Hi James, Thank you for your reply.
If you could create the patch for identifying the words used in the matching that would be great. We understand you have other priorities and will wait until you have time to do it. Thank you for logging the issue with the incorrect chunking as well. Regards, Dennis -----Original Message----- From: Masanz, James J. Sent: Thursday, August 29, 2013 8:38 AM To: '[email protected]' Subject: RE: Concept annotation questions I created JIRA issue CTAKES-231 for this as the code in trunk and in the cTAKES 3.1 branch also get the chunking wrong. https://issues.apache.org/jira/browse/CTAKES-231 Thanks, -- James From: [email protected] [mailto:[email protected]] On Behalf Of Masanz, James J. Sent: Thursday, August 29, 2013 9:19 AM To: '[email protected]' Subject: RE: Concept annotation questions Hi Dennis, Thanks for explaining why you are interested in finding out which words in the original text cause a particular concept to be annotated. We are currently working on getting Apache cTAKES 3.1 out. Depending on your timeline, after that is done, perhaps I could create a patch for you that would help with determining which words from the text matched a dictionary entry, rather than just the begin offset of the first word and the end offset of the last word. As far as the chunking, the fact “liver” and “and” are being tagged as O-chunks explains why the dictionary lookup component is not finding liver cancer or lung cancer in “cancer of colon, liver and lung” I’ll try that sentence with the latest chunker model (which will be in cTAKES 3.1) and see if it assigns correct chunk tags for that sentence. -- James From: [email protected] [mailto:[email protected]] On Behalf Of Dennis Lee Hon Kit Sent: Wednesday, August 28, 2013 2:33 PM To: [email protected] Subject: Re: Concept annotation questions Hi James & Pei, Thank you for your replies and sorry for my late reply as I have been away. Q1 – The longest span could work and is one of the options we are looking at but when there are overlaps it can get complicated. In the following example, the longest would work. We can take start with 01, and ignore 02 and 03 because their start positions overlap the end position of 01, and then continue with 04. But I don’t think it will always be this straight forward as the being/end string positions may not always be a good indicator of what exactly in the original text was coded. 00 Invasive ductal carcinoma of the left breast with bone metastases. 01 Invasive ductal carcinoma of the left breast 408643008|Infiltrating duct carcinoma of breast (disorder)| 02 breast with bone 56873002|Bone structure of sternum (body structure)| 03 breast with bone metastases 94297009|Secondary malignant neoplasm of female breast (disorder)| 04 bone metastases 94222008|Secondary malignant neoplasm of bone (disorder)| Q2 – As we are beginners, we are not at the level where we are comfortable with modifying cTakes or even know where to begin modifying cTakes but that would be an option in the future. Going back to the example of “cancer of liver” and using the begin/end position of the string that was used to identify the concept, the original string would be “cancer of colon, lung and liver.” The CUI that was identified was C0345904, which has 209 (137 unique) descriptions for all languages. Examples of English terms include: • CA - Liver cancer • Cancer of Liver • cancer of the liver • Cancer, Hepatic • CANCER, HEPATOCELLULAR • Malignant hepatic neoplasm • Malignant liver tumor • Malignant liver tumour • Malignant neoplasm of liver • malignant neoplasm of liver (diagnosis) • Malignant neoplasm of liver unspecified • Malignant neoplasm of liver unspecified (disorder) • Malignant neoplasm of liver, not specified as primary or secondary • Malignant neoplasm of liver, NOS • Malignant neoplasm of liver, unspecified • malignant neosplasm of the liver • Malignant tumor of liver • Malignant tumor of liver (disorder) • Malignant tumour of liver It would seem suboptimal to go through each of the descriptions to try and determine which was the UMLS term that was used in the coding. It is important for us to know which part of the string is matched because something like “Invasive ductal carcinoma of the left breast” will be matched to the SNOMED CT concept “408643008|Infiltrating duct carcinoma of breast (disorder)|”, but we would like to know that “left” was not matched and would like to post-coordinate the expression to indicate the left breast, i.e.: 408643008|Infiltrating duct carcinoma of breast (disorder)|:363698007|Finding site (attribute)|=80248007|Left breast structure (body structure)|. When there are other qualifiers like severity, chronicity and episodicity that may be ignored when matching, we would like to capture it at the level of granularity specified in the original text. In terms of the chunking, here is what I see for “cancer of colon, lung and liver”: • NP: cancer of colon, lung and liver • PP: of • NP: colon, lung and liver For “cancer of colon, liver and lung” here is what I see: • NP: cancer of colon, • PP: of • NP: colon • O: liver • O: and • NP: lung Q3 – To answer Pei’s question, we are not looking at the preferred name from the UMLS, just which term was used. Regards, Dennis From: Chen, Pei Sent: Thursday, August 22, 2013 12:27 PM To: [email protected] Subject: RE: Concept annotation questions Also, > 3)… or the exact description that was returned in the UMLS? I presume you mean to save the preferred name from UMLS? If so, this seems to be a common request- see: https://issues.apache.org/jira/browse/CTAKES-224 --Pei From: Masanz, James J. [mailto:[email protected]] Sent: Thursday, August 22, 2013 3:24 PM To: '[email protected]' Subject: RE: Concept annotation questions Welcome to the cTAKES community. Q1 – some people use the longest span. Q2 &Q3 – can you just use the text from the dictionary “Malignant neoplasm of liver (disorder)“. Alternatively you could modify cTAKES to save the text of the words that it matches when it is performing dictionary lookup. I would guess there is a term in the UMLS dictionary with the same code as Malignant neoplasm of liver (disorder) that just has the words “cancer of liver”, but there isn’t anything in cTAKES to give that to you just through a configuration change. For “cancer of colon, liver and lung“, can you look at the chunk tag for liver. If it’s in a separate noun phrase (NP) from “cancer of colon” that would account for why cancer is not getting tied to liver in that case (but wouldn’t account for why the chunker is creating as a separate noun phrase) -- James From: [email protected] [mailto:[email protected]] On Behalf Of Dennis Lee Hon Kit Sent: Wednesday, August 21, 2013 1:10 PM To: [email protected] Subject: Concept annotation questions Hi Everyone, We are new to cTakes so please bear with our questions. We are using cTakes to annotate things like encounter diagnoses and referral notes and are especially interested with the SNOMED CT encodings. But we are not sure how to make sense of all the outputs. Example #1 In the example below, “cancer of colon, lung and liver” has been encoded with SNOMED CT and additional concepts that do not apply have been removed (e.g., general “cancer” concept, lung, colon and liver structures, etc). They have been plotted out by the begin/end positions. If the terms to do not align, its probably because the email only accepts plain text and a mono-spaced font is not the default. cancer of colon, lung and liver cancer of colon, lung and liver 93870000|Malignant neoplasm of liver (disorder)| cancer of colon, lung 363358000|Malignant tumor of lung (disorder)| cancer of colon 363406005|Malignant tumor of colon (disorder)| Question (1) – We had to do quite a bit of post-processing to remove inactive concepts, subtype concepts, concepts that are part of the defining attributes, etc. Are there a set of guidelines to help sort out the CUI or SNOMED CT codes that have been identified? Question (2) – How can we determine that “93870000|Malignant neoplasm of liver (disorder)|” refers to “cancer of liver” as opposed to using the begin/end string, which points to “cancer of colon, lung and liver”? Certainly we can try to do additional parsing but there are a lot of different scenarios to take into account. Question (3) – This relates to question 2, are we able to identify the original terms that were used for the concept matching or the exact description that was returned in the UMLS? While the CUI is helpful, the CUI can refer to tens or even hundreds of descriptions. ________________________________________ Example #2 Switching the position of colon, lung and liver can result in different encodings. Once again, after removing additional concepts not needed (i.e., “cancer” and “colon structure”), we get the following. What happened to liver and lung cancer? cancer of colon, liver and lung cancer of colon 363406005|Malignant tumor of colon (disorder)| lung 39607008|Lung structure (body structure)| We have more questions but will start with these. Thank you in advance. Regards, Dennis
