Academic Commons

Theses Doctoral

Wearable biosensors for mobile health

Colburn, David Alexander

Mobile health (mHealth) promises a paradigm shift towards digital medicine where biomarkers in individuals are continuously monitored with wearable biosensors in decentralized locations to facilitate improved diagnosis and treatment of disease. Despite recent progress, the impact of wearables in health monitoring remains limited due to the lack of devices that measure meaningful health data and are accurate, minimally invasive, and unobtrusive. Therefore, next-generation biosensors must be developed to realize the vision of mHealth. To that end, in this dissertation, we develop wearable biochemical and biophysical sensors for health monitoring that can serve as platforms for future mHealth devices.

First, we developed a skin patch biosensor for minimally invasive quantification of endogenous biochemical analytes in dermal interstitial fluid. The patch consisted of a polyacrylamide hydrogel microfilament array with covalently-tethered fluorescent aptamer sensors. Compared to prior approaches for hydrogel-based sensing, the microfilaments enable in situ sensing without invasive injection or removal. The patch was fabricated via replica molding with high-percentage polyacrylamide that provided high elastic modulus in the dehydrated state and optical transparency in the hydrated state. The microfilaments could penetrate the skin with low pain and without breaking, elicited minimal inflammation upon insertion, and were easily removed from the skin. To enable functional sensing, the polyacrylamide was co-polymerized with acrydite-modified aptamer sensors for phenylalanine that demonstrated reversible sensing with fast response time in vitro. In the future, hydrogel microfilaments could be integrated with a wearable fluorometer to serve as a platform for continuous, minimally invasive monitoring of intradermal biomarkers.

Next, we shift focus to biophysical signals and the required signal processing, particularly towards the development of cuffless blood pressure (BP) monitors. Cuffless BP measurement could enable early detection and treatment of abnormal BP patterns and improved cardiovascular disease risk stratification. However, the accuracy of emerging cuffless monitoring methods is compromised by arm movement due to variations in hydrostatic pressure, limiting their clinical utility. To overcome this limitation, we developed a method to correct hydrostatic pressure errors in noninvasive BP measurements. The method tracks arm position using wearable inertial sensors at the wrist and a deep learning model that estimates parameterized arm-pose coordinates; arm position is then used for analytical hydrostatic pressure compensation. We demonstrated the approach with BP measurements derived from pulse transit time, one of the most well-studied modalities for cuffless BP measurement. Across hand heights of 25 cm above or below the heart, mean errors for diastolic and systolic BP were 0.7 ± 5.7 mmHg and 0.7 ± 4.9 mmHg, respectively, and did not differ significantly across arm positions. This method for correcting hydrostatic pressure may facilitate the development of cuffless devices that can passively monitor BP during everyday activities.

Finally, towards a fully integrated device suitable for ambulatory BP monitoring, we developed a deep learning model for BP prediction from photoplethysmography waveforms acquired at a single measurement site. In contrast to competing methods that require thousands of measurements for adaptation to new users, our proposed approach enables accurate BP prediction following calibration with a single reference measurement. The model uses a convolutional neural network with temporal attention for feature extraction and a Siamese architecture for effective calibration. To prevent overfitting to person-specific variations that fail to generalize, we introduced an adversarial patient classification task to encourage the learning of patient-invariant features. Following calibration, the model accurately predicted diastolic and systolic BP over 24 hours, with mean errors of -0.07 ± 3.86 mmHg and -0.94 ± 7.32 mmHg, respectively, which meets the accuracy criteria for clinical validation. The proposed deep learning model could integrate with wearable photoplethysmography sensors, such as those in smartwatches, to enable cuffless ambulatory BP monitoring.

Underlying this work is the development of minimally invasive biosensors that can integrate with wearable mHealth devices to facilitate passive monitoring of health parameters. The proliferation of mHealth wearables will enable the widespread collection of meaningful health data that provide actionable insights and a more comprehensive understanding of patient health. In a step towards this vision, we leveraged innovations in materials, multi-sensor fusion, and data-driven signal processing to develop sensors for measuring biochemical and biophysical markers. Overall, this work serves as an example of how the adoption of new technologies can facilitate the development of next-generation wearable biosensors.


This item is currently under embargo. It will be available starting 2023-10-05.

More About This Work

Academic Units
Biomedical Engineering
Thesis Advisors
Sia, Samuel K.
Ph.D., Columbia University
Published Here
October 6, 2021