All you need to read in the other general journalsBMJ 2011; 343 doi: https://doi.org/10.1136/bmj.d5481 (Published 31 August 2011) Cite this as: BMJ 2011;343:d5481
Free-text analysis improves detection of postoperative complications
Searches of electronic medical records to detect postoperative complications usually rely on specific administrative data codes assigned at discharge from hospital. On the basis of such codes, 20 “patient safety indicators” have been developed for screening for adverse events. However, free-text searches of narrative clinical notes might complement or even surpass use of patient safety indicators.
To test this, a study used data for nearly 3000 patients who underwent surgery in six US medical centres. Trained nurses reviewed all complications in the 30 days after surgery. The researchers then compared the accuracy of detecting postoperative complications with patient safety indicators against that with a natural language processing approach that analysed free text in discharge summaries, progress and operative notes, microbiology reports, imaging reports, and outpatient visit notes.
Overall, natural language processing showed twofold to 12-fold better sensitivities than safety indicators, which in turn showed 4% to 7% higher specificities.⇑ Natural language processing correctly identified 82% (95% CI 67% to 91%) of cases of acute renal failure, versus 38% (25% to 54%) with patient safety indicators, and also proved better at identifying cases of venous thromboembolism (59% (44% to 72%) v 46% (32% to 60%)), pneumonia (64% (58% to 70%) v 5% (3% to 9%)), sepsis (89% (78% to 94%) v 34% (24% to 47%)), and postoperative myocardial infarction (91% (78% to 97%) v 89% (74% to 96%)).
One of the advantages of the natural language processing approach is the possibility of real-time quality assurance, say the authors. This assumes prospective monitoring and identification of complications while a patient is still in the hospital.