Artificial Intelligence in Medicine

Complex human phenotypes, such as autism, schizophrenia, and anxiety, undoubtedly partially overlap in genomic variations that predispose to or protect against these maladies. (Genetic overlap of complex phenotypes has gained increasing experimental support and is no longer just an ungrounded scientific hypothesis.) Furthermore, as yet largely unknown shared environmental factors likely tend to trigger the manifestation of more than one phenotype. Although it may seem overly ambitious to target multiple phenotypes jointly, we believe we can obtain much more information from existing data and gain new insights into individual phenotypes by modeling phenotypes jointly. My talk sketches two distinct computational approaches to this problem.

Healthcare organizations are facing the challenge of delivering personalized services to their patients in a cost-effective and efficient manner. This, in turn, requires advanced IT support for healthcare processes covering both organizational procedures and knowledge-intensive, dynamic treatment processes. Nowadays, required agility is often hindered by a lack of flexibility in hospital information systems. To overcome this inflexibility a new generation of information systems, denoted as process-aware information systems (PAISs), has emerged. In contrast to data- and function-centered information systems, a PAIS separates process logic from application code and thus provides an additional architectural layer. However, the introduction of process-aware hospital information systems must neither result in rigidity nor restrict staff members in their daily work. This keynote presentation reflects on recent developments from the business process management (BPM) domain, which enable process adaptation, process flexibility, and process evolution. These key features will be illustrated along existing BPM frameworks. Altogether, emerging BPM methods, concepts and technologies will contribute to further enhance IT support for healthcare processes.

The paper describes the process of knowledge elicitation for a neurological decision support system. To alleviate the difficult problem of knowledge elicitation from data and domain experts, we used a recently developed technique called ABML (Argument Based Machine Learning). The paper demonstrates ABML’s advantage in combining machine learning and expert knowledge. ABML guides the expert to explain critical special cases which cannot be handled automatically by machine learning. This very efficiently reduces the expert’s workload, and combines it with automatically learned knowledge. We developed a decision support system to help the neurologists differentiate between three types of tremors: Parkinsonian, essential, and mixed tremor (co-morbidity). The system is intended to act as a second opinion for the neurologists, and most importantly to help them reduce the number of patients in the “gray area” that require a very costly further examination (DaTSCAN).

Helping someone who is depressed can be very important to the depressed person. A number of supportive family members or friends can often make a big difference. This paper addresses how a social support network can be formed, taking the needs of the support recipient and the possibilities of the potential support providers into account. To do so, dynamic models about the preferences and needs of both support providers and support recipients are exploited. The outcome of this is used as input for a configuration process of a support network. In a case study, it is show how such an intelligently formed network results in a reduced long term stress level.

The EIRA system has proved to be successful in the detection of anomalous patient responses to treatments in the Intensive Care Unit (ICU). One weakness of EIRA is the lack of mechanisms to describe to the clinicians, rationales behind the anomalous detections. In this paper, we extend EIRA by providing it with an argumentation-based justification system that formalizes and communicates to the clinicians the reasons why a patient response is anomalous. The implemented justification system uses human-like argumentation techniques and is based on real dialogues between ICU clinicians.

The Sleep Apnea Syndrome is a sleep disorder characterized by frequently repeated respiratory disorders during sleep. It needs the simultaneous recording of many physiological parameters to be diagnosed. The analysis of these curves is a time consuming task made by sleep Physicians. First, they detect some physiological events on each curve and then, they point out links between respiratory events and their consequences. To support the diagnosis, we used symbolic fusion on elementary events, which connects events to their sleep context - sleep-stage and body position - and to the respiratory event responsible of their occurrence. The reference indicator is the Apnea-Hypopnea Index (AHI), defined as the average hourly frequency of arisen of Apneas or Hypopneas while the patient is sleeping. We worked on the polysomnography of 59 patients, that were first completely analyzed by a sleep Physician and then analyzed by our method. We compared the ratio of the AHI got by the automatic analysis and the AHI got by the sleep Physician.

$$\delta=\frac{AHI(automatic analysis)}{AHI(Sleep Physician Analysis)}$$

Globally, we overvalued the count of apneas and hypopneas for the group of patients with

AHI

 ≤ 5, that are considered as healthy patients. In average, for these patients,

δ

 = 2,71. For patients with mild or moderate Sleep Apnea Syndrome we globally found a similar AHI. In average, for these patients,

δ

 = 1,04. For patients with severe Sleep Apnea Syndrome, we undervalued a little the count of respiratory events. In average, for these patients,

δ

 = 0,83. This leads to the same severity class for most of the patients.

Improving physical activity patterns is an important focus in the treatment of chronic illnesses. We describe a system to monitor activity and provide feedback to help patients reach a healthy daily pattern. The system has shown positive effects in trials on patient groups including COPD and obese patients. We describe the design and implementation of a new feedback generation module which improves interaction with the patient by providing personalised dynamic context-aware feedback. The system uses an ontology of messages to find appropriate feedback using context information to prune irrelevant paths. The system adapts using derived probabilities about user preferences for certain message types. We aim to improve patient compliance and user experience.

Classification of different types of dementia commonly involves examination from several perspectives, e.g., medical images, neuropsychological tests, etc. Thus, dementia classification should lend itself to so-called

multi-view learning

. Instead of simply combining several views, we use stacking to make the most of the information from the various views (PET scans, MMSE, CERAD and demographic variables). In the paper, we not only show the performance of stacked multi-view learning on classifying dementia data, we also try to explain the factors contributing to its performance. More specifically, we show that the correlation of views on the base and the meta level should be within certain ranges to facilitate successful stacked multi-view learning.

This work explores the automatic recognition of physical activity intensity patterns from multi-axial accelerometry and heart rate signals. Data collection was carried out in free-living conditions and in three controlled gymnasium circuits, for a total amount of 179.80 h of data divided into: sedentary situations (65.5%), light-to-moderate activity (17.6%) and vigorous exercise (16.9%). The proposed machine learning algorithms comprise the following steps: time-domain feature definition, standardization and PCA projection, unsupervised clustering (by

k

-means and GMM) and a HMM to account for long-term temporal trends. Performance was evaluated by 30 runs of a 10-fold cross-validation. Both

k

-means and GMM-based approaches yielded high overall accuracy (86.97% and 85.03%, respectively) and, given the imbalance of the dataset, meritorious F-measures (up to 77.88%) for non-sedentary cases. Classification errors tended to be concentrated around transients, what constrains their practical impact. Hence, we consider our proposal to be suitable for 24 h-based monitoring of physical activity in ambulatory scenarios and a first step towards intensity-specific energy expenditure estimators.

Understanding the key role that miRNAs play in the regulation of gene expression is one of the most important challenges in modern molecular biology. Standard gene set enrichment analysis (GSEA) is not appropriate in this context, due to the low specificity of the relation between miRNAs and their target genes. We developed alternative strategies to gain better insights in the differences in biological processes involved in different experimental conditions. We here describe a novel method to analyze and interpret miRNA expression data correctly, and demonstrate that annotating miRNA directly to biological processes through their target genes (which is nevertheless the only way possible) is a non-trivial task. We are currently employing the same strategy to relate miRNA expression patterns directly to pathway information, to generate new hypotheses, which may be relevant for the interpretation of their role in the gene expression regulatory processes.

Genome-scale transcription profile is known to be a good reporter of the state of the cell. Much of the early predictive modelling and cell-type clustering relied on this relation and has experimentally confirmed it. We have examined if this also holds for prediction of cell’s staging, and focused on the inference of stage prediction models for stem cell development. We show that the problem relates to rank learning and, from the user’s point of view, to projection of transcription profile data to a single dimension. Our comparison of several state-of-the-art algorithms on 10 data sets from Gene Expression Omnibus shows that rank-learning can be successfully applied to developmental cell staging, and that relatively simple techniques can perform surprisingly well.

Keratoconus identification has become a step of primary importance in the preoperative evaluation for the refractive surgery. With the ophthalmology knowledge improvement, corneal physical parameters were considered important to its evaluation. The Ocular Response Analyzer (ORA) provides some physical parameters using an applanation process to measure cornea biomechanical properties. This paper presents a study of machine learning classifiers in keratoconus diagnosis from ORA examinations. As a first use of machine learning approach with ORA parameters, this research work presents a performance comparison of the main machine learning algorithms. This approach improves ORA parameters’ analysis helping ophthalmologist’s efficiency in clinical diagnosis.

Heart rate variability (HRV) analysis can be successfully applied to automatic classification of cardiac rhythm abnormalities. This paper presents a novel Java-based computer framework for feature extraction from cardiac rhythms. The framework called HRVFrame implements more than 30 HRV linear time domain, frequency domain, time-frequency domain, and nonlinear features. Output of the framework in the form of .arff files enables easier medical knowledge discovery via platforms such as RapidMiner or Weka. The scope of the framework facilitates comparison of models for different cardiac disorders. Some of the features implemented in the framework can also be applied to other biomedical time-series. The thorough approach to feature extraction pursued in this work is also encouraged for other types of biomedical time-series.

A potentially effective IT intervention to implement guidelines and evidence based practice consists of the use of computerized decision support systems (CDSS). CDSSs aim at providing meaningful feedback to professionals in order to positively influence their behavior. Intensive care medicine, with its heavy reliance on information and the advanced information infrastructure in intensive care units (ICUs), is an attractive specialty and environment for applying and investigating CDSSs. In particular, antibiotic prescription, control of the tidal volumes in the lungs, and control of glucose levels in the blood form hot topics in intensive care medicine and provide opportunities for decision support applications. However, issues pertaining to the design, implementation, critical success factors, as well as the evaluation of CDSSs are largely still open, especially in these domains. This work describes important issues learned from designing and implementing CDSSs in these domains based on our literature reviews and lessons learned from conducting various trials in our ICU.

Cardiac rehabilitation is a multidisciplinary therapy aimed at recovery and secondary prevention after hospitalization for cardiac incidents (such as myocardial infarctions) and cardiac interventions (such as heart surgery). To stimulate implementation of the national guidelines, an electronic patient record system with computerised decision support functionalities called CARDSS (cardiac rehabilitation decision support system) was developed, and made available to Dutch rehabilitation clinics. The system was quantitatively evaluated in a cluster randomised trial at 31 clinics, and qualitatively by interviewing 29 users of the system. Computerised decision support was found to improve guideline concordance by increasing professional knowledge of preferred practice, by reducing inertia to previous practice, and by reducing guideline complexity. It was not effective when organizational or procedural changes were required that users considered to be beyond their responsibilities.

Clinical decision support systems (CDSSs) may be appropriate tools to promote the use of clinical practice guidelines (CPGs). However, compliance with CPGs is a multifactorial process that relies on the CPGs to be implemented, the physician(s) in charge of the decision, and the patient to manage. Formal concept analysis (FCA) allows to derive implicit relationships from a set of objects described by their attributes, based on the principle of attribute sharing between objects. We used FCA to elicit patient-based formal concepts related to the non-conformity of multidisciplinary staff meetings (MSMs) decisions with CPGs in the domain of breast cancer management. We developed a strategy for selecting attributes and make lattices manageable. We found that when not using the guideline-based CDSS OncoDoc2, patients with bad prognostic factors were associated with non-compliant decisions. This was corrected when the system was used during MSMs.

This paper describes the process of developing the decision-support system DMSS (Dementia Management and Support System) and some lessons learned. An action research and participatory design approach has been adopted during development, with a strong research focus on optimizing support to physicians in dementia diagnosis assessment, involving a number of physicians and clinics in the process. A stand-alone version is currently used in 11 clinics distributed over four countries. Results from evaluation studies show that the system and the physician comply in 84,6% of the patient cases and that reasons for non-compliance lie primarily in physician’s insufficient knowledge. The impact the system has had on the individual physician’s diagnostic procedure in observation studies, factors identified enabling the integration and obstacles to use are presented and discussed. The system’s support for assessing basic cognitive functions is being improved, primarily as a feature for personalization of a future web-based version of DMSS.

Online delivery of lifestyle intervention programs offers the potential to cost effectively reach large cohorts of users with various information and dietary needs. Unfortunately, online systems can fail to engage users in the long term, affecting their ability to sustain positive lifestyle change. In this work we present the initial analysis of a large scale application study of personalized technologies for lifestyle change. We evaluate the stickiness of an eHealth portal which provides individuals with three personalized tools – meal planner, social network feeds, and social comparison – to make change a reality in their lives. More than 5000 Australians took part in a 12 week study and provided solid empirical evidence for how the inclusion of personalized tools can assist and motivate users. Initial results show that the personalized tools boost user interaction with the portal, simplify information access, and assist in motivating users.

This paper describes progress in a model-based approach to building a decision support system for mechanical ventilation. It highlights that the process of building models promotes generation of ideas and describes three systems resulting from this process, i.e. for assessing pulmonary gas exchange, calculating arterial acid-base status; and optimizing mechanical ventilation. Each system is presented and its current status and impact reviewed.

Much knowledge of human physiology is formalised as systems of differential equations. For example, standard models of pharmacokinetics and pharmacodynamics use systems of differential equations to describe a drug’s movement through the body and its effects. Here, we propose a method for automatically incorporating this existing knowledge into a Dynamic Bayesian Network (DBN) framework. A benefit of recasting a differential equation model as a DBN is that the DBN can be used to individualise the model parameters dynamically, based on real-time evidence. Our approach provides principled handling of data and model uncertainty, and facilitates integration of multiple strands of temporal evidence. We demonstrate our approach with an abstract example and evaluate it in a real-world medical problem, tracking the interaction of insulin and glucose in critically ill patients. We show that it is better able to reason with the data, which is sporadic and has measurement uncertainties.

Managing chronic disease through automated systems has the potential to both benefit the patient and reduce health-care costs. We are developing and evaluating a monitoring system for patients with chronic obstructive pulmonary disease which aims to detect exacerbations and thus help patients manage their disease and prevent hospitalisation. We have carefully drafted a system design consisting of an intelligent device that is able to alert the patient, collect case-specific, subjective and objective, physiological data, offer a patient-specific interpretation of the collected data by means of probabilistic reasoning, and send data to a central server for inspection by health-care professionals. A first pilot with actual COPD patients suggests that an intervention based on this system could be successful.

There is increasing consensus among health-care professionals and patients alike that many disorders can be managed, in principle, much better at home than in an out-patient clinic or hospital. In the paper, we describe a novel temporal Bayesian network model for the at home time-related development of preeclampsia, a common pregnancy-related disorder. The network model drives an android-based smartphone application that offers patients and their doctor insight into whether or not the disorder is developing positively—no clinical intervention required—or negatively—clinical intervention is definitely required. We discuss design considerations of the model and system, and review results obtained with actual patients.

We are interested in retrieving relevant information from biomedical documents according to healthcare professional’s information needs. It is well known that biomedical documents are indexed using conceptual descriptors issued from terminologies for a better retrieval performance. Our attempt to develop a conceptual retrieval framework relies on the hypothesis that there are several broad categories of knowledge that could be captured from different terminologies and processed by retrieval algorithms. With this in mind, we propose a multi-terminology based indexing approach for selecting the best representative concepts for each document. We instantiate this general approach on four terminologies namely MeSH (Medical Subject Headings), SNOMED (Systematized Nomenclature of Medicine), ICD-10 (International Classification of Diseases) and GO (Gene Ontology). Experimental studies were conducted on large and official document test collections of real world clinical queries and associated judgments extracted from MEDLINE scientific collections, namely TREC Genomics 2004 & 2005. The obtained results demonstrate the advantages of our multi-terminology based biomedical information retrieval approach over state-of-the art approaches.

We present a method for creating mappings between the Orphanet terminology of rare diseases and the Unified Medical Language System (UMLS), in particular the SNOMED CT, MeSH, and MedDRA terminologies. Our method is based on: (i) aggressive normalisation of terms specific to the Orphanet terminology on top of standard UMLS normalisation; (ii) semantic ranking of partial candidate mappings in order to group similar mappings and attribute higher ranking to the more informative ones. Our results show that, by using the aggressive normalisation function, we increase the number of exact candidate mappings by 7.1-9.5% compared to a mapping method based on MetaMap. A manual assessment of our results shows a high precision of 94.6%. Our results imply that Orphanet diseases are under-represented in the aforementioned terminologies: SNOMED CT, MeSH, and MedDRA are found to contain only 35%, 42%, and 15% of the Orphanet rare diseases, respectively.

Representing the Foundational Model of Anatomy (FMA) in OWL 2 is essential for semantic interoperability. The paper describes the method and tool used to formalize the FMA in OWL 2. One main strength of the approach is to leverage OWL 2 expressiveness and the naming conventions of the native FMA to make explicit some implicit semantics, meanwhile improving its ontological model and fixing some errors. A second originality is the flexible tool developed. It enables to easily generate a new version for each Protégé FMA update. While it provides one ‘standard’ FMA-OWL version by default, many options allow for producing other variants customized to users applications. To the best of our knowledge, no complete representation of the entire FMA in OWL DL or OWL 2 existed so far.

This work aims at improving information retrieval in a health gateway by meta-modelling multiple terminologies related to medicine. The meta-model is based on meta-terms that gather several terms semantically related. Meta-terms, initially modelled for the MeSH thesaurus, are extended for other terminologies such as IC10 or SNOMED Int. The usefulness of this model and the relevance of information retrieval is evaluated and compared in the case of one and multiple terminologies. The results show that exploiting multiple terminologies contributes to increase recall but lowers precision.

MedDRA is exploited for the indexing of pharmacovigilance spontaneous reports. But since spontaneous reports cover only a small proportion of the existing adverse drug reactions, the exploration of clinical reports is seriously considered. Through the UMLS, the current mapping between MedDRA and SNOMED CT, this last being used for indexing clinical data in many countries, is only 42%. In this work, we propose to improve this mapping through an automatic lexical-based approach. We obtained 308 direct mappings of a MedDRA term to a SNOMED CT concept. After segmenting MedDRA terms, we identified 535 full mappings associating a MedDRA term with one or more SNOMED CT concepts. The direct approach resulted in 199 (64.6%) correct mappings while through segmentation this number raises to 423 (79.1%). On the whole, our method provided interesting and useful results.

This paper presents our work-in-progress on designing and implementing an integrated ontology for widespread knowledge dissemination in the domain of obesity with emphasis on childhood obesity. The COPE ontology aims to support a knowledge-based infrastructure to promote healthy eating habits and lifestyles, analyze children’s behaviors and habits associated with obesity and to prevent or reduce the prevalence of childhood obesity and overweight. By formally integrating and harmonizing multiple knowledge sources across disciplinary boundaries, we will facilitate cross-sectional analysis of the domain of obesity and generate both generic and customized preventive recommendations, which take into consideration several factors, including existing conditions in individuals and communities.

Recently, we devised a method to develop prognostic models incorporating patterns of sequential organ failure to predict the eventual hospital mortality at each day of intensive care stay. In this study, we aimed to understand, using a real world setting, how these models perform compared to physicians, who are exposed to additional information than the models. We found a slightly better discriminative ability for physicians (AUC range over days: 0.73-0.83 vs. 0.70-0.80) and a slightly better accuracy for the models (Brier score range: 0.14-0.19 vs. 0.16-0.19). However when we combined both sources of predictions we arrived at a significantly superior discrimination as well as accuracy (AUC range: 0.81-0.88; Brier score range: 0.11-0.15). Our results show that the models and the physicians draw on complementary information that can be best harnessed by combining both prediction sources. Extensive external validation and impact studies are imperative to further investigate the ability of the combined model.

Condition monitoring of premature babies in intensive care can be carried out using a Factorial Switching Linear Dynamical System (FSLDS) [15]. A crucial part of training the FSLDS is the manual

calibration

stage, where an interval of normality must be identified for each baby that is monitored. In this paper we replace this manual step by using a classifier to predict whether an interval is normal or not. We show that the monitoring results obtained using automated calibration are almost as good as those using manual calibration.

This paper proposes an algorithm for discovering temporal patterns, represented in the Simple Temporal Problem (STP) formalism, that frequently occur in a set of temporal sequences. To focus the search, some initial knowledge can be provided as a seed pattern by a domain expert: the mining process will find those frequent temporal patterns consistent with the seed. The algorithm has been tested on a database of temporal events obtained from polysomnography tests in patients with Sleep Apnea-Hypopnea Syndrome (SAHS).

The Knowledge-Based Temporal Abstraction (KBTA) method is a well-established mechanism for representing and reasoning with temporal information. Implementations to date have been somewhat heavyweight, however, and custom tools are typically required to build abstraction knowledge and query the resulting abstractions. To address this shortcoming, we created a lightweight method that allows users to rapidly specify KBTA-based temporal knowledge and to immediately construct complex temporal queries with it. The approach is built on the Web Ontology Language (OWL), and its associated rule and query languages, SWRL and SQWRL. The method is reusable and can serve as the basis of a KBTA implementation in any OWL-based system.

We report on work in progress on the development of SWEETInfo (Semantic Web-Enabled Exploration of Temporal Information), a tool for querying and visualizing time-oriented clinical data. SWEETInfo is based on an open-source Web-based infrastructure that allows clinical investigators to import data and to perform operations on their temporal dimensions. The architecture combines Semantic Web standards, such as OWL and SWRL, with advanced Web development software, such as the Google Web Toolkit. User interaction with SWEETInfo creates OWL-based specifications of (1) data operations, such as filtering, grouping, and visualization, and (2) data pipelines for data analyses. Both of these can be shared with and adapted by other users via the Web. Our system meets the functional and nonfunctional specifications derived from the use cases. We will demo how SWEETInfoprovides non-technical users the ability to interactively define data pipelines for such complex temporal analyses.

Decision-making, care planning and adaptation of treatment are important aspects of the work of clinicians, that can clearly benefit from IT support. Clinical Practice Guidelines (CPG) languages provide formalisms for specifying knowledge related to such tasks, such as decision criteria and time-oriented aspects of the patient treatment. In these CPG languages, little research has been directed to efficiently deal with the integration of temporal and resource constraints, for the purpose of generating patient tailored treatment plans, i.e. care pathways. This paper presents an AI-based knowledge engineering methodology to develop, model, and operationalize care pathways, providing computer-aided support for the planning, visualization and execution of the patient treatment. This is achieved by translating time-annotated Asbru CPG’s into temporal HTN planning domains. The proposed methodology is illustrated through a case study based on Hodgkin’s disease.

Clinical guidelines contain recommendations based on the best empirical evidence available at the moment. There is a wide consensus about the benefits of guidelines and about the fact that they should be deployed through clinical information systems, making them available during consultation time. However, one of the main obstacles to this integration is still the interaction with the electronic health record. In this paper we present an archetype-based approach to solve the interoperability problems of guideline systems, as well as to enable guideline sharing. We also describe the knowledge requirements for the development of archetype-enabled guideline systems, and then focus on the development of appropriate guideline archetypes and on the connection of these archetypes to the target electronic health record.

For clinical processes, meaningful variations may be related to care performance or even the patient survival. It is imperative that the variations be predicted timely so that the patient care “journey” can be more adaptive and efficient. This study addresses the question of how to predict variations in clinical processes. Given the assumption that a clinical case with low appropriateness between its specific patient state and its’ applied medical intervention is more likely to be a variation than other cases, this paper proposes a method to construct an appropriateness measure model based on historical clinical cases so as to predict such variations in future cases of clinical processes. The proposed method is demonstrated on a real life data set from the Chinese Liberation Army General Hospital. The experimental results confirm the given assumption and indicate the feasibility of the proposed method.

This paper describes a novel methodological approach to identifying inconsistencies when concurrently using multiple clinical practice guidelines. We discuss how to construct a formal guideline model using Constraint Logic Programming, chosen for its ability to handle relationships between patient information, diagnoses, and treatment suggestions. We present methods to identify inconsistencies that are manifested by treatment-treatment and treatment-disease interactions associated with comorbidity. Using an open source constraint programming system (ECLiPSe), we demonstrate the ability of our approach to find treatment given incomplete patient data and to identify possible inconsistencies.

Computer Interpretable Guidelines (CIG) have largely contributed to the simplification and dissemination of clinical guidelines. However, the formalization of CIG contents, especially care actions, is still an open issue. Actually, this information, which is the heart of the guideline, is still expressed as free text and therefore prevents the development of intelligent tools for assisting physicians defining treatments. In this paper, we introduce a framework for formalizing care actions using natural language processing techniques, Semantic Web technologies and medical standards.

Ontology-based modeling of Clinical Practice Guidelines (CPG) is a well-established approach to computerize CPG for execution in clinical decision support systems. Many CPG computerization approaches use the Web Ontology Language (OWL) to represent the CPG’s knowledge, but they do not exploit its reasoning services to execute the CPG. In this paper, we present our CPG execution approach that leverages OWL reasoning services to execute CPG. In this way, both CPG knowledge representation and execution semantics are maintained within the same formalism. We have developed three different OWL-based CPG execution engines using OWL-DL, OWL 2 and SWRL. We evaluate the efficacy of our execution engines by executing an existing OWL based CPG. We also present a comparison of the execution capabilities of our three CPG execution engines.

This paper describes a knowledge acquisition tool for translating a guideline recommendation into a computer-interpretable format. The novelty of the tool is that it is addressed to the domain experts, and it helps them to disambiguate the natural language, by decomposing the recommendation into elements, eliciting tacit and implicit knowledge hidden into a recommendation and its context, mapping patient’s data, available from the electronic record, to standard terms and immediately testing the formalised rule using past cases data.

A difficulty with part-of-speech (POS) tagging of biomedical text is accessing and annotating appropriate training corpora. The latter may result in POS taggers trained on corpora that differ from the tagger’s target biomedical text. In such cases where training and target corpora differ tagging accuracy decreases. We present a POS tagger that is more accurate than two frequently used biomedical POS taggers (Brill and TnT) when trained on a non-biomedical corpus and evaluated on the MedPost corpus (our tagger: 81.0%, Brill: 77.5%, TnT: 78.2%). Our tagger is also significantly faster than the next best tagger (TnT). It estimates a tag’s likelihood for a token by combining prior probabilities (using existing methods) and token probabilities calculated in part using a Naive Bayes classifier. Our results suggest that future work should reexamine POS tagging methods for biomedical text. This differs from the work to date that has focused on retraining existing POS taggers.

Information about medications is critical in supporting decision-making during the prescription process and thus in improving the safety and quality of care. The Summary of Product Characteristics (SPC) represents the basis of information for health professionals on how to use medicines. However, this information is locked in free-text and, as such, cannot be actively accessed and elaborated by computerized applications. In this work, we propose a machine learning based system for the automatic recognition of drug-related entities (active ingredient, interaction effects, etc.) in SPCs, focusing on drug interactions. Our approach learns to classify this information in a structured prediction framework, relying on conditional random fields. The classifier is trained and evaluated using a corpus of a hundred SPCs. They have been hand-annotated with thirteen semantic labels that have been derived from a previously developed domain ontology. Our evaluations show that the model exhibits high overall performance, with an average F

1

-measure of about 90%.

SNOMED is a large description logic based terminology for recording in electronic health records. Often, neither the labels nor the description logic definitions are easy for users to understand. Furthermore, information is increasingly being recorded not just using individual SNOMED concepts but also using complex expressions in the description logic (“post-coordinated” concepts). Such post-coordinated expressions are likely to be even more complex than other definitions, and therefore can have no pre-assigned labels. Automatic verbalisation will be useful both for understanding and quality assurance of SNOMED definitions, and for helping users to understand post-coordinated expressions. OntoVerbal is a system that presents a compositional terminology expressed in OWL as natural language. We describe the application of OntoVerbal to SNOMED-CT, whereby SNOMED classes are presented as textual paragraphs through the use of natural language generation technology.

In literature-based creative knowledge discovery the goal is to identify interesting terms or concepts which relate different domains. We propose to support this cross-context link discovery process by inspecting outlier documents which are not in the mainstream domain literature. We have explored the utility of outlier documents, discovered by combining three classification-based outlier detection methods, in terms of their potential for bridging concept discovery in the migraine-magnesium cross-domain discovery problem and in the autism-calcineurin domain pair. Experimental results prove that outlier documents present a small fraction of a domain pair dataset that is rich on concept bridging terms. Therefore, by exploring only a small subset of documents, where a great majority of bridging terms are present and more frequent, the effort needed for finding cross-domain links can be substantially reduced.

The prediction of diagnosis codes is typically based on free-text entries in clinical documents. Previous attempts to tackle this problem range from strictly rule-based systems to utilizing various classification algorithms, resulting in varying degrees of success. A novel approach is to build a word space model based on a corpus of coded patient records, associating co-occurrences of words and ICD-10 codes. Random Indexing is a computationally efficient implementation of the word space model and may prove an effective means of providing support for the assignment of diagnosis codes. The method is here qualitatively evaluated for its feasibility by a physician on clinical records from two Swedish clinics. The assigned codes were in this initial experiment found among the top 10 generated suggestions in 20% of the cases, but a partial match in 77% demonstrates the potential of the method.