Outlier score as a feature

[pekkatii]

A time series feature for the degree of "outlierness" could be useful to identify sites and/or individual subjects which are outliers, i.e. have few neighbours and are different from most of the other subjects.

[fcollinbi]

One suggestion could be strating with ARIMA modeling as done in tsoutliners package. Alternatively as you have alredy computet the PCA, that could be used to forecast and detection then via deviation from the forecast.

[pekkatii]

One simple option would the local outlier factor (https://en.wikipedia.org/wiki/Local_outlier_factor). I think this was the one I was experimenting with once.