Unsupervised Information Extraction from Behaviour Change Literature.

This paper describes our approach to construct a scalable system for unsupervised information extraction from the behaviour change intervention literature. Due to the many different types of attribute to be extracted, we adopt a passage retrieval based framework that provides the most likely value for an attribute. Our proposed method is capable of addressing variable length passage sizes and different validation criteria for the extracted values corresponding to each attribute to be found. We evaluate our approach by constructing a manually annotated ground-truth from a set of 50 research papers with reported studies on smoking cessation.

The Human Behaviour-Change Project: harnessing the power of artificial intelligence and machine learning for evidence synthesis and interpretation.

BACKGROUND: Behaviour change is key to addressing both the challenges facing human health and wellbeing and to promoting the uptake of research findings in health policy and practice. We need to make better use of the vast amount of accumulating evidence from behaviour change intervention (BCI) evaluations and promote the uptake of that evidence into a wide range of contexts. The scale and complexity of the task of synthesising and interpreting this evidence, and increasing evidence timeliness and accessibility, will require increased computer support. The Human Behaviour-Change Project (HBCP) will use Artificial Intelligence and Machine Learning to (i) develop and evaluate a 'Knowledge System' that automatically extracts, synthesises and interprets findings from BCI evaluation reports to generate new insights about behaviour change and improve prediction of intervention effectiveness and (ii) allow users, such as practitioners, policy makers and researchers, to easily and efficiently query the system to get answers to variants of the question 'What works, compared with what, how well, with what exposure, with what behaviours (for how long), for whom, in what settings and why?'. METHODS: The HBCP will: a) develop an ontology of BCI evaluations and their reports linking effect sizes for given target behaviours with intervention content and delivery and mechanisms of action, as moderated by exposure, populations and settings; b) develop and train an automated feature extraction system to annotate BCI evaluation reports using this ontology; c) develop and train machine learning and reasoning algorithms to use the annotated BCI evaluation reports to predict effect sizes for particular combinations of behaviours, interventions, populations and settings; d) build user and machine interfaces for interrogating and updating the knowledge base; and e) evaluate all the above in terms of performance and utility. DISCUSSION: The HBCP aims to revolutionise our ability to synthesise, interpret and deliver evidence on behaviour change interventions that is up-to-date and tailored to user need and context. This will enhance the usefulness, and support the implementation of, that evidence.

Probability Statements Extraction with Constrained Conditional Random Fields.

This paper investigates how to extract probability statements from academic medical papers. In previous work we have explored traditional classification methods which led to numerous false negatives. This current work focuses on constraining classification output obtained from a Conditional Random Field (CRF) model to allow for domain knowledge constraints. Our experimental results indicate constraining leads to a significant improvement in performance.

Extracting Dependence Relations from Unstructured Medical Text.

Dependence relations among disease and risk factors are a key ingredient in risk modeling and decision support models. Currently such information is either provided by experts (costly and time consuming) or extracted from data (if available). The published medical literature represents a promising source of such knowledge; however its manual processing is practically infeasible. While a number of solutions have been introduced to add structure to biomedical literature, none adequately recover dependence relations. The objective of our research is to build such an automatic dependence extraction solution, based on a sequence of natural language processing steps, which take as input a set of MEDLINE abstracts and provide as output a list of structured dependence statements. This paper presents a hybrid pipeline approach, a combination of rule-based and machine learning algorithms. We found that this approach outperforms a strictly rule-based approach.

Person-Specific Standardized Vulnerability Assessment in Health and Social Care.

We describe an integrated person-specific standardized vulnerability assessment model designed to facilitate patient management in health and social care. Such a system is not meant to replace existing health and social assessment models but rather to complement them by providing a holistic picture of the vulnerabilities faced by a given patient. In fact, it should be seen as a screening tool for health and social care workers. One key aspect of the modeling framework is the ability to provide personalized yet standardized multi-dimensional assessments of risk based on incomplete information about the patient status, as is the case in screening situations. Specifically, we integrate a Markov chain model describing the evolution of patients in and out of vulnerable states over time with a Bayesian network that serves to customize the dynamic model. We present an application in the context of elder care.

Extracting risk modeling information from medical articles.

Risk modeling in healthcare is both ubiquitous and in its infancy. On the one hand, a significant proportion of medical research focuses on determining the factors that influence the incidence, severity and treatment of diseases, which is a form of risk identification. Those studies typically investigate the micro-level of risk modeling, i.e., the existence of dependences between a reduced set of hypothesized (or demonstrated) risk factors and a focus disease or treatment. On the other hand, the macro-level of risk modeling, i.e., articulating how a large number of such risk factors interact to affect diseases and treatments is not widespread, though essential for medical decision support modeling. By exploiting advances in natural language processing, we believe that information contained in unstructured texts such as medical articles could be extracted to facilitate aggregation into macro-level risk models.

Engineering risk analysis of a hospital oxygen supply system.

Reports from the Food and Drug Administration (FDA) and the Joint Commission on Accreditation of Healthcare Organizations (JCAHO) have emphasized the potential for injury to patients caused by failures in oxygen supply systems. This article presents a model of patient risk related to the process of supplying oxygen at a single university hospital. One of the goals of the article is to illustrate how probabilistic risk analysis (PRA) can be used by hospitals to assess and mitigate risk and, therefore, to meet JCAHO requirements. PRA techniques are useful to 1) model the reliability of a complex system and 2) assess the cost-effectiveness of different risk mitigation measures. The authors focus on the risk estimation step, describing in detail their modeling of the oxygen supply system and analysis of the results. For the hospital that the authors study (20,000 admissions yearly), the total expected number of fatalities from oxygen system failure is 44 over a 30-year time horizon. The greatest contribution to the risk (94% of the expected number of fatalities) comes from problems that involve the supply network (e.g., damage to structure and poisoning) as opposed to incidents that occur inside patient rooms. Although the threat to patient safety is not dramatic, health care organizations should be concerned about potential failures of their oxygen system because improving this system could avoid low-probability, high-consequence failures at a low cost.