FIGURE 3 Age group division. The bars represent number of subjects in each age group, while the colours illustrate the within-class gender distribution.
2.3.2 | Classification to age groups with variable age group thresholds
The effect of varying the age group thresholds was further analysed. Age grid of 26 to 63 was used to evaluate different thresholds. Instead of using all the ages between minimum and maximum in the age grid, only the ages appearing in the data were considered as part of the grid. As a result, the best corresponding thresholds were found for each classifier based on the achieved balanced accuracy. For each threshold, stratified K-fold CV was used with K=5. Figure 4 illustrates the age threshold finding process for each test.
In the age threshold test 1, all the data was used and split into two age groups of young and old adults using the aforementioned age grid. For the test 2, two age thresholds were used and the in-between data was excluded from the ML pipeline, while the beyond threshold subjects were sorted in to young and old adult groups. Test 3 used similar method as test 2, but the in-between data was used as a third rejection class. The goal of the tests 2 and 3 was to identify how the exclusion or inclusion of the in-between groups affects the classifiers performance.