Amiajnl4036 549.554

Linguistic approach for identification of
medication names and related information in
clinical narratives
JAMIAdoi: 10.1136/jamia.2010.004036 Updated information and services can be found at: References
This article cites 24 articles, 10 of which can be accessed free at: Email alerting
Receive free email alerts when new articles cite this article. Sign up in the box at the top right corner of the online article.
To order reprints of this article go to: Journal of the American Medical Informatics Association Linguistic approach for identification of medicationnames and related information in clinical narratives prescriptions filled at a given hospital, and not at Background Pharmacotherapy is an integral part of any other places. Nevertheless, in scientific literature and medical care process and plays an important role in the clinical records, information on medication is buried medical history of most patients. Information on in a mass of narrative text. To avoid this information medication is crucial for several tasks such as becoming lost, we need specific tools and methods to pharmacovigilance, medical decision or biomedical Objectives Within a narrative text, medication-related information can be buried within other non-relevant data.
Specific methods, such as those provided by text mining, Natural language processing (NLP) and text mining must be designed for accessing them, and this is the tools allow us to access relevant information within narrative documents. They perform parsing Methods The authors designed a system for analyzing and analysis of unstructured documents in order to Correspondence toDr Thierry Hamon, Laboratoire narrative clinical documents to extract from them localize the data searched for. For instance, medi- medication occurrences and medication-related cation-related information may consist of a drug information. The system also attempts to deduce name, dose, frequency, duration, status and mode medications not covered by the dictionaries used.
of administration. Detection of medication names Results Results provided by the system were evaluated is mostly dictionary-based: a nomenclature of drugs within the framework of the I2B2 NLP challenge held in is used and their occurrences are detected in 2009. The system achieved an F-measure of 0.78 and biomedical literature16e18 or in clinical records.18e22 ranked 7th out of 20 participating teams (the highest It has been observed that the quality of such F-measure was 0.86). The system provided good results nomenclatures must be controlled,19 as it has for the annotation and extraction of medication names, a direct impact on the quality of results. Approxi- their frequency, dosage and mode of administration mate matching was proposed as a method of drug (F-measure over 0.81), while information on duration and name recognition20 and shown to improve extrac- reasons is poorly annotated and extracted (F-measure tion results compared with dictionary-based exact 0.36 and 0.29, respectively). The performance of the matching. Other methods aim to identify new drug system was stable between the training and test sets.
names through naming conventions23 24 orcontextual rules.25 Previous work has also addressedthe extraction of drug-related information. The firststudy of this kind29 focused on extracting drug names, a process improved by considering their Pharmacotherapy is an integral part of any medical context: dosage information allowed disambigua- care process and plays an important role in the tion of medication names. Extraction of drug- medical history of patients. Acquiring accurate related data was also considered separately by medication-related data is an important task. It is research26 30 31 33 and commercial27 28 systems. The useful for improving patient safety and the quality of individual healthcare. Thus, pharmaco- F-measures of 0.27 to 0.90 depending on the cate- vigilance1 2 aims to prevent adverse drug effects.
gory of data: they are difficult to compare, as no Medical3e6 and pharmacological7 decision systems common ‘gold standard’ has been used. Notice that are oriented towards prescription assistance: they applying such methods to database entries32 improve medication reconciliation and reduce significantly improves results (up to F-measure of errors caused by misinterpretation of handwritten 0.98). Common difficulties are related to incom- orders, incorrect doses, etc. With translational pleteness of drug lexica19 20 26 and ambiguous drug medicine, a better connection between clinical healthcare and biomedical research is established,8 9while the scientific literature helps biologistscarrying out research on new drugs.8 10 Knowledge about drugs is thus necessary, and medication- In this work, we proposed to extract medication related information (eg, dosage, mode, time) names and medication-related information, such as provides even more precise knowledge.
those underlined in the excerpt from box 1, from Large-scale observation of data is necessary and narrative discharge summaries. We proposed to go becomes possible through extensive study of beyond the state-of-the-art and to address the scientific literature and patient records. For this, following problems: (1) recognizing new medica- structured data on prescriptions can be exploited,11 12 tion names; (2) disambiguating medication names; but it has been observed that this type of data is often (3) detecting contexts where drug names do not incomplete or out of date13e15 and limited to J Am Med Inform Assoc 2010;17:549e554. doi:10.1136/jamia.2010.004036 (http://www.cdc.gov/nchs/data/nhanes/nhanes_01_02/rxq_rx_ Box 1 Excerpt from a narrative discharge summary with b_doc.pdf). In addition, among the drug names, we distin- medication-related information (underlined) to be guished 108 ambiguous entries that also referred to biologicalcharacteristics of patients (eg, ‘iron’). They were assigned a specific status.
Snomed International37 proved to be an efficient and user- The patient is currently off diuretics at this time. Daily weights friendly source for NLP processing38; we used the 45 898 terms should be checked and if her weight increases by more than 3 from the Diagnosis and Morphology axes for the detection of pounds Dr Bockoven should be notified. The patient was also reasons. A total of 476 terms corresponding to patient problems started on calcitriol given elevation of parathyroid hormone.
in the training set were added to this resource.
Cardiovascular: Rate and rhythm: The patient has a history ofatrial fibrillation with a slow ventricular response. The patient was started on metoprolol 12.5 mg p.o. q.6 h. for rate control, We exploited NegEx (http://www.dbmi.pitt.edu/chapman/NegEx.
however, this dose was decreased to 12.5 mg p.o. twice a day, html) to detect negation and reduce the number of false positives.
given some bradycardia on her telemetry. The patient was also Negation markers consist of pre-negation (eg, ‘deny’, ‘cannot’, started on Flecainide 75 mg p.o. q.12 h. She will continue on ‘without’) and post-negation (eg, ‘free’, ‘was ruled out’). Some these two medications upon discharge.
additional markers were added, making a total of 284 markers.
We also evaluate our results through the common framework Given the very small number of annotated documents available of the I2B2 NLP medication challenge held in 2009. This for tuning the systems (n¼17), we used a rule-based approach: framework allows comparison between several automatic learning algorithms would require a larger training set. The systems and NLP methods. We consider the categories targeted system designed performs information extraction by three main by the challenge (table 1): dosage, frequency, duration, mode of steps: pre-processing, processing and post-processing (figure 1).
administration and reason for prescription, as well as the The processing step is built on the Ogmios platform39 suitable semantic relations between them. The NLP system designed for the processing and annotation of large datasets and tunable exploits nomenclatures and terminologies, contextual rules and to specialized areas. For pre- and post-processing steps, we shallow parsing. Concurrent annotations may be proposed for developed specific modules to disambiguate and select the rele- a given token and then disambiguated.
vant annotations, to compute semantic relations, etc.
Input discharge summaries are full-text documents. To prepare Discharge summaries were provided by Partners Healthcare: them for the NLP tools, we first attempted to split them into they were written in English and were prepared and deidenti- sections and lists through the use of specific parsers and section fied.34 A total of 1249 documents were used, split into training markers (eg, ‘discharge meds’, ‘history of present illness’, ‘family (n¼696) and test (n¼553) sets. Within the training set, only 17 history’, ‘physical examination’). As these markers were not documents were manually annotated and provided as an illus- standardized across the discharge summaries, we supplemented them with contextual heuristics (eg, ‘uppercase characters’, ‘punctuation’). Contextual heuristics were also used for the detection of lists and enumerations. Documents were then We used two types of resource for the annotation (a total of converted into XML format, with section and list tags. This step 290 243 entries): drug nomenclature and pathology terms.
also computed offset data (number of lines and tokens) for the We created a medication list containing 243 869 entries mainly provided by RxNorm.35 36 This list has three main limitations: the entries can be composed entries, common English words are The processing step was dedicated to linguistic and semantic used, and it is not exhaustive. To address the first two limita- annotation: we assigned semantic categories to textual entities tions, entries were split and cleaned up to remove ones such as and provided their semantic contexts. Our system supports ‘golden eye’, ‘ginger’, ‘bermuda’, ‘vital’, ‘Marihuana’ or ‘water’.
concurrent annotations, while semantic contexts allow perfor- As for the third limitation, the list was enriched with drug mance of their disambiguation. The annotation process was names found in the training set. Moreover, we used therapeutic performed through the following main modules: classes and groups of medications, found on the CDC website < The named entity recognizer (NER) identified frequency, dosage, duration and mode of administration. For this, specific Examples of the targeted categories of information on drugs, automata were implemented as regular expressions (box 2).
as extracted from the excerpt given in box 1 (except for the values of the Preliminary disambiguation was performed in order to (1) select the longest match and avoid multiple annotations within nested strings (eg, ‘ten minutes’ was recognized as both Calcitriol, metoprolol, flecainide, these two frequency and duration entities), and (2) merge adjacent named entities of the same semantic type: ‘q6h’ and ‘prn’ were first recognized individually as frequency and then merged.
< Word and sentence segmentation was then performed.
Having this step after the NER module allows the 7-day course, 35 days, # for 7 days, 5 more disambiguation of characters, such as punctuation, dashes, slashes, etc, that are widely used within discharge summaries Elevation of parathyroid hormone, rate control often altering the segmentation process.
J Am Med Inform Assoc 2010;17:549e554. doi:10.1136/jamia.2010.004036 System architecture for the extraction of medication-related information and for establishing dependencies among the annotations. POS, < Term and semantic tagging was used to detect drugs and reasons. The system also performed the longest match and Some medication names (eg, iron) are ambiguous: they can merged adjacent medication terms: in ‘singulair (montelu- correspond to biological characteristics or drugs. They were first kast)’, the two drugs correspond to two separate entities in assigned a specific semantic tag. Then, if they occurred in listings or medication-related sections (box 3, example ii), their tags < Term extraction was performed with YATEA40: it organizes were modified into drug names. Otherwise, they were not the identification of missing medication names and reasons during the post-processing step. Part-of-speech tagging andlemmatization were performed with Genia.41 Detection of negative contexts and allergies Our system deals with several contexts where medication names In charge of several treatments on drugs and related information do not correspond to prescriptions (box 3, examples iiiev). In and computing dependency relations, the post-processing step example iii, drug names are related to allergies: a specific module exploits annotations from the processing step.
detects this relation and such drugs are not extracted. In J Am Med Inform Assoc 2010;17:549e554. doi:10.1136/jamia.2010.004036 Box 2 Excerpts from four regular expressions for the Box 3 Examples of textual data to be processed extraction of mode (1, 2) and frequency (3,4) information i. Heme. Anemia workup. Iron 49, TIBC 256, B12 555, folate Pipe and parentheses allow disjunction of strings, while square normal, ferritin 102, reticulocyte 7.9, and Epogen level 19.
brackets allow disjunction of characters, \n means end of lines ii. HOME MEDS: methadone 20 bid, imdur 120 bid, hydral taking and ? means an optional string or character, and back slash (\) is 25 bid, lasix 20 bid, coumadin, colace, iron, nexium 40 bid, used to despecialize characters. Strings with the $ symbol indi- cate variables: they are described in the second part of this iii. ALLERGY: prednisone, penicillins, tamsulosin, simvastatin, example. The first regular expression detects entities such as to each nostril, under the tongue, by mouth; the second expression iv. . did not require medications for abdominal pain detects nasal, drip, inhaled, subq; the third expression detects v. INR’s will be followed by Coumadin clinic; insulin-dependent once a day, two per day, 2 per day; and the last expression vi. . Methadone 20 bid, Ofloxacin 200 mg p.o. q 12, Insulin 1. ($adv)($sep)?($det|$adj)?($sep)?($anatomy) 2. (subcutaneously|subcutaneous|subcutane|subcu|subquta- vii. . history of atrial flutter controlled on Amiodarone neously|subqutaneous|subqtane|subq|inhaled|inh|iv|intra- viii. . started on calcitriol given elevation of parathyroid venously|intravenous|intraven|neb|drip|injection|inj|im| intramuscularly|intramuscular|intramusc)s? ix. . started on metoprolol 12.5 mg p.o. q.6 h. for rate control 3. ($number)($sep)?(a|per|$det)($sep)?d(ay)? x. . should be switched to Toprol as her blood pressure xi. She was initially diuresed with IV Lasix.
adv ¼ (through|per|by|with|via|in|to|under) xii. packed red blood cells, red blood transfusions, red blood cell, autologous red blood cells, blood, autologous blood, prbcs, xiii. ., HCTZ 25 mg PO QD, Norvasc 10 mg PO QD, Pavachol number ¼ ([0e9]+|once|one|two|twice|three|four) as the majority are not relevant for the reason category.
example iv, drugs occur in negative context, detected with Combining noun phrases with 52 contextual patterns (‘for’, a NegEx-inspired algorithm: it exploits the proximity of pre- and ‘given’, ‘controlled on’, .) allows them to be constrained post-negation markers. In example v, drug names appear in other contexts: within names of diseases and institutions. This situ-ation is processed through an extension of NegEx resources: proximity of terms such as ‘clinic’, ‘dependent’ or ‘deficiency’ Evaluation was performed by organizers of the challenge: allows these drugs to be not detected as prescriptions.
automatically generated results are compared with the 251documents from ground truth according to the protocol described by Uzuner et al.34 The main evaluation measure is the With the rapid evolution of therapeutic research, new drugs F-measure computed for exact and inexact matches.
appear,24 but drug nomenclatures cannot keep pace. We proposea novel method for a more exhaustive identification of newdrugs. The main indication we rely on is that drugs often occur in specific semantic contexts together with medication-related Table 2 presents results for our system in terms of F-measure F, information (box 3, example vi). The corresponding semantic precision P and recall R. The global exact-match F-measure was pattern is: m do mo? f, where medication name, m (‘metha- 0.78. Within the challenge framework, our system ranked 7th done’, ‘ofloxacin’, .), is followed by dosage, do (‘20’, ‘200 mg’, out of 20 participating systems. The system generated good ‘12 units’), possibly followed by administration mode, mo (‘p.o.’, results (F-measure over 0.81) for four categories (drug, dosage, ‘subcu’), and followed by frequency, f (‘twice daily’, ‘q 12’, ‘q p.
frequency, mode). The two remaining categories (duration and m.’). If all entities (do, mo and f) except the first one are reasons) were extracted with lower performance (F-measure 0.36 recognized, we infer that the first entity is a new drug name. We and 0.29, respectively). Exact match performed slightly better additionally check whether this entity is a stopword and than inexact match. Within the interval of medication occur- whether its ending is typical of drug endings (eg, ‘ine’, ‘one’, rences,2 11 6 the mean number of medications per document was ‘ase’, ‘ate’, ‘cin’, ‘rin’).
35.6. Only one document has no mention of drugs.
Reasons are identified by two approaches: (1) the use of termi- As shown in table 3, the performances obtained on the training nological resources; (2) the use of noun phrase extraction (n¼17) and test sets were comparable. Stability of the system together with reason markers. The first approach applies only was a positive result, especially given the very small set of Snomed International terms and patient complaints. The second annotated training data. We assume that the system may be approach allows the sensitivity of this vocabulary to be useful for the processing of other clinical records, or at least can increased through extraction of noun phrases. However, be easily adapted. Overall, it allows processing of narrative exploiting all these noun phrases can be disastrous for precision, clinical documents and extraction of several medication-related J Am Med Inform Assoc 2010;17:549e554. doi:10.1136/jamia.2010.004036 Test set: performance of the system for exact and inexact The most commonly recurring problem is associated with reason detection: in examples xexi (box 3), our system wrongly extracts ‘blood pressure’ as the reason for administration of ‘toprol’ and ‘diuresed’ as the reason for ‘IV Lasix’.
We found several cases of false negatives among drug names: 1. Ambiguous drug names (eg, ‘iron’, ‘statin’, ‘blood’, ‘magne- sium’, ‘glucose’) corresponding to administered products but not occurring in expected positive contexts 2. Terms such as ‘fluids’, ‘agents’ or ‘medication’ that we 3. Some classes of drugs (eg, ‘antianginal therapy’, ‘pressure medications’) missing from our resources 4. New drug names (eg, ‘vp-16’, ‘ducolox’, ‘vasopressor’, data with good performance, making the tedious manual ‘guqifenesin’) that did not occur within expected semantic The core platform for NLP processing relies on standard NLP 5. Misspellings and abbreviations (eg, ‘aspirin325’, ‘hep.’) steps (NER, tokenization, part-of-speech (POS) tagging, 6. Pronominal phrases (eg, ‘these medications’) lemmatization), but also on specific modules designed for this Blood products remain difficult to detect, as they seldom appear within listings but mainly in narrative sections. More- over, their nomenclature is not standardized, and various phrases dallows disambiguation of several cases where punc- tuation does not stand for sentence boundaries. Implementation are used to refer to a blood transfusion (box 3, example xii). An of the tools and modules used within the Ogmios platform also extension of semantic patterns may be helpful: ‘required’ and facilitates communication between them, making the manage- ‘one unit of’ are valuable indicators that ‘blood’ was adminis- ment of linguistic and semantic annotations easier.39 In addi- tered in the phrase ‘required one unit of blood during her tion, the integration of modules with regular expressions is also easy and does not conflict with other modules and tools.
An additional analysis was performed of the module for detec- An analysis of these results was performed on 26 randomly tion of new medication names. It extracted 49 occurrences, 15 of selected discharge summaries from the ground truth (10%).
which are real drug names (precision¼30%), such as ‘pendalol’, Within this set, a total of 729 medication annotations were ‘lithium’, ‘permatol’, ‘levoxine’ or ‘pavachol’ (box 3, example xiii).
analyzed: 380 were identical and 47 overlapped with the refer- The precision is low, but it should be noted that we used it for ence annotations. In the remaining annotations, at least one enriching an already large drug nomenclature (over 240 000 category was different. This difference may correspond to false- entries) and it missed only a few occurrences (such as ‘guqife- positive (n¼70) or false-negative (n¼162) annotations.
nesin’). A more thorough evaluation of this module is ongoing.
We found only 16 (2%) false positives due to the extraction of Other false negatives correspond to missed drug-related wrong medication names, which attests to the quality of the information. It is seldom due to the incompleteness of the drug lexicon. However, a few entries (ie, ‘acute phase reactant’, defined rules, but to wrong computation of dependency rela- ‘haemophilus influenzae’, ‘chewable’) remained that were tions. Syntactic parsing42 43 may be helpful for this.
wrongly considered as drugs. The quality of medication lexica isa common problem19 31: with the original RxNorm, the F-measure falls to 40.73%. Early in our experience, we observed We have described a system developed for the annotation and this fact and manually removed a large number of entries.
extraction of medication-related information from narrative Nevertheless, additional filtering is required. It cannot be done discharge summaries. We looked at this task as an annotation using a vocabulary of common English words, as in Sirohi and and annotation-disambiguation problem. Specific semantic Peissig,19 because nearly all these entries are relevant to the resources were exploited in a rule-based approach. We also medical area: cleaning them up would instead require additional proposed a novel module for detection of new medication names manual work or contextual rules. Another category of noise through the exploitation of semantic patterns. Global perfor- among the extracted drugs is related to ambiguous medication mances of our system (F-measure 0.78) rate it 7th among the 20 names that escaped our attention or for which the context is not participants of the I2B2 challenge. Our system provides an F- indicative of their semantics. False positives within medication- measure of over 0.81 for extraction of medication names, their related information are often due to wrong semantic relations.
frequency, dosage and mode of administration; however, itperforms poorly with duration and reasons, which is also thecase for other participating systems.
Training set: performance of the system for exact and inexact Among the benefits are: improved duration extraction through exploitation of prepositional phrases; improved reason extraction with extended noun phrases; further evaluation of the module for deducing new medications; improved establish-ment of dependency relations between drug names and the Acknowledgments We are grateful to: the organizers of the I2B2 challenge for preparing and providing such an exciting framework for the evaluation of text mining systems; the anonymous reviewers for helpful and constructive comments; and J Am Med Inform Assoc 2010;17:549e554. doi:10.1136/jamia.2010.004036 Provenance and peer review Not commissioned; externally peer reviewed.
Cimino JJ, Bright TJ, Li J. Medication reconciliation using natural languageprocessing and controlled terminologies. Stud Health Technol Inform 2007:679e83.
WHO. The use of stems in the selection of international nonproprietary names (inn) forpharmaceutical substances. Technical report. Geneva: World Health Organization, 2006.
Segura-Bedmar I, Martinez P, Segura-Bedmar M. Drug name recognition and WHO. International drug monitoring: the role of national centers. Geneva, classification in biomedical texts. Drug Safety 2008;13:816e23.
Switzerland: World Health Organization, 1972.
Xu R, Morgan A, Das AK, et al. Investigation of unsupervised pattern learning FDA. Good pharmacovigilance practices and pharmacoepidemiologic assessement.
techniques for bootstrap construction of a medical treatment lexicon. Proceedings of Rockville, MD: Food and Drug Administration, 2005.
Pronovost P, Weast B, Schwarz M, et al. Medication reconciliation: a practical tool Gold S, Elhadad N, Zhu X, et al. Extracting structured medication event information to reduce the risk of medication errors. J Crit Care 2003;18:201 from discharge summaries. AMIA Annu Symp Proc 2008:237 Bates D, Leape L, Cullen D, et al. Effect of computerized physician order entry and Turchin A, Morin L, Semere L, et al. Comparative evaluation of accuracy of a team intervention on prevention of serious medication errors. JAMA extraction of medication information from narrative physician notes by commercial and academic natural language processing software packages. AMIA Annu Symp Teich J, Merchia P, Schmiz J, et al. Effects of computerized physician order entry on prescribing practices. Arch Intern Med 2000;160:2741 Jagannathan V, Mullett C, Arbogast J, et al. Assessment of commercial nlp Oren E, Shaffer E, Guglielmo B. Impact of emerging technologies on medication engines for medication information extraction from dictated clinical notes. Int J Med errors and adverse drug events. Am J Health Syst Pharm 2003;60:1447 Boussadi A, Bousquet C, Sabatier B, et al. Specification of business rules for the Evans DA, Brownlow ND, Hersh WR, et al. Automating concept identification in the development of hospital alarm system: application to the pharmaceutical validation.
electronic medical record: an experiment in extracting dosage information. Proc Imming P, Sinning C, Meyer A. Drugs, their targets and the nature and number of Iglesias JE, Rocks K, Jahanshad N, et al. Tracking medication information across drug targets. Nat Rev Drug Discov 2006;5:821 medical records. Proc AMIA Annu Fall Symp 2009:266 Rader D, Daugherty A. Translating molecular discoveries into new therapies for Xu H, Stenner S, Doan S, et al. MedEx: a medication information extraction system for clinical narratives. J Am Med Inform Assoc 2010;17:19 Hale R. Text mining: getting more value from literature resources. Drug Discov Today Shah AD, Martinez C. An algorithm to derive a numerical daily dose from unstructured text dosage instructions. Pharmacoepidemiol Drug Saf 2006;15:161 McDonald C, Tierney W. The medical gopher Hripcsak G, Friedman C, Alderson P, et al. Unlocking clinical data from narrative find, organize and decide about patient data. West J Med 1986;145:823 reports: a study of natural language processing. Ann Intern Med 1995;122:681 Poon E, Blumenfeld B, Hamann C, et al. Design and implementation of an application ¨, Solti I, Cadag E. Extracting medication information from clinical text.
and associated services to support interdisciplinary medication reconciliation efforts J Am Med Inform Assoc 2010;17:514e8.
at an integrated healthcare delivery network. J Am Med Inform Assoc Simon L, Wei M, Robin M, et al. Rxnorm: prescription for electronic drug information Manley H, Drayer D, McClaran M, et al. Drug record discrepancies in an outpatient RxNorm, a standardized nomenclature for clinical drugs. Technical report, National electronic medical record: frequency, type, and potential impact on patient care at Library of Medicine. Bethesda, Maryland, 2009. http://www.nlm.nih.gov/research/ a hemodialysis center. Pharmacotherapy 2003;23:231 Grant R, Devita N, Singer D, et al. Improving adherence and reducing medication ´ RA, Rothwell DJ, Palotay JL. SNOMED InternationaledThe systematized discrepancies in patients with diabetes. Ann Pharmacother 2003;37:962 nomenclature of human and veterinary medicine. College of American Pathologists – Thomsen L, Winterstein A, Søndergaard B, et al. Systematic review of the incidence American Veterinary Medical Association, Northfield, 1993.
and characteristics of preventable adverse drug events in ambulatory care. Ann ˆte´ RA. The SNOMED model: a knowledge source for the controlled terminology of the computerized patient record. Methods Inf Med Rindflesch T, Tanabe L, Weinstein J, et al. EDGAR: extraction of drugs, genes and relations from the biomedical literature. Pac Symp Biocomput 2008:517 Hamon T, Nazarenko A, Poibeau T, et al. A robust linguistic platform for efficient and ´rik C, Hofmann-Apitius M, Zimmermann M, et al. Identification of new drug domain specific web content analysis. In Proceedings of RIAO 2007 (electronic classification terms in textual resources. Bioinformatics 2007;23:264 Chen E, Hripcsak G, Xu H, et al. Automated acquisition of disease drug knowledge Aubin S, Hamon T. Improving term extraction with terminological resources. In: from biomedical and clinical documents: an initial study. J Am Med Inform Assoc Salakoski T, Ginter F, Pyysalo S, et al, eds. Advances in natural language processing (5th International Conference on NLP, FinTAL 2006), number 4139 in LNAI. Springer, Sirohi E, Peissig P. Stufy of effect of drug lexicons on medication extraction from electronic medical records. Pac Symp Biocomput 2005:308 Tsuruoka Y, Tateishi Y, Kim JD, et al. Developing a robust part-of-speech tagger for Levin M, Krol M, Doshi A, et al. Extraction and mapping of drug names from free biomedical text. LNCS 2005;3746:382e92.
text to a standardized nomenclature. AMIA Annu Symp Proc 2007:438 Charniak E. Immediate-head parsing for language models. In: Proceedings of the Chhieng D, Day T, Gordon G, et al. Use of natural language programming to extract 39th Annual Meeting of the Association for Computational Linguistics. 2001:124e31.
medication from unstructured electronic medical records. AMIA Annu Symp Proc Klein D, Manning CD. Accurate unlexicalized parsing. In: Proceedings of the 41st Meeting of the Association for Computational Linguistics, 2003:423 J Am Med Inform Assoc 2010;17:549e554. doi:10.1136/jamia.2010.004036

Source: http://natalia.grabar.perso.sfr.fr/publications/hamon-JAMIA2010.pdf