Interpretability. While numbers of people screened are easy to interpret, the NBCSP screening pathway may be confusing to some users. Information on the NBCSP is available in all AIHW NBCSP monitoring reports, for example, National bowel cancer screening program monitoring report 2009. Information on ABS data is available on the ABS website. Extensive information is also available on bowel cancer screening and can be located both on the internet and in hard copy.
Interpretability. Besides the performance of the proposed model, the interpretability is crucial in applying the results in practice. The attention layer gives us the possibility to extract the importance of each event in a single input sequence. An example is given in Figure 4: it shows an input sequence of three events (and seven padded non-events) from least recent to most recent. According to the weights, we can conclude that starting a new job is an important event in the life of the participant, regarding his or her activity towards pension. The two (different) offline campaigns have very little influence on this behavior. Furthermore we see that the padded events hold information as well. Apparently the fact that there is only three events in this sequence increases the possibility of activation. This could be due to customers being less and less probable to be activated every time they not respond to outbound communication. Hence the relatively high focus on the non-events preceding the actual events. Figure 4: Example of attention divided over a single sequence leading up to activation. The possibility of extracting the crucial events for a specific client, is what distinguishes this technique from other semi-interpretable machine learning techniques. For example XGBoost is able to provide the user with aggregated variable importance based on information gain (see Table 4). This however is not applicable to a single instance and therefore less insightful in this case, where the pension fund strives for personal and individual contact. Type Content Information gain Offline magazine Status update fund (Nov15) 0.203 Offline magazine Status update fund (Feb16) 0.075 Offline magazine Status update fund (Mar17) 0.064 Life event New job 0.048 Mail Status update fund (Mar17) 0.046 Table 4: Top five events with highest information gain as results of the XGBoost model. However, when we do take a closer look at the average activity for each of the events, we see a clear set of events that have a greater effect on activating the participants (Table 5). Apparently a very specific campaign in 2014, where call agents pro-actively called participants welcoming them to the fund after starting a new job, spiked the activity of those participants. Furthermore, campaigns revolving losing a job, switching jobs, working part-time and retiring part-time are quite effective. These events can be seen as ‘touchpoints’ in the lives of the participants where they are more interested in their ...
Interpretability. Supporting information on the quality and use of the Admitted Patient Care NMDS are published annually in Australian hospital statistics (technical appendixes), available in hard copy or on the AIHW website. Readers are advised to read caveat information to ensure appropriate interpretation of the performance indicator. Supporting information includes discussion of coverage, completeness of coding, the quality of Indigenous data, and changes in service delivery that might affect interpretation of the published data. Metadata information for this indicator are published in the AIHW’s online metadata repository- METeOR. Metadata information for the NMDS for Admitted patient care are also published in METeOR and the National health data dictionary. Information on ABS data is available on the ABS website. Principal diagnoses reported for this indicator were classified, coded and reported to the National Hospital Morbidity Database (NHMD) using the sixth edition of the International statistical classification of diseases and related health problems, 10th revision, Australian modification (ICD-10-AM) (NCCH 2008).
