Dane Bell

Analyzing the language of food on social media

We investigate the predictive power behind the language of food on social media. We collect a corpus of over three million food-related posts from Twitter and demonstrate that many latent population characteristics can be directly predicted from this data: overweight rate, diabetes rate, political leaning, and home geographical location of authors. For all tasks, our language-based models significantly outperform the majority-class baselines. Performance is further improved with more complex natural language processing, such as topic modeling. We analyze which textual features have greatest predictive power for these datasets, providing insight into the connections between the language of food, geographic locale, and community characteristics. Lastly, we design and implement an online system for real-time query and visualization of the dataset. Visualization tools, such as geo-referenced heatmaps and temporal histograms, allow us to discover more complex, global patterns mirrored in the language of food.

A Test of The Risk Perception Attitude Framework as a Message Tailoring Strategy to Promote Diabetes Screening

ABSTRACT The risk perception attitude (RPA) framework was tested as a message tailoring strategy to encourage diabetes screening. Participants (N = 602) were first categorized into one of four RPA groups based on their diabetes risk and efficacy perceptions and then randomly assigned to receive a message that matched their RPA, mismatched their RPA, or a control message. Participants receiving a matched message reported greater intentions to engage in self-protective behavior than participants who received a mismatched message or the control message. The results also showed differences in attitudes and behavioral intentions across the four RPA groups. Participants in the responsive group had more positive attitudes toward diabetes screening than the other three groups, whereas participants in the indifferent group reported the weakest intentions to engage in self-protective behavior.

Large-scale automated machine reading discovers new cancer-driving mechanisms

Abstract PubMed, a repository and search engine for biomedical literature, now indexes >1 million articles each year. This exceeds the processing capacity of human domain experts, limiting our ability to truly understand many diseases. We present Reach, a system for automated, large-scale machine reading of biomedical papers that can extract mechanistic descriptions of biological processes with relatively high precision at high throughput. We demonstrate that combining the extracted pathway fragments with existing biological data analysis algorithms that rely on curated models helps identify and explain a large number of previously unidentified mutually exclusive altered signaling pathways in seven different cancer types. This work shows that combining human-curated ‘big mechanisms’ with extracted ‘big data’ can lead to a causal, predictive understanding of cellular processes and unlock important downstream applications.

A Study of Calorie Estimation in Pictures of Food (Preprint)

Background Software designed to accurately estimate food calories from still images could help users and health professionals identify dietary patterns and food choices associated with health and health risks more effectively. However, calorie estimation from images is difficult, and no publicly available software can do so accurately while minimizing the burden associated with data collection and analysis. Objective The aim of this study was to determine the accuracy of crowdsourced annotations of calorie content in food images and to identify and quantify sources of bias and noise as a function of respondent characteristics and food qualities (eg, energy density). Methods We invited adult social media users to provide calorie estimates for 20 food images (for which ground truth calorie data were known) using a custom-built webpage that administers an online quiz. The images were selected to provide a range of food types and energy density. Participants optionally provided age range, gender, and their height and weight. In addition, 5 nutrition experts provided annotations for the same data to form a basis of comparison. We examined estimated accuracy on the basis of expertise, demographic data, and food qualities using linear mixed-effects models with participant and image index as random variables. We also analyzed the advantage of aggregating nonexpert estimates. Results A total of 2028 respondents agreed to participate in the study (males: 770/2028, 37.97%, mean body mass index: 27.5 kg/m2). Average accuracy was 5 out of 20 correct guesses, where “correct” was defined as a number within 20% of the ground truth. Even a small crowd of 10 individuals achieved an accuracy of 7, exceeding the average individual and expert annotator’s accuracy of 5. Women were more accurate than men (P<.001), and younger people were more accurate than older people (P<.001). The calorie content of energy-dense foods was overestimated (P=.02). Participants performed worse when images contained reference objects, such as credit cards, for scale (P=.01). Conclusions Our findings provide new information about how calories are estimated from food images, which can inform the design of related software and analyses.Background: Software designed to accurately estimate food calories from still images could help users and health professionals more efficiently identify dietary patterns and food choices associated with health and health risks. However, calorie estimation from images is difficult, and no publicly available software can do so accurately while minimizing the burden associated with data collection and analysis. Objective: The aim of this study is to determine the accuracy of crowdsourced annotations of calorie content in food images, and to identify and quantify sources of bias and noise as a function of respondent characteristics and food qualities (e.g., energy density). Methods: We invited adult social media users to provide calorie estimates for 20 food images (for which ground truth calorie data were known) using a custom-built webpage that administers an online quiz. The images were selected to provide a range of food types and energy density. Participants optionally provided age range, gender, and their height and weight. Additionally, five nutrition experts provided annotations for the same data to form a basis of comparison. We examined estimate accuracy on the basis of expertise, demographic data, and food qualities using linear mixed effects models with participant and image index as random variables. We also analyzed the advantage of aggregating nonexpert estimates. Results: 2028 respondents agreed to participate in the study (males: 770 [38%], mean body mass index: 27.5). Average accuracy was 5 out of 20 correct guesses, where “correct” was defined as a number within 20% of the ground truth. Even a small crowd of 10 individuals achieved an accuracy of 7, exceeding the average individuals and expert annotator’s accuracy of 5. Women were more accurate than men (P<.001), and younger people were more accurate than older people (P<.001). The calorie content of energy-dense foods was overestimated (P=.024). Participants performed worse when images contained reference objects, such as credit cards, for scale (P=.014). Conclusions: Our findings provide new information about how calories are estimated from food images, which can inform the design of related software and analyses.

This before That: Causal Precedence in the Biomedical Domain.

Causal precedence between biochemical interactions is crucial in the biomedical domain, because it transforms collections of individual interactions, e.g., bindings and phosphorylations, into the causal mechanisms needed to inform meaningful search and inference. Here, we analyze causal precedence in the biomedical domain as distinct from open-domain, temporal precedence. First, we describe a novel, hand-annotated text corpus of causal precedence in the biomedical domain. Second, we use this corpus to investigate a battery of models of precedence, covering rule-based, feature-based, and latent representation models. The highest-performing individual model achieved a micro F1 of 43 points, approaching the best performers on the simpler temporal-only precedence tasks. Feature-based and latent representation models each outperform the rule-based models, but their performance is complementary to one another. We apply a sieve-based architecture to capitalize on this lack of overlap, achieving a micro F1 score of 46 points.