Leader: Zaccaria Del Prete (Sapienza); Other collaborator(s): CNR-IMM
Development of both hardware and AI-based algorithms of the local mesuring hub of the cloud based system. The hub will provide an “age-friendly” interface for filling in anamnestic questionnaire and performing guided measurements of physiological parameters. To this purpose, the hub will integrate:
• New optical sensors for Oxygen Saturation and blood pressure
• Analog front end for Biopotential (ECG, R-to-R) and Bioimpedance (plethysmography).
• New microfluidic sensors for breath monitoring.
Low cost IMU-based or insole pressure sensor-based devices will be developed for continuous monitoring of the risk of fall, to feed specific AI-algorithms implemented in the hub.
Considering the development of new sensor typologies and novel AI algorithms, starting TRL is equal to 1-2 while target TRL will be equal to 4.
Brief description of the activities and of the intermediate results:
In the reference period (November 2023 to March 2024), we developed the first prototype of a hub for ECG, RR, and PPG measurements, along with related data processing. This included the segmentation of ECG and PPG traces to identify any pathologies and/or correlations with the age of the subjects.
Main policy, industrial and scientific implications:
The results emerging from the research conducted during the reference period (November 2023 to March 2024) could potentially have a significant innovative impact on the design of small, reliable multi-sensor local hubs for at-home applications.
Please see the next reporting period.
During the July-September quarter, our focus was on two tasks: evaluating the accuracy of ECG wave feature detection and segmentation and developing and validating algorithms for calibrating photoplethysmography (PPG) sensors.
The first task began with an in-depth review of existing methods for ECG wave feature detection, including algorithms based on machine learning, signal processing, and heuristic approaches. We then utilized an external ECG signal synthesizer to simulate both healthy and unhealthy ECG signals under varying noise conditions. This simulation aimed to assess the performance of a specially designed algorithm in detecting key features within ECG signals, specifically the P, QRS, and T waves. In the second phase, we acquired ECG signals from two cohorts of healthy subjects (young and adult) to evaluate the sensitivity and accuracy of our algorithm using real signals. Initial findings indicate varied accuracy levels among different algorithms. Traditional methods like Pan-Tompkins exhibit competitive performance in detecting QRS complexes but struggle with noise and non-stationary signals. In contrast, machine learning-based methods demonstrate improved segmentation, particularly in noisy environments. These results highlight the need for hybrid approaches that combine traditional signal processing with modern machine learning techniques. The accuracy evaluation revealed that the integrated circuit's performance demonstrates a normalized root mean square error lower than 10% in the worst-case scenario (signals with power line noise) and an average of 2.5% in the best-case scenario (no noise). Repeatability and reproducibility were generally higher for typical ECG signals, with power line noise minimally affecting repeatability, while white noise had the most significant attenuation impact. Our algorithm achieved a 95% detection rate for P, R, and S peaks, and 98% precise segmentation for the QRS complex under no-noise conditions. In noisy environments, segmentation rates slightly decreased, achieving 87% for the QRS complex and 91% for P and T peaks. Preliminary results from healthy subjects aligned with those from synthesized signals, indicating promising segmentation performance.
The second task involved developing and validating an algorithm for calibrating PPG sensors, focusing on accuracy and reliability. The design phase included an in-depth literature analysis to create an algorithm that adjusts for calibration errors, compensating for variables like sensor placement, skin type, and ambient light interference. Preliminary calibration utilized a dataset of raw PPG signals from tested sensors and reference devices compliant with the ISO/CD-80601-2-61 standard. The performance of the calibration algorithm was assessed based on bias, precision, accuracy, and RMSE. The performance of the calibration algorithm is assessed from a metrological point of view focusing on bias, precision, accuracy, and RMSE. Preliminary results indicate that the calibration algorithm significantly reduces measurement error in PPG signals, improving accuracy by up to 15% compared to uncalibrated devices. Moreover, with the proposed algorithm an experimental calibration curve correlating the average PPG ratio with SpO2 levels by collecting signals from a cohort of fifty healthy subjects was assed, showing a linear calibration equation with an R2=0.997, an RMSE=0.11 and a maximum uncertainty of 3.7%.
Tests on a cohort of unhealthy subjects are currently underway.
Scientific publications
Dissemination events