Combining Syntactic Information and domain-specific Lexical Patterns to Extract Drug-Drug Interactions from Biomedical texts

Editor's Notes

• #2: Ok, I am the next speaker. My name is Isabel Segura Bedmar from the University Carlos III of Madrid. I'm going to talk about an approach for the extraction of drug-drug interactions from biomedical texts. This approach combines syntactic information and a set of lexical patterns to extract these interactions from text. Of course, the method is not novel, but it is the problem of extraction drug interactions. Maybe after you have listened to me, you may are interested in working in this novel domain.
• #3: Ok, this is the typical outline /autline/. Firstly, I am going to explain /iks'plein/ why the extraction of DDI /didiai/ from texts is important. Secondly, I'm going to talk about the main approaches in biomedical relation extraction, Thirdly, I'm going to describe our hybrid approach and finally, I am going to present the results /'risolts/ and main conclusions.
• #4: So, what is a drug-drug interactiong? A DDI occurs /a'curs/ when a drug influences /in'fluensis/ the level or the activity of another drug. In other words /wuerds/, it's a reaction between two or more drugs taken concurrently.
• #5: Some interactions can be beneficial, in fact, it's very common to combine several drugs in order to obtain more effective drugs. For example, the combination of retrovirals like ritonavir and lopinavir achieves a more potent /po'tnt/ antiretrovial. /Sam interakshons can be benefishol, in fact, it's very kamon to com-bain sevral drags in order to obtein more iffektive drags. For exampol, the combinesion of re'trovirals like ritonavir and lopinavir achieves a more potent antiretroviral/.
• #6: But unfortunatly, many interactions are very dangerous. For example, aspirin and heparin taken concurrently can cause a bleeding. /Bat anfortunatli, meni interakshons are very den'yeras/. For igsam'pol, asprin and heparin teikin cancurrentli can co:s a bli:din/ Or, Aspirin and Acetazolamide can cause the death of the patient. Or, asprin and acetazolamide can co:s the dez of the peishent.
• #7: In fact, several studies have shown that drug interactions are a serious problem for the patient safety. In fact, sevral stadis hav shon dat drag interakshons are a sirias problem for the peishent seifti. In addition, drug interactions increase the healthcare costs. In adision, drug interakshons incri:s the helzker costs. Therefore, it's very important to develop resources for healthcare professionals in order to detect these drug interactions. Derefor, its very important to dive'lop risoursis for helzker professionals in order to ditekt thiis drug interakshons. j
• #8: So, hau do helzker proffesionasl avoid /'avoid/ drug-drug interarkshons
• #9: There are several resources that contain information about drug-drug interactions (for example, micromedex or drugbank databases /deitabesis/), but unfortunatly, they are not comprehensive, because many interactions are only reported in medical journal or in drug safety reports. Therefore, healthcare professionals have to spend a long time in reading them. /There are sevral risoursis that contein informashon abaut drag-drag interakshons, bat anfortunatli, dei are not compre'jensive, because meni interakshons are only reportid in medical yurnal or riports. Helzker professionals have to spend a long time in ri:din them/
• #10: For this reason, we think that Information Extraction can help to improve the early detection of drug interactions and to reduce the time spent by healthcare proffessionals on reading the drug safety reports and medical journals, /for dis ri'son, wi think that informeison extrakson can help to impruv the erli detekshon of drag interakshons and to ridus the taim spent by helzker profesionals on riiding the medical journals and riports/. For example, you can see this sentence “A possible interaction resulting in acute renal failure has been reported in a few subjects when indomethacin, a nonsteroidal anti-inflammatory agent, was given with triamterene” /for igsampol, yu can sii dat dis sentens “a posibol interakshon risalting in akiut rinal feilier has biin riportid in a fiu sabjects when indometazin, a nonsteroidal anti-inflamatori agent, was given with triamteren/ The final goal of our method is to identify the drugs (indomethacin and tramteren) and to detect the interaction between them. /the fainal gol of auer mezod is to identifai de drags (indometazin and tramteren) and to ditect the interakshon bituin dem/
• #11: In more detail /di'teil/, the objectives of this work are: First, we must study the main approaches for biomedical relation /rileshon/ extraction. Second, we need to build an annotated /a'noteitid/ corpus to evaluate /ivalueti/ the approach. Third, we are combining the resolution /risolusion/ of complex syntactic constructions /canstrakshons/ and a set of lexical patterns defined /difainid/ by a pharmacist in order to extract drug interactions.
• #12: The goal /gol/ of biomedical relation /rileishon/ extraction is to detect /di'tekt/ occurrences /a'kurensis/ of a predefined /pridifainid/ type of relationship between a pair /per/ of given entity types (like genes, proteins or drugs) in text. During the last years /yirs/, many works /wurks/ have addressed /a-drest/ this task, in particular, the extraction of protein interactions, however /hau'var/, no approach has been proposed /propoust/ to solve /solf/ the problem of extracting DDIs in biomedical texts.
• #13: Current /ka-rent/ approaches to biomedical relation extraction may be classified in four main categories: Linguistic approaches are based /beist/ on the application of linguistic technology /teknology/. Patterns approaches apply rules to detect the relationships. Approaches that use machine learning techniques /tekniks/. Of course, many works combine these approaches. In general, the approaches show poor results. In fact, the best system for PPI extraction in BioCreative workshop in last year achieve an f-measure of 30%. Few approaches have dealt with the complexity /'campleksity/ of biomedical sentences. However, language structures /strakchurs/ such as appositions /a'posishons/, coordinations and complex sentences are very common in the biomedical literature /'literachur/. In our case, for this first approximation for DDI extraction, we decided to combine shallow parsing and pattern matching.
• #14: Ok, Now, I'm goint to describe our proposal. First of all, we built an annotated corpus with DDI to evaluate our method.
• #15: First, I'm going to describe the creation of the first corpus annotated with DDI, and then, I'm goint to present our system for DDI extraction.
• #16: We are talking about the construction of a corpus for drug interactions and the developtment of a system for drug drug interaction extraction.
• #17: For building the corpus, we used the pharmacological database DrugBank that contains information about almost five thousand drugs. For bildin the corpus, wiyus the pharmacological databeis DragBank dat conteins informations about olmost faiv zausand drags.
• #18: In particular, for each drug, DrugBank also contains a text document with information about its drug interactions. For example, this text shows interactions for Heparin. In the last paragraph, the sentence: Digitalis, tetracyclines, nicotine, or antihistamines may partially counteract the anticoagulant action of heparin sodium. Contains several interactions with heparin sodium. We have used these text documents to build our corpus. /En particular, for ich drag, DragBank olso conteins a text document with information about its drug interakshons/ For example, this text shous the interakshons for heparin. In the last paragraf, the sentens: Digitalis, tetracyclines, nicotine, or antihistamines mei parshali kaunterakt di anticoagualnt acsion of heparin sodium. Contains sevral interakshons wiz heparin sodium/ Wi haev ju:st thi:s text documents to bild aur corpus.
• #19: Then, We used the MetaMap tool for parsing the texts. This tool provides shallow syntactic information and semantic information from the biomedical UMLS ontology. You can see an example:
• #20: First, MetaMap splits the text into sentences.
• #21: Then, MetaMap makes shallow parsing.
• #22: Also, Metamap performs tokenization
• #23: Finally, for each phrase, MetaMap tries to look for a concept in the UMLS ontology. For example, for the phrase 'Aspirin'. MetaMap found the UMLS concept with the identifier C0004057. Also, MetaMap provides semantic types to classify the phrases, for example, the phrase 'Aspirin' was classified by MetaMap as a pharmacological substance and as organic chemical.
• #24: Once /guans/ we have collected /kolektid/ and processed the corpus, we annotated the documents with the assistant of a pharmacist For example, in this case, we annotated the three interactions between Aspirin and these three drugs. Each drug interaction is represened in a DDI element that contains attributes for the name of the drugs that interact and its phrase id.
• #25: The corpus consists of almost six hundred documents and contains around of fifteen thousands drugs, and three thousand drug interactions. In a future, we would like to work with more pharmacists in order to assess the agreement among /aman/ several annotators /anoteitars/
• #26: Ok, I have just described the process of construction of the corpus. And, now, I'm going to explain our hybrid approach for the extraction of drug-drug interactions. .
• #27: You can see our system that consits of four main modules. I have just describe the first process for obtaining the syntactic and semantic information by the MetaMap tool. The second process tries ecognizes and classifies the drug names. The third process is an anaphora resolver. But in this presentation, I focuns on the fourth process for the extraction of DDI.
• #28: Regarding the drug name recognition, we used the semantic information provided by MetaMap in order to identify drug names.
• #29: The third process tries to resolve the pronominal and drug nominal anaphors occuring in the texts by a set of linguistic heuristics.
• #30: Now, we are going to explain the most important process to extract DDIs from texts.
• #31: The basic idea is to detect linguistic complex structures such as coordinations, appositions and clauses /closis/ becuase they are very common in biomedical texts. The method proposes a set of syntactic patterns to detect the structures and then applies a set of rules for sentence simplification. Finally, relations are extracted from simple sentences by a pattern matching
• #32: Despite the richness of natural language expressions, in practice, DDI are often expressed by a limited number of constructions. This fact favors the use of patterns as an excellent method for their extraction. Based on her professional experience and the corpus observation, our pharmacist defined a set of lexical patterns (see Table 6) to capture the various language constructions used to express DDI in pharmacological texts. Moreover, the pharmacist provided a set of synonyms for the verbs that can indicate a possible DDI.
• #33: Now, I am going to explain by an example why the resolution of these structures is important. For example, you can see the following sentence:
• #34: Our method, first, is able to identify the coordination by this pattern.
• #38: In a similar way, our method identifies appositions and applies pattern matching to detect DDIs.
• #39: For the clause splitting, we use a set of clues (for example, the use of relative pronouns, subordinators, coordinators, commas) to detect the clause boundaries /baundris/ Once the bondaries /baundris/ have been detected, new simple sentences are generated from these clauses by a set of simplification rules
• #40: Now, I'm going to describe the experiments that we have performed.
• #41: Our baseline experiment considers every pair of drugs as a DDI. This baseline achives the maximum recall but a low precision.
• #42: In our 2nd experiment, we have only applied the lexical pattern on the sentences in the corpus DrugDDI, without performing syntactic processing.
• #43: In our 3rd experiment, we first resolve coordinatives and appositions by a set of syntactic patterns, as I have just described, and then, we applied the lexical patterns on the sentences in order to detect the DDIs.
• #44: In our last experiment, we also detected the clause boundaries /clos baundris/ by a clause splitting algorithm and split the complex and compound /com-pound/ sentences into simple sentences. Finally, we applied lexical pattern matching /maching/ on these simple sentences.
• #45: This graph compares the three experiments. You can see that the baseline achieves a very low precision (only 11%).(the green bars) Llexical patterns achieve the best precision (67%), however, obtain a low recall (14%). (the blue bars) The detection of coordinatives and appositions helps to improve the recall to 26%, but the precision decreases to 48%. (the red bars) Finally, you can see that the clause splitting and sentence simplification do not achieve our second experiment. (the yellow bars) This is mainly due to two reasons: first, our method for clause splitting is too much basic and is not able to detect the boundaries, but on the other hand, many interactions often span several clauses and our lexical patterns are not able to capture them.
• #46: So to conclude...
• #47: Our main contributions are: We have created the first annotated corpus with drug-drug interactions. and we have developed the first approximation for DDI extraction based on the use of syntactic information and lexical patterns. Also, we have performed several experiments. We can conclude that the resolution of complex linguistic constructions and pattern matching can help to detect DDI, although we know that there is lot of work to do yet.
• #48: Future directions include to improve the DrugDDI corpus. Also, we will improve the resolution of the syntactic constructions by the handling of the mistakes made by Metamap. Other important point, is the integration of the drug anaphora resolver in order to extract DDIs that span several sentences.
• #49: Of course, we must improve our clause splitting process and the treatment of negation and modality. We will apply a bootstraping method called Spindel to acquire new patterns. Also, we will consider to apply machine learning techniques.
• #50: We will also detect relevant information such as the severity, mechanism /mekanisim/, dosage drugs /dosich drags/, time course that will help to improve the quality of drug interaction resources. Also, we think that it is very important to carry out an user-oriented evaluation.
• #53: Debo decir que este proyecto se ha desarrollado en el marco de dos proyectos de investigación: el proyecto ISSE sobre interoperabilidad semántica en el dominio de la salud y el proyecto BRAVO. También agradecer al consorcio MAVIR su apoyo durante este trabajo.
• #54: And also, we applied a set of affixes to identify and classify drugs that are not detected by MetaMap.
• #55: You can see a more detailed /diteld/ description of this drug name recogniton taks in our paper “Drug name recognition and classification in biomedical texts” \
• #56: We have developed two different approaches for anaphora resolution. We believe that they are the fisrt works for the pharmacological domain. The first approach uses syntactic and semantic features to score the antecedents. This approach has been published in NLDB conference, held in saarbrucken. The second approach uses a set of linguistic rules (based on the centering theory) in order to find the antecedents. This work has been submitted to the DTMBIO workshop. \textcolor{blue}{\textit{ We have di-vilop two different a-ppro-ches for anaphora ri-so-lusson. We bi-liv that they are the first works for the pharmacological domein. The first approach uses syn-tac-tic and si-man-tic features to score the antecedents. This approch has been published in the NLDB (en-el-di-bi) conference, held in saabrucken, this year.} The second approach ius a set of linguistic ruuls based, on the sente-ring zeo-ri, in order to faind the antecedents}
• #57: For this reason, we think that the information extraction can improve the early detection of drug interactions. We are developing a system that provides a more effective access to this information. This is a very simple sentence 'Aspirin may decrease the effects of probenecid, sulfinpyrazone and phenylbutazone'. It contains three interactions. So, the goal of our system is to identify the drug names and extract the interactions between them.
• #58: DrugBank is an online and free database. You can seach /serch/ any drug, for example, heparin. DragBank isan online and fri databeis. Yu can serch ani drug, for igsampol, heparin, and then, drugbank shows pharmacological and kemical information about this drug.