Detecting occurrence and timing of grizzly bear parturition based on anomalies in accelerometer data
Documenting natality of brown (grizzly) bears ( Ursus arctos ) is an important component of many population monitoring and research programs. Brown bears give birth during hibernation, so observation of litters is generally not feasible until females exit dens, and detection of litters can be compromised by poor observability. Using triaxial accelerometer activity data stored on radiocollars, we developed a technique that predicts births by the presence of brief upsurges in activity likely associated with postnatal maternal behaviors. We developed presence criteria with a training sample of known parturient females ( n = 22) to detect anomalies within time-series of daily motion-counts (i.e., activity readings >0 measured at 10-min intervals) during 25 December–7 March. To test performance, we applied criteria to a blinded sample of activity data obtained from female grizzly bears (captured and collared during 2012–2022) in 4 populations in interior North America ( n = 295). We assigned predicted status and compared assignments to reproductive status at first visual observation. The true positive rate was 91% ( n = 47 females observed with cubs of the year) and the false-positive rate was 17% ( n = 65 females observed with older offspring). Births were predicted for 50% of females with unknown reproductive status ( n = 48), 21% of females observed without offspring ( n = 114), and 10% of subadult females considered too young to reproduce ( n = 21). Dates of predicted births varied from 27 December to 28 February. Our anomaly detection technique was successful at estimating parturition events, and despite some error, indicated that a number of litters were born but not observed, presumably due to mortality in or shortly after den emergence. This technique provides an additional tool for supplementing visual observations for natality estimation and population modeling, provided that potential biases stemming from increased detection rates are considered.