Decision making is one of those extremely complex things that humans can do with relative ease most of the time. Healthcare providers do this hundreds and thousands of times per day, and do an amazing job given their various levels of expertise and the resources available to them. The Elsa Health Assistant is a set of tools and technologies that leverage advances in Artificial Intelligence and causal modeling to augment the capacity of lower cadre healthcare providers and support optimal and consistent decision making. Here we will share the challenges, failures and successes of the technologies and the team.

Consumer wearable sensors are estimated to be used by one in five Americans for tracking fitness and other personal health. Recently, they have been touted as low-cost vehicles for frequent healthcare monitoring and have received approval as diagnostic devices to detect conditions such as atrial fibrillation. Common fitness tracker measurements such as heart rate or steps can be used to implicate underlying causes. One application of interest is to anticipate or detect influenza-like illness (ILI). However, a timely detection of influenza is a challenge as the virus can be transmitted prior to symptom onset (pre-symptomatic), or by individuals who harbour the virus, but do not experience symptoms (asymptomatic). Similarly, 44\% of viral shedding of COVID-19, another disease which causes ILI, in symptomatic individuals happens prior to the onset of symptoms. We investigate if ILI (as caused by influenza, COVID-19, and other diseases) can be detected by wearable sensors, and if possible, how early we can anticipate the onset of symptoms. Having a system to warn users that they are about to become ill can reduce viral transmission -- mitigating the spread of seasonal influenza and suppressing the COVID-19 epidemic.

ILI symptoms can be detected from wearable sensors. For example, temperature covaries with cardiac rhythm. The associated increase of resting heart rate (RHR) during ILI has been demonstrated in previous studies and has been used to estimate the incidence of influenza at a population level, using data collected from wearable sensors. Yet, individual-level ILI predictions from wearable features have been elusive, though research is actively underway. Rigorous work is required to evaluate the sensitivity of models that anticipate ILI onset prior to experiencing symptoms. In this paper we expose potential pitfalls in building ILI prediction models. Specifically, we compare the performance of a model trained and evaluated retrospectively with a held-out set of subjects versus prospectively on a held-out future week of data, mimicking actual deployment scenarios. We show that when the design is focused on deployment, though the performance may drop, it is still improved over naive baselines, indicating potential real-world applications.

Modern wearable devices are embedded with a range of noninvasive biomarker sensors that hold promise for improving detection and treatment of disease. One such sensor is the single-lead electrocardiogram (ECG) which measures electrical signals in the heart. The benefits of the sheer volume of ECG measurements with rich longitudinal structure made possible by wearables come at the price of potentially noisier measurements compared to clinical ECGs, e.g., due to movement. In this work, we develop a statistical model to simulate a structured noise process in ECGs derived from a wearable sensor, design a beat-to-beat representation that is conducive for analyzing variation, and devise a factor analysis-based method to denoise the ECG. We study synthetic data generated using a realistic ECG simulator and a structured noise model. At varying levels of signal-to-noise, we quantitatively measure an upper bound on performance and compare estimates from linear and non-linear models. Finally, we apply our method to a set of ECGs collected by wearables in a mobile health study.

The COVID-19 outbreak demonstrates the need for measurement of hand hygiene behaviors such as handwashing and face touching to prevent the spread of infectious diseases. Wearable technologies and machine-learning-based algorithms can be used to automatically detect these behaviors. In this work, we demonstrate a recurrent neural network with a set of local-extrema-based features for detecting hand hygiene behaviors (handwashing and face touching activities simultaneously) using data from inertial sensors (i.e., accelerometer, magnetometer, and gyroscope) on the wrist(s). The training and validation dataset were gathered from ten individuals; each person provided 60 min of data (sampled at 100 Hz) while performing 12 steps of handwashing, 8 variations of face touching, and 7 variations of other face-to-head gestures across six sessions. With 10 min of person-specific training data, the real-time algorithm achieved its best performance (F1-score of 0.88 for handwashing steps and 0.80 for face touching) using leave-one-session-out validation. We also describe a pilot evaluation on six-hour, free-living waking-day datasets of two participants annotated via front-facing video.

Reinforcement Learning provides an attractive suite of online learning methods for personalizing interventions in a Digital Health. However after an reinforcement learning algorithm has been run in a clinical study, how do we assess whether personalization occurred? We might find users for whom it appears that the algorithm has indeed learned in which contexts the user is more responsive to a particular intervention. But could this have happened completely by chance? We discuss some first approaches to addressing these questions.

Mobile health seeks to provide in-the-moment support to individuals in need. In this talk, I will discuss the challenges associated with behavior and interventions that are based on language. Language is high-dimensional and complex, but is a critical component of many health and support interactions. Specifically, I will describe how we can measure empathy in mental health peer support and how we can give feedback in order to empower peer supporters to increase expressed levels of empathy, using large-scale neural transformer architectures and reinforcement learning.

Mobile health (mHealth) apps such as menstrual trackers provide a rich source of self-tracked health observations that can be leveraged for statistical modeling. However, such data streams are notoriously unreliable since they hinge on user adherence to the app. Thus, it is crucial for machine learning models to account for self-tracking artifacts like skipped self-tracking. In this abstract, we propose and evaluate a hierarchical, generative model for predicting next cycle length based on previously tracked cycle lengths that accounts explicitly for the possibility of users forgetting to track their period. Our model offers several advantages: 1) accounting explicitly for self-tracking artifacts yields better prediction accuracy as likelihood of skipping increases; 2) as a generative model, predictions can be updated online as a given cycle evolves; and 3) its hierarchical nature enables modeling of an individual's cycle length history while incorporating population-level information. Our experiments using real mHealth cycle length data from 5,000 menstruators show that our method yields state-of-the-art performance against neural network-based and summary statistic-based baselines.

Depression and post-traumatic stress disorder (PTSD) are psychiatric conditions commonly associated with experiencing a traumatic event. Estimating mental health status through non-invasive techniques such as activity-based algorithms can help to identify successful early interventions. In this work, we used locomotor activity captured from 1113 individuals who wore a research grade smartwatch post-trauma. A convolutional variational autoencoder (VAE) architecture was used for unsupervised feature extraction from four weeks of actigraphy data. By using VAE latent variables and the participant’s pre-trauma physical health status as features, a logistic regression classifier achieved an area under the receiver operating characteristic curve (AUC) of 0.64 to estimate mental health outcomes. The results indicate that the VAE model is a promising approach for actigraphy data analysis for mental health outcomes in long-term studies.

Users can be supported to adopt healthy behaviors, such as regular physical activity, via relevant and timely suggestions on their mobile devices. Recently, reinforcement learning algorithms have been found to be effective for learning the optimal context under which to provide suggestions. However, these algorithms are not necessarily designed for the constraints posed by mobile health (mHealth) settings, that they be efficient, domain-informed and computationally affordable. We propose an algorithm for providing physical activity suggestions in mHealth settings. Using domain-science, we formulate a contextual bandit algorithm which makes use of a linear mixed effects model. We then introduce a procedure to efficiently perform hyper-parameter updating, using far less computational resources than competing approaches. Not only is our approach computationally efficient, it is also easily implemented with closed form matrix algebraic updates and we show improvements over state of the art approaches both in speed and accuracy of up to 99% and 56% respectively.