Abstract
OBJECTIVES: Explanations help to understand why anomaly detection algorithms identify data as anomalous. This study evaluates whether robustly standardized explanation scores correctly identify the implausible variables that make cancer data anomalous. MATERIALS AND METHODS: The dataset analyzed consists of 18 587 truncated real-world cancer registry records containing 8 categorical variables describing patients diagnosed with bladder and lung tumors. We identified 800 anomalous records using an autoencoder's per-record reconstruction error, which is a common neural network-based anomaly detection approach. For each variable of a record, we determined a robust explanation score, which indicates how anomalous the variable is. A variable's robust explanation score is the autoencoder's per-variable reconstruction error measured by cross-entropy and robustly standardized across records; that is, large reconstruction errors have a small effect on standardization. To evaluate the explanation scores, medical coders identified the implausible variables of the anomalous records. We then compare the explanation scores to the medical coders' validation in a classification and ranking setting. As baselines, we identified anomalous variables using the raw autoencoder's per-variable reconstruction error, the non-robustly standardized per-variable reconstruction error, the empirical frequency of implausible variables according to the medical coders' validation, and random selection or ranking of variables. RESULTS: When we sort the variables by their robust explanation scores, on average, the 2.37 highest-ranked variables contain all implausible variables. For the baselines, on average, the 2.84, 2.98, 3.27, and 4.91 highest-ranked variables contain all the variables that made a record implausible. DISCUSSION: We found that explanations based on robust explanation scores were better than or as good as the baseline explanations examined in the classification and ranking settings. Due to the international standardization of cancer data coding, we expect our results to generalize to other cancer types and registries. As we anticipate different magnitudes of per-variable autoencoder reconstruction errors in data from other medical registries and domains, these may also benefit from robustly standardizing the reconstruction errors per variable. Future work could explore methods to identify subsets of anomalous variables, addressing whether individual variables or their combinations contribute to anomalies. This direction aims to improve the interpretability and utility of anomaly detection systems. CONCLUSIONS: Robust explanation scores can improve explanations for identifying implausible variables in cancer data.