Policy on sharing summary statistics

[kartik727]

Hi,

We are working on using MIMIC-III to create a Markov decision process (MDP) to model the evolution of the status of sepsis patients over time in response to different clinical interventions, similar to the AI clinician paper that also uses MIMIC-III (GitHub link).

The state of each patient at any given time step is clustered into 750 states, and two additional states corresponding to whether the patients survived or not, for a total of 752 states, and the clinical interventions are clustered into 25 actions. We then aggregate the number of transitions that take place from each initial state to each final state for each action and normalize them to create valid probability distributions in a 752 x 25 x 752 transition matrix. We similarly aggregate the number of times each state was seen as the first recorded state of a patient in the initial-state distribution, which is a vector of length 752.

Our question is, can we make the resulting transition matrix open to the general public? This matrix only includes summary statistics about how patients falling into 752 different clusters tend to respond to treatments falling into 25 different clusters. Although this doesn't mean that we will be directly publishing MIMIC-III data, it would be summary statistics from that data, and so we wanted to ensure that posting our resulting MDP/transition matrix would not violate the Data Use Agreement (DUA), which says:

I will not share access to PhysioNet restricted data with anyone else.

[alistairewj]

Always hard to draw the line between publishing findings and sharing the data, but this sounds like publishing a model from MIMIC, which has a lot of precedent. Just make sure there isn't some trivial way of regenerating individual level data (e.g. states of n=1).