Presentation
A User-Centered Approach for Reducing Unsuccessful Heart Rate Measurements in PPG-Based Blood Pressure Monitoring Devices by Prompting User Behavior
SessionPoster Session 2
DescriptionPhotoplethysmography (PPG) detects changes in blood volume in the microvasculature under the skin using optical techniques. This is integrated into a wearable device to enable continuous monitoring of cardiovascular health metrics such as heart rate, blood oxygen saturation, and, most recently, high blood pressure. This technology provides significant clinical benefits by enabling remote patient monitoring, continuous data collection, early detection of cardiovascular irregularities, and effective management of chronic diseases. These advantages are especially valuable for patients or healthcare providers who encounter limited resources or obstacles to conventional care methods. Nonetheless, its reliability is often compromised when users are moving, adopting poor posture, or measuring in low-light environments.
Companies have traditionally solved this issue by filtering poor quality signals as “unsuccessful measurement”, giving general information with different possibilities for the reason why the measurement was unsuccessful, which does not give users good information on how to improve their next measurements.
The need is especially urgent as these tools are becoming more mainstream. For example, in September 2025, Apple announced a high blood pressure screening tool, which works by taking the data from the optical heart sensor from the watch to an algorithm that can detect potential hypertension by analyzing how the blood vessels respond to beats of the heart over a period of thirty days. Users are then notified of the possibility of having hypertension.
The accuracy of this PPG blood pressure measurement will rely completely on the quality of the signals that the sensor captures. Poor signal quality may trigger unnecessary warnings that amplify user stress or reduce adherence to the tool.
Our work takes a human factors approach to the design of AI, with a focus on generating actionable model outputs to improve input signal quality. To show our approach is feasible, the artificial intelligence pipeline is presented, from preprocessing and feature extraction from the waveforms to the training of a comprehensive number of models with stratified splits to ensure balanced class representation. Standard classification metrics were used to determine the best model.
Then, to enhance transparency, two explainable AI methods, SHAP and LIME, were employed to interpret model outputs and identify the influential features that determine whether a signal is of poor quality, leading to an unsuccessful reading. Llama 3 was then integrated into the pipeline to explain in plain language why a signal is classified as poor, and suggest corrective actions (e.g., adjust lighting, reduce hand motion, stop speaking). Test cases were randomly selected and fed into the pipeline to obtain the corrective action.
Usability testing with prototypes that simulate the measurement failure was conducted. Using a think-aloud protocol, the research focuses on how interface feedback with corrective actions influences comprehension, willingness to act, and emotional response.
By shifting the focus from reactive algorithmic filtering to proactive, behaviorally informed interface design, we aim to increase the proportion of successful readings and thereby support accurate hypertension notifications and overall trust in the technology. The key takeaway: wearable reliability depends not only on sensors and algorithms but on interfaces that empower users to succeed. Developing these interfaces successfully requires specific AI features to be used in the AI pipeline. By aligning design with user motivation and behavioral science, this work helps ensure that hypertension notifications and other critical health features are presented to the user, often patients or potential patients, in an accurate and effective way.
Companies have traditionally solved this issue by filtering poor quality signals as “unsuccessful measurement”, giving general information with different possibilities for the reason why the measurement was unsuccessful, which does not give users good information on how to improve their next measurements.
The need is especially urgent as these tools are becoming more mainstream. For example, in September 2025, Apple announced a high blood pressure screening tool, which works by taking the data from the optical heart sensor from the watch to an algorithm that can detect potential hypertension by analyzing how the blood vessels respond to beats of the heart over a period of thirty days. Users are then notified of the possibility of having hypertension.
The accuracy of this PPG blood pressure measurement will rely completely on the quality of the signals that the sensor captures. Poor signal quality may trigger unnecessary warnings that amplify user stress or reduce adherence to the tool.
Our work takes a human factors approach to the design of AI, with a focus on generating actionable model outputs to improve input signal quality. To show our approach is feasible, the artificial intelligence pipeline is presented, from preprocessing and feature extraction from the waveforms to the training of a comprehensive number of models with stratified splits to ensure balanced class representation. Standard classification metrics were used to determine the best model.
Then, to enhance transparency, two explainable AI methods, SHAP and LIME, were employed to interpret model outputs and identify the influential features that determine whether a signal is of poor quality, leading to an unsuccessful reading. Llama 3 was then integrated into the pipeline to explain in plain language why a signal is classified as poor, and suggest corrective actions (e.g., adjust lighting, reduce hand motion, stop speaking). Test cases were randomly selected and fed into the pipeline to obtain the corrective action.
Usability testing with prototypes that simulate the measurement failure was conducted. Using a think-aloud protocol, the research focuses on how interface feedback with corrective actions influences comprehension, willingness to act, and emotional response.
By shifting the focus from reactive algorithmic filtering to proactive, behaviorally informed interface design, we aim to increase the proportion of successful readings and thereby support accurate hypertension notifications and overall trust in the technology. The key takeaway: wearable reliability depends not only on sensors and algorithms but on interfaces that empower users to succeed. Developing these interfaces successfully requires specific AI features to be used in the AI pipeline. By aligning design with user motivation and behavioral science, this work helps ensure that hypertension notifications and other critical health features are presented to the user, often patients or potential patients, in an accurate and effective way.
Event Type
Poster Presentation
TimeTuesday, March 244:45pm - 6:15pm EDT
LocationRhinelander Gallery
Digital Health