This report summarizes the presentations and discussions from a workshop convened at the National Science Foundation (NSF) on March 7-8, 2019 in Alexandria, VA. The focus of this multi-phased workshop was to determine future strategies for advancing the fundamental understanding and engineering of reconfigurable sensor systems by integrating hardware with data harnessing, real-time learning, and artificial intelligence capabilities. Specifically, this workshop addressed the changing application requirements, resources, and future challenges facing the research and development of reconfigurable sensor systems for clinical applications in the fields of medicine, human performance, and behavioral psychology.

The workshop enabled the broader engineering community to discuss and highlight issues confronting the development and application of reconfigurable sensor system for medicine. Through keynote talks, panels, and breakout discussions, researchers, clinicians, and representatives of Government agencies from different disciplines identified challenges and produces a set of recommendations to advance next-generation sensor hardware. The workshop presenters and discussion participants highlighted the shifts in hardware needs to meet the expectation of data scientists in the areas of artificial intelligence and machine learning, with a broad base of applications and resource needs that have more thorough validation and reliability than that provided by current demonstration of reconfigurable sensor system hardware. In summary, the healthcare community's needs are evolving rapidly, and advanced data science capabilities are more pervasive, emphasizing the need for researchers to accelerate the development of validated, user-friendly designs and platforms to support wider application of multimodal sensing hardware.

We summarize the findings by the participants into key aspects as below:

Putting All the Sensors Together: The advances made in flexible electronics integrating multiple types of substrates (i.e., silicon, flexible electrodes, etc.) have been the greatest breakthrough because of the way in which it has extended the reach of sensor technology, allowing different sensor types to be integrated and to interface with soft or flexible biological systems. The greatest barriers still lie in the available bio-recognition elements; accuracy and reliability; in "real-world" operation; packaging; electrical interconnection between flexible and rigid component. The application of data analytics techniques may be able to improve accuracy and reliability. The development of hybrid fabrication strategies, combining 3-D printing with standard microfabrication techniques for on-demand material delivery, can be considered as the next-generation approach for building reconfigurable sensors.

Reliability and Reproducibility of Multimodal Sensing: With miniaturized and multimodal biosensing systems, lack of accuracy and precision makes quantitative analysis difficult. Sensitivity and reproducibility are multiplied in the multimodal systems. This is often the result of (1) each sensor modality having its own failure modes and limitations; (2) testing or demonstration in non-physiologically relevant experimental conditions.  Simulation or actual in vivo testing are expensive to conduct and non-standard. We have to understand how sensors will operate in real conditions. Standardization of sensor testing protocols, both in vitro and in vivo, will enable the use of advanced data analyses and provide for accurate comparisons across sensor modalities.

 Designing for Data: Machine learning and big data analytics are now an essential part of the scientific discovery process, complementary to and increasingly integrated with hardware design. Considerations for the needs of machine learning and big data analytics should be included in the initial design and simulation phased of engineering reconfigurable hardware, i.e. being able to build enough reliable sensors to collect the necessary volume of data for machine learning and artificial intelligence application. Accordingly, the first identified challenge was the absence of collection of data to train machine-learning algorithms to enable an artificial intelligence approach to identifying digital biomarkers. An important resource to achieve this is the databanks that made available by federal agencies such as National Institute of Health (NIH) and Centers for Disease Control and Prevention (CDC), but they are only available for a limited set of sensor modalities.

Medicine Requires Human-Centered Design: Reconfigurable sensor technologies are not readily adopted by the end users, including patients, physicians and nurses. For example, wearables tend to suffer from low patient compliance and digital data from new medical technology can be difficult to interpret by clinical staff. Operator-in-the-loop methods are needed to bridge interface capability gaps, ensure interoperability of the hardware, and reduce hazardous situations.

Organization. The workshop was held at the NSF Headquarters on March 7-8, 2019 in Alexandria, VA. The workshop included approximately 75 participants, drawn from academia, industry, healthcare systems, Federal laboratories, and other Government agencies. A participant list is provided in Appendix A.

Last Modified: 08/21/2019
Modified by: Michael Daniele