ML Inference Function 1: Use ECG/PPG-derived HRV(heart rate variability) and irregularity features to classify AFib vs. normal rhythm using a lightweight classifier on RR(beat-to-beat intervals) sequences. Function 2: Compute mean HR and rhythm regularity from ECG/PPG to flag sinus bradycardia when < 60 bpm using simple threshold logic (no training needed). Function 3: Combine PTT (ECG→PPG latency) and PPG features to classify Normal/Elevated/Hypertensive risk levels using a classifier trained with cuff labels. Components: - Compute: PC(Laptop) for computation. - ML Libraries: PyTorch, scikit-learn, XGBoost, MLflow - Preprocessing module: filtering, motion gating, feature extraction (RR, PP, PTT, HRV, PPG morphology). - Models: AFib classifier, Brady threshold logic, BP-risk classifier. - Calibration & Decision Layer: Produces JSON outputs with labels, confidence, and window duration # Criterion For Success ## Functional Affordable Prototype: A working hardware prototype integrating both ECG and PPG on the same board, and collecting, processing, and displaying data in real time. Target total cost remains < $70. ## Signal Processing with Baseline Accuracy: The system will have noise filtering and feature extraction (heart rate and rhythm patterns) on baseline ECG and PPG datasets from a sampled group of individuals. Success criteria include: Accuracy Benchmarks: Heart rate estimated within ±5 bpm of reference values in ≥80% of trials, and rhythm classification achieving ≥75% agreement with reference annotations. Literature-based Targets: Based on literature, we should aim for AFib detection AUC ≥ 0.95 (high ability to differentiate AFib from normal rhythm), Sinus Bradycardia detection F1 ≥ 0.95 (recall for slow, regular rhythms), and hypertension classification Macro-F1 ≥ 0.80 with one calibration reading (accuracy across Normal/Elevated/Hypertensive categories when compared to a real reference). Dataset Training & Validation: We will validate performance through repeated training and testing, comparing our outputs against certified standards for consistency. Evaluation will be conducted using publicly available medical datasets such as MIT-BIH (ECG) and PPG-DaLiA (PPG) to test reliability across diverse conditions. Noise Mitigation: Integration of accelerometer inputs for motion-based denoising. ## ML Interpretation: Backend ML models must correctly interpret ECG/PPG readings to achieve at least 85% classification accuracy on validation datasets for AFib detection and hypertension risk classification, showing that the models can reliably separate normal from abnormal conditions when tested against certified public medical datasets. ## Repeatability & Sensor Placement: The same individual tested five times in a row should yield similar results, with differences within a 10% margin for all key metrics (heart rate, rhythm classification, BP risk). For standard ECG lead & sensor placement, we will use standard placement (as shown in LITFL ECG Lead Positioning) to reduce variability caused by electrode misplacement. # Resources: ECG Lead positioning: https://litfl.com/ecg-lead-positioning/ Atrial Fibrillation: https://pmc.ncbi.nlm.nih.gov/articles/PMC11262392/ Sinus Bradycardia: https://www.researchgate.net/publication/377965437_Detecting_Sinus_Bradycardia_From_ECG_Signals_Using_Signal_Processing_And_Machine_Learning Hypertension: https://pmc.ncbi.nlm.nih.gov/articles/PMC11904724/#S10 |
|||||