A machine-learned analysis suggests non-redundant diagnostic information in olfactory subtests

Background: The functional performance of the human sense of smell can be approached via assessment of the olfactory threshold, the ability to discriminate odors or the ability to identify odors. Contemporary clinical test batteries include all or a selection of these components, with some dissent about the required number and choice. Methods: Olfactory thresholds, odor discrimination and odor identification scores were available from 10,714 subjects (3662 with anomia, 4299 with hyposmia, and 2752 with normal olfactory function). To assess, whether the olfactory subtests confer the same information or each subtest confers at least partly non-redundant information relevant to the olfactory diagnosis, we compared the diagnostic accuracy of supervised machine learning algorithms trained with the complete information from all three subtests with that obtained when performing the training with the information of only two or one subtests. Results: The training of machine-learned algorithms with the full information about olfactory thresholds, odor discrimination and odor identification from 2/3 of the cases, resulted in a balanced olfactory diagnostic accuracy of 98% or better in the 1/3 remaining cases. The most pronounced decrease in the balanced accuracy, to approximately 85%, was observed when omitting olfactory thresholds from the training, whereas omitting odor discrimination or identification was associated with smaller decreases (balanced accuracies approximately 90%). Conclusions: Results support partly non-redundant contributions of each olfactory subtest to the clinical olfactory diagnosis. Olfactory thresholds provided the largest amount of non-redundant information to the olfactory diagnosis