Temporal Information Systems in Medicine

Free download. Book file PDF easily for everyone and every device. You can download and read online Temporal Information Systems in Medicine file PDF Book only if you are registered here. And also you can download or read online all Book PDF file that related with Temporal Information Systems in Medicine book. Happy reading Temporal Information Systems in Medicine Bookeveryone. Download file Free Book PDF Temporal Information Systems in Medicine at Complete PDF Library. This Book have some digital formats such us :paperbook, ebook, kindle, epub, fb2 and another formats. Here is The CompletePDF Book Library. It's free to register here to get Book file PDF Temporal Information Systems in Medicine Pocket Guide.

HPA s. Humatics s. MedBrains s.

Temporal Information Systems In Medicine 1st Edition

In order to assess a diagnosis, in fact, the physician finds out the clinical history of the patient, composed usually by previous pathologies, therapies, and symptoms, the patient narrates; this information completes data collected directly from the patient, i. In supporting the computer-based storage and retrieval of this information, we need to represent the temporal dimension of clinical data. The considered temporal dimension is usually the valid time, i. Different units of measure allow one to represent different granularities [Snodgrass95,Bettini98a]. Granularity is present in several application domains as, for example, geographic information systems, planning and scheduling, medical information systems, office information systems, real-time systems, natural language processing [Brusoni99,Chittaro00,Combi97, Koubarakis99,Maiocchi92,Montanari96,Snodgrass95,Staab99].

In general, the temporal dimension of information can be expressed with different time granularities — e. Moreover, in querying the system about the stored temporal information it is usual to use different granularities with no relation to the ones used when storing data. Supporting different granularities involves several research topics, as the representation of multiple granularities and the modeling and querying of temporal data given at different granularities. Among them, some works propose different frameworks allowing the formal definition of multiple granularities and of relationships among them [Bettini98a, Clifford88, Goralwalla01, Montanari92, Montanari96].

Other research efforts focus on granularity and calendars [Snodgrass95]. Mainly, a granularity is represented as the partitioning of the basic time line. Clifford and Rao [Clifford88] introduce a general structure for time domains called "temporal universe" which consists of a totally ordered set of granularities e. Operations are defined on a temporal universe, which basically convert different anchored times to a common finer granularity before carrying out the operation.

Wang et al. Montanari et al. In [Goralwalla01], granularity is modeled as a special kind of unanchored temporal primitive that can be used as a unit of time. Granularities are accommodated within the context of calendars and granularity conversions are presented and discussed in terms of unanchored durations of time.

Bettini et al. The Artificial Intelligence in Medicine AIM community is interested in the broader issues involved in reasoning about time-oriented data: in particular, the implications of temporal reasoning on clinical tasks such as diagnosis, monitoring, and longitudinal guideline-based care. Temporal Information Systems in Medicine introduces the engineering of information systems for medically-related problems and applications. The chapters are self-contained with pointers to other relevant chapters or sections in this book when necessary.

Time is of central importance and is a key component of the engineering process for information systems. We noticed that the operation date can be inconsistently treated as either postoperation day POD 0 or POD 1 in different notes.


We used a heuristic to determine how the operation date is recorded in a specific note. For example, if the patient is discharged on POD 7 and we know the discharge date, then it is clear how the operation date is recorded. The detailed rules are categorized and described with examples in table 1 B. We outlined some of the rules that extract TLINK events in a single sentence, which were designed after manually examining the annotations in the training corpus. We also show an example sentence that matches each of the rules.

We also implemented rules to relate and classify events and time expressions, which occur beyond sentence boundaries. These rules are based on either knowledge derived from the inherent structure of the clinical note or on sound linguistic discourse principles such as co-reference.

Spatial Temporal Information Systems: An Ontological Approach using STK®

These rules significantly improve the recall of the system. We briefly discuss some of the rules below:. Thus, it is impractical to train a machine learner on all the potential pairs. Also, the imbalanced class distribution due to the preponderance of negative pairs would degrade the classification performance. To avoid this problem, we used a simple approach to select potential TLINK pairs: 1 pairs from our previous rules in Rule-based framework ; 2 any possible pairs within the same sentence. The union of 1 and 2 will be our potential pairs. These potential pairs were used in both the training and test phases.

We use three sources for feature generation. Second, attributes extracted from the Clinical Narrative Temporal Relation Ontology CNTRO 27 , 28 —that is, concept definitions about time eg, temporal instant, temporal interval and relations between the time concepts eg, prior to, before, earlier, after, until, subsequent, overlap, during, etc. CNTRO, as domain ontologies, have been used in information technology for providing semantic definitions of a particular domain, which enable automated agents to perform queries intelligently and infer new knowledge.

Third, features listed in the following were also found to be discriminative for detecting temporal relations.

Section head : admission, history, hospital course, discharge. Adjacent events : events close to involved events, including their types, POS and event names. Relative temporal relations : temporal relations found among current events and other events. For example, most events in the history section occur before admission, and events in the hospital course occur after admission and before discharge. Normalized time stamps make temporal expressions more consistent. POS can help distinguish events of different types.

About this book

The collection reader is designed to read data. The aggregate rule generator generates diverse rules and relevant annotations. Table 2 summarizes the results of the evaluation on the test dataset. We note that the recall of the system was far lower than the precision. It may be possible to further improve the F-measure by optimizing the recall—precision trade-off. Table 3 shows the evaluation performance of all runs on the test set.

All three runs produced similar F-measures but run 2 had a slightly higher F-measure wing to higher recall. The accuracy scores for all runs were similar. We report the runs for the hybrid and the rule-based system in table 4. Events used in this i2b2 temporal resolution challenge were broader than conventional event definitions in the i2b2 medical concept definition.

In order to capture conventional events ie, clinical concepts and other types of events, well-defined expert knowledge and ontologies might be useful features in a machine learning model to further improve performance. We experimented with a variety of features to improve the performance of the system. Lexical features, semantic types, and POS tags, were useful as compared with other features.

We did not use post-processing rules to refine the system output, which might have further improved performance. Also, there was a considerable room to optimize for the F-measure by improving recall—precision trade-off, since the recall was far lower than the precision. We believe that manually curated comprehensive rules to cover most TIMEX3 expressions, heuristic rules to catch relative date and time, and the systematic value normalization process produce this high performance.

The explicit time expressions were easy to find, but relative time expressions, which require inference or reasoning, were often difficult to catch. Inconsistency of relative time annotation in the gold standard makes the problem even more difficult. The hybrid approach produced balanced precision and recall while the rule-based system produced comparatively high precision but low recall.

The rule-based approach identifies specific sets of TLINK pairs, leading to higher precision but much lower recall than the hybrid approach.

Upcoming Events

However, added coreference resolution did not show significant improvement, perhaps owing to the prevalence of false-positive coreferences links. POS and tense features also did not contribute much to performance enhancement. A few observations have been noticed for future improvements. TLINKs generated by the system can be expanded. Finally, assigning different weights to the rule-based features in machine learning training might improve the precision. A comprehensive temporal information extraction and classification system was designed and tested in the i2b2 NLP challenge.

A rule-based approach could be effectively applied to extract time expression patterns. The machine learning approach was attempted for Events and TLINK tasks, as these annotations require a variety of factors. We could also take advantage of knowledge engineering into a machine learning part, such as simple TLINK rules as part of machine learning features.

Overview of the functions of the cerebral cortex (video) | Khan Academy

All authors conceived and designed the experiments. All authors contributed to the writing of the manuscript and revisions. All authors reviewed and approved the final manuscript. Oxford University Press is a department of the University of Oxford. It furthers the University's objective of excellence in research, scholarship, and education by publishing worldwide.

Sign In or Create an Account. Sign In. Advanced Search. Article Navigation.