This talk introduces Emerald, a novel MIT technology for in-home non-intrusive patient monitoring. The Emerald device is a WiFi-like box that runs customized machine learning algorithms to learn digital biomarkers from the wireless signals in the patient's home. It can remotely monitor the patient’s gait speed, falls, respiratory signal, heart rate, and sleep quality and stages. The sensing is completely passive – i.e., the patient can go about her normal life without having to wear any sensors on her body, write a diary, or actively measure herself. This talk will discuss the technology and the results from pilot studies in various therapeutic areas.
Breast cancer is global problem with over 500,000 women dying from the disease every year, yet all of our decisions and insights are based on only a fraction of the information that exists at both the patient and population level. In this talk, we explore a machine learning approach to cancer that integrates rich patient information at population scale, and discuss the type of tools this enables. We have developed A.I systems for automatically reading mammograms, performing personalized risk assessment and mining medical records and implemented them clinically at Massachusetts General Hospital.
Although process data analytics is a valuable tool for improving the manufacturing of biologic drugs, selection of the best method requires a substantial level of expertise. This talk describes a robust and automated approach for process data analytics tool selection that allows the user to focus on goals rather than methods. The approach first applies tools to automatically interrogate the data to ascertain its characteristics, e.g., nonlinearity, correlation, dynamics. This information is then used to select a best-in-class process data analytics tool. The approach is demonstrated for industrial data for the manufacturing of a monoclonal antibody.
What are the prospects for applying AI to improve healthcare? Three types of problems that AI can address in healthcare will be outlined, the most challenging of which is the development of new therapeutics. To address this challenge, we leverage recent advances in machine learning and high-throughput experimentation to apply the engineering cycle to drug discovery. The engineering cycle is based on iteratively measuring a system, modeling it computationally, and manipulating it. Each time the cycle is completed, the results improve. This iterative approach is fundamental to all engineering design but, until now, has had limited impact on drug discovery. Progress on unpublished projects relating to these efforts will be described, including a collaborative, multi-institutional project called Answer ALS.
Contact
As a strategy to save the cost of expensive substrates in semiconductor processing, the technique called “layer-transfer” has been developed. In order to achieve real cost-reduction via the “layer-transfer”, the following needs to be insured: (1) Reusability of the expensive substrate, (2) Minimal substrate refurbishment step after the layer release, (3) Fast release rate, and (4) Precise control of a released interface. Although a number of layer transfer methods have been developed including chemical lift-off, optical lift-off, and mechanical lift-off, none of those three methods fully satisfies conditions listed above. In this talk, we will discuss our recent development in a “graphene-based layer-transfer” process that could fully satisfy the above requirements, where epitaxial graphene can serve as a universal seed layer to grow single-crystalline GaN, III-V, II-VI and IV semiconductor films and a release layer that allows precise and repeatable release at the graphene surface. We will further discuss about cost-effective, defect-free heterointergration of semiconductors using graphene-based layer transfers.
Lastly, I will introduce our new research activities in developing advanced RRAM devices. Resistive switching devices have attracted tremendous attention due to their high endurance, sub-nanosecond switching, long retention, scalability, low power consumption, and CMOS compatibility. RRAMs have also emerged as a promising candidate for non-Von Neumann computing architectures based on neuromorphic and machine learning systems to deal with “big data” problems such as pattern recognition from large amounts of data sets. However, currently reported RRAM devices have not shown uniform switching behaviors across the devices with high on-off ratio which holds up commercialization of RRAM-based data storages as well as demonstration of large-scale neuromorphic functions. Recently, we redesigned RRAM devices and this new device structure exhibits most of functions required for large-array memories and neuromorphic computing, which are (1) excellent retention with high endurance, (2) excellent device uniformity, (3) high on/off current ratio, and (4) current suppression in low voltage regime. I will discuss about the characterization results of this new RRAM device.