Natural Language Processing’s Role in the Analysis of Electronic Medical Records

As artificial intelligence algorithms become more refined, electronic medical records (EMRs) could become a frontier for effective and relevant applications of these new technologies. EMRs represent a massive collection of information with which computers can perform several helpful tasks. Algorithms could more efficiently recognize data that are legible in each individual record but are not codified – for instance, a medical professional could likely recognize asthma in an EMR, but the EMR may not have an "asthma" label attached to it. Also, computational algorithms can detect potentially salient patterns in the data that would be very time- and effort-intensive to recognize manually. 

However, machines can have difficulty interpreting notes that are written by and for people – nonstandard abbreviations, nomenclature, and formatting of notes can obfuscate important information (Cai et al., 2018). Additionally, simple human errors made in the transcription of notes into EMR systems – sometimes just typos – can create faulty data, which jeopardizes the merit of any analyses that might be performed (Tian et al., 2019).  

Natural language processing (NLP) is a branch of artificial intelligence which uses "computational linguistics that provides parsing and semantic interpretation of text, which allows systems to learn, analyze, and understand human language" ("Build Apps with Natural Language Processing," 2020). As such, NLP is instrumental in the interpretation of such sets of text as EMRs (one such set of text is known as a "corpus;" plural: "corpora"). NLP applications in analyzing EMRs could help progress research, improve diagnostic and prognostic accuracy, and effectively anticipate adverse reactions to treatments.   

NLP algorithms can be broken down into three basic categories. First are rule-based methods, in which handwritten rules are used to interpret text and determine which items are important. Second are machine learning (ML)-based methods, in which self-correcting algorithms train on the corpus and compare their conclusions with optimal results to "learn" how to make effective inferences based on the data available (Jurafsky & Martin, 2019, pp. 57). Last are hybridized methods, in which ML algorithms are fine-tuned with the addition of rules. 

NLP applications in analyzing EMRs can be as simple as effectively sorting patients: one study produced an algorithm for recognizing the prognosis of asthma patients based on EMRs. The algorithm achieved F measures of .96 and .95 using macro-averaged and micro-averaged methods, respectively, on analyses of 36 patients (Sohn et al., 2018). F measure, or F-score, is a useful metric of predictive algorithms' effectiveness. It's equal to the harmonic mean (a form of averaging) of precision and recall – reported true positives divided by reported positives, and reported true positives divided by total true positives, respectively. (Manning et al., 2009, pp. 155). Such a successful algorithm for recognizing asthma patient prognosis could help check research findings against clinical data, automate the review of medical charts, and identify at-risk populations by observing the correlation of asthma prognosis with patients' belonging to certain demographics (Sohn et al., 2018). 

Another study proved promising for researchers of multiple sclerosis. Not only did a rule-based NLP algorithm identify which patients suffered from MS with very high precision (p values below .001) – such a successful identification allowed researchers to cross-compare carefully curated and annotated research data with the massive corpus of available data collected in a clinical setting. Researchers found that their data were largely in agreement with clinical data; such verification could help researchers build cases for conclusions drawn in future studies (Damotte et al., 2018). 

Applications can also focus on revealing undiscovered patterns. Researchers investigating the recognition of drug-adverse event (AE) relations in EMRs had success with a rule-based NLP algorithm. The algorithm included allowances for bad entries, utilizing so-called "fuzzy logic" to account for minor typos and phonetic misspellings in uncommon words. The drug-AE relation extraction had an F measure of .534 and about a 60% report rate. This may seem low, but it was obtained on a previously-unseen corpus containing different writing styles, and could still provide a dramatic improvement over the 20% average report rate of drug-AE relations (Tang et al., 2019).  

NLP was also helpful in recognizing which pregnant women were likely to develop suicidal behaviors in a study of over 275,000 women's EMRs. With the help of NLP mining patient data, 11 times more women were identified to be at risk than when medical professionals relied only on diagnostic codes (Zhong et al., 2018). 

In a 2019 review of NLP's effectiveness in oncology, researchers concluded that ML methods require significant effort in labeling the vast datasets necessary to effectively train, and once trained, translate poorly to different fields of medicine and different human languages. Researchers also questioned the medical ethics of training ML algorithms on corpora of data which may not represent "the whole of human diversity" (Savova et al., 2019, pp. 5467). Furthermore, researchers comparing the effectiveness of rule-based and ML-dependent hybrid approaches even found that the rule-based approach was slightly more effective, with F measures .03 to .04 higher in the extraction of information from hepatectomy operation notes (Chen et al., 2019).  

The path forward for ML-based methods of NLP on EMRs could involve addressing these and more challenges before ML-based NLP becomes standard practice. As such, it's important that more sophisticated and effective rule-based approaches continue to be developed, as they can yield helpful results for medical professionals until such a time as the challenges facing ML-based methods are overcome. 

References 

Build Apps with Natural Language Processing (NLP) | IBM Watson. Ibm.com. (2020). Retrieved 30 April 2020, from https://www.ibm.com/watson/natural-language-processing?p1=Search&p4=p50290112947&p5=e&cm_mmc=Search_Google-_-1S_1S-_-WW_NA-_-natural language processing_e&cm_mmca7=71700000060771761&cm_mmca8=kwd-35559750&cm_mmca9=EAIaIQobChMI5JKpx6SL6QIVmo7ICh1oeA_jEAAYASAAEgLUbPD_BwE&cm_mmca10=405867665754&cm_mmca11=e&gclid=EAIaIQobChMI5JKpx6SL6QIVmo7ICh1oeA_jEAAYASAAEgLUbPD_BwE&gclsrc=aw.ds. 

Cai, X., Dong, S., & Hu, J. (2019). A deep learning model incorporating part of speech and self-matching attention for named entity recognition of Chinese electronic medical records. BMC Medical Informatics And Decision Making, 19(S2). https://doi.org/10.1186/s12911-019-0762-7 

Chen, L., Song, L., Shao, Y., Li, D., & Ding, K. (2019). Using natural language processing to extract clinically useful information from Chinese electronic medical records. International Journal Of Medical Informatics, 124, 6-12. https://doi.org/10.1016/j.ijmedinf.2019.01.004 

Damotte, V., Lizée, A., Tremblay, M., Agrawal, A., Khankhanian, P., & Santaniello, A. et al. (2018). Harnessing electronic medical records to advance research on multiple sclerosis. Multiple Sclerosis Journal, 25(3), 408-418. https://doi.org/10.1177/1352458517747407 

Jurafsky, D., & Martin, J. (2019). Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. Web.stanford.edu. Retrieved 30 April 2020, from https://web.stanford.edu/~jurafsky/slp3/edbook_oct162019.pdf. 

Manning, C., Raghavan, P., & Schütze, H. (2020). An Introduction to Information Retrieval. Nlp.stanford.edu. Retrieved 30 April 2020, from https://nlp.stanford.edu/IR-book/pdf/irbookonlinereading.pdf. 

Savova, G., Danciu, I., Alamudun, F., Miller, T., Lin, C., & Bitterman, D. et al. (2019). Use of Natural Language Processing to Extract Clinical Cancer Phenotypes from Electronic Medical Records. Cancer Research, 79(21), 5463-5470. https://doi.org/10.1158/0008-5472.can-19-0579 

Sohn, S., Wi, C., Wu, S., Liu, H., Ryu, E., & Krusemark, E. et al. (2018). Ascertainment of asthma prognosis using natural language processing from electronic medical records. Journal Of Allergy And Clinical Immunology, 141(6), 2292-2294.e3. https://doi.org/10.1016/j.jaci.2017.12.1003 

Tang, Y., Yang, J., Ang, P., Dorajoo, S., Foo, B., & Soh, S. et al. (2019). Detecting adverse drug reactions in discharge summaries of electronic medical records using Readpeer. International Journal Of Medical Informatics, 128, 62-70. https://doi.org/10.1016/j.ijmedinf.2019.04.017 

Tian, Q., Liu, M., Min, L., An, J., Lu, X., & Duan, H. (2019). An automated data verification approach for improving data quality in a clinical registry [Abstract]. Computer Methods And Programs In Biomedicine, 181, 104840. https://doi.org/10.1016/j.cmpb.2019.01.012 

Zhong, Q., Mittal, L., Nathan, M., Brown, K., Knudson González, D., & Cai, T. et al. (2018). Use of natural language processing in electronic medical records to identify pregnant women with suicidal behavior: towards a solution to the complex classification problem [Abstract]. European Journal Of Epidemiology, 34(2), 153-162. https://doi.org/10.1007/s10654-018-0470-0