A record split gave 89.15% mean balanced accuracy for permuted person labels in the UCI Parkinsons data.

On the UCI Parkinsons voice table, a one-neighbor classifier gave 89.1485615% mean person-weighted balanced accuracy when person labels were permuted and records were split across training and test sets. With a person split, the mean was 48.5252976%. Both results use 1,000 permutations and one fixed split. A radial-basis-function support-vector machine gave means of 62.2105159% and 48.6061508%, respectively. In record splits, training and test sets shared person codes; in person splits, they did not. The high record-split score is a method check on this table, not evidence of disease classification in new people. UCI reports 31 people and 23 disease cases, while the analyzed file contains 32 name prefixes and 24 disease-label prefixes. We cannot verify that each prefix identifies one person. The analysis does not establish clinical performance and does not claim that the known identity-confounding problem is new.

Parkinsons · voice measurements · identity confounding · person-label permutation · record split · person split · balanced accuracy · nearest neighbor · support-vector machine · UCI dataset