Search engines that help people make better health decisions

Abstract

The majority of Internet users access health information online using commercial search engines; sometimes preferring search over consultation with doctors and nurses. While the general public trust online health advice, and believe they are effective searchers, studies have reported that most people actually fail to subsequently make correct health decisions based on their search activities. This is concerning as less than two-thirds of those who search online to make health decisions then go on to consult a health professional with regard to their health condition.This thesis aims to create and validate methods to help the general public searching for health advice online (we refer to this search activity as consumer health search). Specifically, we investigate methods to assist people in formulating better queries and more effective appraisal of search results.In consumer health search, people tend to formulate ineffective queries that lead to non-relevant search results. We investigate methods that mitigate this problem by automatically clarifying the initial query with the use of knowledge-base resources. Our empirical evaluation shows that the proposed methods improve people's queries, leading to a higher number of relevant documents being presented more prominently. However, we also find that these improvements are highly subjective to the specific settings of the knowledge-base. To this end, we further investigate the pay-offs and pitfalls of using knowledge-bases for consumer health search.We then investigate methods to assist the appraisal of health search results, considering the impact of health cards - a specific type of entity card which presents information about a specific health concept in an enhanced and easily digestible way. Our empirical investigation with real users shows that presenting health cards increases the chances that less knowledgable people identify the correct health information. This is even more so in well-defined health search tasks such as searching for more information about a known health condition, rather than in exploratory tasks such as self-diagnosis.Subsequently, we investigate the impact of presenting health cards for people undertaking self-diagnosis tasks to assist in (1) determining the diagnosis; and (2) determining the follow-up action that should be taken (level of urgency). Motivated by the intuition underlying the differential diagnosis method\footnote{Differential diagnosis is the process of determining the most likely health condition among multiple health conditions that share similar symptoms.} in clinical practice, we propose a multi-card interface, which displays, side-by-side, several health cards deemed relevant to the person's query. Through empirical evaluation, we find that people interact more with the multi-cards interface than with a single health card interface, although it does not improve decision correctness, and that health card correctness is the main factor influencing the decision correctness. In addition, search activities tend to make people under-estimate the level of urgency of a health condition. This is of particular concern because while an incorrect diagnosis can be rectified by a health professional, this would not happen if medical attention is not sought.Finally, we investigate the effectiveness of general methods used for entity retrieval in ranking health cards given an explorative, self-diagnostic consumer health query. This study is the first quantitative evaluation of entity retrieval techniques in this context. Our empirical evaluation demonstrates that current entity retrieval models are not effective across a range of self-diagnostic, often underspecified queries - they are effective instead when queries that specifically mention the target health condition are stated.This thesis provides the following contributions: (1) We present methods to help people undertake health search tasks. These methods are important because they better ensure that appropriate, understandable information is accessed by people searching for health information online, thus enabling better decisions; (2) We divide the abstract problem of improving query formulation into a set of concrete choices which allow future work to focus on improving the best setting of each choice; (3) We classify health search intents and evaluate the effectiveness of search result interfaces based on the presentation of health cards, across search intents; (4) We propose a novel health card based interface to help people undertake difficult health search tasks for which current search interfaces provide limited support; (5) We create and publicly release the first health card collection which opens up an array of future avenues for studying ranking methods for health card retrieval.