GIGI is a learning software. It is an affordable & quickly deploy-able technology that leverages predictive algorithms to decrease the demand on your human resources when evaluating claim requests for Utilization Review.
In 2019, EHS leadership committed to create a third generation medical triage technology. The vision was to create an AI supported, data-driven med-auth algorithm to eliminate many of the issues incumbent in the current system and further reduce the need for human intervention in the process. Having spent a year visiting with clients about the potential of the concept, we contracted with Dr. Laura Gardner, MD, MPH, PhD, FACPM, of Persimmon AI, to serve as our data scientist. Dr. Gardner's mission was to work with us to develop an algorithm that would predict the likelihood of a medical request being certified via the UR process. Our goal being to eliminate the time, expense and errors associated with the current triage processes for medical request authorization.
The predictive algorithm was developed using a logistic regression model to predict the likelihood of certification. The model was deliberately constructed using only those variables that would be known to the user when the treatment request is first received. The analysis was first performed on a split sample; that is, the model was built on a random half subset of the data and then the parameter estimates from the first subset were used to score the holdout data set. Subsequently, the model was rerun on the full data set to create the predictive weights for the algorithm.
The split-sample results show a relatively high accuracy rate of 76% and in the full sample the accuracy rate is 79%. However, digging into the split- sample results we found that among the requests that were actually certified, 92% were predicted to be certified and only 8% were false negatives.
Additionally, the split sample revealed that 19% of the requests were certified by the model but not actually certified (false positives). When we investigated the types of procedures that fell into that group, we found that most of the requests were either for pharmacy or physical medicine. Fewer than 8% of the group were for surgery, either major or minor. In actual practice we are confident these types of requests can be handled effectively via business rules specific to the handling of such requests.