Over the years, Backblaze has published several interesting reports and statistics on their harddrive performance and failures. In their latest post, they look at which SMART metrics they monitor and react to. They say, "when the RAW value for one of these five attributes is greater than zero, we have a reason to investigate".

SMART 5 Reallocated Sectors Count
SMART 187 Reported Uncorrectable Errors
SMART 188 Command Timeout
SMART 197 Current Pending Sector Count
SMART 198 Uncorrectable Sector Count

They go on to point out, that although a single one of those metrics might only indicate 30% to 40% probability of a failed drive, when combined the probability increases to 76%. Furthermore, looking at multiple failed values, and their rate of increase might also help determine if a drive is about to fail. A higher rate is usually worse. Finally, they note that the the SMART 189 - High Fly Writes is another good candidate to consider, and again when looking at the rate.

On a related note, here's a note on Google's data on their SSD failures.