Or how about work in general? If doing something 6 hours a week is borderline what is doing something 40-60 hours a week? Assuming most people work for money, isn't this just an extreme cultural addition to money?
Depending on the application, you could keep a limited number of samples in memory and "below" or "above" those samples would simply be a counter of how many samples fall into either range. This would only work for applications where you expected the median to only vary in a limited fashion, and only after the system had been "primed".
For example, at any point in time, only store 100-1000 samples in memory, tree'd into "greater than median" and "less than median", additionally in each side of the tree store the number of samples that are in that tree but were purged do to the samples in memory requirement. When each new sample comes in, balance the tree to find the median.
I think this would work as long as a the median didn't have drastic movements up or down e.g. 1000 samples in a row < median, if that happened, it would exhaust your in memory sample set and you'd run out of samples to use as the median.
Even if you did expect wide variations you could use a similar approach but instead of purging samples into a counter you could serialize them to disk, or even have 3 layers, memory -> disk -> purge to counter.
EDIT: now that I think about this for more than 10 seconds, you may need to track the first derivative of median movement to determine how many samples to keep on a given side also.