Longitudinal analysis of pain in cancer patients by natural language processing of medical record text
Personalized-medicine researchers have established the feasibility of using computerized analysis of the free text of clinical records to track longitudinal patterns of pain in cancer patients. The results could point the way to automatic pain monitoring and better pain management using electronic medical record systems, and could help researchers better understand the link between cancer biology and pain.
The researchers used computerized Natural Language Processing (NLP) to analyze text from 23,887 pages of clinical records from 33 men with metastatic prostate cancer. The records spanned the time from diagnosis to death from prostate cancer for each man, which varied from 1 to 15 years in this study cohort.
Using NLP algorithms, the researchers were able to detect four pain phenotypes: no pain, pain, controlled pain, and severe pain, and each pain phenotype instance was resolved to a specific date. Comparison of the NLP findings in these subjects both to known clinical correlates of pain provided validation of the approach, and generalizability of the method was shown by using a version of the NLP algorithm in a set of publicly available clinical records of patients with various illnesses.
Pain management in cancer, and particularly metastatic prostate cancer, is difficult, and severe pain was common in these 33 men, especially in the last year of life. Better methods of monitoring pain status could allow better pain management methods to be tested in the clinical setting. Surprisingly, two of the 33 subjects in the study did not have severe or controlled pain, suggesting that “outlier” phenotypes useful for probing the molecular basis of cancer pain may exist.
Pain was chosen as the subject of the current study because it is a primary concern of the patient and caregivers, and discussion of pain in clinical encounter records is common. Future research using NLP will test whether the free text of clinical records can be used to better understand disease processes on a more general basis. Patients and clinicians often manage each perceived problem in short bursts, based on clinical encounters. Routine re-examination of the text of a full series of clinical encounters for important patterns is beyond the capacity of patients or busy clinicians. In theory at least, “well-trained” computers could provide a novel stream of useful, and traceable information automatically.
The research team was led by Dr. G. Steven Bova, Finnish Distinguished Professor at Institute of Biomedical Technology and BioMediTech, University of Tampere, and included researchers from Johns Hopkins University, Lockheed Martin Corporation, and Sage Analytica. It is published today in JAMIA, currently the top medical informatics journal as measured by publication impact factor.
Contact Information:
g.steven.bova@uta.fi
Institute of Biomedical Technology and BioMediTech
PELICAN-Personalized Cancer Medicine Group
FI-33014 University of Tampere
Tampere, Finland
