match between an expression and a note that may contain the expression (e.g., “hastroublegoingtosleep” and “hasahistoryofalcohol,” as presented as examples in Figures 5 and 6, respectively). In traditional machine-learning approaches for text classification, a human expert is required to label phrases or entire notes, and then a supervised-learning algorithm attempts to generalize the associations and apply them to new data. In contrast, using non-negated distinct expressions eliminates the need for an additional computational method to achieve generalizability, as the expressions have consistently been found highly prevalent across multiple clinical conditions by considering more than 10 million clinical narrative notes. TN thus provides distinct classifications and is thereby expected to provide robust results.

TN yields a high performance for the determination of a variety of clinical conditions using rapid processing (i.e., approximately 1 millisecond per note on average); however, no benchmark for processing performance comparison has yet been reported for similar tasks. We also extended TN for uses beyond extracting smoking status.

For instance, we used TN to extract family history of coronary artery disease, classify patients with sleep disorders, improve the accuracy of the Framingham risk score for patients with nonalcoholic fatty liver disease, and classify nonadherence to T2DM (see past projects: http:// researcher.ibm.com/researcher/ view_person_pubs.php?person=ibm-Uri.Kartoun).

Further, TN could be used to enhance the standard regular expression pattern language (in which a sequence of characters defines a search pattern). Applying standard regular expressions relies on knowing a priori the patterns to search for, and this is exactly what TN’s human-in-the-loop step addresses (Figure 2). The regular expression pattern language can benefit from TN’s initial identification of a collection of phrases to match.

An additional advantage of TN
is that it is not sensitive to negations
and is capable of ignoring them
efficiently (e.g., “no longer smokes,”
“not a cigarette smoker,” and “has not
smoked since”). Further, TN does not
require setting a priori, subjectively
selected, or data-dependent
configuration parameters such as those
required when using SVMs, which was
the most popular approach used at the
i2b2 smoking status natural language
processing challenge [ 6]. Another
advantage unique to TN is its potential
applicability to languages other than
English because the human-in-the-
loop procedure is not tied to a specific
language. For instance, the phrase
“smokes two packages per day” would
be “smokestwopackagesperday”
in English, “ ”
in Hebrew, “ ” in Chinese,
and “fumadoispacotespordia”
in Portuguese.

TN has several limitations. First, TN requires that a human dedicate time to identifying candidate expressions. While TN requires only a few human hours for the task of classifying smoking status, performing more complex tasks (e.g., identifying complications after a surgery) would require additional time. However, when this effort is complete, the identified expressions can be generalizable and could be deployed on any database and used by the research community. In addition, the language describing an individual’s smoking status might be quite diverse in various places. Per Zipf’s law, English has an infinite number of possible expressions, so one cannot enumerate all the ways to describe smoking status. However, across varied conditions, the results demonstrate that the tail of the distributions can be ignored (as seen in Figures 4–6).

Humans are the ones who created letters and languages, and therefore we are capable of accurately identifying highly descriptive non-negated expressions. Similar to my Ph. D. dissertation in which I described how a collaboration between a human and a robot can expedite a learning task [ 4], my research on TN demonstrated that an interactive human-in-the-loop extension, applied here on large collections of clinical narrative

DOI: 10.1145/3139488 © 2017 ACM 1072-5520/17/11 $15.00

notes, can produce high classification accuracy across distinct medical conditions. In conclusion, TN is a rapid, accurate, and easily adaptable method of identifying patients’ clinical descriptors by interacting with clinical narrative notes. The use of TN allows for accurate and comprehensive identification of many medical conditions, which will improve precision and recall values in studies that rely on textual data.

ACKNOWLEDGMENTS

I would like to thank Professor Peter Szolovits of MIT’s Department of Electrical Engineering and Computer Science for his constructive feedback regarding Text Nailing. I further wish to express my deepest gratitude to Drs. Stanley Shaw and Kathleen Corey for their priceless guidance during my fellowship at MGH.

Endnotes

1. Kartoun, U. The man who had them all. ACM Interactions 24, 4 (July–Aug. 2017), 22–23.

2. Kartoun, U., Kumar, V., Cheng, S.C., Yu, S., Liao, K., Karlson, E., Ananthakrishnan, A., Xia, Z., Gainer, V., Cagan, A., Savova, G., Chen, P., Murphy, S., Churchill, S., Kohane, I., Szolovits, P., Cai, T., and Shaw, S. Demonstrating the advantages of applying data mining techniques on time-dependent electronic medical records. Proc of AMIA 2015 Annual Symposium.

3. Savova, G.K., Ogren, P.V., Duffy, P.H., Buntrock, J.D., and Chute, C.G. Mayo clinic NLP system for patient smoking status identification. Journal of the American Medical Informatics Association 15, 1 (2008), 25–28.

4. Kartoun, U. Human–robot collaborative learning methods. Ph. D. Dissertation. Ben-Gurion University of the Negev, Israel. 2008.

5. Kartoun, U.*, Beam, A.*, Pai, J., Chatterjee, A., Fitzgerald, T., Kohane, I.*, and Shaw, S*. The spectrum of insomnia-associated comorbidities in an electronic medical records cohort. Proc. of AMIA 2016 Annual Symposium. *Contributed equally.

6. Uzuner, O., Goldstein, I., Luo, Y., and Kohane, I. Identifying patient smoking status from medical discharge records. Journal of the American Medical Informatics Association 15, 1 (2008), 14–24.

 

Uri Kartoun is a research staff member at
IBM Research in Cambridge, MA. Previously
he was a research fellow at Harvard Medical
School/Massachusetts General Hospital. His
Ph.D. from Ben-Gurion University of the Negev,
Israel, focused on human-robot collaboration.
→ [e-mail]
INTERACTIONS.ACM.ORG NOVEMBER–DECEMBER 2017 INTERACTIONS 49