many important nonstandard indicators of patient care-can easily be extended to hand hygiene surveillance. Because logins linked to computers in patient rooms correlate with actual visits by HCWs vis-a-vis hand hygiene opportunities arising from HCW/patient contact, they can also be used to estimate temporal patterns appropriate for effectively monitoring hand hygiene activity levels.To validate our new method, we used 660 days of UIHC MICU log-in data (September 1, 2006, through June 21, 2008, restricted to those log-ins linked to patient rooms (a total of 1,757 unique users). We then counted the number of unique users who log in for every hour for each day in our data set. For each day and night shift, we then rank ordered each hour on the basis of the number of unique individual log-ins observed (the choice of this metric was motivated by our simulations of hand hygiene compliance, which show that methodologies that favor observing more unique individuals rather than more events results in a better overall estimate of unit compliance, in essence by reducing sample bias in population selection). The resulting distribution of each hour's respective rank is then calculated across the entire data set and validated against a similar rank-order statistic derived from the sensor-mote data, where each hour in the shift is ranked by median number of captured hand hygiene events.Overall, we found that the observed hourly, unique HCW log-ins in our sensor data set are highly correlated with the same measure in our log-in data, with a Spearman's p of 0.86 (P<.001).Choosing a single observation hour on the basis of log-in rank results in selecting the single best hour 22% and 29% of the time for the day and night shift, respectively, a 2.5-3-fold improvement over simply selecting an hour at random. If we instead select the top quartile of hours per shift, it will contain the best hour 47% and 60% of the time for the day and night shift, respectively, a 1.5-2.5-fold improvement over uniform random selection. More generally, because this approach calculates qualitative rankings for each hour, we can both identify alternate candidate hours for observation (eg, the second-best hour to observe) and identify candidates that are consistently the best hours for observation compared with other, more variable candidate hours. By examining the variability (ie, entropy) of a given hour's rankings, we find that the best hours for observation tend to be those with lower variability for a shift-that is, their rankings are more stable across any given day. As a last point, all these results assume a static choice of observation hour across the entire data set. An obvious improvement would be a dynamic schedule in which an algorithm uses a window of recent log-in data to propose different candidate hours, providing real-time guidance to observers on which hours to monitor. We leave this idea for exploration in future work.Data-driven approaches are being applied to problems in health care with increasing regularity. Improving hand hygien...