Publications

Hemorrhage is a leading cause of trauma death, particularly in prehospital environments when evacuation is delayed. Obtaining central vascular access to a deep artery or vein is important for administration of emergency drugs and analgesics, and rapid replacement of blood volume, as well as invasive sensing and emerging life-saving interventions. However, central access is normally performed by highly experienced critical care physicians in a hospital setting. We developed a handheld AI-enabled interventional device, AI-GUIDE (Artificial Intelligence Guided Ultrasound Interventional Device), capable of directing users with no ultrasound or interventional expertise to catheterize a deep blood vessel, with an initial focus on the femoral vein. AI-GUIDE integrates with widely available commercial portable ultrasound systems and guides a user in ultrasound probe localization, venous puncture-point localization, and needle insertion. The system performs vascular puncture robotically and incorporates a preloaded guidewire to facilitate the Seldinger technique of catheter insertion. Results from tissue-mimicking phantom and porcine studies under normotensive and hypotensive conditions provide evidence of the technique's robustness, with key performance metrics in a live porcine model including: a mean time to acquire femoral vein insertion point of 53 plus or minus 36 s (5 users with varying experience, in 20 trials), a total time to insert catheter of 80 plus or minus 30 s (1 user, in 6 trials), and a mean number of 1.1 (normotensive, 39 trials) and 1.3 (hypotensive, 55 trials) needle insertion attempts (1 user). These performance metrics in a porcine model are consistent with those for experienced medical providers performing central vascular access on humans in a hospital.

Estimations of human joint torques can provide clinically valuable information to inform patient care, plan therapy, and assess the design of wearable robotic devices. Predicting joint torques into the future can also be useful for anticipatory robot control design. In this work, we present a method of mapping joint torque estimates and sequences of torque predictions from motion capture and ground reaction forces to wearable sensor data using several modern types of neural networks. We use dense feedforward, convolutional, neural ordinary differential equation, and long short-term memory neural networks to learn the mapping for ankle plantarflexion and dorsiflexion torque during standing,walking, running, and sprinting, and consider both single-point torque estimation, as well as the prediction of a sequence of future torques. Our results show that long short-term memory neural networks, which consider incoming data sequentially, outperform dense feedforward, neural ordinary differential equation networks, and convolutional neural networks. Predictions of future ankle torques up to 0.4 s ahead also showed strong positive correlations with the actual torques. The proposed method relies on learning from a motion capture dataset, but once the model is built, the method uses wearable sensors that enable torque estimation without the motion capture data.

Future wearable technology may provide for enhanced communication in noisy environments and for the ability to pick out a single talker of interest in a crowded room simply by the listener shifting their attentional focus. Such a system relies on two components, speaker separation and decoding the listener's attention to acoustic streams in the environment. To address the former, we present a system for joint speaker separation and noise suppression, referred to as the Binaural Enhancement via Attention Masking Network (BEAMNET). The BEAMNET system is an end-to-end neural network architecture based on self-attention. Binaural input waveforms are mapped to a joint embedding space via a learned encoder, and separate multiplicative masking mechanisms are included for noise suppression and speaker separation. Pairs of output binaural waveforms are then synthesized using learned decoders, each capturing a separated speaker while maintaining spatial cues. A key contribution of BEAMNET is that the architecture contains a separation path, an enhancement path, and an autoencoder path. This paper proposes a novel loss function which simultaneously trains these paths, so that disabling the masking mechanisms during inference causes BEAMNET to reconstruct the input speech signals. This allows dynamic control of the level of suppression applied by BEAMNET via a minimum gain level, which is not possible in other state-of-the-art approaches to end-to-end speaker separation. This paper also proposes a perceptually-motivated waveform distance measure. Using objective speech quality metrics, the proposed system is demonstrated to perform well at separating two equal-energy talkers, even in high levels of background noise. Subjective testing shows an improvement in speech intelligibility across a range of noise levels, for signals with artificially added head-related transfer functions and background noise. Finally, when used as part of an auditory attention decoder (AAD) system using existing electroencephalogram (EEG) data, BEAMNET is found to maintain the decoding accuracy achieved with ideal speaker separation, even in severe acoustic conditions. These results suggest that this enhancement system is highly effective at decoding auditory attention in realistic noise environments, and could possibly lead to improved speech perception in a cognitively controlled hearing aid.

Precise estimation of uncertainty in predictions for AI systems is a critical factor in ensuring trust and safety. Deep neural networks trained with a conventional method are prone to over-confident predictions. In contrast to Bayesian neural networks that learn approximate distributions on weights to infer prediction confidence, we propose a novel method, Information Aware Dirichlet networks, that learn an explicit Dirichlet prior distribution on predictive distributions by minimizing a bound on the expected max norm of the prediction error and penalizing information associated with incorrect outcomes. Properties of the new cost function are derived to indicate how improved uncertainty estimation is achieved. Experiments using real datasets show that our technique outperforms, by a large margin, state-of-the-art neural networks for estimating within-distribution and out-of-distribution uncertainty, and detecting adversarial examples.

Finding the most likely path to a set of failure states is important to the analysis of safety critical systems that operate over a sequence of time steps, such as aircraft collision avoidance systems and autonomous cars. In many applications such as autonomous driving, failures cannot be completely eliminated due to the complex stochastic environment in which the system operates. As a result, safety validation is not only concerned about whether a failure can occur, but also discovering which failures are most likely to occur. This article presents adaptive stress testing (AST), a framework for finding the most likely path to a failure event in simulation. We consider a general black box setting for partially observable and continuous-valued systems operating in an environment with stochastic disturbances. We formulate the problem as a Markov decision process and use reinforcement learning to optimize it. The approach is simulation-based and does not require internal knowledge of the system, making it suitable for black-box testing of large systems. We present different formulations depending on whether the state is fully observable or partially observable. In the latter case, we present a modified Monte Carlo tree search algorithm that only requires access to the pseudorandom number generator of the simulator to overcome partial observability. We also present an extension of the framework, called differential adaptive stress testing (DAST), that can find failures that occur in one system but not in another. This type of differential analysis is useful in applications such as regression testing, where we are concerned with finding areas of relative weakness compared to a baseline. We demonstrate the effectiveness of the approach on an aircraft collision avoidance application, where a prototype aircraft collision avoidance system is stress tested to find the most likely scenarios of near mid-air collision.

Compared to other major medical imaging modalities such as X-ray, computed tomography (CT), and magnetic resonance imaging, medical ultrasound (US) has unique attributes that make it the preferred modality for many clinical applications. In particular, US is nonionizing, portable, and provides real-time imaging, with adequate spatial and depth resolution to visualize tissue dynamics. The ability to measure Doppler information is also important, particularly for measuring blood flows. The small size of US transducers is a key attribute for intravascular applications. In addition, accessibility has been increased with the use of portable US, which continues to move toward a smaller footprint and lower cost. Nowadays, some US probes can even be directly connected to a phone or tablet. On the other hand, US also has unique challenges, particularly in that image quality is highly dependent on the operator’s skill in acquiring images based on the proper position, orientation, and probe pressure. Additional challenges that further require operator skill include the presence of noise, artifacts, limited field of view, difficulty in imaging structures behind bone and air, and device variability across manufacturers. Sonographers become highly proficient through extensive training and long experience, but high intra- and interobserver variability remains. This skill dependence has limited the wider use of US by healthcare providers who are not US imaging specialists. Recent advances in machine learning (ML) have been increasingly applied to medical US (Brattain, Telfer, Dhyani, Grajo, & Samir, 2018), with a goal of reducing intra- and interobserver variability as well as interpretation time. As progress toward these goals is made, US use by nonspecialists is expected to proliferate, including nurses at the bedside or medics in the field. The acceleration in ML applications for medical US can be seen from the increasing number of publications (Fig. 8.1) and Food and Drug Administration (FDA) approvals (Table 8.1) in the past few years. Fig. 8.1 shows that cardiovascular applications (spanning the heart, brain and vessels) have received the most attention, compared to other organs. Table 8.1 shows that pace of US FDA-cleared artificial intelligence (AI) products that combine AI and ultrasound is accelerating. Of note, many of the products have been approved over the last couple of years. Companies such as Butterfly Network (Guilford, CT) have also demonstrated AI-driven applications for portable ultrasound and more FDA clearances are expected to be published. The goals of this chapter are to highlight the recent progress, as well as the current challenges and future opportunities. Specifically, this chapter addresses topics such as the following: (1) what is the current state of machine learning for medical US application, both in research and commercially; (2) what applications are receiving the most attention and have performance improvements been quantified; (3) how do ML solutions fit in an overall workflow; and (4) what open-source datasets are available for the broader community to contribute to progress in this field. The focus is on cardiovascular applications (Section Cardiovascular/echocardiography), but common themes and differences for other applications for medical US are also summarized (Section Breast, liver, and thyroid ultrasound). A discussion is offered in Discussion and outlook section.

Early detection for sepsis, a high-mortality clinical condition, is important for improving patient outcomes. The performance of conventional deep learning methods degrades quickly as predictions are made several hours prior to the clinical definition. We adopt recurrent neural networks (RNNs) to improve early prediction of the onset of sepsis using times series of physiological measurements. Furthermore, physiological data is often missing and imputation is necessary. Absence of data might arise due to decisions made by clinical professionals which carries information. Using the missing data patterns into the learning process can further guide how much trust to place on imputed values. A new multi-task LSTM model is proposed that takes informative missingness into account during training that effectively attributes trust to temporal measurements. Experimental results demonstrate our method outperforms conventional CNN and LSTM models on the PhysioNet-2019 CiC early sepsis prediction challenge in terms of area under receiver-operating curve and precision-recall curve, and further improves upon calibration of prediction scores.

The MIT/IEEE/Amazon GraphChallenge.org encourages community approaches to developing new solutions for analyzing graphs and sparse data. Sparse AI analytics present unique scalability difficulties. The Sparse Deep Neural Network (DNN) Challenge draws upon prior challenges from machine learning, high performance computing, and visual analytics to create a challenge that is reflective of emerging sparse AI systems. The sparse DNN challenge is based on a mathematically well-defined DNN inference computation and can be implemented in any programming environment. In 2019 several sparse DNN challenge submissions were received from a wide range of authors and organizations. This paper presents a performance analysis of the best performers of these submissions. These submissions show that their state-of-the-art sparse DNN execution time, TDNN, is a strong function of the number of DNN operations performed, Nop. The sparse DNN challenge provides a clear picture of current sparse DNN systems and underscores the need for new innovations to achieve high performance on very large sparse DNNs.

In this paper we propose an integrated system concept for autonomously surveying and planning emergency response for areas impacted by natural disasters. Referred to as AASAPS-HADR, this system is composed of a network of ground stations and autonomous aerial vehicles interconnected by an ad hoc emergency communication network. The system objectives are three-fold: to provide situational awareness of the evolving disaster event, to generate dispatch and routing plans for emergency vehicles, and to provide continuous communication networks which augment pre-existing communication infrastructure that may have been damaged or destroyed. Lacking development in previous literature, we give particular emphasis to the situational awareness objective of disaster response by proposing an autonomous aerial survey that is tasked with assessing damage to existing road networks, detecting and locating human victims, and providing a cursory assessment of casualty types that can be used to inform medical response priorities. In this paper we provide a high-level system design concept, identify existing AI perception and planning algorithms that most closely suit our purposes as well as technology gaps within those algorithms, and provide initial experimental results for non-contact health monitoring using real-time pose recognition algorithms running on a Nvidia Jetson TX2 mounted on board a quadrotor UAV. Finally we provide technology development recommendations for future phases of the AASAPS-HADR system.

The AI revolution is data driven. AI "data wrangling" is the process by which unusable data is transformed to support AI algorithm development (training) and deployment (inference). Significant time is devoted to translating diverse data representations supporting the many query and analysis steps found in an AI pipeline. Rigorous mathematical representations of these data enables data translation and analysis optimization within and across steps. Associative array algebra provides a mathematical foundation that naturally describes the tabular structures and set mathematics that are the basis of databases. Likewise, the matrix operations and corresponding inference/training calculations used by neural networks are also well described by associative arrays. More surprisingly, a general denormalized form of hierarchical formats, such as XML and JSON, can be readily constructed. Finally, pivot tables, which are among the most widely used data analysis tools, naturally emerge from associative array constructors. A common foundation in associative arrays provides interoperability guarantees, proving that their operations are linear systems with rigorous mathematical properties, such as, associativity, commutativity, and distributivity that are critical to reordering optimizations.