Machine learning foretells how dangerous a patient fall could be

Patients falls during hospitalization remains one of the most prevalent patient-safety problems facing hospitals, especially falls resulting in injuries.  In the United States alone, hundreds of thousands of patient falls occur in hospitals yearly, with about 30-50 percent resulting in injuries. Falls with injuries lead to increased medical costs, prolonged hospital stays, litigations, and even death.  One study reported that a fall with injury added 6.3 days to the hospital stay on the average. The average cost for a fall with injury is about an additional $14,000 of hospital costs.  The first effort to prevent and reduce patient falls is to use a standardized assessment tool to identify fall, then customized interventions can be offered to individual patients identified.  Although a considerable body of fall assessment tools exists, surprisingly little research has been done to assess the severity of injury caused by the falls. 

One may argue that the patient who is at high risk of falling identified by one of the fall assessment tools will also be more likely to have severe injury when they fall. That is equivalent to say, fall assessment tools should not only predict the fall, but also foretell the severity of resulted injury.  But is that true?  We tested one of such commercial assessment tools, the Hester Davis scale for fall risk assessment, used in Houston Methodist Hospital, and the result indicates that even though the Hester Davis scale has the capability to predict the patients’ fall risk, the scale has nothing to do with the severity of the injuries resulting from the falls.

In this paper, we developed and validated a predictive model for severity of injuries of falls based on multi-view ensemble learning with missing values (MELMV), an integrative machine learning method. Without a priori consideration of the patients will fall or not, we tried to find the unique factors that will contribute to patients’ injuries, we collected fall patient’s demographic characteristics, diagnoses, procedural data, and bone density measurements from different data sources or reports. Because it is difficult to merge the heterogeneous medical data from different resources into one frame, an ensemble method is employed. First, certain sub-models were trained independently from each source of medical data, then a final model is generated by assembling all the sub-models together. The ensemble learning will ensure that the final model has better performance than any sub-model can obtain.

The MELMV model is implemented in a web application, SOFRA (Severity Of Patient Falls Risk Assessment) to incorporate the severity risk score into the clinical workflow via the electronic medical record (EMR) to alert care providers. We validated the app on all the fall cases from May 2016 to December 2016 in HMH, the AUC is 81% We also validated our app on the patients with High Hester Davis score (patients have high risk to fall), the AUC is 86%, showed that among the high risk of fall patients, this model accurately predicted which of these fall patients would have severe injuries. Houston Methodist Hospital is integrating SOFRA app into our EMR, EPIC for daily operations.