Natural Language Processing with Electronic Medical Record Improves Identification of VA Post-Operative Complications

BACKGROUND:
One method for identifying patient safety concerns is to analyze administrative data for specific codes that suggest a medical injury or complication. For example, the Agency for Healthcare Research and Quality (AHRQ) developed a set of 20 patient safety indicator measures that use administrative data to screen for potential adverse events that occur during hospitalization. However, the quality of this approach is limited by the well-documented variable quality of administrative data. A potential alternative is to apply automated natural language processing (NLP) methods to textual medical documents to extract specific medical concepts that are independent of discharge codes and offer a powerful alternative to labor-intensive medical chart reviews. This cross-sectional study evaluated a NLP search approach to detect post-operative surgical complications within VA’s electronic medical record (EMR). Investigators identified 2,974 Veterans who underwent inpatient surgical procedures at six VAMCs from 1999 to 2006. Specifically, post-operative occurrences of acute renal failure requiring dialysis, deep vein thrombosis, pulmonary embolism, sepsis, pneumonia, and myocardial infarction (MI) were identified through patient chart review. The NLP approach to identify these complications was then compared with the usual approach using administrative discharge information.

FINDINGS:

• Among Veterans undergoing inpatient VA surgery, NLP with the EMR greatly improved the identification of post-operative complications compared to the administrative-code based algorithm.
• NLP correctly identified 82% of acute renal failure cases compared with 38% for patient safety indicators; 59% vs. 46% for venous thromboembolism; 64% vs. 5% for pneumonia; 89% vs. 34% for sepsis; and 91% vs. 89% for post-operative MI.
• An accompanying Editorial states that NLP has the potential to greatly enhance the EMR with new applications, such as automated quality assessment to assist in the performance of comparative effectiveness research. However, while the study by Murff, et al suggests that these benefits may be closer than ever, it will require considerable new investment in research and development.

LIMITATIONS:

• Administrative code-based algorithms used in this study were not originally designed for VA data.
• Only a small minority of healthcare systems currently have integrated EMRs. Thus, some of the NLP query strategies used here are not feasible at non-VA institutions.

AUTHOR/FUNDING INFORMATION:
This study was funded by HSR&D (SAF 03-223). Drs. Murff, FitzHenry, Matheny, and Speroff are part of the VA Tennessee Valley Healthcare System.

Murff H, FitzHenry F, Matheny M, et al… and Speroff T. Automated Identification of Post-Operative Complications within an Electronic Medical Record Using Natural Language Processing. JAMA August 24/31, 2011;306(8):848-855.

Jha A. The Promise of Electronic Records: Around the Corner or Down the Road? Editorial JAMA August 24/31, 2011;306(8):880-881.