Thinking about Hail Mary Passes, Reducing Rehospitalizations, and Skewed Risk Reward

The football playoff season is quickly approaching!? To get into the spirit, we were browsing YouTube recently and found this great play from January, 2012.

The 2011 New York Giants had struggled all season long, making the playoffs on the final day of the regular season with a record of 9-7.? But on a 1st and 10 play with 6 seconds to go in the 2nd quarter of the Divisional Playoff game at Lambeau Field against the Green Bay Packers, Eli Manning completed a pass to Hakeem Nicks that put the Giants ahead 20 - 10. This play (video), coming at the end of the first half of the game, completely demoralized the Packers and propelled the Giants to a stupendous Super Bowl victory several weeks later.

Eli Manning and coach Tom Coughlin were later hailed as brilliant strategists for deciding to call such a bold play right before the half.? But upon reflection, the call doesn’t seem “gutsy” at all.? In fact, it seems like an obvious thing to do.??

Here’s why: In the best case, which admittedly was unlikely, the pass is complete and the Giants go to the locker-room on an emotional high, leading the game 20 - 10.? The most likely outcome was that the pass would be incomplete, time would run out, and the score would remain 13 - 10.? The Giants would be no worse off in this case.? There was also a chance that the Packers would intercept the pass, but that wouldn’t matter because time would expire before they had a chance to run another play.? Again, the Giants would be no worse off in this situation.? In fact, the ONLY way the Giants would be worse off was the very slim chance that Manning’s pass would be intercepted AND run all the way back for a touchdown.? Very unlikely.

Mathematicians call these kinds of decisions “skewed risk-reward” situations - when you take Risk and Reward into account, you’re much better off taking one action than another one.? This reminds me of the decision about whether to use Machine Learning to predict rehospitalizations among nursing home patients.??

Despite the fact that nursing home patients are often very sick and frail, it still turns out that rehospitalizations among nursing home patients are relatively rare.? In any particular week, in a home of 100 patients, you might have 2 rehospitalizations on the low side and 7-8 on the high side.? With this data, if someone asked you to predict who is going to the hospital in the next three days , you’re better off predicting that no one will be rehospitalized.? You’ll be 99% accurate!??

Our company SAIVA builds Machine Learning models and systems that predict re-hospitalizations within nursing homes. Every day our system provides? nurses a list of the 10 - 15 patients who it calculates are at greatest risk of going to the hospital over the next 3 days.? We know from analyzing our data? that 80%+ of patients who do ultimately? get rehospitalized were on our list 1-3 days BEFORE they were rehospitalized.??

Ok - so, how should we think about this???

On average, it costs a nursing home about $2,500 every time a Medicare patient goes back to the hospital.? So, if the Saiva AI System can prevent even only 2 of these per week, that’s $5,000 per week.? With 52 weeks in a year, that’s? $260,000 per year in added revenue to the nursing home.

Essentially we’re paying less than $500/month? to save $260,000+ a year!? Boom!? And of course, there are plenty of non-monetary benefits like reducing stress on mostly elderly, frail patients. Boom, again!

This is a classic skewed risk-reward situation.? There is very little downside to having a nurse check on patients who won’t ultimately rehospitalize.? But there is a HUGE upside to having a targeted list of patients the nurse should check on? who WILL most likely rehospitalize, assuming the nurse will take action to treat these patients in-place and save the patient from rehospitalizing.

Using Machine Learning to reduce nursing home rehospitalizations may seem like a gutsy call.? But just as in football, when looked at through the lens of risk-reward, a well timed Hail Mary play can be logical, smart, and game winning!???

Be sure to check out this video of thrilling Hail Mary passes over the years.?