Repetition features reduced later odor intensity prediction error compared with a fixed time and exposure model in 24 ds003487 people

We tested OpenNeuro ds003487 version 2.0.0. Two fixed models trained on the first 27 trials and predicted the last 27 trials for the same 24 people. The baseline used person-condition means, elapsed time, and total prior exposure. Adding repetition features reduced combined chocolate and neutral mean squared error from 339.24 to 271.84 squared rating points, a 19.87% gain. A 20,000-draw person bootstrap refitted both models; its 95% relative-gain interval was 2.09% to 32.77%, under independent and exchangeable people. All planned checks retained a positive point gain. The source paper already reported a repetition effect; this work adds a temporal prediction test. An exploratory check found that early person-condition means alone had less error than the primary baseline. Adding repetition to that simple model reduced error by 12.85%. These results do not select an optimal model or establish a cause or neural mechanism. Training and testing used the same people, and transfer to other tasks, people, odors, or brain estimates requires a new test.

odor intensity · repetition · prediction · sensory history · ridge regression · bootstrap · OpenNeuro · temporal validation