AI-Based Signal Processing in Diagnostic Devices

Abstract

Diagnostic medical devices -- encompassing electrocardiographs, pulse oximeters, spirometers, electroencephalographs, and point-of-care ultrasound systems -- rely on signal processing algorithms to transform raw physiological signals into clinically actionable measurements. Traditional digital signal processing (DSP) pipelines employ fixed filter banks, threshold-based feature extraction, and rule-based classification that were designed decades ago for controlled laboratory conditions but struggle with the noise, artefacts, and patient variability encountered in real-world clinical use. AI-based signal processing offers transformative improvements through learned representations that adapt to signal characteristics, reject artefacts contextually, and extract diagnostic features invisible to hand-crafted algorithms. This study presents the AI Signal Processing for Diagnostics Assessment Framework (AISPDAF), evaluating five signal processing approaches -- conventional DSP pipelines, adaptive filtering with Kalman estimation, convolutional neural network (CNN) feature extraction, recurrent neural network (RNN) temporal modelling, and transformer-based foundation models -- across four diagnostic device categories: electrocardiography, pulse oximetry, spirometry, and electroencephalography. Our AI Signal Processing Performance Score (ASPPS) integrates signal-to-noise improvement, diagnostic accuracy enhancement, artefact rejection capability, computational efficiency, and cross-device generalisability. Transformer-based foundation models achieved the highest ASPPS (0.928) through self-supervised pre-training on 48 million unlabelled physiological signal segments that learned universal signal representations transferable across all four device categories, while CNN feature extraction achieved the highest diagnostic accuracy for single-device applications (0.955 for ECG arrhythmia classification)