Table 4. Example of mismatched therapies and comparison between true and AI assisted prescriptions using strategy 1.

For a given antimicrobial drug and inpatient, data contains prescription records (first column—1 if the drug was prescribed during admission, 0 otherwise), and later culture test (sixth column—“True outcome”, S or R if the isolate was found susceptible or resistant to the drug, respectively). In this example, 5 inpatients had resistant cultures, which defines n in the algorithm. Drug treatment corresponds to a physician’s prediction (second column—a physician prescribes a drug if they believe that isolates are susceptible to the drug, while nothing can be said if a drug is not prescribed). An AI system such as the GBDT model returns the probability that an isolate is found resistant (third column—“AI outcome”). The binary prediction is obtained by taking the n = 5 cases with the highest probability and assigning a label accordingly (AI prediction—fourth column). This suggests not to administer a drug if resistance is predicted (AI prescription—fifth column), thus lowering the total exposure to the drug (from 6 to 4 inpatients under treatment). By comparing the true prescription or the AI prescription (columns 1 and 5) with the culture test results (column 6) the number of mismatches can be computed (7th and 8th columns for the physician and AI assisted prediction, respectively).

True prescr.	Physician predic.	AI outcome	AI predic.	AI prescr.	True outcome	Physician mism.	AI mism.
1	S	0.1	S	1	S	0	0
1	S	0.2	S	1	S	0	0
1	S	0.7	R	0	R	1	0
1	S	0.4	S	1	S	0	0
1	S	0.6	R	0	R	1	0
1	S	0.5	S	1	R	1	1
0	?	0.1	S	0	S	-	-
0	?	0.6	R	0	S	-	-
0	?	0.8	R	0	R	-	-
0	?	0.9	R	0	R	-	-