Predicting Health Outcomes With Machine Learning

Posted on 25th October 2013 in health

Eric Horvitz, MD, PhD co-director, Microsoft Research

  • from data to predictive models to decision-making
  • costly challenge of re-admissions: they worked with Washington Medical Center, which had been accumulating lots of data on ER visits. Built a predictive model of readmissions: identified relevant evidence out of 25,000 indicators: if “fluid” is in the record, that’s an indicator the person will be readmitted.
  • Created tool called “Readmissions Manager.
  • running this program worldwide
  • example of Congestive Heart Failure. $35 b. annual cost. Should you invest in an intervention program?
  • another big challenge is hospital-associated infection. $20 b.  a year cost. In top 10 causes of deaths in the US. Beyond looking at EMR, they introduced analysis by space and time. Another factor is the patient’s path through various parts of the hospital: could predict what % of the patient population got infections in what part of the hospital. Analyzed various factors, such as which unit of hospital they were in, who was the attending doc.
  • New kinds of predictions: “surprise models” — looked at people who were re-admitteed in 72 hrs. for a condition that wasn’t on chart before.
  • new set of data: patients searching web. For example, nutritional content in logs of downloaded recipes. (Maine people downloaded a lot of recipes with high carb content). This was because they found a lot of older ppl in DC area had downloaded recipes with high sodium content in holiday season.
  • looked at Twitter feeds about birth of child — overlaid lots of people who’d just had babies: about 12% of women showed signs of crashing in terms of mood (post-partu m depression is underreported). Suggests you might be able to intervene in advance to help them.">Stephenson blogs on Internet of Things Internet of Things strategy, breakthroughs and management