Hi there, 

I've been doing some extensive research on the potential value of conceptually applying the heritage algorithm in a healthcare provider setting (hospitals, etc). From a data analytics perspective, the scope for developing this algorithm undoubtedly exists. However, various similar algorithms (with the purpose of identifying patients at risk of admission or preventing 'unnecessary hospitalizations') have been developed and implemented across the globe for a number of years now. Due to the fact that some of these algorithms are disease based, they tend to have much higher accuracies as opposed to the heritage 'accuracy threshold' of 0.4.

How does the heritage algorithm then differ, if at all, from these algorithms?

Regards

Two things:

1) I don't think there is a heritage algo per se - they want to see if we can make one using their data. I believe they said in the past they don't have one, but not 100% sure.
2) 0.4 isn't an "accuracy" score as typically used. I am not sure if you knew that or not. In this case lower is better. It would be easy to get an accuracy score in the high .80s just by submitting zeros for every one. I don't believe most algos use RMSLE as a metric - so you can't really compare - at least not without doing some math - and even then I don't think you could compare exactly(but could probably make some good guesses).

Thank you for the reply Chris, its a fascinating subject in healthcare and I have great admiration for the work being done here! But I'm still a bit dubious;

From a data analyst's point of view, these algorithms (existing vs heritage health prize) differ substantially due to the 'mechanics' of the models and the data used for developing these algorithms. However, from a healthcare provider perspective, all of these algorithms aim to predict patients at risk of admission to hospital...

Sure, I understand that various vehicles are built for the same purpose of transportation and that each performs a different purpose, but are we not replicating work already done...surely to "change healthcare as we know it" the answer does not lie within the development of a predictive model for every health insurance group in the world? Will we not reach a stage where we also build "cars" that are closely matched in performance, but dressed in different shells???

I'm playing devil's advocate here, forgive me, but I come from a health system's engineering background, with little exposure to data mining...I'm trying to align the theory of building this model with the practical application thereof :)

Regards

Surely any serious method of predicting hospitalisation would use the patient's weight, Body Mass Index and smoking history as well as age and sex.

They are trying to strike a balance between privacy and Accuracy. Most of the techniques mentioned in the papers could easily be expanded to include those other variables. I'm all for more data, we just aren't going to be getting any. I don't know of another competition that has made more health care data available.

BMI would be cool to have - although got enough issues dealing with pregnant women - don't relish the idea of having it for women.
Actually I would prefer weight and height as who knows - we might be able to improve upon BMI itself. It is designed to be a simple index.

Also we don't really know what are happening to these patients - I'd want to know for sure they were still alive and not entered in a hospital through a different system.

Weird things happen - airbags INCREASED medical bills in some cases. Why? Because it kept people alive with expensive lower body injuries - while they would have normally died and had no hospital bills. I am not sure days in hospital is the best indicator of quality - or what we want to really measure. Certainly it isn't when that stay actually fixes something that another doctor would have missed and then the patient dies.

But we have to measure something as the endpoint - people do get sicker FROM hospitals - so in a way it isn't a bad thing to measure.