The advent of electronic medical records improved how readily clinical staff can access patients’ health information and allows admitting physicians to view records of new patients while they are still being worked up in the emergency department. The record contains much more data than the clinical notes entered by doctors or allied health. The number and frequency distribution of past admissions or referrals, changes in baseline vital signs, patterns of pathology or radiology reports, and ward transfers are all points of data available for analysis in electronic medical records. Could an algorithm use these data to predict the length of stay for a given patient?or predict 7-day mortality?
Researchers in New South Wales and Victoria, Australia used electronic health record data to predict mortality, readmission and length of stay in real time with good accuracy. Data from over 32,000 patients between 2008 and 2010 were used to build a Bayesian predictive model, which was tested on retrospective data from 2011. Predictions gave probabilities of specific outcomes (discharged, still admitted, or dead) over each of the next 7 days. Predictions had an 86% accuracy within 24 hours of prediction with an average daily accuracy of 80% by the 7th day. Death could be predicted with 93% accuracy. The algorithm therefore gave a reasonably accurate trajectory for any given patient’s admission.
Each new entry to patients’ prompts a recalculation. Providing information to patients and their families can help give an idea of the duration of stay that can be anticipated. Predictions of long stays could be used to focus efforts on best treatment strategies and bed planning. An increased probability of 7-day mortality can trigger counselling services and discussion of end-of-life care. Such prediction models would also be greatly valuable in emergency departments.
Benefits:
- Accurate predictions of patient trajectories over 7 days
- Uses existing data in patients’ existing medical records
- Encourages effective allocation of resources
Limitations:
- Possibility for less experienced physicians becoming dependent on predictions
- Risk of patient distress if information is misinterpreted
For more information:
- See the original article in the Journal of the American Medical Informatics Association
