Table 3 Number of selected and implausible records overall and per tumor localization in each sample. FindFPOF and the autoencoder had a higher precision overall and for each tumor localization than the baseline. In the random sample, \(8\%\) of all records and \(2\%\) of the breast records were implausible. For the autoencoder sample, \(28\%\) of all records and \(10\%\) of the breast records were implausible. FindFPOF and the autoencoder selected more records from those localizations that had a higher percentage of implausible records in the random sample. For the random sample, \(18\%\) of the colorectal records were implausible, while only \(2\%\) of the breast records were implausible. Thus, the autoencoder and FindFPOF returned more colorectal records (approximately two-thirds of the records are colorectal) that had a higher percentage of implausible records (approximately one-third). In contrast, the samples returned by the autoencoder and FindFPOF contain a lower percentage of breast records (\(28\%\) and \(14\%\), respectively) than both the random sample and the full dataset (\(54\%\) and \(58\%\), respectively). For each sample, the tumor localizations with the highest number of selected and implausible records are highlighted

From: Unsupervised anomaly detection of implausible electronic health records: a real-world evaluation in cancer registries