Bastiaan Sallevelt

123 Conversion of STOPP/START version 2 into coded algorithms During the (first) functionality check, 23 (68%) of 34 START criteria were correctly triggered, 5 (15%) could be improved and 6 (17%) did not show up within the CDSS. Regarding STOPP criteria, 41 (51%) were triggered accurately during first evaluation. Eleven (14%) could be improved and 28 (35%) did not showup. The reasons for incorrect triggering (both false positive and false negative) varied from simply dots instead of commas in the algorithms (logical error) to non-present ATC code for a specific medication in the algorithm (content). For all algorithms that were not triggered when expected or that could be improved, the logic was reevaluated on errors and the content adjusted, as depicted in Figure 2, until all algorithms were functional and correct. The algorithms From a total of 114 criteria, we were able to code all 34 START criteria and 76 out of 80 STOPP criteria, corresponding with 96% of all criteria. The final 110 algorithms are attached as Supplementary Information SI1. All ICPC 1, ICPC 2, ICD9-, ICD10- and ATC codes used to convert individual STOPP/START criteria are listed as a code dictionary in Supplementary Information SI2. Technical aspects From the initial draft onwards, the algorithms were described using decision tables, a commonly used approach to modeling inference rules [26]. Decision tables have the advantage of being easily understandable for domain experts while being logically unambiguous. We created a colorized domain-specific decision table format to optimize the readability as much as possible. All criteria were modeled using this format. A (simplified) example of the decision table format for START criterion C3 is shown in Table 5. 2