Predicting the Effect of Proton Beam Therapy Technology on Pulmonary Toxicities for Patients With Locally Advanced Lung Cancer Enrolled in the Proton Collaborative Group Prospective Clinical Trial

PURPOSE

This study aimed to predict the probability of grade ≥2 pneumonitis or dyspnea within 12 months of receiving conventionally fractionated or mildly hypofractionated proton beam therapy for locally advanced lung cancer using machine learning.

METHODS AND MATERIALS

Demographic and treatment characteristics were analyzed for 965 consecutive patients treated for lung cancer with conventionally fractionated or mildly hypofractionated (2.2-3 Gy/fraction) proton beam therapy across 12 institutions. Three machine learning models (gradient boosting, additive tree, and logistic regression with lasso regularization) were implemented to predict Common Terminology Criteria for Adverse Events version 4 grade ≥2 pulmonary toxicities using double 10-fold cross-validation for parameter hyper-tuning without leak of information. Balanced accuracy and area under the curve were calculated, and 95% confidence intervals were obtained using bootstrap sampling.

RESULTS

The median age of the patients was 70 years (range, 20-97), and they had predominantly stage IIIA or IIIB disease. They received a median dose of 60 Gy in 2 Gy/fraction, and 46.4% received concurrent chemotherapy. In total, 250 (25.9%) had grade ≥2 pulmonary toxicity. The probability of pulmonary toxicity was 0.08 for patients treated with pencil beam scanning and 0.34 for those treated with other techniques (P = 8.97e-13). Use of abdominal compression and breath hold were highly significant predictors of less toxicity (P = 2.88e-08). Higher total radiation delivered dose (P = .0182) and higher average dose to the ipsilateral lung (P = .0035) increased the likelihood of pulmonary toxicities. The gradient boosting model performed the best of the models tested, and when demographic and dosimetric features were combined, the area under the curve and balanced accuracy were 0.75 ± 0.02 and 0.67 ± 0.02, respectively. After analyzing performance versus the number of data points used for training, we observed that accuracy was limited by the number of observations.

CONCLUSIONS

In the largest analysis of prospectively enrolled patients with lung cancer assessing pulmonary toxicities from proton therapy to date, advanced machine learning methods revealed that pencil beam scanning, abdominal compression, and lower normal lung doses can lead to significantly lower probability of developing grade ≥2 pneumonitis or dyspnea.

Tapping into alcohol use during COVID: Drinking correlates among bartenders and servers

The COVID pandemic placed a spotlight on alcohol use and the hardships of working within the food and beverage industry, with millions left jobless. Following previous studies that have found elevated rates of alcohol problems among bartenders and servers, here we studied the alcohol use of bartenders and servers who were employed during COVID. From February 12-June 16, 2021, in the midst of the U.S. COVID national emergency declaration, survey data from 1,010 employed bartender and servers were analyzed to quantify rates of excessive or hazardous drinking along with regression predictors of alcohol use as assessed by the 10-item Alcohol Use Disorders Identification Test (AUDIT). Findings indicate that more than 2 out of 5 (44%) people surveyed reported moderate or high rates of alcohol problem severity (i.e., AUDIT scores of 8 or higher)-a rate 4 to 6 times that of the heavy alcohol use rate reported pre- or mid-pandemic by adults within and outside the industry. Person-level factors (gender, substance use, mood) along with the drinking habits of one's core social group were significantly associated with alcohol use. Bartenders and servers reported surprisingly high rates of alcohol problem severity and experienced risk factors for hazardous drinking at multiple ecological levels. Being a highly vulnerable and understudied population, more studies on bartenders and servers are needed to assess and manage the true toll of alcohol consumption for industry employees.

Key language markers of depression on social media depend on race

Depression has robust natural language correlates and can increasingly be measured in language using predictive models. However, despite evidence that language use varies as a function of individual demographic features (e.g., age, gender), previous work has not systematically examined whether and how depression's association with language varies by race. We examine how race moderates the relationship between language features (i.e., first-person pronouns and negative emotions) from social media posts and self-reported depression, in a matched sample of Black and White English speakers in the United States. Our findings reveal moderating effects of race: While depression severity predicts I-usage in White individuals, it does not in Black individuals. White individuals use more belongingness and self-deprecation-related negative emotions. Machine learning models trained on similar amounts of data to predict depression severity performed poorly when tested on Black individuals, even when they were trained exclusively using the language of Black individuals. In contrast, analogous models tested on White individuals performed relatively well. Our study reveals surprising race-based differences in the expression of depression in natural language and highlights the need to understand these effects better, especially before language-based models for detecting psychological phenomena are integrated into clinical practice.

Large language models could change the future of behavioral healthcare: a proposal for responsible development and evaluation

Large language models (LLMs) such as Open AI's GPT-4 (which power ChatGPT) and Google's Gemini, built on artificial intelligence, hold immense potential to support, augment, or even eventually automate psychotherapy. Enthusiasm about such applications is mounting in the field as well as industry. These developments promise to address insufficient mental healthcare system capacity and scale individual access to personalized treatments. However, clinical psychology is an uncommonly high stakes application domain for AI systems, as responsible and evidence-based therapy requires nuanced expertise. This paper provides a roadmap for the ambitious yet responsible application of clinical LLMs in psychotherapy. First, a technical overview of clinical LLMs is presented. Second, the stages of integration of LLMs into psychotherapy are discussed while highlighting parallels to the development of autonomous vehicle technology. Third, potential applications of LLMs in clinical care, training, and research are discussed, highlighting areas of risk given the complex nature of psychotherapy. Fourth, recommendations for the responsible development and evaluation of clinical LLMs are provided, which include centering clinical science, involving robust interdisciplinary collaboration, and attending to issues like assessment, risk detection, transparency, and bias. Lastly, a vision is outlined for how LLMs might enable a new generation of studies of evidence-based interventions at scale, and how these studies may challenge assumptions about psychotherapy.

Differential temporal utility of passively sensed smartphone features for depression and anxiety symptom prediction: a longitudinal cohort study

While studies show links between smartphone data and affective symptoms, we lack clarity on the temporal scale, specificity (e.g., to depression vs. anxiety), and person-specific (vs. group-level) nature of these associations. We conducted a large-scale (n = 1013) smartphone-based passive sensing study to identify within- and between-person digital markers of depression and anxiety symptoms over time. Participants (74.6% female; M age = 40.9) downloaded the LifeSense app, which facilitated continuous passive data collection (e.g., GPS, app and device use, communication) across 16 weeks. Hierarchical linear regression models tested the within- and between-person associations of 2-week windows of passively sensed data with depression (PHQ-8) or generalized anxiety (GAD-7). We used a shifting window to understand the time scale at which sensed features relate to mental health symptoms, predicting symptoms 2 weeks in the future (distal prediction), 1 week in the future (medial prediction), and 0 weeks in the future (proximal prediction). Spending more time at home relative to one's average was an early signal of PHQ-8 severity (distal β = 0.219, p = 0.012) and continued to relate to PHQ-8 at medial (β = 0.198, p = 0.022) and proximal (β = 0.183, p = 0.045) windows. In contrast, circadian movement was proximally related to (β = -0.131, p = 0.035) but did not predict (distal β = 0.034, p = 0.577; medial β = -0.089, p = 0.138) PHQ-8. Distinct communication features (i.e., call/text or app-based messaging) related to PHQ-8 and GAD-7. Findings have implications for identifying novel treatment targets, personalizing digital mental health interventions, and enhancing traditional patient-provider interactions. Certain features (e.g., circadian movement) may represent correlates but not true prospective indicators of affective symptoms. Conversely, other features like home duration may be such early signals of intra-individual symptom change, indicating the potential utility of prophylactic intervention (e.g., behavioral activation) in response to person-specific increases in these signals.

Smartphone sensor data estimate alcohol craving in a cohort of patients with alcohol-associated liver disease and alcohol use disorder

BACKGROUND

Sensors within smartphones, such as accelerometer and location, can describe longitudinal markers of behavior as represented through devices in a method called digital phenotyping. This study aimed to assess the feasibility of digital phenotyping for patients with alcohol-associated liver disease and alcohol use disorder, determine correlations between smartphone data and alcohol craving, and establish power assessment for future studies to prognosticate clinical outcomes.

METHODS

A total of 24 individuals with alcohol-associated liver disease and alcohol use disorder were instructed to download the AWARE application to collect continuous sensor data and complete daily ecological momentary assessments on alcohol craving and mood for up to 30 days. Data from sensor streams were processed into features like accelerometer magnitude, number of calls, and location entropy, which were used for statistical analysis. We used repeated measures correlation for longitudinal data to evaluate associations between sensors and ecological momentary assessments and standard Pearson correlation to evaluate within-individual relationships between sensors and craving.

RESULTS

Alcohol craving significantly correlated with mood obtained from ecological momentary assessments. Across all sensors, features associated with craving were also significantly correlated with all moods (eg, loneliness and stress) except boredom. Individual-level analysis revealed significant relationships between craving and features of location entropy and average accelerometer magnitude.

CONCLUSIONS

Smartphone sensors may serve as markers for alcohol craving and mood in alcohol-associated liver disease and alcohol use disorder. Findings suggest that location-based and accelerometer-based features may be associated with alcohol craving. However, data missingness and low participant retention remain challenges. Future studies are needed for further digital phenotyping of relapse risk and progression of liver disease.

Specific associations of passively sensed smartphone data with future symptoms of avoidance, fear, and physiological distress in social anxiety

Background

Prior literature links passively sensed information about a person's location, movement, and communication with social anxiety. These findings hold promise for identifying novel treatment targets, informing clinical care, and personalizing digital mental health interventions. However, social anxiety symptoms are heterogeneous; to identify more precise targets and tailor treatments, there is a need for personal sensing studies aimed at understanding differential predictors of the distinct subdomains of social anxiety. Our objective was to conduct a large-scale smartphone-based sensing study of fear, avoidance, and physiological symptoms in the context of trait social anxiety over time.

Methods

Participants (n = 1013; 74.6 % female; M age = 40.9) downloaded the LifeSense app, which collected continuous passive data (e.g., GPS, communication, app and device use) over 16 weeks. We tested a series of multilevel linear regression models to understand within- and between-person associations of 2-week windows of passively sensed smartphone data with fear, avoidance, and physiological distress on the self-reported Social Phobia Inventory (SPIN). A shifting sensor lag was applied to examine how smartphone features related to SPIN subdomains 2 weeks in the future (distal prediction), 1 week in the future (medial prediction), and 0 weeks in the future (proximal prediction).

Results

A decrease in time visiting novel places was a strong between-person predictor of social avoidance over time (distal β = -0.886, p = .002; medial β = -0.647, p = .029; proximal β = -0.818, p = .007). Reductions in call- and text-based communications were associated with social avoidance at both the between- (distal β = -0.882, p = .002; medial β = -0.932, p = .001; proximal β = -0.918, p = .001) and within- (distal β = -0.191, p = .046; medial β = -0.213, p = .028) person levels, as well as between-person fear of social situations (distal β = -0.860, p < .001; medial β = -0.892, p < .001; proximal β = -0.886, p < .001) over time. There were fewer significant associations of sensed data with physiological distress. Across the three subscales, smartphone data explained 9-12 % of the variance in social anxiety.

Conclusion

Findings have implications for understanding how social anxiety manifests in daily life, and for personalizing treatments. For example, a signal that someone is likely to begin avoiding social situations may suggest a need for alternative types of exposure-based interventions compared to a signal that someone is likely to begin experiencing increased physiological distress. Our results suggest that as a prophylactic means of targeting social avoidance, it may be helpful to deploy interventions involving social exposures in response to decreases in time spent visiting novel places.

Comparison of wellbeing structures based on survey responses and social media language: A network analysis

Wellbeing is predominantly measured through surveys but is increasingly measured by analysing individuals' language on social media platforms using social media text mining (SMTM). To investigate whether the structure of wellbeing is similar across both data collection methods, we compared networks derived from survey items and social media language features collected from the same participants. The dataset was split into an independent exploration (n = 1169) and a final subset (n = 1000). After estimating exploration networks, redundant survey items and language topics were eliminated. Final networks were then estimated using exploratory graph analysis (EGA). The networks of survey items and those from language topics were similar, both consisting of five wellbeing dimensions. The dimensions in the survey- and SMTM-based assessment of wellbeing showed convergent structures congruent with theories of wellbeing. Specific dimensions found in each network reflected the unique aspects of each type of data (survey and social media language). Networks derived from both language features and survey items show similar structures. Survey and SMTM methods may provide complementary methods to understand differences in human wellbeing.

Filling in the white space: Spatial interpolation with Gaussian processes and social media data

Full national coverage below the state level is difficult to attain through survey-based data collection. Even the largest survey-based data collections, such as the CDC's Behavioral Risk Factor Surveillance System or the Gallup-Healthways Well-being Index (both with more than 300,000 responses p.a.) only allow for the estimation of annual averages for about 260 out of roughly U.S. 3,000 counties when a threshold of 300 responses per county is used. Using a relatively high threshold of 300 responses gives substantially higher convergent validity-higher correlations with health variables-than lower thresholds but covers a reduced and biased sample of the population. We present principled methods to interpolate spatial estimates and show that including large-scale geotagged social media data can increase interpolation accuracy. In this work, we focus on Gallup-reported life satisfaction, a widely-used measure of subjective well-being. We use Gaussian Processes (GP), a formal Bayesian model, to interpolate life satisfaction, which we optimally combine with estimates from low-count data. We interpolate over several spaces (geographic and socioeconomic) and extend these evaluations to the space created by variables encoding language frequencies of approximately 6 million geotagged Twitter users. We find that Twitter language use can serve as a rough aggregate measure of socioeconomic and cultural similarity, and improves upon estimates derived from a wide variety of socioeconomic, demographic, and geographic similarity measures. We show that applying Gaussian Processes to the limited Gallup data allows us to generate estimates for a much larger number of counties while maintaining the same level of convergent validity with external criteria (i.e., N = 1,133 vs. 2,954 counties). This work suggests that spatial coverage of psychological variables can be reliably extended through Bayesian techniques while maintaining out-of-sample prediction accuracy and that Twitter language adds important information about cultural similarity over and above traditional socio-demographic and geographic similarity measures. Finally, to facilitate the adoption of these methods, we have also open-sourced an online tool that researchers can freely use to interpolate their data across geographies.

Analyzing text message linguistic features: Do people with depression communicate differently with their close and non-close contacts?

BACKGROUND

Relatively little is known about how communication changes as a function of depression severity and interpersonal closeness. We examined the linguistic features of outgoing text messages among individuals with depression and their close- and non-close contacts.

METHODS

419 participants were included in this 16-week-long observational study. Participants regularly completed the PHQ-8 and rated subjective closeness to their contacts. Text messages were processed to count frequencies of word usage in the LIWC 2015 libraries. A linear mixed modeling approach was used to estimate linguistic feature scores of outgoing text messages.

RESULTS

Regardless of closeness, people with higher PHQ-8 scores tended to use more differentiation words. When texting with close contacts, individuals with higher PHQ-8 scores used more first-person singular, filler, sexual, anger, and negative emotion words. When texting with non-close contacts these participants used more conjunctions, tentative, and sadness-related words and fewer first-person plural words.

CONCLUSION

Word classes used in text messages, when combined with symptom severity and subjective social closeness data, may be indicative of underlying interpersonal processes. These data may hold promise as potential treatment targets to address interpersonal drivers of depression.

Predicting U.S. county opioid poisoning mortality from multi-modal social media and psychological self-report data

Opioid poisoning mortality is a substantial public health crisis in the United States, with opioids involved in approximately 75% of the nearly 1 million drug related deaths since 1999. Research suggests that the epidemic is driven by both over-prescribing and social and psychological determinants such as economic stability, hopelessness, and isolation. Hindering this research is a lack of measurements of these social and psychological constructs at fine-grained spatial and temporal resolutions. To address this issue, we use a multi-modal data set consisting of natural language from Twitter, psychometric self-reports of depression and well-being, and traditional area-based measures of socio-demographics and health-related risk factors. Unlike previous work using social media data, we do not rely on opioid or substance related keywords to track community poisonings. Instead, we leverage a large, open vocabulary of thousands of words in order to fully characterize communities suffering from opioid poisoning, using a sample of 1.5 billion tweets from 6 million U.S. county mapped Twitter users. Results show that Twitter language predicted opioid poisoning mortality better than factors relating to socio-demographics, access to healthcare, physical pain, and psychological well-being. Additionally, risk factors revealed by the Twitter language analysis included negative emotions, discussions of long work hours, and boredom, whereas protective factors included resilience, travel/leisure, and positive emotions, dovetailing with results from the psychometric self-report data. The results show that natural language from public social media can be used as a surveillance tool for both predicting community opioid poisonings and understanding the dynamic social and psychological nature of the epidemic.

Opioid death projections with AI-based forecasts using social media language

Targeting of location-specific aid for the U.S. opioid epidemic is difficult due to our inability to accurately predict changes in opioid mortality across heterogeneous communities. AI-based language analyses, having recently shown promise in cross-sectional (between-community) well-being assessments, may offer a way to more accurately longitudinally predict community-level overdose mortality. Here, we develop and evaluate, TROP (Transformer for Opiod Prediction), a model for community-specific trend projection that uses community-specific social media language along with past opioid-related mortality data to predict future changes in opioid-related deaths. TOP builds on recent advances in sequence modeling, namely transformer networks, to use changes in yearly language on Twitter and past mortality to project the following year's mortality rates by county. Trained over five years and evaluated over the next two years TROP demonstrated state-of-the-art accuracy in predicting future county-specific opioid trends. A model built using linear auto-regression and traditional socioeconomic data gave 7% error (MAPE) or within 2.93 deaths per 100,000 people on average; our proposed architecture was able to forecast yearly death rates with less than half that error: 3% MAPE and within 1.15 per 100,000 people.

Prospective associations of text-message-based sentiment with symptoms of depression, generalized anxiety, and social anxiety

OBJECTIVE

Language patterns may elucidate mechanisms of mental health conditions. To inform underlying theory and risk models, we evaluated prospective associations between in vivo text messaging language and differential symptoms of depression, generalized anxiety, and social anxiety.

METHODS

Over 16 weeks, we collected outgoing text messages from 335 adults. Using Linguistic Inquiry and Word Count (LIWC), NRC Emotion Lexicon, and previously established depression and stress dictionaries, we evaluated the degree to which language features predict symptoms of depression, generalized anxiety, or social anxiety the following week using hierarchical linear models. To isolate the specificity of language effects, we also controlled for the effects of the two other symptom types.

RESULTS

We found significant relationships of language features, including personal pronouns, negative emotion, cognitive and biological processes, and informal language, with common mental health conditions, including depression, generalized anxiety, and social anxiety (ps < .05). There was substantial overlap between language features and the three mental health outcomes. However, after controlling for other symptoms in the models, depressive symptoms were uniquely negatively associated with language about anticipation, trust, social processes, and affiliation (βs: -.10 to -.09, ps < .05), whereas generalized anxiety symptoms were positively linked with these same language features (βs: .12-.13, ps < .001). Social anxiety symptoms were uniquely associated with anger, sexual language, and swearing (βs: .12-.13, ps < .05).

CONCLUSION

Language that confers both common (e.g., personal pronouns and negative emotion) and specific (e.g., affiliation, anticipation, trust, and anger) risk for affective disorders is perceptible in prior week text messages, holding promise for understanding cognitive-behavioral mechanisms and tailoring digital interventions.

Head versus heart: social media reveals differential language of loneliness from depression

We study the language differentially associated with loneliness and depression using 3.4-million Facebook posts from 2986 individuals, and uncover the statistical associations of survey-based depression and loneliness with both dictionary-based (Linguistic Inquiry Word Count 2015) and open-vocabulary linguistic features (words, phrases, and topics). Loneliness and depression were found to have highly overlapping language profiles, including sickness, pain, and negative emotions as (cross-sectional) risk factors, and social relationships and activities as protective factors. Compared to depression, the language associated with loneliness reflects a stronger cognitive focus, including more references to cognitive processes (i.e., differentiation and tentative language, thoughts, and the observation of irregularities), and cognitive activities like reading and writing. As might be expected, less lonely users were more likely to reference social relationships (e.g., friends and family, romantic relationships), and use first-person plural pronouns. Our findings suggest that the mechanisms of loneliness include self-oriented cognitive activities (i.e., reading) and an overattention to the interpretation of information in the environment. These data-driven ecological findings suggest interventions for loneliness that target maladaptive social cognitions (e.g., through reframing the perception of social environments), strengthen social relationships, and treat other affective distress (i.e., depression).

The association of language style matching in text messages with mood and anxiety symptoms

Context

Impairment in social functioning is a feature and consequence of depression and anxiety disorders. For example, in depression, anhedonia and negative feelings about the self may impact relationships; in anxiety, fear of negative evaluation may interfere with getting close to others. It is unknown whether social impairment associated with depression and anxiety symptoms is reflected in day-to-day language exchanges with others, such as through reduced language style matching (LSM).

Methods

Over 16 weeks, we collected text message data from 458 adults and evaluated differences in LSM between people with average scores above/below the clinical cutoff for depression, generalized anxiety, and social anxiety in text message conversations. Text message sentiment scores were computed across 73 Linguistic Inquiry and Word Count (LIWC) categories for each participant. T-tests were used to compare LSM across two groups (average scores above/below clinical cutoff) for each of the 3 diagnostic categories (depression, generalized anxiety, social anxiety), and each of the 73 LIWC categories, with correction for multiple comparisons.

Results

We found reduced LSM of function words (namely, prepositions [t=-2.82, p=.032], articles [t=-5.26, p<.001], and auxiliary verbs [t=-2.64, p=.046]) in people with average scores above the clinical cutoff for generalized anxiety, and reduced LSM of prepositions (t=-4.26, p<.001) and articles (t=-3.39, p=.010) in people with average scores above the clinical cutoff for social anxiety. There were no significant differences in LSM of function words between people with average scores above and below the clinical cutoff for depression. Across all symptom categories, elevated affective psychopathology was associated with being more likely to style match on formality, including netspeak (generalized anxiety, t=5.77, p<.001; social anxiety, t=4.14, p<.001; depression, t=3.13, p=.021) and informal language (generalized anxiety, t=6.65, p<.001; social anxiety, t=5.14, p>.001; depression, t=3.20, p=.020).We also observed content-specific LSM differences across the three groups.

Conclusions

Reduced LSM of function words among patients reporting elevated anxiety symptoms suggests that anxiety-related psychosocial difficulties may be perceptible in subtle cues from day-to-day language. Conversely, the absence of differences in the LSM of function words among people with average scores above and below the clinical cutoff for depression indicates a potentially distinct mechanism of social impairment.

Implications

Results point to potential markers of psychosocial difficulties in daily conversations, particularly among those experiencing heightened anxiety symptoms. Future studies may consider the degree to which LSM is associated with self-reported psychosocial impairment, with the promise of informing cognitive-behavioral mechanisms and tailoring digital interventions for social skills.

Linguistic predictors from Facebook postings of substance use disorder treatment retention versus discontinuation

Background: Early indicators of who will remain in - or leave - treatment for substance use disorder (SUD) can drive targeted interventions to support long-term recovery.Objectives: To conduct a comprehensive study of linguistic markers of SUD treatment outcomes, the current study integrated features produced by machine learning models known to have social-psychology relevance.Methods: We extracted and analyzed linguistic features from participants' Facebook posts (N = 206, 39.32% female; 55,415 postings) over the two years before they entered a SUD treatment program. Exploratory features produced by both Linguistic Inquiry and Word Count (LIWC) and Latent Dirichlet Allocation (LDA) topic modeling and the features from theoretical domains of religiosity, affect, and temporal orientation via established AI-based linguistic models were utilized.Results: Patients who stayed in the SUD treatment for over 90 days used more words associated with religion, positive emotions, family, affiliations, and the present, and used more first-person singular pronouns (Cohen's d values: [-0.39, -0.57]). Patients who discontinued their treatment before 90 days discussed more diverse topics, focused on the past, and used more articles (Cohen's d values: [0.44, 0.57]). All ps < .05 with Benjamini-Hochberg False Discovery Rate correction.Conclusions: We confirmed the literature on protective and risk social-psychological factors linking to SUD treatment in language analysis, showing that Facebook language before treatment entry could be used to identify the markers of SUD treatment outcomes. This reflects the importance of taking these linguistic features and markers into consideration when designing and recommending SUD treatment plans.

Mobile footprinting: linking individual distinctiveness in mobility patterns to mood, sleep, and brain functional connectivity

Mapping individual differences in behavior is fundamental to personalized neuroscience, but quantifying complex behavior in real world settings remains a challenge. While mobility patterns captured by smartphones have increasingly been linked to a range of psychiatric symptoms, existing research has not specifically examined whether individuals have person-specific mobility patterns. We collected over 3000 days of mobility data from a sample of 41 adolescents and young adults (age 17-30 years, 28 female) with affective instability. We extracted summary mobility metrics from GPS and accelerometer data and used their covariance structures to identify individuals and calculated the individual identification accuracy-i.e., their "footprint distinctiveness". We found that statistical patterns of smartphone-based mobility features represented unique "footprints" that allow individual identification (p < 0.001). Critically, mobility footprints exhibited varying levels of person-specific distinctiveness (4-99%), which was associated with age and sex. Furthermore, reduced individual footprint distinctiveness was associated with instability in affect (p < 0.05) and circadian patterns (p < 0.05) as measured by environmental momentary assessment. Finally, brain functional connectivity, especially those in the somatomotor network, was linked to individual differences in mobility patterns (p < 0.05). Together, these results suggest that real-world mobility patterns may provide individual-specific signatures relevant for studies of development, sleep, and psychopathology.

Feasibility of Mobile Health and Social Media–Based Interventions for Young Adults With Early Psychosis and Clinical Risk for Psychosis: Survey Study

Background

Digital technology, the internet, and social media are increasingly investigated as promising means for monitoring symptoms and delivering mental health treatment. These apps and interventions have demonstrated preliminary acceptability and feasibility, but previous reports suggest that access to technology may still be limited among individuals with psychotic disorders relative to the general population.

Objective

We evaluated and compared access to and use of technology and social media in young adults with psychotic disorders (PD), young adults with clinical risk for psychosis (CR), and psychosis-free youths (PF).

Methods

Participants were recruited through a coordinated specialty care clinic dedicated toward early psychosis as well as ongoing studies. We surveyed 21 PD, 23 CR, and 15 PF participants regarding access to technology and use of social media, specifically Facebook and Twitter. Statistical analyses were conducted in R. Categorical variables were compared among groups using Fisher exact test, continuous variables were compared using 1-way ANOVA, and multiple linear regressions were used to evaluate for covariates.

Results

Access to technology and social media were similar among PD, CR, and PF participants. Individuals with PD, but not CR, were less likely to post at a weekly or higher frequency compared to PF individuals. We found that decreased active social media posting was unique to psychotic disorders and did not occur with other psychiatric diagnoses or demographic variables. Additionally, variation in age, sex, and White versus non-White race did not affect posting frequency.

Conclusions

For young people with psychosis spectrum disorders, there appears to be no “technology gap” limiting the implementation of digital and mobile health interventions. Active posting to social media was reduced for individuals with psychosis, which may be related to negative symptoms or impairment in social functioning.

Correcting Sociodemographic Selection Biases for Population Prediction from Social Media

Social media is increasingly used for large-scale population predictions, such as estimating community health statistics. However, social media users are not typically a representative sample of the intended population - a "selection bias". Within the social sciences, such a bias is typically addressed with restratification techniques, where observations are reweighted according to how under- or over-sampled their socio-demographic groups are. Yet, restratifaction is rarely evaluated for improving prediction. In this two-part study, we first evaluate standard, "out-of-the-box" restratification techniques, finding they provide no improvement and often even degraded prediction accuracies across four tasks of esimating U.S. county population health statistics from Twitter. The core reasons for degraded performance seem to be tied to their reliance on either sparse or shrunken estimates of each population's socio-demographics. In the second part of our study, we develop and evaluate Robust Poststratification, which consists of three methods to address these problems: (1) estimator redistribution to account for shrinking, as well as (2) adaptive binning and (3) informed smoothing to handle sparse socio-demographic estimates. We show that each of these methods leads to significant improvement in prediction accuracies over the standard restratification approaches. Taken together, Robust Poststratification enables state-of-the-art prediction accuracies, yielding a 53.0% increase in variance explained (R ²) in the case of surveyed life satisfaction, and a 17.8% average increase across all tasks.

Using Facebook language to predict and describe excessive alcohol use

BACKGROUND

Assessing risk for excessive alcohol use is important for applications ranging from recruitment into research studies to targeted public health messaging. Social media language provides an ecologically embedded source of information for assessing individuals who may be at risk for harmful drinking.

METHODS

Using data collected on 3664 respondents from the general population, we examine how accurately language used on social media classifies individuals as at-risk for alcohol problems based on Alcohol Use Disorder Identification Test-Consumption score benchmarks.

RESULTS

We find that social media language is moderately accurate (area under the curve = 0.75) at identifying individuals at risk for alcohol problems (i.e., hazardous drinking/alcohol use disorders) when used with models based on contextual word embeddings. High-risk alcohol use was predicted by individuals' usage of words related to alcohol, partying, informal expressions, swearing, and anger. Low-risk alcohol use was predicted by individuals' usage of social, affiliative, and faith-based words.

CONCLUSIONS

The use of social media data to study drinking behavior in the general public is promising and could eventually support primary and secondary prevention efforts among Americans whose at-risk drinking may have otherwise gone "under the radar."

The relationship between text message sentiment and self-reported depression

BACKGROUND

Personal sensing has shown promise for detecting behavioral correlates of depression, but there is little work examining personal sensing of cognitive and affective states. Digital language, particularly through personal text messages, is one source that can measure these markers.

METHODS

We correlated privacy-preserving sentiment analysis of text messages with self-reported depression symptom severity. We enrolled 219 U.S. adults in a 16 week longitudinal observational study. Participants installed a personal sensing app on their phones, which administered self-report PHQ-8 assessments of their depression severity, collected phone sensor data, and computed anonymized language sentiment scores from their text messages. We also trained machine learning models for predicting end-of-study self-reported depression status using on blocks of phone sensor and text features.

RESULTS

In correlation analyses, we find that degrees of depression, emotional, and personal pronoun language categories correlate most strongly with self-reported depression, validating prior literature. Our classification models which predict binary depression status achieve a leave-one-out AUC of 0.72 when only considering text features and 0.76 when combining text with other networked smartphone sensors.

LIMITATIONS

Participants were recruited from a panel that over-represented women, caucasians, and individuals with self-reported depression at baseline. As language use differs across demographic factors, generalizability beyond this population may be limited. The study period also coincided with the initial COVID-19 outbreak in the United States, which may have affected smartphone sensor data quality.

CONCLUSIONS

Effective depression prediction through text message sentiment, especially when combined with other personal sensors, could enable comprehensive mental health monitoring and intervention.

Patient Experience and Satisfaction in Online Reviews of Obstetric Care: Observational Study

Background

The quality of care in labor and delivery is traditionally measured through the Hospital Consumer Assessment of Healthcare Providers and Systems but less is known about the experiences of care reported by patients and caregivers on online sites that are more easily accessed by the public.

Objective

The aim of this study was to generate insight into the labor and delivery experience using hospital reviews on Yelp.

Methods

We identified all Yelp reviews of US hospitals posted online from May 2005 to March 2017. We used a machine learning tool, latent Dirichlet allocation, to identify 100 topics or themes within these reviews and used Pearson r to identify statistically significant correlations between topics and high (5-star) and low (1-star) ratings.

Results

A total of 1569 hospitals listed in the American Hospital Association directory had at least one Yelp posting, contributing a total of 41,095 Yelp reviews. Among those hospitals, 919 (59%) had at least one Yelp rating for labor and delivery services (median of 9 reviews), contributing a total of 6523 labor and delivery reviews. Reviews concentrated among 5-star (n=2643, 41%) and 1-star reviews (n=1934, 30%). Themes strongly associated with favorable ratings included the following: top-notch care (r=0.45, P<.001), describing staff as comforting (r=0.52, P<.001), the delivery experience (r=0.46, P<.001), modern and clean facilities (r=0.44, P<.001), and hospital food (r=0.38, P<.001). Themes strongly correlated with 1-star labor and delivery reviews included complaints to management (r=0.30, P<.001), a lack of agency among patients (r=0.47, P<.001), and issues with discharging from the hospital (r=0.32, P<.001).

Conclusions

Online review content about labor and delivery can provide meaningful information about patient satisfaction and experiences. Narratives from these reviews that are not otherwise captured in traditional surveys can direct efforts to improve the experience of obstetrical care.

Word embeddings trained on published case reports are lightweight, effective for clinical tasks, and free of protected health information

OBJECTIVE

Quantify tradeoffs in performance, reproducibility, and resource demands across several strategies for developing clinically relevant word embeddings.

MATERIALS AND METHODS

We trained separate embeddings on all full-text manuscripts in the Pubmed Central (PMC) Open Access subset, case reports therein, the English Wikipedia corpus, the Medical Information Mart for Intensive Care (MIMIC) III dataset, and all notes in the University of Pennsylvania Health System (UPHS) electronic health record. We tested embeddings in six clinically relevant tasks including mortality prediction and de-identification, and assessed performance using the scaled Brier score (SBS) and the proportion of notes successfully de-identified, respectively.

RESULTS

Embeddings from UPHS notes best predicted mortality (SBS 0.30, 95% CI 0.15 to 0.45) while Wikipedia embeddings performed worst (SBS 0.12, 95% CI -0.05 to 0.28). Wikipedia embeddings most consistently (78% of notes) and the full PMC corpus embeddings least consistently (48%) de-identified notes. Across all six tasks, the full PMC corpus demonstrated the most consistent performance, and the Wikipedia corpus the least. Corpus size ranged from 49 million tokens (PMC case reports) to 10 billion (UPHS).

DISCUSSION

Embeddings trained on published case reports performed as least as well as embeddings trained on other corpora in most tasks, and clinical corpora consistently outperformed non-clinical corpora. No single corpus produced a strictly dominant set of embeddings across all tasks and so the optimal training corpus depends on intended use.

CONCLUSION

Embeddings trained on published case reports performed comparably on most clinical tasks to embeddings trained on larger corpora. Open access corpora allow training of clinically relevant, effective, and reproducible embeddings.

Development and validation of a prediction model for actionable aspects of frailty in the text of clinicians' encounter notes

OBJECTIVE

Frailty is a prevalent risk factor for adverse outcomes among patients with chronic lung disease. However, identifying frail patients who may benefit from interventions is challenging using standard data sources. We therefore sought to identify phrases in clinical notes in the electronic health record (EHR) that describe actionable frailty syndromes.

MATERIALS AND METHODS

We used an active learning strategy to select notes from the EHR and annotated each sentence for 4 actionable aspects of frailty: respiratory impairment, musculoskeletal problems, fall risk, and nutritional deficiencies. We compared the performance of regression, tree-based, and neural network models to predict the labels for each sentence. We evaluated performance with the scaled Brier score (SBS), where 1 is perfect and 0 is uninformative, and the positive predictive value (PPV).

RESULTS

We manually annotated 155 952 sentences from 326 patients. Elastic net regression had the best performance across all 4 frailty aspects (SBS 0.52, 95% confidence interval [CI] 0.49-0.54) followed by random forests (SBS 0.49, 95% CI 0.47-0.51), and multi-task neural networks (SBS 0.39, 95% CI 0.37-0.42). For the elastic net model, the PPV for identifying the presence of respiratory impairment was 54.8% (95% CI 53.3%-56.6%) at a sensitivity of 80%.

DISCUSSION

Classification models using EHR notes can effectively identify actionable aspects of frailty among patients living with chronic lung disease. Regression performed better than random forest and neural network models.

CONCLUSIONS

NLP-based models offer promising support to population health management programs that seek to identify and refer community-dwelling patients with frailty for evidence-based interventions.

Closed- and open-vocabulary approaches to text analysis: A review, quantitative comparison, and recommendations

Technology now makes it possible to understand efficiently and at large scale how people use language to reveal their everyday thoughts, behaviors, and emotions. Written text has been analyzed through both theory-based, closed-vocabulary methods from the social sciences as well as data-driven, open-vocabulary methods from computer science, but these approaches have not been comprehensively compared. To provide guidance on best practices for automatically analyzing written text, this narrative review and quantitative synthesis compares five predominant closed- and open-vocabulary methods: Linguistic Inquiry and Word Count (LIWC), the General Inquirer, DICTION, Latent Dirichlet Allocation, and Differential Language Analysis. We compare the linguistic features associated with gender, age, and personality across the five methods using an existing dataset of Facebook status updates and self-reported survey data from 65,896 users. Results are fairly consistent across methods. The closed-vocabulary approaches efficiently summarize concepts and are helpful for understanding how people think, with LIWC2015 yielding the strongest, most parsimonious results. Open-vocabulary approaches reveal more specific and concrete patterns across a broad range of content domains, better address ambiguous word senses, and are less prone to misinterpretation, suggesting that they are well-suited for capturing the nuances of everyday psychological processes. We detail several errors that can occur in closed-vocabulary analyses, the impact of sample size, number of words per user and number of topics included in open-vocabulary analyses, and implications of different analytical decisions. We conclude with recommendations for researchers, advocating for a complementary approach that combines closed- and open-vocabulary methods. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

Effect of Posting on Social Media on Systolic Blood Pressure and Management of Hypertension: A Randomized Controlled Trial

Background

Online platforms are used to manage aspects of our lives including health outside clinical settings. Little is known about the effectiveness of using online platforms to manage hypertension. We assessed effects of tweeting/retweeting cardiovascular health content by individuals with poorly controlled hypertension on systolic blood pressure (SBP) and patient activation.

Methods and Results

We conducted this 2‐arm randomized controlled trial. Eligibility included diagnosis of hypertension; SBP >140 mm Hg; and an existing Twitter account or willingness to create one to follow study Twitter account. Intervention arm was asked to tweet/retweet health content 2×/week using a specific hashtag for study duration (6 months). The main measures include primary outcome change in SBP; secondary outcome point change in Patient Activation Measure (PAM). We remotely recruited and enrolled 611 participants, mean age 52 (SD, 11.7). Mean baseline SBP for the intervention group was 155.8 and for control was 155.6. At 6 months, mean SBP for intervention group was 137.6 and for control was 135.7. Mean change in SBP from baseline to 6 months for the intervention group was −18.5 and for control was −19.8 (P=0.48). Mean PAM at baseline for the intervention group was 70.3 for control was 72.7. At 6 months, mean PAM scores were 71.1 (intervention) and 75.6 (control). Mean change in PAM score for the intervention group was 0.0 and for control was 3.3 (P=0.12).

Conclusions

Recruiting and engaging patients and collecting outcome measures remotely are feasible using Twitter. Encouraging patients with poorly controlled hypertension to tweet or retweet health content on Twitter did not improve SBP or PAM score at 6 months.

Registration

URL: https://www.clinicaltrials.gov. Unique identifier: NCT02622256.

Regional personality assessment through social media language

OBJECTIVE

We explore the personality of counties as assessed through linguistic patterns on social media. Such studies were previously limited by the cost and feasibility of large-scale surveys; however, language-based computational models applied to large social media datasets now allow for large-scale personality assessment.

METHOD

We applied a language-based assessment of the five factor model of personality to 6,064,267 U.S. Twitter users. We aggregated the Twitter-based personality scores to 2,041 counties and compared to political, economic, social, and health outcomes measured through surveys and by government agencies.

RESULTS

There was significant personality variation across counties. Openness to experience was higher on the coasts, conscientiousness was uniformly spread, extraversion was higher in southern states, agreeableness was higher in western states, and emotional stability was highest in the south. Across 13 outcomes, language-based personality estimates replicated patterns that have been observed in individual-level and geographic studies. This includes higher Republican vote share in less agreeable counties and increased life satisfaction in more conscientious counties.

CONCLUSIONS

Results suggest that regions vary in their personality and that these differences can be studied through computational linguistic analysis of social media. Furthermore, these methods may be used to explore other psychological constructs across geographies.

Evaluation of Changes in Depression, Anxiety, and Social Anxiety Using Smartphone Sensor Features: Longitudinal Cohort Study

BACKGROUND

The assessment of behaviors related to mental health typically relies on self-report data. Networked sensors embedded in smartphones can measure some behaviors objectively and continuously, with no ongoing effort.

OBJECTIVE

This study aims to evaluate whether changes in phone sensor-derived behavioral features were associated with subsequent changes in mental health symptoms.

METHODS

This longitudinal cohort study examined continuously collected phone sensor data and symptom severity data, collected every 3 weeks, over 16 weeks. The participants were recruited through national research registries. Primary outcomes included depression (8-item Patient Health Questionnaire), generalized anxiety (Generalized Anxiety Disorder 7-item scale), and social anxiety (Social Phobia Inventory) severity. Participants were adults who owned Android smartphones. Participants clustered into 4 groups: multiple comorbidities, depression and generalized anxiety, depression and social anxiety, and minimal symptoms.

RESULTS

A total of 282 participants were aged 19-69 years (mean 38.9, SD 11.9 years), and the majority were female (223/282, 79.1%) and White participants (226/282, 80.1%). Among the multiple comorbidities group, depression changes were preceded by changes in GPS features (Time: r=-0.23, P=.02; Locations: r=-0.36, P<.001), exercise duration (r=0.39; P=.03) and use of active apps (r=-0.31; P<.001). Among the depression and anxiety groups, changes in depression were preceded by changes in GPS features for Locations (r=-0.20; P=.03) and Transitions (r=-0.21; P=.03). Depression changes were not related to subsequent sensor-derived features. The minimal symptoms group showed no significant relationships. There were no associations between sensor-based features and anxiety and minimal associations between sensor-based features and social anxiety.

CONCLUSIONS

Changes in sensor-derived behavioral features are associated with subsequent depression changes, but not vice versa, suggesting a directional relationship in which changes in sensed behaviors are associated with subsequent changes in symptoms.

A Simulated Prospective Evaluation of a Deep Learning Model for Real-Time Prediction of Clinical Deterioration Among Ward Patients

OBJECTIVES

The National Early Warning Score, Modified Early Warning Score, and quick Sepsis-related Organ Failure Assessment can predict clinical deterioration. These scores exhibit only moderate performance and are often evaluated using aggregated measures over time. A simulated prospective validation strategy that assesses multiple predictions per patient-day would provide the best pragmatic evaluation. We developed a deep recurrent neural network deterioration model and conducted a simulated prospective evaluation.

DESIGN

Retrospective cohort study.

SETTING

Four hospitals in Pennsylvania.

PATIENTS

Inpatient adults discharged between July 1, 2017, and June 30, 2019.

INTERVENTIONS

None.

MEASUREMENTS AND MAIN RESULTS

We trained a deep recurrent neural network and logistic regression model using data from electronic health records to predict hourly the 24-hour composite outcome of transfer to ICU or death. We analyzed 146,446 hospitalizations with 16.75 million patient-hours. The hourly event rate was 1.6% (12,842 transfers or deaths, corresponding to 260,295 patient-hours within the predictive horizon). On a hold-out dataset, the deep recurrent neural network achieved an area under the precision-recall curve of 0.042 (95% CI, 0.04-0.043), comparable with logistic regression model (0.043; 95% CI 0.041 to 0.045), and outperformed National Early Warning Score (0.034; 95% CI, 0.032-0.035), Modified Early Warning Score (0.028; 95% CI, 0.027- 0.03), and quick Sepsis-related Organ Failure Assessment (0.021; 95% CI, 0.021-0.022). For a fixed sensitivity of 50%, the deep recurrent neural network achieved a positive predictive value of 3.4% (95% CI, 3.4-3.5) and outperformed logistic regression model (3.1%; 95% CI 3.1-3.2), National Early Warning Score (2.0%; 95% CI, 2.0-2.0), Modified Early Warning Score (1.5%; 95% CI, 1.5-1.5), and quick Sepsis-related Organ Failure Assessment (1.5%; 95% CI, 1.5-1.5).

CONCLUSIONS

Commonly used early warning scores for clinical decompensation, along with a logistic regression model and a deep recurrent neural network model, show very poor performance characteristics when assessed using a simulated prospective validation. None of these models may be suitable for real-time deployment.

Studying social media language changes associated with pregnancy status, trimester, and parity from medical records

We sought to evaluate whether there was variability in language used on social media across different time points of pregnancy (before, during, and after pregnancy, as well as by trimester and parity). Consenting patients shared access to their individual Facebook posts and electronic medical records. Random forest models trained on Facebook posts could differentiate first trimester of pregnancy from 3 months before pregnancy (F1 score = .63) and from a random 3-month time period (F1 score = .64). Posts during pregnancy were more likely to include themes about family (β = .22), food craving (β = .14), and date/times (β = .13), while posts 3 months prior to pregnancy included themes about social life (β = .30), sleep (β = .31), and curse words (β = .27), and 3 months post-pregnancy included themes of gratitude (β = .17), health appointments (β = .21), and religiosity (β = .18). Users who were pregnant for the first time were more likely to post about lack of sleep (β = .15), activities of daily living (β = .09), and communication (β = .08) compared with those who were pregnant after having a child who posted about others’ birthdays (β = .16) and life events (.12). A better understanding about social media timelines can provide insight into lifestyle choices that are specific to pregnancy.

Social media language of healthcare super-utilizers

An understanding of healthcare super-utilizers' online behaviors could better identify experiences to inform interventions. In this retrospective case-control study, we analyzed patients' social media posts to better understand their day-to-day behaviors and emotions expressed online. Patients included those receiving care in an urban academic emergency department who consented to share access to their historical Facebook posts and electronic health records. Super-utilizers were defined as patients with more than six visits to the Emergency Department (ED) in a year. We compared posts by super-utilizers with a matched group using propensity scoring based on age, gender and Charlson comorbidity index. Super-utilizers were more likely to post about confusion and negativity (D = .65, 95% CI-[.38, .95]), self-reflection (D = .63 [.35, .91]), avoidance (D = .62 [.34, .90]), swearing (D = .52 [.24, .79]), sleep (D = .60 [.32, .88]), seeking help and attention (D = .61 [.33, .89]), psychosomatic symptoms, (D = .49 [.22, .77]), self-agency (D = .56 [.29, .85]), anger (D = .51, [.24, .79]), stress (D = .46, [.19, .73]), and lonely expressions (D = .44, [.17, .71]). Insights from this study can potentially supplement offline community care services with online social support interventions considering the high engagement of super-utilizers on social media.

Predicting Cardiovascular Risk Using Social Media Data: Performance Evaluation of Machine-Learning Models

Background

Current atherosclerotic cardiovascular disease (ASCVD) predictive models have limitations; thus, efforts are underway to improve the discriminatory power of ASCVD models.

Objective

We sought to evaluate the discriminatory power of social media posts to predict the 10-year risk for ASCVD as compared to that of pooled cohort risk equations (PCEs).

Methods

We consented patients receiving care in an urban academic emergency department to share access to their Facebook posts and electronic medical records (EMRs). We retrieved Facebook status updates up to 5 years prior to study enrollment for all consenting patients. We identified patients (N=181) without a prior history of coronary heart disease, an ASCVD score in their EMR, and more than 200 words in their Facebook posts. Using Facebook posts from these patients, we applied a machine-learning model to predict 10-year ASCVD risk scores. Using a machine-learning model and a psycholinguistic dictionary, Linguistic Inquiry and Word Count, we evaluated if language from posts alone could predict differences in risk scores and the association of certain words with risk categories, respectively.

Results

The machine-learning model predicted the 10-year ASCVD risk scores for the categories <5%, 5%-7.4%, 7.5%-9.9%, and ≥10% with area under the curve (AUC) values of 0.78, 0.57, 0.72, and 0.61, respectively. The machine-learning model distinguished between low risk (<10%) and high risk (>10%) with an AUC of 0.69. Additionally, the machine-learning model predicted the ASCVD risk score with Pearson r=0.26. Using Linguistic Inquiry and Word Count, patients with higher ASCVD scores were more likely to use words associated with sadness (r=0.32).

Conclusions

Language used on social media can provide insights about an individual’s ASCVD risk and inform approaches to risk modification.

Well-Being Depends on Social Comparison: Hierarchical Models of Twitter Language Suggest That Richer Neighbors Make You Less Happy

Psychological research has shown that subjective well-being is sensitive to social comparison effects; individuals report decreased happiness when their neighbors earn more than they do. In this work, we use Twitter language to estimate the well-being of users, and model both individual and neighborhood income using hierarchical modeling across counties in the United States (US). We show that language-based estimates from a sample of 5.8 million Twitter users replicate results obtained from large-scale well-being surveys - relatively richer neighbors leads to lower well-being, even when controlling for absolute income. Furthermore, predicting individual-level happiness using hierarchical models (i.e., individuals within their communities) out-predicts standard baselines. We also explore language associated with relative income differences and find that individuals with lower income than their community tend to swear (f*ck, sh*t, b*tch), express anger (pissed, bullsh*t, wtf), hesitation (don't, anymore, idk, confused) and acts of social deviance (weed, blunt, drunk). These results suggest that social comparison robustly affects reported well-being, and that Twitter language analyses can be used to both measure these effects and shed light on their underlying psychological dynamics.

(Un)happiness and voting in U.S. presidential elections

A rapidly growing literature has attempted to explain Donald Trump's success in the 2016 U.S. presidential election as a result of a wide variety of differences in individual characteristics, attitudes, and social processes. We propose that the economic and psychological processes previously established have in common that they generated or electorally capitalized on unhappiness in the electorate, which emerges as a powerful high-level predictor of the 2016 electoral outcome. Drawing on a large dataset covering over 2 million individual surveys, which we aggregated to the county level, we find that low levels of evaluative, experienced, and eudaemonic subjective well-being (SWB) are strongly predictive of Trump's victory, accounting for an extensive list of demographic, ideological, and socioeconomic covariates and robustness checks. County-level future life evaluation alone correlates with the Trump vote share over Republican baselines at r = -.78 in the raw data, a magnitude rarely seen in the social sciences. We show similar findings when examining the association between individual-level life satisfaction and Trump voting. Low levels of SWB also predict anti-incumbent voting at the 2012 election, both at the county and individual level. The findings suggest that SWB is a powerful high-level marker of (dis)content and that SWB should be routinely considered alongside economic explanations of electoral choice. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

COVID-19 pandemic: every day feels like a weekday to most

The COVID-19 outbreak has clear clinical and economic impacts, but also affects behaviors e.g. through social distancing, and may increase stress and anxiety. However, while case numbers are tracked daily, we know little about the psychological effects of the outbreak on individuals in the moment. Here we examine the psychological and behavioral shifts over the initial stages of the outbreak in the United States in an observational longitudinal study. Through GPS phone data we find that homestay is increasing, while being at work dropped precipitously. Using regular real-time experiential surveys we observe an overall increase in stress and mood levels which is similar in size to the weekend vs. weekday differences. As there is a significant difference between weekday and weekend mood and stress levels, this is an important decrease in wellbeing. For some, especially those affected by job loss, the mental health impact is severe.

Estimating geographic subjective well-being from Twitter: A comparison of dictionary and data-driven language methods

Researchers and policy makers worldwide are interested in measuring the subjective well-being of populations. When users post on social media, they leave behind digital traces that reflect their thoughts and feelings. Aggregation of such digital traces may make it possible to monitor well-being at large scale. However, social media-based methods need to be robust to regional effects if they are to produce reliable estimates. Using a sample of 1.53 billion geotagged English tweets, we provide a systematic evaluation of word-level and data-driven methods for text analysis for generating well-being estimates for 1,208 US counties. We compared Twitter-based county-level estimates with well-being measurements provided by the Gallup-Sharecare Well-Being Index survey through 1.73 million phone surveys. We find that word-level methods (e.g., Linguistic Inquiry and Word Count [LIWC] 2015 and Language Assessment by Mechanical Turk [LabMT]) yielded inconsistent county-level well-being measurements due to regional, cultural, and socioeconomic differences in language use. However, removing as few as three of the most frequent words led to notable improvements in well-being prediction. Data-driven methods provided robust estimates, approximating the Gallup data at up to r = 0.64. We show that the findings generalized to county socioeconomic and health outcomes and were robust when poststratifying the samples to be more representative of the general US population. Regional well-being estimation from social media data seems to be robust when supervised data-driven methods are used.

Political depression? A big-data, multimethod investigation of Americans' emotional response to the Trump presidency

Previous studies suggested that the 2016 presidential elections gave rise to pathological levels of election-related distress in liberal Americans; however, it has also been suggested that the public discourse and the professional discourse have increasingly overgeneralized concepts of trauma and psychopathology. In light of this, in the current research, we utilized an array of big data measures and asked whether a political loss in a participatory democracy can indeed lead to psychopathology. We observed that liberals report being more depressed when asked directly about the effects of the election; however, more indirect measures show a short-lived or nonexistent effect. We examined self-report measures of clinical depression with and without a reference to the election (Studies 1A & 1B), analyzed Twitter discourse and measured users' levels of depression using a machine-learning-based model (Study 2), conducted time-series analysis of depression-related search behavior on Google (Study 3), examined the proportion of antidepressants consumption in Medicaid data (Study 4), and analyzed daily surveys of hundreds of thousands of Americans (Study 5), and saw that at the aggregate level, empirical data reject the accounts of "Trump Depression." We discuss possible interpretations of the discrepancies between the direct and indirect measures. The current investigation demonstrates how big-data sources can provide an unprecedented view of the psychological consequences of political events and sheds light on the complex relationship between the political and the personal spheres. (PsycInfo Database Record (c) 2020 APA, all rights reserved).

Machine learning highlights the deficiency of conventional dosimetric constraints for prevention of high-grade radiation esophagitis in non-small cell lung cancer treated with chemoradiation

•

A large cohort to predict radiation esophagitis in lung cancer patients was used.

•

Modern machine learning models were implemented to predict radiation esophagitis.

•

Previously published predictors of grade ≥ 3 radiation esophagitis may be unreliable.

Background and Purpose

Radiation esophagitis is a clinically important toxicity seen with treatment for locally-advanced non-small cell lung cancer. There is considerable disagreement among prior studies in identifying predictors of radiation esophagitis. We apply machine learning algorithms to identify factors contributing to the development of radiation esophagitis to uncover previously unidentified criteria and more robust dosimetric factors.

Materials and Methods

We used machine learning approaches to identify predictors of grade ≥ 3 radiation esophagitis in a cohort of 202 consecutive locally-advanced non-small cell lung cancer patients treated with definitive chemoradiation from 2008 to 2016. We evaluated 35 clinical features per patient grouped into risk factors, comorbidities, imaging, stage, histology, radiotherapy, chemotherapy and dosimetry. Univariate and multivariate analyses were performed using a panel of 11 machine learning algorithms combined with predictive power assessments.

Results

All patients were treated to a median dose of 66.6 Gy at 1.8 Gy per fraction using photon (89.6%) and proton (10.4%) beam therapy, most often with concurrent chemotherapy (86.6%). 11.4% of patients developed grade ≥ 3 radiation esophagitis. On univariate analysis, no individual feature was found to predict radiation esophagitis (AUC range 0.45–0.55, p ≥ 0.07). In multivariate analysis, all machine learning algorithms exhibited poor predictive performance (AUC range 0.46–0.56, p ≥ 0.07).

Conclusions

Contemporary machine learning algorithms applied to our modern, relatively large institutional cohort could not identify any reliable predictors of grade ≥ 3 radiation esophagitis. Additional patients are needed, and novel patient-specific and treatment characteristics should be investigated to develop clinically meaningful methods to mitigate this survival altering toxicity.

Variability in Language used on Social Media prior to Hospital Visits

Forecasting healthcare utilization has the potential to anticipate care needs, either accelerating needed care or redirecting patients toward care most appropriate to their needs. While prior research has utilized clinical information to forecast readmissions, analyzing digital footprints from social media can inform our understanding of individuals' behaviors, thoughts, and motivations preceding a healthcare visit. We evaluate how language patterns on social media change prior to emergency department (ED) visits and inpatient hospital admissions in this case-crossover study of adult patients visiting a large urban academic hospital system who consented to share access to their history of Facebook statuses and electronic medical records. An ensemble machine learning model forecasted ED visits and inpatient admissions with out-of-sample cross-validated AUCs of 0.64 and 0.70 respectively. Prior to an ED visit, there was a significant increase in depressed language (Cohen's d = 0.238), and a decrease in informal language (d = 0.345). Facebook posts prior to an inpatient admission showed significant increase in expressions of somatic pain (d = 0.267) and decrease in extraverted/social language (d = 0.357). These results are a first step in developing methods to utilize user-generated content to characterize patient care-seeking context which could ultimately enable better allocation of resources and potentially early interventions to reduce unplanned visits.

Examining the Phenomenon of Quarter-Life Crisis Through Artificial Intelligence and the Language of Twitter

Quarter-life crisis (QLC) is a popular term for developmental crisis episodes that occur during early adulthood (18–30). Our aim was to explore what linguistic themes are associated with this phenomenon as discussed on social media. We analyzed 1.5 million tweets written by over 1,400 users from the United Kingdom and United States that referred to QLC, comparing their posts to those used by a control set of users who were matched by age, gender and period of activity. Logistic regression was used to uncover significant associations between words, topics, and sentiments of users and QLC, controlling for demographics. Users who refer to a QLC were found to post more about feeling mixed emotions, feeling stuck, wanting change, career, illness, school, and family. Their language tended to be focused on the future. Of 20 terms selected according to early adult crisis theory, 16 were mentioned by the QLC group more than the control group. The insights from this study could be used by clinicians and coaches to better understand the developmental challenges faced by young adults and how these are portrayed naturalistically in the language of social media.

Cultural Differences in Tweeting about Drinking Across the US

Excessive alcohol use in the US contributes to over 88,000 deaths per year and costs over $250 billion annually. While previous studies have shown that excessive alcohol use can be detected from general patterns of social media engagement, we characterized how drinking-specific language varies across regions and cultures in the US. From a database of 38 billion public tweets, we selected those mentioning “drunk”, found the words and phrases distinctive of drinking posts, and then clustered these into topics and sets of semantically related words. We identified geolocated “drunk” tweets and correlated their language with the prevalence of self-reported excessive alcohol consumption (Behavioral Risk Factor Surveillance System; BRFSS). We then identified linguistic markers associated with excessive drinking in different regions and cultural communities as identified by the American Community Project. “Drunk” tweet frequency (of the 3.3 million geolocated “drunk” tweets) correlated with excessive alcohol consumption at both the county and state levels (r = 0.26 and 0.45, respectively, p < 0.01). Topic analyses revealed that excessive alcohol consumption was most correlated with references to drinking with friends (r = 0.20), family (r = 0.15), and driving under the influence (r = 0.14). Using the American Community Project classification, we found a number of cultural markers of drinking: religious communities had a high frequency of anti-drunk driving tweets, Hispanic centers discussed family members drinking, and college towns discussed sexual behavior. This study shows that Twitter can be used to explore the specific sociocultural contexts in which excessive alcohol use occurs within particular regions and communities. These findings can inform more targeted public health messaging and help to better understand cultural determinants of substance abuse.

Language of ADHD in Adults on Social Media

Objective: We computationally analyze the language of social media users diagnosed with ADHD to understand what they talk about, and how their language is correlated with users' characteristics such as personality and temporal orientation. Method: We analyzed approximately 1.3 million tweets written by 1,399 Twitter users with self-reported diagnoses of ADHD, comparing their posts with those used by a control set matched by age, gender, and period of activity. Results: Users with ADHD are found to be less agreeable, more open, to post more often, and to use more negations, hedging, and swear words. Posts are suggestive of themes of emotional dysregulation, self-criticism, substance abuse, and exhaustion. A machine learning model can predict which of these Twitter users has ADHD with an out-of-sample AUC of .836. Conclusion: Based on this emerging technology, conjectures of future uses of social media by researchers and clinicians to better understand the naturalistic manifestations and sequelae of ADHD.

Evaluating the predictability of medical conditions from social media posts

We studied whether medical conditions across 21 broad categories were predictable from social media content across approximately 20 million words written by 999 consenting patients. Facebook language significantly improved upon the prediction accuracy of demographic variables for 18 of the 21 disease categories; it was particularly effective at predicting diabetes and mental health conditions including anxiety, depression and psychoses. Social media data are a quantifiable link into the otherwise elusive daily lives of patients, providing an avenue for study and assessment of behavioral and environmental disease risk factors. Analogous to the genome, social media data linked to medical diagnoses can be banked with patients’ consent, and an encoding of social media language can be used as markers of disease risk, serve as a screening tool, and elucidate disease epidemiology. In what we believe to be the first report linking electronic medical record data with social media data from consenting patients, we identified that patients’ Facebook status updates can predict many health conditions, suggesting opportunities to use social media data to determine disease onset or exacerbation and to conduct social media-based health interventions.

The language of character strengths: Predicting morally valued traits on social media

OBJECTIVE

Social media is increasingly being used to study psychological constructs. This study is the first to use Twitter language to investigate the 24 Values in Action Inventory of Character Strengths, which have been shown to predict important life domains such as well-being.

METHOD

We use both a top-down closed-vocabulary (Linguistic Inquiry and Word Count) and a data-driven open-vocabulary (Differential Language Analysis) approach to analyze 3,937,768 tweets from 4,423 participants (64.3% female), who answered a 240-item survey on character strengths.

RESULTS

We present the language profiles of (a) a global positivity factor accounting for 36% of the variances in the strengths, and (b) each of the 24 individual strengths, for which we find largely face-valid language associations. Machine learning models trained on language data to predict character strengths reach out-of-sample prediction accuracies comparable to previous work on personality (r_median = 0.28, ranging from 0.13 to 0.51).

CONCLUSIONS

The findings suggest that Twitter can be used to characterize and predict character strengths. This technique could be used to measure the character strengths of large populations unobtrusively and cost-effectively.

(Not) hearing happiness: Predicting fluctuations in happy mood from acoustic cues using machine learning

Recent popular claims surrounding virtual assistants suggest that computers will soon be able to hear our emotions. Supporting this possibility, promising work has harnessed big data and emergent technologies to automatically predict stable levels of one specific emotion, happiness, at the community (e.g., counties) and trait (i.e., people) levels. Furthermore, research in affective science has shown that nonverbal vocal bursts (e.g., sighs, gasps) and specific acoustic features (e.g., pitch, energy) can differentiate between distinct emotions (e.g., anger, happiness) and that machine-learning algorithms can detect these differences. Yet, to our knowledge, no work has tested whether computers can automatically detect normal, everyday, within-person fluctuations in one emotional state from acoustic analysis. To address this issue in the context of happy mood, across 3 studies (total N = 20,197), we asked participants to repeatedly report their state happy mood and to provide audio recordings-including both direct speech and ambient sounds-from which we extracted acoustic features. Using three different machine learning algorithms (neural networks, random forests, and support vector machines) and two sets of acoustic features, we found that acoustic features yielded minimal predictive insight into happy mood above chance. Neither multilevel modeling analyses nor human coders provided additional insight into state happy mood. These findings suggest that it is not yet possible to automatically assess fluctuations in one emotional state (i.e., happy mood) from acoustic analysis, pointing to a critical future direction for affective scientists interested in acoustic analysis of emotion and automated emotion detection. (PsycInfo Database Record (c) 2020 APA, all rights reserved).