Automatic Classification of Users' Health Information Need Context: Logistic Regression Analysis of Mouse-Click and Eye-Tracker Data.

BACKGROUND: Users searching for health information on the Internet may be searching for their own health issue, searching for someone else's health issue, or browsing with no particular health issue in mind. Previous research has found that these three categories of users focus on different types of health information. However, most health information websites provide static content for all users. If the three types of user health information need contexts can be identified by the Web application, the search results or information offered to the user can be customized to increase its relevance or usefulness to the user. OBJECTIVE: The aim of this study was to investigate the possibility of identifying the three user health information contexts (searching for self, searching for others, or browsing with no particular health issue in mind) using just hyperlink clicking behavior; using eye-tracking information; and using a combination of eye-tracking, demographic, and urgency information. Predictive models are developed using multinomial logistic regression. METHODS: A total of 74 participants (39 females and 35 males) who were mainly staff and students of a university were asked to browse a health discussion forum, Healthboards.com. An eye tracker recorded their examining (eye fixation) and skimming (quick eye movement) behaviors on 2 types of screens: summary result screen displaying a list of post headers, and detailed post screen. The following three types of predictive models were developed using logistic regression analysis: model 1 used only the time spent in scanning the summary result screen and reading the detailed post screen, which can be determined from the user's mouse clicks; model 2 used the examining and skimming durations on each screen, recorded by an eye tracker; and model 3 added user demographic and urgency information to model 2. RESULTS: An analysis of variance (ANOVA) analysis found that users' browsing durations were significantly different for the three health information contexts (P<.001). The logistic regression model 3 was able to predict the user's type of health information context with a 10-fold cross validation mean accuracy of 84% (62/74), followed by model 2 at 73% (54/74) and model 1 at 71% (52/78). In addition, correlation analysis found that particular browsing durations were highly correlated with users' age, education level, and the urgency of their information need. CONCLUSIONS: A user's type of health information need context (ie, searching for self, for others, or with no health issue in mind) can be identified with reasonable accuracy using just user mouse clicks that can easily be detected by Web applications. Higher accuracy can be obtained using Google glass or future computing devices with eye tracking function.

Feasibility of 30-day hospital readmission prediction modeling based on health information exchange data.

INTRODUCTION: Unplanned 30-day hospital readmission account for roughly $17 billion in annual Medicare spending. Many factors contribute to unplanned hospital readmissions and multiple models have been developed over the years to predict them. Most researchers have used insurance claims or administrative data to train and operationalize their Readmission Risk Prediction Models (RRPMs). Some RRPM developers have also used electronic health records data; however, using health informatics exchange data has been uncommon among such predictive models and can be beneficial in its ability to provide real-time alerts to providers at the point of care. METHODS: We conducted a semi-systematic review of readmission predictive factors published prior to March 2013. Then, we extracted and merged all significant variables listed in those articles for RRPMs. Finally, we matched these variables with common HL7 messages transmitted by a sample of health information exchange organizations (HIO). RESULTS: The semi-systematic review resulted in identification of 32 articles and 297 predictive variables. The mapping of these variables with common HL7 segments resulted in an 89.2% total coverage, with the DG1 (diagnosis) segment having the highest coverage of 39.4%. The PID (patient identification) and OBX (observation results) segments cover 13.9% and 9.1% of the variables. Evaluating the same coverage in three sample HIOs showed data incompleteness. DISCUSSION: HIOs can utilize HL7 messages to develop unique RRPMs for their stakeholders; however, data completeness of exchanged messages should meet certain thresholds. If data quality standards are met by stakeholders, HIOs would be able to provide real-time RRPMs that not only predict intra-hospital readmissions but also inter-hospital cases. CONCLUSION: A RRPM derived using HIO data exchanged through may prove to be a useful method to prevent unplanned hospital readmissions. In order for the RRPM derived from HIO data to be effective, hospitals must actively exchange clinical information through the HIO and develop actionable methods that integrate into the workflow of providers to ensure that patients at high-risk for readmission receive the care they need.

Understanding user intents in online health forums.

Online health forums provide a convenient way for patients to obtain medical information and connect with physicians and peers outside of clinical settings. However, large quantities of unstructured and diversified content generated on these forums make it difficult for users to digest and extract useful information. Understanding user intents would enable forums to find and recommend relevant information to users by filtering out threads that do not match particular intents. In this paper, we derive a taxonomy of intents to capture user information needs in online health forums and propose novel pattern-based features for use with a multiclass support vector machine (SVM) classifier to classify original thread posts according to their underlying intents. Since no dataset existed for this task, we employ three annotators to manually label a dataset of 1192 HealthBoards posts spanning four forum topics. Experimental results show that a SVM using pattern-based features is highly capable of identifying user intents in forum posts, reaching a maximum precision of 75%, and that a SVM-based hierarchical classifier using both pattern and word features outperforms its SVM counterpart that uses only word features. Furthermore, comparable classification performance can be achieved by training and testing on posts from different forum topics.