Win $3M (if you have nothing better to do)

Heritage Provider Network is offering a cool $3 millions in prize money for the development of an algorithm that can best predict how often people are likely to be sent to the hospital. Jonathan Gluck — senior executive at Heritage — said the goal of the competition is to create a model that can “identify people who can benefit from additional services,” such as nurse visits and preventive care. Such additional services could reduce health care spending and cut back on excessive hospitalizations, Gluck said.

The algorithm contest, the largest of its kind so far, is an attempt (also see Slate article here) to help find the best answers to complicated data-analysis questions. Previous known was the $1 million Netflix Inc. prize awarded in 2009 for a model to better predict what movies people would like. In 2009, a global team of seven members consisting of statisticians, machine-learning experts and computer engineers was awarded the $1 Million contest prize and Netflix replaced its legacy recommendation system with the team’s new algorithm (2nd Netflix’s competition was stopped by FTC and lawyers). I personally think that this time Data Visualization will be a large part of winning solution.

The competition – which will be run by Australian startup firm Kaggle — begins on April 4 and will be open for about two years. Contestants will have access to de-identified insurance claims data to help them develop a system for predicting the number of days an individual is likely to spend in a hospital in one year. Kaggle spent months streamlining claims data and removing potentially identifying information, such as names, addresses, treatment dates and diagnostic codes. Teams will have access to three years of non-identifiable healthcare data for thousands of patients.

The data will include outpatient visits, hospitalizations, medication claims and outpatient laboratory visits, including some test results. The data for each de-identified patient will be organized into two sections: “Historical Data” and “Admission Data.” Historical Data will represent three years of past claims data. This section of the dataset will be used to predict if that patient is going to be admitted during the Admission Data period. Admission Data represents previous claims data and will contain whether or not a hospital admission occurred for that patient; it will be a binary flag.

The training dataset includes several thousand anonymized patients and will be made available, securely and in full, to any registered team for the purpose of developing effective screening algorithms. The quiz/test dataset is a smaller set of anonymized patients. Teams will only receive the Historical Data section of these datasets and the two datasets will be mixed together so that teams will not be aware of which de-identified patients are in which set.

Teams will make predictions based on these data sets and submit their predictions to HPN through the official Heritage Health Prize web site. HPN will use the Quiz Dataset for the initial assessment of the Team’s algorithms. HPN will evaluate and report back scores to the teams through the prize website’s leader board.

Scores from the final Test Dataset will not be made available to teams until the accuracy thresholds are passed. The test dataset will be used in the final judging and results will be kept hidden. These scores are used to preserve the integrity of scoring and to help validate the predictive algorithms. You can find more about Online Testing and Judging here.

The American Hospital Association estimates that more than 71 million people are admitted to the hospital each year, and that $30 Billion is spent on unnecessary admissions.