Christopher M. Danforth’s research while affiliated with University of Vermont and other places

Although nearly 600,000 people experience homelessness in the United States (U.S) every year, efforts to address this public health crisis are limited by the underperformance of standard methods to estimate localized and nationwide homelessness. Recent studies suggest social media activity can function as a proxy for measures of state-level public health, detectable through straightforward applications of natural language processing. We present the results of our efforts to apply this approach to estimate homelessness at the state level throughout the US during the period 2010-2019 and 2022 using a dataset of roughly 1 million geotagged tweets containing the substring “homeless.” Correlations between homelessness-related tweet counts and ranked per capita homelessness volume, but not general-population densities, suggest a relationship between the likelihood of Twitter users to personally encounter or observe homelessness in their everyday lives and their likelihood to communicate about it online. An increase in the log-odds of the word “homeless” appearing in an English-language tweet, as well as an acceleration in the increase in average tweet sentiment, suggest that tweets about homelessness are also affected by trends at the nation-scale. Additionally, changes to the lexical content of tweets over time suggest that reversals to the polarity of national or state-level trends may be detectable through an increase in political or service-sector language over the semantics of charity or direct appeals. While a computational approach to social media analysis may provide a low-cost, real-time dataset rich with information about nationwide and localized impacts of homelessness and homelessness policy, we find that practical issues abound, limiting the potential of social media as a proxy to complement other measures of homelessness.

To optimize interventions for improving wellness, it is essential to understand habits, which wearable devices can measure with greater precision. Using high temporal resolution biometric data taken from the Oura Gen3 ring, we examine daily and weekly sleep and activity patterns of a cohort of young adults (N=582) in their first semester of college. A high compliance rate is observed for both daily and nightly wear, with slight dips in wear compliance observed shortly after waking up and also in the evening. Most students have a late-night chronotype with a median midpoint of sleep at 5AM, with males and those with mental health impairment having more delayed sleep periods. Social jetlag, or the difference in sleep times between free days and school days, is prevalent in our sample. While sleep periods generally shift earlier on weekdays and later on weekends, sleep duration on both weekdays and weekends is shorter than during prolonged school breaks, suggesting chronic sleep debt when school is in session. Synchronized spikes in activity consistent with class schedules are also observed, suggesting that walking in between classes is a widespread behavior in our sample that substantially contributes to physical activity. Lower active calorie expenditure is associated with weekends and a delayed but longer sleep period the night before, suggesting that for our cohort, active calorie expenditure is affected less by deviations from natural circadian rhythms and more by the timing associated with activities. Our study shows that regular sleep and activity routines may be inferred from consumer wearable devices if high temporal resolution and long data collection periods are available.

Tokenization is a necessary component within the current architecture of many language models, including the transformer-based large language models (LLMs) of Generative AI, yet its impact on the model's cognition is often overlooked. We argue that LLMs demonstrate that the Distributional Hypothesis (DM) is sufficient for reasonably human-like language performance, and that the emergence of human-meaningful linguistic units among tokens motivates linguistically-informed interventions in existing, linguistically-agnostic tokenization techniques, particularly with respect to their roles as (1) semantic primitives and as (2) vehicles for conveying salient distributional patterns from human language to the model. We explore tokenizations from a BPE tokenizer; extant model vocabularies obtained from Hugging Face and tiktoken; and the information in exemplar token vectors as they move through the layers of a RoBERTa (large) model. Besides creating sub-optimal semantic building blocks and obscuring the model's access to the necessary distributional patterns, we describe how tokenization pretraining can be a backdoor for bias and other unwanted content, which current alignment practices may not remediate. Additionally, we relay evidence that the tokenization algorithm's objective function impacts the LLM's cognition, despite being meaningfully insulated from the main system intelligence.

Objectives Despite the development of efficacious wellness interventions, sustainable wellness behavior change remains challenging. To optimize engagement, initiating small behaviors that build upon existing practices congruent with individuals’ lifestyles may promote sustainable wellness behavior change. In this study, we crowd-sourced helpful, flexible, and engaging wellness practices to identify a list of those commonly used for improving sleep, productivity, and physical, emotional, and social wellness from participants who felt they had been successful in these dimensions. Method We recruited a representative sample of 992 U.S. residents to survey the wellness dimensions in which they had achieved success and their specific wellness practices. Results Responses were aggregated across demographic, health, lifestyle factors, and wellness dimension. Exploration of these data revealed that there was little overlap in preferred practices across wellness dimensions. Within wellness dimensions, preferred practices were similar across demographic factors, especially within the top 3–4 most selected practices. Interestingly, daily wellness practices differ from those typically recommended as efficacious by research studies and seem to be impacted by health status (e.g., depression, cardiovascular disease). Additionally, we developed and provide for public use a web dashboard that visualizes and enables exploration of the study results. Conclusions Findings identify personalized, sustainable wellness practices targeted at specific wellness dimensions. Future studies could leverage tailored practices as recommendations for optimizing the development of healthier behaviors.

The common phrase 'representation matters' asserts that media has a measurable and important impact on civic society's perception of self and others. The representation of health in media, in particular, may reflect and perpetuate a society's disease burden. Here, for the top 10 major causes of death in the United States, we examine how cinematic representation of overall and by-gender mortality diverges from reality. Using crowd-sourced data on film deaths from Cinemorgue Wiki, we employ natural language processing (NLP) techniques to analyze shifts in representation of deaths in movies versus the 2021 National Vital Statistic Survey (NVSS) top ten mortality causes. Overall, movies strongly overrepresent suicide and, to a lesser degree, accidents. In terms of gender, movies overrepresent men and underrepresent women for nearly every major mortality cause, including heart disease and cerebrovascular disease. The two exceptions for which women are overrepresented are suicide and accidents. We discuss the implications of under- and over-representing causes of death overall and by gender, as well as areas of future research.

The history of fair housing policy in the United States has fallen short of two of its primary objectives: Deconcentration of poverty and community desegregation. We compare distributions of Housing Choice Voucher (HCV) households at various scales to broader US distributions to better understand the program's successes and shortcomings. We find that HCV households at the nation-scale have been consistently more highly distributed among lower-Area Median Income (AMI) census tracts and racially or ethnically concentrated areas of poverty (RECAPs). We specifically consider the potential of affirmatively furthering fair housing within the HCV program to simultaneously boost overall distributions in higher-AMI census tracts and to decrease RECAP concentrations of HCV households, observing that targeted deployment of mobility-based interventions can significantly decrease the proportion of HCV households residing in RECAPs. We further consider the impact of small area fair market rents (SAFMRs) and state-level source-of-income protections on HCV household distributions, finding that both interventions may boost HCV households into higher-AMI census tracts.

Introduction Wearable devices are rapidly improving our ability to observe health-related processes for extended durations in an unintrusive manner. In this study, we use wearable devices to understand how the shape of the heart rate curve during sleep relates to mental health. Methods As part of the Lived Experiences Measured Using Rings Study (LEMURS), we collected heart rate measurements using the Oura ring (Gen3) for over 25,000 sleep periods and self-reported mental health indicators from roughly 600 first-year university students in the USA during the fall semester of 2022. Using clustering techniques, we find that the sleeping heart rate curves can be broadly separated into two categories that are mainly differentiated by how far along the sleep period the lowest heart rate is reached. Results Sleep periods characterized by reaching the lowest heart rate later during sleep are also associated with shorter deep and REM sleep and longer light sleep, but not a difference in total sleep duration. Aggregating sleep periods at the individual level, we find that consistently reaching the lowest heart rate later during sleep is a significant predictor of (1) self-reported impairment due to anxiety or depression, (2) a prior mental health diagnosis, and (3) firsthand experience in traumatic events. This association is more pronounced among females. Conclusion Our results show that the shape of the sleeping heart rate curve, which is only weakly correlated with descriptive statistics such as the average or the minimum heart rate, is a viable but mostly overlooked metric that can help quantify the relationship between sleep and mental health.

Consumer wearables have been successful at measuring sleep and may be useful in predicting changes in mental health measures such as stress. A key challenge remains in quantifying the relationship between sleep measures associated with physiologic stress and a user’s experience of stress. Students from a public university enrolled in the Lived Experiences Measured Using Rings Study (LEMURS) provided continuous biometric data and answered weekly surveys during their first semester of college between October-December 2022. We analyzed weekly associations between estimated sleep measures and perceived stress for participants (N = 525). Through mixed-effects regression models, we identified consistent associations between perceived stress scores and average nightly total sleep time (TST), resting heart rate (RHR), heart rate variability (HRV), and respiratory rate (ARR). These effects persisted after controlling for gender and week of the semester. Specifically, for every additional hour of TST, the odds of experiencing moderate-to-high stress decreased by 0.617 or by 38.3% (p<0.01). For each 1 beat per minute increase in RHR, the odds of experiencing moderate-to-high stress increased by 1.036 or by 3.6% (p<0.01). For each 1 millisecond increase in HRV, the odds of experiencing moderate-to-high stress decreased by 0.988 or by 1.2% (p<0.05). For each additional breath per minute increase in ARR, the odds of experiencing moderate-to-high stress increased by 1.230 or by 23.0% (p<0.01). Consistent with previous research, participants who did not identify as male (i.e., female, nonbinary, and transgender participants) had significantly higher self-reported stress throughout the study. The week of the semester was also a significant predictor of stress. Sleep data from wearable devices may help us understand and to better predict stress, a strong signal of the ongoing mental health epidemic among college students.

Objective: Panic attacks are an impairing mental health problem that affects 11% of adults every year. Current criteria describe them as occurring without warning, despite evidence suggesting individuals can often identify attack triggers. We aimed to prospectively explore qualitative and quantitative factors associated with the onset of panic attacks. Results: Of 87 participants, 95% retrospectively identified a trigger for their panic attacks. Worse individually reported mood and state-level mood, as indicated by Twitter ratings, were related to greater likelihood of next-day panic attack. In a subsample of participants who uploaded their wearable sensor data (n=32), louder ambient noise and higher resting heart rate were related to greater likelihood of next-day panic attack. Conclusions: These promising results suggest that individuals who experience panic attacks may be able to anticipate their next attack which could be used to inform future prevention and intervention efforts.