Emily Spahn, one of ProKarma’s data scientists, will be presenting at the Strata Business Summit at Strata + Hadoop World, which runs March 13 to 16 at the San Jose Convention Center in California. She’ll be discussing ProKarma’s successful collaboration with Intel and how medical organizations can apply machine learning to improve patient care.
The healthcare industry is eager to tap into the latest AI capabilities, and it is well-positioned for the shakeup: The stakes are high, both in terms of cost and life, and healthcare is producing vast amounts of data. ProKarma partnered recently with Intel to develop a proof of concept that healthcare groups can use to create their own inpatient risk scoring system.
The challenge was to develop a patient risk score:
- From a model predicting imminent health decline
- For inpatient hospital populations
- Using data from electronic medical records
To do this, they examined hospital Rapid Response Teams (RRTs), which intervene when a patient experiences a sudden decline. Though these RRT events only occur in less than 0.5 percent of patients, they represent a critical moment for lifesaving efforts. Any model able to indicate whether a patient would suffer an RRT event could save lives and increase hospital efficiency.
To develop the solution, ProKarma’s team worked with 1.5 years of medical data. That data was placed in a Cloudera Hadoop cluster, which provided a sandbox environment for the project. ProKarma then used Impala to query the data. Given the rarity of the RRT events, the team was able to use smaller data tools instead of a larger distributed system to model the data.
Two thousand events were matched with 2,000 control patients who had similar profiles but hadn’t experienced an RRT event. Open source Python machine-learning tools were used to identify patterns within the data, and the teams used these patterns to develop a model that could predict the likelihood a patient would require support from the RRT in the next hour, based on the last 12 hours of data for that patient. Among the most predictive characteristics, the team learned, were respiratory rate, pulse, blood pressure and patient age.
The model produces a simple patient risk score on a scale of 0 to 10. By highlighting patients at risk, the hope is that early intervention can help healthcare providers avoid an RRT event altogether.
The patient risk score has two applications for healthcare organizations:
- Identification and prioritization of sick patients. Being able to predict who is most likely to experience an RRT event within an hour allows a healthcare organization to put additional resources in place. This score can also be useful during shift changes, when medical professionals do not always have a sense of immediate patient status.
- Identification of the healthiest patients. This will allow healthcare organizations to more effectively help triage populations when serious events occur, resulting in an influx of patients into the hospital.
The process captured many of the challenges organizations pursuing machine learning can expect. For example, given the rarity of RRT events, there was too little data to delve into the implications of specific causes and symptoms. In addition, the data that was available was not always conclusive: Many RRT events listed multiple explanations, the most frequent of which was “staff concern/worried about patient,” – a reflection of the medical team’s gut feeling or observations about a patient that weren’t always accompanied by usable data. In addition, the data was not automatically entered into the database; it required a nurse or staff member to type it in for it to be included in the health records.
As a collaborative project, careful but adaptive planning was vital. Regular meetings modeled agile processes and kept lines of communication open throughout the 10-week project. New subject matter experts identified during the course of the work were incorporated into the process.
The results of this proof-of-concept project represent a model other medical organizations can apply to their own data. To hear Emily’s presentation about the potential of machine learning, register for the Strata + Hadoop conference here.