Table 3:

Data utility for original, sifted and naive suppressing CDC text (n=86,666) according to the constructed BERT model (f(·)) that maps the CDC injury records to 5-class OIICS labels. The label prediction accuracy records $\frac{1}{n}{\sum }_{i=1}^{n}I[f(W_i^{\mathrm{*}})=L_i]$ and the label prediction agreement records $\frac{1}{n}{\sum }_{i=1}^{n}I[f(W_i)=f(W_i^{\mathrm{*}})]$.

Obfuscation Level	DataSifterText Accuracy	Naive Suppressing Accuracy	DataSifterText Agreement
Original	98.5%	-	-
Sifted low obfuscation	86.5%	85.7%	87.3%
Sifted medium obfuscation	74.5%	71.1%	75.0%
Sifted high obfuscation	60.9%	48.8%	61.1%