1) The oversampling of the training set makes it uncalibrated when applied to the general population (to recalibrate you need an estimate of the prevalence to begin with, which sort of defeats the purpose).
2) Online posts are not a random sample of the population. (Perhaps this is solveable with some poststratification of the estimates, although requires demographic data on the poster.) If you take self reports that the researchers used to define disorders at face value, those would make more sense than using this model.
These text based models are so superficial, when applied to mass datasets with low prevalence of the underlying condition, they will ultimately result in very low positive predictive values (e.g. flag 100 people, if the model is good will only get 5/100 as actual mental health problems).
As version_five asks, it is hard to imagine any reasonable use of the model given such low positive predictive values (which imply incredibly high false positive rates).